diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index f30ba53..ee5e9e8 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -66,9 +66,10 @@ jobs:
dvidia-observe --help
dvidia-observation-studio --help
dvidia-motion --help
+ dvidia-transition --help
dvidia-training-doctor --json
- name: Test footage, visual training and loopback studio
- run: python -m unittest tests.test_footage_data tests.test_footage_train tests.test_footage_pipeline tests.test_footage_benchmark tests.test_training_studio tests.test_doctor tests.test_coverage tests.test_episode_tape tests.test_reaction tests.test_reaction_metrics tests.test_reaction_convergence tests.test_reaction_benchmark tests.test_motor_memory tests.test_motor_metrics tests.test_motor_benchmark tests.test_observations tests.test_observation_models tests.test_motion tests.test_motion_benchmark -v
+ run: python -m unittest tests.test_footage_data tests.test_footage_train tests.test_footage_pipeline tests.test_footage_benchmark tests.test_training_studio tests.test_doctor tests.test_coverage tests.test_episode_tape tests.test_reaction tests.test_reaction_metrics tests.test_reaction_convergence tests.test_reaction_benchmark tests.test_motor_memory tests.test_motor_metrics tests.test_motor_benchmark tests.test_observations tests.test_observation_models tests.test_motion tests.test_motion_benchmark tests.test_transition_model tests.test_transition_pilot tests.test_transition_cli -v
motion:
name: Causal image motion / Linux / Python 3.12
diff --git a/README.md b/README.md
index f9109ad..f2111e9 100644
--- a/README.md
+++ b/README.md
@@ -83,6 +83,40 @@ unqualified. This is one component toward [the full Observation Lab v0.2](docs/o
See the [measured authored benchmark](benchmarks/motion-alpha-20261010/README.md),
including conditional errors, failed tracking and forecast tails.
+## Learn a compact command-to-movement model
+
+The [laptop skill pilot](docs/laptop-skill-pilot.md) connects episode tapes to a
+small action-conditioned predictor. It learns the next 20 ms of joint and jaw
+movement from current state and issued commands, using CPU ridge regression.
+Separate simulation groups test it against no movement, current velocity,
+training-mean movement and a matched model with commands removed. A local,
+offline report exposes prediction errors, artifact size and model-only timing.
+
+```sh
+.venv/bin/python -m pip install '.[arm]'
+.venv/bin/dvidia-transition corpus --output runs/actuator-corpus
+.venv/bin/dvidia-transition run runs/actuator-corpus/corpus.json --output runs/actuator-pilot
+.venv/bin/dvidia-transition inspect runs/actuator-pilot
+```
+
+Use the `feat/laptop-skill-pilot-v0.1` candidate source branch for these commands,
+then open `runs/actuator-pilot/report.html`. Generation uses the existing native
+rigid-box simulator; fitting and inference need no language model, GPU or remote
+service. Simulator measurements are privileged, even though the selected inputs
+resemble signals a robot could expose. The predictor issues no robot commands
+and establishes neither grasp success nor shoe organization.
+
+The [retained CPU benchmark](benchmarks/actuator-pilot-20261011/README.md) includes
+the trained 6.3 KiB predictor, frozen protocol and complete attempt receipts.
+Its warm model computation measured 0.207 ms at p95 on one laptop; sensing and
+actuation are excluded. Command inputs reduce test RMSE, while some stationary
+gripper and worst-error results regress. Read all error measures before reusing it.
+
+The [first shoe collection plan](examples/shoe-organizing-coverage.json) defines
+one shoe, one marked slot and complete attempts. Its benchmark remains blocked
+until a validated shoe environment and compatible physical adapter exist.
+Recording quotas are proposed study sizes, not known sufficiency thresholds.
+
## Try the pipeline
Generate an explicitly synthetic example, then train and verify its output:
diff --git a/benchmarks/actuator-pilot-20261011/LICENSE b/benchmarks/actuator-pilot-20261011/LICENSE
new file mode 100644
index 0000000..a7056f2
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 DVIDIA contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/benchmarks/actuator-pilot-20261011/README.md b/benchmarks/actuator-pilot-20261011/README.md
new file mode 100644
index 0000000..9a74509
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/README.md
@@ -0,0 +1,207 @@
+# DVIDIA compact actuator pilot — frozen evidence, 11 October 2026
+
+This CPU experiment learned a seven-output, next-20-ms actuator transition from
+native simulation telemetry. Issued commands improved joint prediction and jaw
+RMSE relative to state-only and simple baselines, but important jaw errors remain.
+It is a research predictor, not a robot controller or a learned shoe skill.
+
+The exact final run artifacts are retained without modification. These original
+DVIDIA model artifacts, evidence and documentation use the [MIT license](LICENSE).
+The source candidate is
+[`feat/laptop-skill-pilot-v0.1`](https://github.com/Dvidia-Inference/dvidia-training/tree/feat/laptop-skill-pilot-v0.1).
+The measured source is pinned to [commit `0d3f05f`](https://github.com/Dvidia-Inference/dvidia-training/tree/0d3f05f186423e18b15ed8362d772b7d74ec0207);
+[source.json](source.json) retains supplemental code hashes without modifying
+the frozen receipt bundle.
+See the [pilot guide](../../docs/laptop-skill-pilot.md) for installation and the
+[episode-tape contract](../../docs/episode-tape.md) for signal provenance.
+
+## Corpus and evaluation boundary
+
+The authored MuJoCo 3.15.0 task is rigid-box placement with a six-joint arm and a
+parallel-jaw gripper. The 12 declared seed groups, 311000–311011, each contain
+three fresh sibling recordings: nominal, open jaw and release after lift.
+All siblings stay together: eight training groups, two development groups and
+two test groups. These are related native simulation resets, not independently
+captured physical scenes. The same authored controller generates every case.
+
+The capture plan declares scene jitter and bounded mass/friction variation before
+collection. Group mass spans 0.035–0.045 kg and friction 0.70–0.90; the ordered
+split holds out the final groups. Configuration, seed, group, case, phase,
+terminal outcome, object/contact state and after-interval applied controls are
+not predictor inputs. The models receive contemporaneous native joint positions,
+joint velocities and jaw gap, with issued joint/jaw targets added for the action
+model. This is privileged simulator telemetry in sensor-shaped fields; it was
+not extracted from egocentric footage.
+
+| Split | Groups | Recordings | Attempted control steps | Retained 20-ms transitions | Excluded partial boundaries |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Training | 8 | 24 | 18,244 | 18,228 | 16 |
+| Development | 2 | 6 | 4,559 | 4,555 | 4 |
+| Test | 2 | 6 | 4,557 | 4,553 | 4 |
+| Total | 12 | 36 | 27,360 | 27,336 | 24 |
+
+All 36 recording attempts completed and none were discarded. The authored
+controller reached simulation success in 24 recordings; the 12 open-jaw
+recordings reached their horizon without success. These outcomes describe the
+recording controller and are not learned-model success rates. Each case has
+12 recording attempts. The test contains 1,171 nominal, 1,800 open-jaw and
+1,582 release-after-lift transitions. The 24 exclusions retain explicit reasons
+rather than inventing a full 20-ms target from a partial terminal interval.
+
+Both ridge models fit normalization only on training rows, with equal total
+weight per retained episode. Each selects alpha from `[0.0001, 0.01, 1, 100]`
+using the fixed development score `joint_rmse_rad / 0.01 + jaw_rmse_m / 0.001`.
+Both selected `0.0001`. Model files and [freeze.json](freeze.json) were persisted
+before test-tape parsing or transition-target extraction. The manifest's test
+identities, hashes and recording outcomes were already available for preflight;
+`held_out_transition_targets_read: false` refers specifically to the transition
+targets at freeze time. Test results did not choose the models.
+
+## Held-out prediction errors
+
+All five predictors cover the same 4,553 test rows. Joint statistics flatten the
+six joint errors per transition; worst joint error is the largest absolute
+component across all rows. Jaw statistics concern the measured gap delta.
+Radians and millimetres are separate quantities; there is no mixed-unit accuracy
+percentage. JSON receipts retain jaw values in metres; tables convert to mm.
+
+| Predictor | RMSE (rad) | MAE (rad) | Worst absolute error (rad) |
+| --- | ---: | ---: | ---: |
+| Action-conditioned ridge | 3.63056194e-05 | 7.24671574e-06 | 0.00162278889 |
+| State-only ridge | 0.000410349272 | 7.96971967e-05 | 0.0131272116 |
+| Persistence | 0.00260724634 | 0.0015489858 | 0.0141029159 |
+| Constant joint velocity | 0.000577719745 | 0.000265731457 | 0.0120383224 |
+| Training-mean delta | 0.00254929171 | 0.00169018742 | 0.0141142779 |
+
+| Predictor | RMSE (mm) | MAE (mm) | Worst absolute error (mm) |
+| --- | ---: | ---: | ---: |
+| Action-conditioned ridge | 0.34160938 | 0.114485012 | 3.4245926 |
+| State-only ridge | 0.551877626 | 0.188971488 | 3.55103368 |
+| Persistence | 0.564417684 | 0.108018148 | 3.20004021 |
+| Constant joint velocity | 0.564417684 | 0.108018148 | 3.20004021 |
+| Training-mean delta | 0.564417684 | 0.108021865 | 3.20004778 |
+
+Persistence predicts zero delta. Constant velocity predicts `joint_velocity ×
+0.02` for the joints and zero jaw delta because jaw velocity is not an input.
+Training mean uses the fixed weighted training delta. Exact case/group/slice
+results remain in [report.json](report.json).
+
+**The jaw result is mixed.** The action model improves aggregate jaw RMSE
+(0.341609 mm versus 0.564418 mm for persistence), but its jaw MAE is worse
+(0.114485 versus 0.108018 mm), and its worst absolute jaw error is worse
+(3.424593 versus 3.200040 mm). On the 1,800 open-jaw test transitions, action-model
+jaw RMSE is 0.057758 mm versus 0.002761 mm for persistence. A lower average squared
+error does not imply universally better predictions or safe actuator behavior.
+
+The predeclared jaw-changing slice, `abs(measured gap delta) > 0.01 mm`, contains
+296 test transitions. Its action-model jaw RMSE is 1.295277 mm, MAE 0.957763 mm,
+and worst error 3.424593 mm; persistence gives 2.213605, 1.647262 and 3.200040 mm.
+The moving-joint slice, any `abs(measured joint delta) > 0.0001 rad`, contains
+3,398 transitions. Its action-model joint RMSE is 0.0000418183 rad versus
+0.000668735 rad for constant velocity. These slices are evaluation summaries,
+not selection inputs.
+
+## Matched intervention exposes an opening error
+
+The first retained nominal/release pair with identical 13 before-state inputs
+and identical issued joint targets occurs at reset-010, index 290, around 5.8 s.
+The issued jaw target changes from 33.537947 mm to 80 mm.
+
+| Next-20-ms jaw delta | Nominal | Release intervention |
+| --- | ---: | ---: |
+| Native observation (mm) | −0.000001286 | +0.010414352 |
+| Action-model prediction (mm) | −0.073450757 | +3.429849401 |
+| State-model prediction (mm) | −0.108892870 | −0.108892870 |
+
+The action model responds to the changed command, while the state-only model
+necessarily gives the same prediction for both inputs. However, the action model
+substantially overpredicts the opening in this contact-bearing transition:
+3.429849 mm predicted versus 0.010414 mm observed. This demonstrates command
+sensitivity and also a serious transient error. It does not validate contact
+physics, gripper opening or an accurate action-conditioned world model.
+
+## Measured laptop costs
+
+The final fit/evaluation run used macOS 26.5.2 on arm64, with ten logical CPUs,
+Python 3.12.9 and NumPy 2.5.3. The receipt does not identify a specific Apple
+chip model. No GPU or remote inference service was used. Fitting uses NumPy;
+saved-model inference uses the Python standard library.
+
+| Warm compute-only predictor | Median | p95 | Worst |
+| --- | ---: | ---: | ---: |
+| Action model | 0.198250 ms | 0.206752 ms | 0.231667 ms |
+| State model | 0.144833 ms | 0.151550 ms | 0.184917 ms |
+
+Each model received 100 warmup calls followed by 1,000 measured calls over
+pre-extracted test feature vectors. The timed `predict_features` call includes
+model validation and arithmetic. This excludes camera capture, perception,
+feature extraction, transport and actuator response; it is not a measured robot
+reaction time or an operating-system deadline guarantee. Cold saved-model
+loading was not separately timed. Timing differences between repeated runs can
+reflect scheduling and process history.
+
+The original corpus collection took 41.129711 s for
+547.09 simulated seconds. The final fit/evaluation pipeline took
+19.292200 s before the final receipt write,
+including 8.799433 s for fitting/development selection.
+Tape load/validation and feature/target extraction are reported separately:
+
+| Split | Tape load and validation | Feature/target extraction |
+| --- | ---: | ---: |
+| Training | 1.653895 s | 2.772606 s |
+| Development | 0.445649 s | 0.709693 s |
+| Test | 0.442433 s | 0.730566 s |
+
+Process-lifetime peak RSS was 184,614,912 bytes
+(176.062 MiB). This is the process high-water mark,
+which includes earlier work in that process; it is not isolated learner memory
+or a minimum hardware requirement. The two serialized model files total
+11,445 bytes: 6,438 for action and 5,007 for state.
+
+## Artifact use and reproducibility
+
+| Artifact | Purpose |
+| --- | --- |
+| [report.json](report.json) | Exact final results, timings, exclusions and limitations |
+| [corpus-manifest.json](corpus-manifest.json) | Exact manifest bytes, grouped provenance and original tape hashes; tapes omitted |
+| [freeze.json](freeze.json) | Selected models, source hashes and transition-target boundary declaration |
+| [action.model.json](action.model.json) | Original MIT action-conditioned 20-input, seven-output ridge artifact |
+| [state.model.json](state.model.json) | Original MIT state-only baseline with 13 active inputs |
+| [integrity.json](integrity.json) | Hash/byte receipts for the five retained artifacts |
+
+The bundle was checked for private absolute paths, credential-shaped values and
+non-native recording sources before publication. No raw tapes, private footage,
+third-party pretrained model weights or credentials are included. Original
+DVIDIA coefficients and native-simulation evidence are covered by [LICENSE](LICENSE).
+MuJoCo and other dependencies retain their own licenses.
+
+With the candidate source installed, inspect this bundle locally:
+
+```sh
+dvidia-transition inspect benchmarks/actuator-pilot-20261011
+```
+
+Inspection verifies internal hashes, model contracts, frozen selection and
+source/count consistency. It cannot authenticate the origin of declared sources,
+recompute these metrics without the omitted tapes, or establish physical accuracy.
+The model contract explicitly records `is_controller: false`,
+`predicts_task_success: false` and `physical_robot_ready: false`.
+Intended use is offline predictor comparison, simulator research and API
+experimentation. Hardware command generation, stopping guarantees, grasp success
+and shoe organization require separate validated implementations and evidence.
+
+The recorded generator originally emitted `manifest.json`; a byte-identical
+`corpus.json` copy supplied the final CLI contract. The current generator emits
+`corpus.json` directly. This filename correction did not regenerate the corpus
+or change any tape bytes. The final fit reused that corpus and corrected the
+freeze wording to distinguish transition targets from already-read manifest
+outcomes. Both model SHA256 values and all prediction metrics are identical to
+the first fit; timings and final receipt hashes reflect the final rerun.
+Reproduction on another machine can change timings and numerical serialization;
+compare source/configuration contracts and metrics rather than assuming every
+native simulation or fit will be byte-identical.
+
+The next research gap is calibrated actuator/contact prediction and measured
+hardware feedback, followed by a collected shoe task with synchronized robot
+actions and a learned policy. These results establish a local training/evaluation
+building block for Skillspace; they do not show footage-to-shoe-control learning.
diff --git a/benchmarks/actuator-pilot-20261011/action.model.json b/benchmarks/actuator-pilot-20261011/action.model.json
new file mode 100644
index 0000000..f87659c
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/action.model.json
@@ -0,0 +1 @@
+{"alpha":0.0001,"coefficients":[[-0.023413911279235672,0.00013407676756472406,0.0003203766229395209,3.385448333636372e-06,0.00017321301275480667,0.020132376225264318,0.0006013207924994154,4.430150692606901e-07,7.886716196884903e-06,3.4779427934558955e-06,1.803514766483e-06,0.00023824704406615052,5.407164943187078e-07,0.022032663129652248,0.0006466113856412709,0.0010816819321729197,-7.6104386552289184e-06,0.000801525286877072,-0.020931580895070762,2.506335241894005e-07],[0.028897060212908034,-0.03075628405115735,0.018075803960955157,2.7692359613556942e-05,0.011004291463574851,0.028813033274637183,0.0005214717235201334,0.0005487569321548179,0.00042473480840509765,-1.6987682400190732e-05,0.00015540088408312448,0.00024105210598963366,-3.6941762690573377e-06,0.0010797254969632498,0.037538966792016605,-0.0065010137849883885,-1.309353702264536e-05,-0.002867445897627287,0.001017151473169434,-7.55676049845229e-06],[-0.06837042685742636,0.00636055965476416,-0.037459901610317384,5.8408945891694886e-05,0.02154120182035249,-0.06304671798033504,-0.001650886909583609,0.00013064025919859378,0.0021642579818388612,-3.429730369525898e-06,0.0006796452382266949,-0.0008810232082383357,-1.0784527092206725e-05,-0.0022562537849201386,-0.02560019480716129,0.0033559491718872757,-6.563433354691982e-05,-0.045386537137997916,-0.00713874394944704,7.93037189129509e-06],[0.0005810621511276375,-6.609704904846584e-06,-9.555805711105124e-05,-2.8225119969198456e-06,-4.624575739576662e-05,0.00034809326159791147,1.627695604684937e-05,-2.7975119521258152e-08,-3.465369326916786e-06,1.1245225381633713e-06,-1.3938233543710135e-06,5.963562340117197e-06,-8.741771910459407e-08,-8.026010259530823e-05,2.9552246239712374e-05,0.00013758714242834038,2.0800472826216233e-06,7.490805308325208e-05,0.00014652955206523736,3.3369971356306223e-08],[0.04204767706576623,0.004941037866092892,-8.614408755212916e-05,3.533611689534423e-05,-0.0730657021495212,0.040280083782873705,0.0008785599184403115,8.389381136526186e-06,-0.00037785324562798886,1.3038922856525207e-05,0.00025267696419823294,0.00044293823354058057,-6.8952941412634566e-06,0.00039925799818446654,-0.016675829604721082,-0.020783407916172357,-3.270449344164002e-05,0.05846580976024862,0.0019056080571623136,-1.209474061041307e-05],[0.027111811902137423,-1.9314636573421795e-05,-0.0010017447613202527,-1.138432906652123e-05,-0.0005011040107069056,-0.04091116580119544,0.0005658641452908233,2.0310824715503414e-06,-3.4169253489535234e-05,-8.38926729722339e-06,-1.4136569114327172e-05,0.00040498859001560517,-2.782182568318342e-07,-0.033938523136976304,-0.001197370813522306,-0.0011661682083346757,1.1436829824493793e-05,-0.0010107838521480428,0.03321016529468509,4.4133603330231703e-07],[0.26599189884719415,0.052313045664778034,0.08494490416660601,5.187975935088541e-05,0.05556362572906438,0.25770663545560346,0.00528639711577106,0.0013277000745836854,0.002225432003594018,-0.00019222466598737875,0.001107456667416646,0.0025443590514423083,-0.0013900792209215517,-0.004751127625655068,-0.0035234684087305235,0.0024226896146674446,0.00012666435597677748,0.005646375222438801,0.002180006869283331,0.0016143535524403483]],"control_dt_seconds":0.02,"feature_contract":{"acquisition":"required signals contemporaneous with boundary; no delayed-sensor propagation","availability":"before decision plus issued command; no future or applied control","names":["before.joint_position[0]","before.joint_position[1]","before.joint_position[2]","before.joint_position[3]","before.joint_position[4]","before.joint_position[5]","before.joint_velocity[0]","before.joint_velocity[1]","before.joint_velocity[2]","before.joint_velocity[3]","before.joint_velocity[4]","before.joint_velocity[5]","before.gripper_width","command.issued.joint_targets[0]","command.issued.joint_targets[1]","command.issued.joint_targets[2]","command.issued.joint_targets[3]","command.issued.joint_targets[4]","command.issued.joint_targets[5]","command.issued.gripper_width"],"origin":"sensor-shaped privileged native simulation; hardware sensors unqualified","units":["rad","rad","rad","rad","rad","rad","rad/s","rad/s","rad/s","rad/s","rad/s","rad/s","m","rad","rad","rad","rad","rad","rad","m"]},"feature_mode":"action","intercept":[0.0009610293150821192,0.0002868151938799142,1.1362007785179062e-05,1.1868488215456403e-10,-0.00029294613886360224,-0.0009610293636950372,-7.577548372409316e-09],"is_controller":false,"kind":"dvidia.compact-action-transition-model","license":"MIT","model_sha256":"c63cf993c733687a02217e5b060c711c09878e32d0526feeb44abfa0170e29f3","normalization":{"constant_features":[false,false,false,false,false,false,false,false,false,false,false,false,false,false,false,false,false,false,false,false],"means":[-0.0899905108472486,-1.1519101433125067,1.3351944186598976,2.0685882094698694e-06,-0.18334919247409012,0.08884156095454729,0.04256845907588669,0.01272885406120272,0.0012493047510298584,0.00011965114235991223,-0.012338512856367419,-0.038178765041471746,0.06767697721595069,-0.08650386989124388,-1.1509821105807483,1.3351471309301313,1.146019022818662e-06,-0.1841626660141949,0.08651220655980689,0.06582770080989356],"scales":[0.17860326146676422,0.1767063627527591,0.31658147968533146,7.945667011746564e-06,0.22167432531872702,0.1777062994467045,0.06463297032762196,0.08931124694281727,0.1797614962369242,0.00022410177231653518,0.15581823686916052,0.05904040544125759,0.018225447281794958,0.1757749114495912,0.17814145082896937,0.31831002701662303,6.997590745910784e-06,0.22144346050485844,0.17578204958614485,0.021410149727423857],"scope":"fit-training-only"},"output_contract":{"acquisition":"required after signals contemporaneous with the after boundary","contact_and_task_outcomes":"not predicted","definition":"after minus before at exactly one 20 ms control boundary","names":["delta.joint_position[0]","delta.joint_position[1]","delta.joint_position[2]","delta.joint_position[3]","delta.joint_position[4]","delta.joint_position[5]","delta.gripper_width"],"units":["rad","rad","rad","rad","rad","rad","m"]},"physical_robot_ready":false,"predicts_task_success":false,"schema_version":1,"scope":"simulation-only","sensor_origin":"privileged-native-simulation","training":{"positive_weight_rows":18228,"rows":18228,"rows_sha256":"7bff06c153c5071b3b162eec08ea23eaeeee032daf7d1fb0861c6d9cfc674248","sample_weights_sha256":"7787fd55dcedecc0ac29b9030d915ca9036afaa6247c5484333ac4dd9bfc8596","weight_sum":18228.0}}
diff --git a/benchmarks/actuator-pilot-20261011/corpus-manifest.json b/benchmarks/actuator-pilot-20261011/corpus-manifest.json
new file mode 100644
index 0000000..bbbd516
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/corpus-manifest.json
@@ -0,0 +1 @@
+{"collection_elapsed_seconds":41.12971108290367,"groups":[{"attempts":[{"bytes":3997721,"case":"nominal","episode_id":"reset-000-nominal","error":null,"outcome":{"control_steps":588,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.755999999998924},"path":"tapes/reset-000-nominal.json","provenance":{"case":"nominal","configuration_sha256":"95b84f92d27d8030901c6ae901d20b0d4f67a9ad5183879bbeea8712ef989d3f","engine":"mujoco-3.15.0","episode_id":"reset-000-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"f176451537ee7f071dcea04dabb8b4f3b81011004839235439df3e46de16af19","seed":311000,"session_id":"session-311000","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311000-nominal"},"recording_elapsed_seconds":1.0564580000936985,"sha256":"2d9633992eeb1f6c6cc7763d6dffd8b14f0d2eac2c10d895cc0979a22f56c608","source_recording_id":"recording-311000-nominal","status":"completed"},{"bytes":6028529,"case":"open-jaw","episode_id":"reset-000-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-000-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"95b84f92d27d8030901c6ae901d20b0d4f67a9ad5183879bbeea8712ef989d3f","engine":"mujoco-3.15.0","episode_id":"reset-000-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"f176451537ee7f071dcea04dabb8b4f3b81011004839235439df3e46de16af19","seed":311000,"session_id":"session-311000","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311000-open-jaw"},"recording_elapsed_seconds":1.2717415411025286,"sha256":"a20f122ff1670959b22ec9d5dfbb4729d583c8899e9a2ba4a1eb0d644b8432ab","source_recording_id":"recording-311000-open-jaw","status":"completed"},{"bytes":5395144,"case":"release-after-lift","episode_id":"reset-000-release-after-lift","error":null,"outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.855999999996651},"path":"tapes/reset-000-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"95b84f92d27d8030901c6ae901d20b0d4f67a9ad5183879bbeea8712ef989d3f","engine":"mujoco-3.15.0","episode_id":"reset-000-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"f176451537ee7f071dcea04dabb8b4f3b81011004839235439df3e46de16af19","seed":311000,"session_id":"session-311000","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311000-release-after-lift"},"recording_elapsed_seconds":1.2134074580390006,"sha256":"89741ba25f14fa62877940be2325e72e672835922dc2377258dc77af49464315","source_recording_id":"recording-311000-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.7,"object_half_size":[0.02,0.02,0.02],"object_mass":0.035,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"95b84f92d27d8030901c6ae901d20b0d4f67a9ad5183879bbeea8712ef989d3f","group_id":"reset-000","reset_lineage_id":"native-reset-311000","seed":311000,"session_id":"session-311000","split":"train"},{"attempts":[{"bytes":3990695,"case":"nominal","episode_id":"reset-001-nominal","error":null,"outcome":{"control_steps":587,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.735999999998935},"path":"tapes/reset-001-nominal.json","provenance":{"case":"nominal","configuration_sha256":"97b22ed0341322076b1ffe961da00317814dc818fb8e481bfca6ee83c1ffafd9","engine":"mujoco-3.15.0","episode_id":"reset-001-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"029448f23d9823a1ebe42b9567c10a85a2dfa468c1b58a1b7de9b4da9a0a2e24","seed":311001,"session_id":"session-311001","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311001-nominal"},"recording_elapsed_seconds":0.9069484579376876,"sha256":"bd0009e7e6955b2314080e6739951ebd1d965a0276e8a0e9e8c81ca707c56bbe","source_recording_id":"recording-311001-nominal","status":"completed"},{"bytes":6032277,"case":"open-jaw","episode_id":"reset-001-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-001-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"97b22ed0341322076b1ffe961da00317814dc818fb8e481bfca6ee83c1ffafd9","engine":"mujoco-3.15.0","episode_id":"reset-001-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"029448f23d9823a1ebe42b9567c10a85a2dfa468c1b58a1b7de9b4da9a0a2e24","seed":311001,"session_id":"session-311001","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311001-open-jaw"},"recording_elapsed_seconds":1.2672740002162755,"sha256":"896e1fe59923cc89e1bf86437c527e0a8e2faca9d1c2d46471af66e2f83d8841","source_recording_id":"recording-311001-open-jaw","status":"completed"},{"bytes":5396900,"case":"release-after-lift","episode_id":"reset-001-release-after-lift","error":null,"outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.855999999996651},"path":"tapes/reset-001-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"97b22ed0341322076b1ffe961da00317814dc818fb8e481bfca6ee83c1ffafd9","engine":"mujoco-3.15.0","episode_id":"reset-001-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"029448f23d9823a1ebe42b9567c10a85a2dfa468c1b58a1b7de9b4da9a0a2e24","seed":311001,"session_id":"session-311001","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311001-release-after-lift"},"recording_elapsed_seconds":1.2264115000143647,"sha256":"6b6f7dac245b8a538037f41981b5b16fec6e0b193b0eb3c2180e7a25c50b1000","source_recording_id":"recording-311001-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.7181818181818181,"object_half_size":[0.02,0.02,0.02],"object_mass":0.03590909090909091,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"97b22ed0341322076b1ffe961da00317814dc818fb8e481bfca6ee83c1ffafd9","group_id":"reset-001","reset_lineage_id":"native-reset-311001","seed":311001,"session_id":"session-311001","split":"train"},{"attempts":[{"bytes":4005185,"case":"nominal","episode_id":"reset-002-nominal","error":null,"outcome":{"control_steps":588,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.754999999998924},"path":"tapes/reset-002-nominal.json","provenance":{"case":"nominal","configuration_sha256":"9aa27203fc1cfc0eebeecd7e3a7743f83e037b362707c51b78f2c576be2abacc","engine":"mujoco-3.15.0","episode_id":"reset-002-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"caac749e8d168e1b9cf048735937923f88386dc5ce20195d3360b2422406db47","seed":311002,"session_id":"session-311002","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311002-nominal"},"recording_elapsed_seconds":0.8954987078905106,"sha256":"aeace5b052485035cfa2e12b9825e99ca2acfc3ce1a675dca8fb25353437472a","source_recording_id":"recording-311002-nominal","status":"completed"},{"bytes":6051492,"case":"open-jaw","episode_id":"reset-002-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-002-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"9aa27203fc1cfc0eebeecd7e3a7743f83e037b362707c51b78f2c576be2abacc","engine":"mujoco-3.15.0","episode_id":"reset-002-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"caac749e8d168e1b9cf048735937923f88386dc5ce20195d3360b2422406db47","seed":311002,"session_id":"session-311002","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311002-open-jaw"},"recording_elapsed_seconds":1.2493529580533504,"sha256":"cc8ca7c9924faf4e9e608b05fdbc2b4f3e999d25cad2aee883c9863a57640d3f","source_recording_id":"recording-311002-open-jaw","status":"completed"},{"bytes":5403671,"case":"release-after-lift","episode_id":"reset-002-release-after-lift","error":null,"outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.854999999996652},"path":"tapes/reset-002-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"9aa27203fc1cfc0eebeecd7e3a7743f83e037b362707c51b78f2c576be2abacc","engine":"mujoco-3.15.0","episode_id":"reset-002-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"caac749e8d168e1b9cf048735937923f88386dc5ce20195d3360b2422406db47","seed":311002,"session_id":"session-311002","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311002-release-after-lift"},"recording_elapsed_seconds":1.2024170418735594,"sha256":"06ed87cb5511bd0f4fe5e6a8d46eedf8c78aa150a8a3c58d0ffb19b932521015","source_recording_id":"recording-311002-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.7363636363636363,"object_half_size":[0.02,0.02,0.02],"object_mass":0.03681818181818182,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"9aa27203fc1cfc0eebeecd7e3a7743f83e037b362707c51b78f2c576be2abacc","group_id":"reset-002","reset_lineage_id":"native-reset-311002","seed":311002,"session_id":"session-311002","split":"train"},{"attempts":[{"bytes":4006111,"case":"nominal","episode_id":"reset-003-nominal","error":null,"outcome":{"control_steps":589,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.775999999998913},"path":"tapes/reset-003-nominal.json","provenance":{"case":"nominal","configuration_sha256":"9143c18b47087e3807a1bf6043e94748470dc53c96891db49166da1d3a8bef0c","engine":"mujoco-3.15.0","episode_id":"reset-003-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"aa3ef94698dc3ddd23fe26625af50c7b96649af428c7364a239be651a8773c20","seed":311003,"session_id":"session-311003","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311003-nominal"},"recording_elapsed_seconds":0.8993888329714537,"sha256":"6987a0db09d9c4b654511ce5e28f099e87284184a76260679b224e763d6572d0","source_recording_id":"recording-311003-nominal","status":"completed"},{"bytes":6024502,"case":"open-jaw","episode_id":"reset-003-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-003-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"9143c18b47087e3807a1bf6043e94748470dc53c96891db49166da1d3a8bef0c","engine":"mujoco-3.15.0","episode_id":"reset-003-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"aa3ef94698dc3ddd23fe26625af50c7b96649af428c7364a239be651a8773c20","seed":311003,"session_id":"session-311003","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311003-open-jaw"},"recording_elapsed_seconds":1.2608026249799877,"sha256":"17191e2a1b18baa594b91f87b505feece360ec32233e124ec0981139d23e75e9","source_recording_id":"recording-311003-open-jaw","status":"completed"},{"bytes":5404990,"case":"release-after-lift","episode_id":"reset-003-release-after-lift","error":null,"outcome":{"control_steps":794,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.87599999999664},"path":"tapes/reset-003-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"9143c18b47087e3807a1bf6043e94748470dc53c96891db49166da1d3a8bef0c","engine":"mujoco-3.15.0","episode_id":"reset-003-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"aa3ef94698dc3ddd23fe26625af50c7b96649af428c7364a239be651a8773c20","seed":311003,"session_id":"session-311003","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311003-release-after-lift"},"recording_elapsed_seconds":1.2027311669662595,"sha256":"42a2cb8ad32c11f9ddeb551c1eb848d8b5e1863bb9929a47f2fb5c9370b27c4a","source_recording_id":"recording-311003-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.7545454545454545,"object_half_size":[0.02,0.02,0.02],"object_mass":0.03772727272727273,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"9143c18b47087e3807a1bf6043e94748470dc53c96891db49166da1d3a8bef0c","group_id":"reset-003","reset_lineage_id":"native-reset-311003","seed":311003,"session_id":"session-311003","split":"train"},{"attempts":[{"bytes":4011544,"case":"nominal","episode_id":"reset-004-nominal","error":null,"outcome":{"control_steps":590,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.793999999998903},"path":"tapes/reset-004-nominal.json","provenance":{"case":"nominal","configuration_sha256":"1160182c60173800bd24a1915885ec51df280af77aafba3370542344b3c2a9bc","engine":"mujoco-3.15.0","episode_id":"reset-004-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"9a6a332b1e518a049567d2f405a483cd5ac8f19847cd6f1a87aee2649d3f397a","seed":311004,"session_id":"session-311004","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311004-nominal"},"recording_elapsed_seconds":0.9105272910092026,"sha256":"8eb88237fc8cd736d303381da800b09ae0c2d05ac72a724c7f43a11421df91ce","source_recording_id":"recording-311004-nominal","status":"completed"},{"bytes":6029913,"case":"open-jaw","episode_id":"reset-004-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-004-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"1160182c60173800bd24a1915885ec51df280af77aafba3370542344b3c2a9bc","engine":"mujoco-3.15.0","episode_id":"reset-004-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"9a6a332b1e518a049567d2f405a483cd5ac8f19847cd6f1a87aee2649d3f397a","seed":311004,"session_id":"session-311004","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311004-open-jaw"},"recording_elapsed_seconds":1.2703030840493739,"sha256":"179de1f97df949dc251950b490bbfcdeb36c2b87eb89e78a5a04880626ead245","source_recording_id":"recording-311004-open-jaw","status":"completed"},{"bytes":5409678,"case":"release-after-lift","episode_id":"reset-004-release-after-lift","error":null,"outcome":{"control_steps":795,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.89399999999663},"path":"tapes/reset-004-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"1160182c60173800bd24a1915885ec51df280af77aafba3370542344b3c2a9bc","engine":"mujoco-3.15.0","episode_id":"reset-004-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"9a6a332b1e518a049567d2f405a483cd5ac8f19847cd6f1a87aee2649d3f397a","seed":311004,"session_id":"session-311004","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311004-release-after-lift"},"recording_elapsed_seconds":1.2041952088475227,"sha256":"ac5035f5f3857a1df2eeb25174347fb8e7b9a4703d9b964f7110fb8496719905","source_recording_id":"recording-311004-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.7727272727272727,"object_half_size":[0.02,0.02,0.02],"object_mass":0.03863636363636364,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"1160182c60173800bd24a1915885ec51df280af77aafba3370542344b3c2a9bc","group_id":"reset-004","reset_lineage_id":"native-reset-311004","seed":311004,"session_id":"session-311004","split":"train"},{"attempts":[{"bytes":3971961,"case":"nominal","episode_id":"reset-005-nominal","error":null,"outcome":{"control_steps":584,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.675999999998968},"path":"tapes/reset-005-nominal.json","provenance":{"case":"nominal","configuration_sha256":"eb7c5f3d168babeb7610a53cc9abb0809a53dd4197c76a9b4fa577eb64b64e91","engine":"mujoco-3.15.0","episode_id":"reset-005-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"981124b1948e89670f41a284674f285912229aef1626e524a7b6769b5ff793cd","seed":311005,"session_id":"session-311005","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311005-nominal"},"recording_elapsed_seconds":0.9100652921479195,"sha256":"d8bcc607869c1be91911974808210351cfff17e39f6b14848267e2b49758be2d","source_recording_id":"recording-311005-nominal","status":"completed"},{"bytes":6025266,"case":"open-jaw","episode_id":"reset-005-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-005-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"eb7c5f3d168babeb7610a53cc9abb0809a53dd4197c76a9b4fa577eb64b64e91","engine":"mujoco-3.15.0","episode_id":"reset-005-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"981124b1948e89670f41a284674f285912229aef1626e524a7b6769b5ff793cd","seed":311005,"session_id":"session-311005","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311005-open-jaw"},"recording_elapsed_seconds":1.2651432079728693,"sha256":"d32e9e45337caadd9e4b6554298311f3d3b8bd4a471fae5ce00ebba2149d64b3","source_recording_id":"recording-311005-open-jaw","status":"completed"},{"bytes":5378550,"case":"release-after-lift","episode_id":"reset-005-release-after-lift","error":null,"outcome":{"control_steps":790,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.795999999996685},"path":"tapes/reset-005-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"eb7c5f3d168babeb7610a53cc9abb0809a53dd4197c76a9b4fa577eb64b64e91","engine":"mujoco-3.15.0","episode_id":"reset-005-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"981124b1948e89670f41a284674f285912229aef1626e524a7b6769b5ff793cd","seed":311005,"session_id":"session-311005","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311005-release-after-lift"},"recording_elapsed_seconds":1.2388457080814987,"sha256":"a50c4540a275ccbc94e10db3f101ee6255d01b0abb508c1197873a266b0fade5","source_recording_id":"recording-311005-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.7909090909090909,"object_half_size":[0.02,0.02,0.02],"object_mass":0.03954545454545455,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"eb7c5f3d168babeb7610a53cc9abb0809a53dd4197c76a9b4fa577eb64b64e91","group_id":"reset-005","reset_lineage_id":"native-reset-311005","seed":311005,"session_id":"session-311005","split":"train"},{"attempts":[{"bytes":3992747,"case":"nominal","episode_id":"reset-006-nominal","error":null,"outcome":{"control_steps":586,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.715999999998946},"path":"tapes/reset-006-nominal.json","provenance":{"case":"nominal","configuration_sha256":"10df99a68dae9f3f88c7f68723fcccf892f23a6cf267a188044bbb2e9590a4f5","engine":"mujoco-3.15.0","episode_id":"reset-006-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"364310942727e3c26f4243ba7605d764ceef7da3977dd310e55d9787f788afea","seed":311006,"session_id":"session-311006","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311006-nominal"},"recording_elapsed_seconds":0.9010972499381751,"sha256":"1794570e615a0aeb2f219f7375e120dd6d53ee629e739e22c462d6c19efe3dff","source_recording_id":"recording-311006-nominal","status":"completed"},{"bytes":6051719,"case":"open-jaw","episode_id":"reset-006-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-006-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"10df99a68dae9f3f88c7f68723fcccf892f23a6cf267a188044bbb2e9590a4f5","engine":"mujoco-3.15.0","episode_id":"reset-006-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"364310942727e3c26f4243ba7605d764ceef7da3977dd310e55d9787f788afea","seed":311006,"session_id":"session-311006","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311006-open-jaw"},"recording_elapsed_seconds":1.278741457965225,"sha256":"1063dd31fda3922b29438ab8c5ab824f94ae52b6e5e9f5404c561171a07a175b","source_recording_id":"recording-311006-open-jaw","status":"completed"},{"bytes":5399092,"case":"release-after-lift","episode_id":"reset-006-release-after-lift","error":null,"outcome":{"control_steps":792,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.835999999996663},"path":"tapes/reset-006-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"10df99a68dae9f3f88c7f68723fcccf892f23a6cf267a188044bbb2e9590a4f5","engine":"mujoco-3.15.0","episode_id":"reset-006-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"364310942727e3c26f4243ba7605d764ceef7da3977dd310e55d9787f788afea","seed":311006,"session_id":"session-311006","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311006-release-after-lift"},"recording_elapsed_seconds":1.2412830418907106,"sha256":"13b18f7285c77558c29d5762d28917737dec73bcadde655b0e6c1935fc8da0d0","source_recording_id":"recording-311006-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.8090909090909091,"object_half_size":[0.02,0.02,0.02],"object_mass":0.04045454545454546,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"10df99a68dae9f3f88c7f68723fcccf892f23a6cf267a188044bbb2e9590a4f5","group_id":"reset-006","reset_lineage_id":"native-reset-311006","seed":311006,"session_id":"session-311006","split":"train"},{"attempts":[{"bytes":4003645,"case":"nominal","episode_id":"reset-007-nominal","error":null,"outcome":{"control_steps":588,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.753999999998925},"path":"tapes/reset-007-nominal.json","provenance":{"case":"nominal","configuration_sha256":"988917e73f1a168937ebde9eba7a0624241ac3725932c52fa02d6474dc9d4040","engine":"mujoco-3.15.0","episode_id":"reset-007-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"d5576282278d7c3147c90499b78802119c975740d8997d5a525b5f00a2e8afd2","seed":311007,"session_id":"session-311007","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311007-nominal"},"recording_elapsed_seconds":0.9346421668305993,"sha256":"5861f0f3ba283f51990ef0302c7bac21c1aee9c273b37d1b4dc3a08652589a55","source_recording_id":"recording-311007-nominal","status":"completed"},{"bytes":6058072,"case":"open-jaw","episode_id":"reset-007-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-007-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"988917e73f1a168937ebde9eba7a0624241ac3725932c52fa02d6474dc9d4040","engine":"mujoco-3.15.0","episode_id":"reset-007-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"d5576282278d7c3147c90499b78802119c975740d8997d5a525b5f00a2e8afd2","seed":311007,"session_id":"session-311007","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311007-open-jaw"},"recording_elapsed_seconds":1.2638327919412404,"sha256":"a7dcc569db29df4197c651f66ec8a8eab2195e5c1aad5967a923e4e9e69a54a7","source_recording_id":"recording-311007-open-jaw","status":"completed"},{"bytes":5408408,"case":"release-after-lift","episode_id":"reset-007-release-after-lift","error":null,"outcome":{"control_steps":794,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.873999999996641},"path":"tapes/reset-007-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"988917e73f1a168937ebde9eba7a0624241ac3725932c52fa02d6474dc9d4040","engine":"mujoco-3.15.0","episode_id":"reset-007-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"d5576282278d7c3147c90499b78802119c975740d8997d5a525b5f00a2e8afd2","seed":311007,"session_id":"session-311007","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311007-release-after-lift"},"recording_elapsed_seconds":1.2749216670636088,"sha256":"71dd9913d0ac2d036b8ec008c710d958ec6438d430b22720f1db67aa0bb77342","source_recording_id":"recording-311007-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.8272727272727273,"object_half_size":[0.02,0.02,0.02],"object_mass":0.041363636363636366,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"988917e73f1a168937ebde9eba7a0624241ac3725932c52fa02d6474dc9d4040","group_id":"reset-007","reset_lineage_id":"native-reset-311007","seed":311007,"session_id":"session-311007","split":"train"},{"attempts":[{"bytes":3999522,"case":"nominal","episode_id":"reset-008-nominal","error":null,"outcome":{"control_steps":587,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.734999999998935},"path":"tapes/reset-008-nominal.json","provenance":{"case":"nominal","configuration_sha256":"403c9ff95151f3e78d1e921471b543dac30fb128048a623a878d67fd8ad0c504","engine":"mujoco-3.15.0","episode_id":"reset-008-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"525ce67456ed32038cb2a2e48a39a27c0a9212861fdabf8b4085ab0851960c10","seed":311008,"session_id":"session-311008","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311008-nominal"},"recording_elapsed_seconds":0.9376537918578833,"sha256":"e38baf531fff25447ab3eefa4bc50ceeb62197a7ecdd498cdc418501eb2d139e","source_recording_id":"recording-311008-nominal","status":"completed"},{"bytes":6057713,"case":"open-jaw","episode_id":"reset-008-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-008-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"403c9ff95151f3e78d1e921471b543dac30fb128048a623a878d67fd8ad0c504","engine":"mujoco-3.15.0","episode_id":"reset-008-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"525ce67456ed32038cb2a2e48a39a27c0a9212861fdabf8b4085ab0851960c10","seed":311008,"session_id":"session-311008","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311008-open-jaw"},"recording_elapsed_seconds":1.262716457946226,"sha256":"14542d6ca731322640822f163ae1a77a717a3768cb79751cc412cc6258288b18","source_recording_id":"recording-311008-open-jaw","status":"completed"},{"bytes":5398745,"case":"release-after-lift","episode_id":"reset-008-release-after-lift","error":null,"outcome":{"control_steps":792,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.834999999996663},"path":"tapes/reset-008-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"403c9ff95151f3e78d1e921471b543dac30fb128048a623a878d67fd8ad0c504","engine":"mujoco-3.15.0","episode_id":"reset-008-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"525ce67456ed32038cb2a2e48a39a27c0a9212861fdabf8b4085ab0851960c10","seed":311008,"session_id":"session-311008","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311008-release-after-lift"},"recording_elapsed_seconds":1.2447435420472175,"sha256":"51241d48d4c754e818529f6b8341e5f5b99e82a6e6c2b04a7b72b0d3df1fe472","source_recording_id":"recording-311008-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.8454545454545455,"object_half_size":[0.02,0.02,0.02],"object_mass":0.042272727272727274,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"403c9ff95151f3e78d1e921471b543dac30fb128048a623a878d67fd8ad0c504","group_id":"reset-008","reset_lineage_id":"native-reset-311008","seed":311008,"session_id":"session-311008","split":"development"},{"attempts":[{"bytes":3992744,"case":"nominal","episode_id":"reset-009-nominal","error":null,"outcome":{"control_steps":587,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.735999999998935},"path":"tapes/reset-009-nominal.json","provenance":{"case":"nominal","configuration_sha256":"c9599f1b520ba6dbe0abc62501e7a7a4b7eb90585533cc736af724bd93fc738c","engine":"mujoco-3.15.0","episode_id":"reset-009-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"ad6281b7791d27efc070cfb761cf0a96b22f9577586eef457a4fd25e8726ffe4","seed":311009,"session_id":"session-311009","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311009-nominal"},"recording_elapsed_seconds":0.9133894168771803,"sha256":"66073f1a0d01e7dfe9199e547308e07022d93b42e7a1b0110ad124d0b0e407a3","source_recording_id":"recording-311009-nominal","status":"completed"},{"bytes":6027695,"case":"open-jaw","episode_id":"reset-009-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-009-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"c9599f1b520ba6dbe0abc62501e7a7a4b7eb90585533cc736af724bd93fc738c","engine":"mujoco-3.15.0","episode_id":"reset-009-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"ad6281b7791d27efc070cfb761cf0a96b22f9577586eef457a4fd25e8726ffe4","seed":311009,"session_id":"session-311009","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311009-open-jaw"},"recording_elapsed_seconds":1.2751613750588149,"sha256":"5b82cf51e811f4c3b16f2550dae78270c71aa8cba2d0c5acd6a28f35ff75fe0e","source_recording_id":"recording-311009-open-jaw","status":"completed"},{"bytes":5398357,"case":"release-after-lift","episode_id":"reset-009-release-after-lift","error":null,"outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.855999999996651},"path":"tapes/reset-009-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"c9599f1b520ba6dbe0abc62501e7a7a4b7eb90585533cc736af724bd93fc738c","engine":"mujoco-3.15.0","episode_id":"reset-009-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"ad6281b7791d27efc070cfb761cf0a96b22f9577586eef457a4fd25e8726ffe4","seed":311009,"session_id":"session-311009","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311009-release-after-lift"},"recording_elapsed_seconds":1.231245334027335,"sha256":"8012c87829bad7b8e106d41b2d43767833bcf27fb7ffdc9fa9b93520e3441e87","source_recording_id":"recording-311009-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.8636363636363635,"object_half_size":[0.02,0.02,0.02],"object_mass":0.04318181818181818,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"c9599f1b520ba6dbe0abc62501e7a7a4b7eb90585533cc736af724bd93fc738c","group_id":"reset-009","reset_lineage_id":"native-reset-311009","seed":311009,"session_id":"session-311009","split":"development"},{"attempts":[{"bytes":3978851,"case":"nominal","episode_id":"reset-010-nominal","error":null,"outcome":{"control_steps":584,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.676999999998968},"path":"tapes/reset-010-nominal.json","provenance":{"case":"nominal","configuration_sha256":"67f2d0eb5b9b0fe3805647bd62494c4020cf8cac11fd7e5eeb79ada55523307e","engine":"mujoco-3.15.0","episode_id":"reset-010-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"d2c21afebdbdec95538055f3e9b14803841043a1d7cb900e50e972c7fb7b48c9","seed":311010,"session_id":"session-311010","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311010-nominal"},"recording_elapsed_seconds":0.9260092920158058,"sha256":"c1510a2f6e20bb2c762fb297fe9588ba683214f9a653531fad1ed6dc193fde71","source_recording_id":"recording-311010-nominal","status":"completed"},{"bytes":6056759,"case":"open-jaw","episode_id":"reset-010-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-010-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"67f2d0eb5b9b0fe3805647bd62494c4020cf8cac11fd7e5eeb79ada55523307e","engine":"mujoco-3.15.0","episode_id":"reset-010-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"d2c21afebdbdec95538055f3e9b14803841043a1d7cb900e50e972c7fb7b48c9","seed":311010,"session_id":"session-311010","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311010-open-jaw"},"recording_elapsed_seconds":1.2951547501143068,"sha256":"06a0fba35a56ba532f47897b03c6e8256d831d642d415ba4589b06acd87e30ec","source_recording_id":"recording-311010-open-jaw","status":"completed"},{"bytes":5377928,"case":"release-after-lift","episode_id":"reset-010-release-after-lift","error":null,"outcome":{"control_steps":789,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.776999999996695},"path":"tapes/reset-010-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"67f2d0eb5b9b0fe3805647bd62494c4020cf8cac11fd7e5eeb79ada55523307e","engine":"mujoco-3.15.0","episode_id":"reset-010-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"d2c21afebdbdec95538055f3e9b14803841043a1d7cb900e50e972c7fb7b48c9","seed":311010,"session_id":"session-311010","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311010-release-after-lift"},"recording_elapsed_seconds":1.2305635001976043,"sha256":"4907d07843f746bbd1d49614ea5b6418687f5dc5896366d105c76353b51df88f","source_recording_id":"recording-311010-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.8818181818181818,"object_half_size":[0.02,0.02,0.02],"object_mass":0.0440909090909091,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"67f2d0eb5b9b0fe3805647bd62494c4020cf8cac11fd7e5eeb79ada55523307e","group_id":"reset-010","reset_lineage_id":"native-reset-311010","seed":311010,"session_id":"session-311010","split":"test"},{"attempts":[{"bytes":4005540,"case":"nominal","episode_id":"reset-011-nominal","error":null,"outcome":{"control_steps":589,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.773999999998914},"path":"tapes/reset-011-nominal.json","provenance":{"case":"nominal","configuration_sha256":"614396595ad1fc049845f7ae4fcbfa44a2c5cee7b0bc51a1e2e0db38bafe1e33","engine":"mujoco-3.15.0","episode_id":"reset-011-nominal","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"2cc9ee103c443a7b75de33f6553ae1359844962c64f8b6d017a133c51f854679","seed":311011,"session_id":"session-311011","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311011-nominal"},"recording_elapsed_seconds":0.9308858329895884,"sha256":"3ba1498cb3159f7d3d766f73807c3702ccf377e3ea336f9b0889f2e4e15e9261","source_recording_id":"recording-311011-nominal","status":"completed"},{"bytes":6038757,"case":"open-jaw","episode_id":"reset-011-open-jaw","error":null,"outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"path":"tapes/reset-011-open-jaw.json","provenance":{"case":"open-jaw","configuration_sha256":"614396595ad1fc049845f7ae4fcbfa44a2c5cee7b0bc51a1e2e0db38bafe1e33","engine":"mujoco-3.15.0","episode_id":"reset-011-open-jaw","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"2cc9ee103c443a7b75de33f6553ae1359844962c64f8b6d017a133c51f854679","seed":311011,"session_id":"session-311011","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311011-open-jaw"},"recording_elapsed_seconds":1.2852909581270069,"sha256":"cac131cb2149aaf1958c652ec98a64e633a35340608b34524527c44ee2bcd1ed","source_recording_id":"recording-311011-open-jaw","status":"completed"},{"bytes":5410990,"case":"release-after-lift","episode_id":"reset-011-release-after-lift","error":null,"outcome":{"control_steps":795,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.89399999999663},"path":"tapes/reset-011-release-after-lift.json","provenance":{"case":"release-after-lift","configuration_sha256":"614396595ad1fc049845f7ae4fcbfa44a2c5cee7b0bc51a1e2e0db38bafe1e33","engine":"mujoco-3.15.0","episode_id":"reset-011-release-after-lift","policy_origin":"authored controller","policy_revision":"dimension-aware-contact-rates-recovery-round2","scene_sha256":"2cc9ee103c443a7b75de33f6553ae1359844962c64f8b6d017a133c51f854679","seed":311011,"session_id":"session-311011","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"source_kind":"native-simulation-telemetry","source_recording_id":"recording-311011-release-after-lift"},"recording_elapsed_seconds":1.2508349160198122,"sha256":"21d25db5acf787e133bc61a95845bbae9c65f11b9fdace09e93a1329fcbf7a06","source_recording_id":"recording-311011-release-after-lift","status":"completed"}],"configuration":{"control_dt":0.02,"dwell_seconds":0.25,"gravity":9.81,"horizon":18.0,"jaw_force_limit":15.0,"jaw_width_rate_limit":0.16,"joint_target_rate_limit":3.0,"joint_torque_limit":40.0,"object_friction":0.8999999999999999,"object_half_size":[0.02,0.02,0.02],"object_mass":0.045000000000000005,"object_position":[0.42,-0.04,0.31],"position_tolerance":0.025,"scene_jitter":0.006,"settle_time":0.4,"table_height":0.29,"target_position":[0.54,0.1,0.31],"timestep":0.001},"configuration_sha256":"614396595ad1fc049845f7ae4fcbfa44a2c5cee7b0bc51a1e2e0db38bafe1e33","group_id":"reset-011","reset_lineage_id":"native-reset-311011","seed":311011,"session_id":"session-311011","split":"test"}],"kind":"dvidia.actuator-transition-corpus.v1","physical_robot_ready":false,"protocol":{"cases":["nominal","open-jaw","release-after-lift"],"revision":"grouped-native-actuator-20ms-v1","source_files_sha256":{"arm_env.py":"e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00","arm_policy.py":"d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3","episode_tape.py":"abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"},"split_counts":{"development":2,"test":2,"train":8},"split_unit":"predeclared reset group; all sibling interventions stay together","variation":"seeded scene jitter plus predeclared bounded group mass/friction; neither is a model feature"},"schema_version":1,"scope":"simulation-only","shoe_skill_ready":false,"status":"completed"}
diff --git a/benchmarks/actuator-pilot-20261011/freeze.json b/benchmarks/actuator-pilot-20261011/freeze.json
new file mode 100644
index 0000000..3636ffd
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/freeze.json
@@ -0,0 +1 @@
+{"corpus_manifest_sha256":"c529bf61405d3d8d46a6a96ebf56a871436a991a6e11abb9a264c9d880ff6624","development_sources":["e38baf531fff25447ab3eefa4bc50ceeb62197a7ecdd498cdc418501eb2d139e","14542d6ca731322640822f163ae1a77a717a3768cb79751cc412cc6258288b18","51241d48d4c754e818529f6b8341e5f5b99e82a6e6c2b04a7b72b0d3df1fe472","66073f1a0d01e7dfe9199e547308e07022d93b42e7a1b0110ad124d0b0e407a3","5b82cf51e811f4c3b16f2550dae78270c71aa8cba2d0c5acd6a28f35ff75fe0e","8012c87829bad7b8e106d41b2d43767833bcf27fb7ffdc9fa9b93520e3441e87"],"elapsed_seconds_at_freeze":14.39856925001368,"held_out_transition_targets_read":false,"kind":"dvidia.actuator-transition-freeze.v1","models":{"action":{"bytes":6438,"file":"action.model.json","sha256":"611880e00116a6c52d474fe2d08c446bc5475564e4f3d550dd2bb88bb2631045"},"state":{"bytes":5007,"file":"state.model.json","sha256":"0c1aca45dcf66816f587311cf413a298e9076c9f282cd3e4aa20f68510814fe8"}},"schema_version":1,"selection":{"action":{"alpha":0.0001,"trials":[{"alpha":0.0001,"development_score":0.34359880923257363,"metrics":{"jaw_mae_m":0.00011408709850354386,"jaw_max_abs_error_m":0.0034112555979426496,"jaw_rmse_m":0.000340067832374854,"joint_mae_rad":7.1141354587507735e-06,"joint_max_abs_error_rad":0.0012005909356011786,"joint_rmse_rad":3.530976857719615e-05,"rows":4555}},{"alpha":0.01,"development_score":0.3488523445371724,"metrics":{"jaw_mae_m":0.00011030944787074467,"jaw_max_abs_error_m":0.003269044603257664,"jaw_rmse_m":0.000345160387948776,"joint_mae_rad":8.181816234241458e-06,"joint_max_abs_error_rad":0.0012538429794856495,"joint_rmse_rad":3.691956588396459e-05,"rows":4555}},{"alpha":1.0,"development_score":0.37446811496940613,"metrics":{"jaw_mae_m":0.0001131313932677029,"jaw_max_abs_error_m":0.0031182654411507395,"jaw_rmse_m":0.0003555391951989955,"joint_mae_rad":4.7180629138626286e-05,"joint_max_abs_error_rad":0.004918626541840294,"joint_rmse_rad":0.00018928919770410597,"rows":4555}},{"alpha":100.0,"development_score":0.40049833398550805,"metrics":{"jaw_mae_m":0.00011524902086540773,"jaw_max_abs_error_m":0.0028229383383668533,"jaw_rmse_m":0.0003606345536169485,"joint_mae_rad":8.85840301824777e-05,"joint_max_abs_error_rad":0.012589703024714939,"joint_rmse_rad":0.00039863780368559564,"rows":4555}}]},"state":{"alpha":0.0001,"trials":[{"alpha":0.0001,"development_score":0.5927615229709541,"metrics":{"jaw_mae_m":0.00018672971727035493,"jaw_max_abs_error_m":0.0035440017197499595,"jaw_rmse_m":0.0005518057723999874,"joint_mae_rad":8.045498027119069e-05,"joint_max_abs_error_rad":0.012901945226615082,"joint_rmse_rad":0.0004095575057096664,"rows":4555}},{"alpha":0.01,"development_score":0.5931159026759466,"metrics":{"jaw_mae_m":0.00018541009658040093,"jaw_max_abs_error_m":0.0035009574428947242,"jaw_rmse_m":0.000551959882687009,"joint_mae_rad":8.210262251786841e-05,"joint_max_abs_error_rad":0.012939254286617633,"joint_rmse_rad":0.0004115601998893769,"rows":4555}},{"alpha":1.0,"development_score":0.5939541313602504,"metrics":{"jaw_mae_m":0.00018440188004749438,"jaw_max_abs_error_m":0.003446910566375889,"jaw_rmse_m":0.0005527103992896407,"joint_mae_rad":8.404463578661915e-05,"joint_max_abs_error_rad":0.01294704832590305,"joint_rmse_rad":0.0004124373207060971,"rows":4555}},{"alpha":100.0,"development_score":0.5946172957204006,"metrics":{"jaw_mae_m":0.0001823020328579418,"jaw_max_abs_error_m":0.0033305991007897048,"jaw_rmse_m":0.0005528614861221444,"joint_mae_rad":8.993230981100541e-05,"joint_max_abs_error_rad":0.012868733472039574,"joint_rmse_rad":0.0004175580959825623,"rows":4555}}]}},"test_sources":["c1510a2f6e20bb2c762fb297fe9588ba683214f9a653531fad1ed6dc193fde71","06a0fba35a56ba532f47897b03c6e8256d831d642d415ba4589b06acd87e30ec","4907d07843f746bbd1d49614ea5b6418687f5dc5896366d105c76353b51df88f","3ba1498cb3159f7d3d766f73807c3702ccf377e3ea336f9b0889f2e4e15e9261","cac131cb2149aaf1958c652ec98a64e633a35340608b34524527c44ee2bcd1ed","21d25db5acf787e133bc61a95845bbae9c65f11b9fdace09e93a1329fcbf7a06"],"training_mean_delta":[0.0009610293150821221,0.00028681519387991386,1.1362007785179132e-05,1.1868488215456264e-10,-0.0002929461388636014,-0.0009610293636950312,-7.577548372403187e-09],"training_sources":["2d9633992eeb1f6c6cc7763d6dffd8b14f0d2eac2c10d895cc0979a22f56c608","a20f122ff1670959b22ec9d5dfbb4729d583c8899e9a2ba4a1eb0d644b8432ab","89741ba25f14fa62877940be2325e72e672835922dc2377258dc77af49464315","bd0009e7e6955b2314080e6739951ebd1d965a0276e8a0e9e8c81ca707c56bbe","896e1fe59923cc89e1bf86437c527e0a8e2faca9d1c2d46471af66e2f83d8841","6b6f7dac245b8a538037f41981b5b16fec6e0b193b0eb3c2180e7a25c50b1000","aeace5b052485035cfa2e12b9825e99ca2acfc3ce1a675dca8fb25353437472a","cc8ca7c9924faf4e9e608b05fdbc2b4f3e999d25cad2aee883c9863a57640d3f","06ed87cb5511bd0f4fe5e6a8d46eedf8c78aa150a8a3c58d0ffb19b932521015","6987a0db09d9c4b654511ce5e28f099e87284184a76260679b224e763d6572d0","17191e2a1b18baa594b91f87b505feece360ec32233e124ec0981139d23e75e9","42a2cb8ad32c11f9ddeb551c1eb848d8b5e1863bb9929a47f2fb5c9370b27c4a","8eb88237fc8cd736d303381da800b09ae0c2d05ac72a724c7f43a11421df91ce","179de1f97df949dc251950b490bbfcdeb36c2b87eb89e78a5a04880626ead245","ac5035f5f3857a1df2eeb25174347fb8e7b9a4703d9b964f7110fb8496719905","d8bcc607869c1be91911974808210351cfff17e39f6b14848267e2b49758be2d","d32e9e45337caadd9e4b6554298311f3d3b8bd4a471fae5ce00ebba2149d64b3","a50c4540a275ccbc94e10db3f101ee6255d01b0abb508c1197873a266b0fade5","1794570e615a0aeb2f219f7375e120dd6d53ee629e739e22c462d6c19efe3dff","1063dd31fda3922b29438ab8c5ab824f94ae52b6e5e9f5404c561171a07a175b","13b18f7285c77558c29d5762d28917737dec73bcadde655b0e6c1935fc8da0d0","5861f0f3ba283f51990ef0302c7bac21c1aee9c273b37d1b4dc3a08652589a55","a7dcc569db29df4197c651f66ec8a8eab2195e5c1aad5967a923e4e9e69a54a7","71dd9913d0ac2d036b8ec008c710d958ec6438d430b22720f1db67aa0bb77342"]}
diff --git a/benchmarks/actuator-pilot-20261011/integrity.json b/benchmarks/actuator-pilot-20261011/integrity.json
new file mode 100644
index 0000000..184ac0a
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/integrity.json
@@ -0,0 +1 @@
+{"accuracy_verified":false,"artifacts":{"action.model.json":{"bytes":6438,"file":"action.model.json","sha256":"611880e00116a6c52d474fe2d08c446bc5475564e4f3d550dd2bb88bb2631045"},"corpus-manifest.json":{"bytes":57320,"file":"corpus-manifest.json","sha256":"c529bf61405d3d8d46a6a96ebf56a871436a991a6e11abb9a264c9d880ff6624"},"freeze.json":{"bytes":5786,"file":"freeze.json","sha256":"98b2d59affd3cc9a2098594f3d56f1502df9d9fb0f2cbc0bcc4450e65b25205c"},"report.json":{"bytes":43176,"file":"report.json","sha256":"aa605b7e1a09c6a71d8720b5458e2cd2559be5b85dcff0806bbbb7a64742b6d5"},"state.model.json":{"bytes":5007,"file":"state.model.json","sha256":"0c1aca45dcf66816f587311cf413a298e9076c9f282cd3e4aa20f68510814fe8"}},"authenticity_verified":false,"kind":"dvidia.actuator-transition-integrity.v1","schema_version":1}
diff --git a/benchmarks/actuator-pilot-20261011/report.json b/benchmarks/actuator-pilot-20261011/report.json
new file mode 100644
index 0000000..72a46ea
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/report.json
@@ -0,0 +1 @@
+{"artifact_bytes":11445,"attempts":[{"attempted_steps":588,"case":"nominal","episode_id":"reset-000-nominal","exclusions":[{"episode_id":"reset-000-nominal","index":587,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-000","recording_error":null,"recording_status":"completed","rows":587,"simulation_outcome":{"control_steps":588,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.755999999998924},"source_bytes":3997721,"source_sha256":"2d9633992eeb1f6c6cc7763d6dffd8b14f0d2eac2c10d895cc0979a22f56c608","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-000-open-jaw","exclusions":[],"group_id":"reset-000","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6028529,"source_sha256":"a20f122ff1670959b22ec9d5dfbb4729d583c8899e9a2ba4a1eb0d644b8432ab","split":"train","validation_status":"validated"},{"attempted_steps":793,"case":"release-after-lift","episode_id":"reset-000-release-after-lift","exclusions":[{"episode_id":"reset-000-release-after-lift","index":792,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-000","recording_error":null,"recording_status":"completed","rows":792,"simulation_outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.855999999996651},"source_bytes":5395144,"source_sha256":"89741ba25f14fa62877940be2325e72e672835922dc2377258dc77af49464315","split":"train","validation_status":"validated"},{"attempted_steps":587,"case":"nominal","episode_id":"reset-001-nominal","exclusions":[{"episode_id":"reset-001-nominal","index":586,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-001","recording_error":null,"recording_status":"completed","rows":586,"simulation_outcome":{"control_steps":587,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.735999999998935},"source_bytes":3990695,"source_sha256":"bd0009e7e6955b2314080e6739951ebd1d965a0276e8a0e9e8c81ca707c56bbe","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-001-open-jaw","exclusions":[],"group_id":"reset-001","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6032277,"source_sha256":"896e1fe59923cc89e1bf86437c527e0a8e2faca9d1c2d46471af66e2f83d8841","split":"train","validation_status":"validated"},{"attempted_steps":793,"case":"release-after-lift","episode_id":"reset-001-release-after-lift","exclusions":[{"episode_id":"reset-001-release-after-lift","index":792,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-001","recording_error":null,"recording_status":"completed","rows":792,"simulation_outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.855999999996651},"source_bytes":5396900,"source_sha256":"6b6f7dac245b8a538037f41981b5b16fec6e0b193b0eb3c2180e7a25c50b1000","split":"train","validation_status":"validated"},{"attempted_steps":588,"case":"nominal","episode_id":"reset-002-nominal","exclusions":[{"episode_id":"reset-002-nominal","index":587,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-002","recording_error":null,"recording_status":"completed","rows":587,"simulation_outcome":{"control_steps":588,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.754999999998924},"source_bytes":4005185,"source_sha256":"aeace5b052485035cfa2e12b9825e99ca2acfc3ce1a675dca8fb25353437472a","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-002-open-jaw","exclusions":[],"group_id":"reset-002","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6051492,"source_sha256":"cc8ca7c9924faf4e9e608b05fdbc2b4f3e999d25cad2aee883c9863a57640d3f","split":"train","validation_status":"validated"},{"attempted_steps":793,"case":"release-after-lift","episode_id":"reset-002-release-after-lift","exclusions":[{"episode_id":"reset-002-release-after-lift","index":792,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-002","recording_error":null,"recording_status":"completed","rows":792,"simulation_outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.854999999996652},"source_bytes":5403671,"source_sha256":"06ed87cb5511bd0f4fe5e6a8d46eedf8c78aa150a8a3c58d0ffb19b932521015","split":"train","validation_status":"validated"},{"attempted_steps":589,"case":"nominal","episode_id":"reset-003-nominal","exclusions":[{"episode_id":"reset-003-nominal","index":588,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-003","recording_error":null,"recording_status":"completed","rows":588,"simulation_outcome":{"control_steps":589,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.775999999998913},"source_bytes":4006111,"source_sha256":"6987a0db09d9c4b654511ce5e28f099e87284184a76260679b224e763d6572d0","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-003-open-jaw","exclusions":[],"group_id":"reset-003","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6024502,"source_sha256":"17191e2a1b18baa594b91f87b505feece360ec32233e124ec0981139d23e75e9","split":"train","validation_status":"validated"},{"attempted_steps":794,"case":"release-after-lift","episode_id":"reset-003-release-after-lift","exclusions":[{"episode_id":"reset-003-release-after-lift","index":793,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-003","recording_error":null,"recording_status":"completed","rows":793,"simulation_outcome":{"control_steps":794,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.87599999999664},"source_bytes":5404990,"source_sha256":"42a2cb8ad32c11f9ddeb551c1eb848d8b5e1863bb9929a47f2fb5c9370b27c4a","split":"train","validation_status":"validated"},{"attempted_steps":590,"case":"nominal","episode_id":"reset-004-nominal","exclusions":[{"episode_id":"reset-004-nominal","index":589,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-004","recording_error":null,"recording_status":"completed","rows":589,"simulation_outcome":{"control_steps":590,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.793999999998903},"source_bytes":4011544,"source_sha256":"8eb88237fc8cd736d303381da800b09ae0c2d05ac72a724c7f43a11421df91ce","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-004-open-jaw","exclusions":[],"group_id":"reset-004","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6029913,"source_sha256":"179de1f97df949dc251950b490bbfcdeb36c2b87eb89e78a5a04880626ead245","split":"train","validation_status":"validated"},{"attempted_steps":795,"case":"release-after-lift","episode_id":"reset-004-release-after-lift","exclusions":[{"episode_id":"reset-004-release-after-lift","index":794,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-004","recording_error":null,"recording_status":"completed","rows":794,"simulation_outcome":{"control_steps":795,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.89399999999663},"source_bytes":5409678,"source_sha256":"ac5035f5f3857a1df2eeb25174347fb8e7b9a4703d9b964f7110fb8496719905","split":"train","validation_status":"validated"},{"attempted_steps":584,"case":"nominal","episode_id":"reset-005-nominal","exclusions":[{"episode_id":"reset-005-nominal","index":583,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-005","recording_error":null,"recording_status":"completed","rows":583,"simulation_outcome":{"control_steps":584,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.675999999998968},"source_bytes":3971961,"source_sha256":"d8bcc607869c1be91911974808210351cfff17e39f6b14848267e2b49758be2d","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-005-open-jaw","exclusions":[],"group_id":"reset-005","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6025266,"source_sha256":"d32e9e45337caadd9e4b6554298311f3d3b8bd4a471fae5ce00ebba2149d64b3","split":"train","validation_status":"validated"},{"attempted_steps":790,"case":"release-after-lift","episode_id":"reset-005-release-after-lift","exclusions":[{"episode_id":"reset-005-release-after-lift","index":789,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-005","recording_error":null,"recording_status":"completed","rows":789,"simulation_outcome":{"control_steps":790,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.795999999996685},"source_bytes":5378550,"source_sha256":"a50c4540a275ccbc94e10db3f101ee6255d01b0abb508c1197873a266b0fade5","split":"train","validation_status":"validated"},{"attempted_steps":586,"case":"nominal","episode_id":"reset-006-nominal","exclusions":[{"episode_id":"reset-006-nominal","index":585,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-006","recording_error":null,"recording_status":"completed","rows":585,"simulation_outcome":{"control_steps":586,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.715999999998946},"source_bytes":3992747,"source_sha256":"1794570e615a0aeb2f219f7375e120dd6d53ee629e739e22c462d6c19efe3dff","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-006-open-jaw","exclusions":[],"group_id":"reset-006","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6051719,"source_sha256":"1063dd31fda3922b29438ab8c5ab824f94ae52b6e5e9f5404c561171a07a175b","split":"train","validation_status":"validated"},{"attempted_steps":792,"case":"release-after-lift","episode_id":"reset-006-release-after-lift","exclusions":[{"episode_id":"reset-006-release-after-lift","index":791,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-006","recording_error":null,"recording_status":"completed","rows":791,"simulation_outcome":{"control_steps":792,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.835999999996663},"source_bytes":5399092,"source_sha256":"13b18f7285c77558c29d5762d28917737dec73bcadde655b0e6c1935fc8da0d0","split":"train","validation_status":"validated"},{"attempted_steps":588,"case":"nominal","episode_id":"reset-007-nominal","exclusions":[{"episode_id":"reset-007-nominal","index":587,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-007","recording_error":null,"recording_status":"completed","rows":587,"simulation_outcome":{"control_steps":588,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.753999999998925},"source_bytes":4003645,"source_sha256":"5861f0f3ba283f51990ef0302c7bac21c1aee9c273b37d1b4dc3a08652589a55","split":"train","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-007-open-jaw","exclusions":[],"group_id":"reset-007","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6058072,"source_sha256":"a7dcc569db29df4197c651f66ec8a8eab2195e5c1aad5967a923e4e9e69a54a7","split":"train","validation_status":"validated"},{"attempted_steps":794,"case":"release-after-lift","episode_id":"reset-007-release-after-lift","exclusions":[{"episode_id":"reset-007-release-after-lift","index":793,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-007","recording_error":null,"recording_status":"completed","rows":793,"simulation_outcome":{"control_steps":794,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.873999999996641},"source_bytes":5408408,"source_sha256":"71dd9913d0ac2d036b8ec008c710d958ec6438d430b22720f1db67aa0bb77342","split":"train","validation_status":"validated"},{"attempted_steps":587,"case":"nominal","episode_id":"reset-008-nominal","exclusions":[{"episode_id":"reset-008-nominal","index":586,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-008","recording_error":null,"recording_status":"completed","rows":586,"simulation_outcome":{"control_steps":587,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.734999999998935},"source_bytes":3999522,"source_sha256":"e38baf531fff25447ab3eefa4bc50ceeb62197a7ecdd498cdc418501eb2d139e","split":"development","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-008-open-jaw","exclusions":[],"group_id":"reset-008","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6057713,"source_sha256":"14542d6ca731322640822f163ae1a77a717a3768cb79751cc412cc6258288b18","split":"development","validation_status":"validated"},{"attempted_steps":792,"case":"release-after-lift","episode_id":"reset-008-release-after-lift","exclusions":[{"episode_id":"reset-008-release-after-lift","index":791,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-008","recording_error":null,"recording_status":"completed","rows":791,"simulation_outcome":{"control_steps":792,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.834999999996663},"source_bytes":5398745,"source_sha256":"51241d48d4c754e818529f6b8341e5f5b99e82a6e6c2b04a7b72b0d3df1fe472","split":"development","validation_status":"validated"},{"attempted_steps":587,"case":"nominal","episode_id":"reset-009-nominal","exclusions":[{"episode_id":"reset-009-nominal","index":586,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-009","recording_error":null,"recording_status":"completed","rows":586,"simulation_outcome":{"control_steps":587,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.735999999998935},"source_bytes":3992744,"source_sha256":"66073f1a0d01e7dfe9199e547308e07022d93b42e7a1b0110ad124d0b0e407a3","split":"development","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-009-open-jaw","exclusions":[],"group_id":"reset-009","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6027695,"source_sha256":"5b82cf51e811f4c3b16f2550dae78270c71aa8cba2d0c5acd6a28f35ff75fe0e","split":"development","validation_status":"validated"},{"attempted_steps":793,"case":"release-after-lift","episode_id":"reset-009-release-after-lift","exclusions":[{"episode_id":"reset-009-release-after-lift","index":792,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-009","recording_error":null,"recording_status":"completed","rows":792,"simulation_outcome":{"control_steps":793,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.855999999996651},"source_bytes":5398357,"source_sha256":"8012c87829bad7b8e106d41b2d43767833bcf27fb7ffdc9fa9b93520e3441e87","split":"development","validation_status":"validated"},{"attempted_steps":584,"case":"nominal","episode_id":"reset-010-nominal","exclusions":[{"episode_id":"reset-010-nominal","index":583,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-010","recording_error":null,"recording_status":"completed","rows":583,"simulation_outcome":{"control_steps":584,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.676999999998968},"source_bytes":3978851,"source_sha256":"c1510a2f6e20bb2c762fb297fe9588ba683214f9a653531fad1ed6dc193fde71","split":"test","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-010-open-jaw","exclusions":[],"group_id":"reset-010","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6056759,"source_sha256":"06a0fba35a56ba532f47897b03c6e8256d831d642d415ba4589b06acd87e30ec","split":"test","validation_status":"validated"},{"attempted_steps":789,"case":"release-after-lift","episode_id":"reset-010-release-after-lift","exclusions":[{"episode_id":"reset-010-release-after-lift","index":788,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-010","recording_error":null,"recording_status":"completed","rows":788,"simulation_outcome":{"control_steps":789,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.776999999996695},"source_bytes":5377928,"source_sha256":"4907d07843f746bbd1d49614ea5b6418687f5dc5896366d105c76353b51df88f","split":"test","validation_status":"validated"},{"attempted_steps":589,"case":"nominal","episode_id":"reset-011-nominal","exclusions":[{"episode_id":"reset-011-nominal","index":588,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-011","recording_error":null,"recording_status":"completed","rows":588,"simulation_outcome":{"control_steps":589,"policy_failure_reason":null,"recovery_count":0,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":11.773999999998914},"source_bytes":4005540,"source_sha256":"3ba1498cb3159f7d3d766f73807c3702ccf377e3ea336f9b0889f2e4e15e9261","split":"test","validation_status":"validated"},{"attempted_steps":900,"case":"open-jaw","episode_id":"reset-011-open-jaw","exclusions":[],"group_id":"reset-011","recording_error":null,"recording_status":"completed","rows":900,"simulation_outcome":{"control_steps":900,"policy_failure_reason":"phase_timeout:close","recovery_count":1,"result":{"reason":"horizon","success":false,"terminated":false,"truncated":true,"valid":true},"simulation_seconds":17.999999999999016},"source_bytes":6038757,"source_sha256":"cac131cb2149aaf1958c652ec98a64e633a35340608b34524527c44ee2bcd1ed","split":"test","validation_status":"validated"},{"attempted_steps":795,"case":"release-after-lift","episode_id":"reset-011-release-after-lift","exclusions":[{"episode_id":"reset-011-release-after-lift","index":794,"reason":"incomplete_20ms_boundary"}],"group_id":"reset-011","recording_error":null,"recording_status":"completed","rows":794,"simulation_outcome":{"control_steps":795,"policy_failure_reason":null,"recovery_count":1,"result":{"reason":"success","success":true,"terminated":true,"truncated":false,"valid":true},"simulation_seconds":15.89399999999663},"source_bytes":5410990,"source_sha256":"21d25db5acf787e133bc61a95845bbae9c65f11b9fdace09e93a1329fcbf7a06","split":"test","validation_status":"validated"}],"corpus_manifest_sha256":"c529bf61405d3d8d46a6a96ebf56a871436a991a6e11abb9a264c9d880ff6624","counts":{"attempted_steps":{"development":4559,"test":4557,"train":18244},"attempts":36,"case_attempts":{"nominal":12,"open-jaw":12,"release-after-lift":12},"case_rows":{"nominal":1171,"open-jaw":1800,"release-after-lift":1582},"completed_recordings":36,"excluded_steps":{"development":4,"test":4,"train":16},"failed_recordings":0,"groups":{"development":2,"test":2,"train":8},"rows":{"development":4555,"test":4553,"train":18228},"simulation_successes":24,"simulation_unsuccessful":12},"freeze":{"bytes":5786,"file":"freeze.json","sha256":"98b2d59affd3cc9a2098594f3d56f1502df9d9fb0f2cbc0bcc4450e65b25205c"},"hardware":{"gpu_used":false,"logical_cpu_count":10,"machine":"arm64","memory_scope":"process-lifetime high-water mark; includes earlier work in this process; not isolated learner memory","numpy":"2.5.3","platform":"macOS-26.5.2-arm64-arm-64bit","process_lifetime_peak_rss_bytes":184614912,"processor":"arm","python":"3.12.9"},"kind":"dvidia.actuator-transition-pilot.v1","limitations":["Authored rigid-box simulator with privileged joint telemetry; no shoe footage or shoe physics.","Predicts the next 20 ms joint-position and jaw-gap deltas, not object motion, contact force or a control policy.","Held-out results are a small grouped research test, not independent physical robot qualification.","Models use before-state and optionally issued actuator commands; phases, cases and simulator configuration are not inputs.","Compute-only inference timings exclude sensing, transport, actuator response and operating-system deadline guarantees.","Hashes verify internal consistency, not provenance authenticity or measured physical accuracy."],"matched_branch":{"before_state_identical":true,"before_time_seconds":5.800000000000272,"group_id":"reset-010","index":290,"issued_jaw_gap_m":{"nominal":0.03353794662489178,"release":0.08},"joint_command_identical":true,"observed_delta":{"nominal":[-1.0518214868526776e-08,-0.0015201613567266126,-0.0035697291750693694,1.682435840189377e-10,0.005087947866465992,8.422609290570371e-09,-1.2862975268257415e-09],"release":[-1.047361926354018e-08,-0.0015201612527868669,-0.003569729412420841,1.8661526137799222e-10,0.00508794753085684,8.422617853165448e-09,1.0414352173984676e-05]},"prediction_delta":{"action":{"nominal":[7.713487948723166e-07,-0.0015331085578998772,-0.0035750972769338363,-2.4198221895276824e-08,0.005071986847504324,-1.0889586383777747e-06,-7.345075712337937e-05],"release":[1.315247288191287e-06,-0.0015495074441333665,-0.0035578876185807305,4.821777752655842e-08,0.005045740114485017,-1.312176325133535e-07,0.0034298494007314263]},"state":{"nominal":[4.195064692613019e-05,-0.0015200774303030767,-0.003526471758662089,1.1000058922538273e-07,0.004961283063797254,-6.556976093816764e-05,-0.0001088928697452801],"release":[4.195064692613019e-05,-0.0015200774303030767,-0.003526471758662089,1.1000058922538273e-07,0.004961283063797254,-6.556976093816764e-05,-0.0001088928697452801]}},"status":"matched"},"metrics":{"development":{"action":{"jaw_mae_m":0.00011408709850354386,"jaw_max_abs_error_m":0.0034112555979426496,"jaw_rmse_m":0.000340067832374854,"joint_mae_rad":7.1141354587507735e-06,"joint_max_abs_error_rad":0.0012005909356011786,"joint_rmse_rad":3.530976857719615e-05,"rows":4555},"constant_velocity":{"jaw_mae_m":0.00010794690518584453,"jaw_max_abs_error_m":0.003200039935075838,"jaw_rmse_m":0.0005643217213877972,"joint_mae_rad":0.0002646450964939215,"joint_max_abs_error_rad":0.01172399342172524,"joint_rmse_rad":0.0005763532444979818,"rows":4555},"persistence":{"jaw_mae_m":0.00010794690518584453,"jaw_max_abs_error_m":0.003200039935075838,"jaw_rmse_m":0.0005643217213877972,"joint_mae_rad":0.0015407149157630703,"joint_max_abs_error_rad":0.013258501435516545,"joint_rmse_rad":0.0025975452010896204,"rows":4555},"state":{"jaw_mae_m":0.00018672971727035493,"jaw_max_abs_error_m":0.0035440017197499595,"jaw_rmse_m":0.0005518057723999874,"joint_mae_rad":8.045498027119069e-05,"joint_max_abs_error_rad":0.012901945226615082,"joint_rmse_rad":0.0004095575057096664,"rows":4555},"training_mean":{"jaw_mae_m":0.00010795062942042845,"jaw_max_abs_error_m":0.00320004751262421,"jaw_rmse_m":0.0005643217213502803,"joint_mae_rad":0.0016820923351936023,"joint_max_abs_error_rad":0.013269863443301725,"joint_rmse_rad":0.00253897181574975,"rows":4555}},"per_case":{"nominal":{"action":{"jaw_mae_m":0.00014043254672261001,"jaw_max_abs_error_m":0.0026121612687166835,"jaw_rmse_m":0.00037778399228510735,"joint_mae_rad":9.067762386158603e-06,"joint_max_abs_error_rad":0.0014851922180525038,"joint_rmse_rad":4.115871673672732e-05,"rows":1171},"constant_velocity":{"jaw_mae_m":0.0001399527385531288,"jaw_max_abs_error_m":0.0032000401574693166,"jaw_rmse_m":0.0006424120230747426,"joint_mae_rad":0.0003586484709497719,"joint_max_abs_error_rad":0.011968814817229181,"joint_rmse_rad":0.0006598393089054227,"rows":1171},"persistence":{"jaw_mae_m":0.0001399527385531288,"jaw_max_abs_error_m":0.0032000401574693166,"jaw_rmse_m":0.0006424120230747426,"joint_mae_rad":0.0020977425697452983,"joint_max_abs_error_rad":0.013927230299878413,"joint_rmse_rad":0.0030786717336183816,"rows":1171},"state":{"jaw_mae_m":0.00022392693414671568,"jaw_max_abs_error_m":0.003099747038606016,"jaw_rmse_m":0.0006154577647997416,"joint_mae_rad":9.64218977979969e-05,"joint_max_abs_error_rad":0.01308326016295802,"joint_rmse_rad":0.000452931262834069,"rows":1171},"training_mean":{"jaw_mae_m":0.00013995574941613344,"jaw_max_abs_error_m":0.0032000477350176888,"jaw_rmse_m":0.0006424120229601376,"joint_mae_rad":0.002119138665845942,"joint_max_abs_error_rad":0.013938592307663592,"joint_rmse_rad":0.0029838191497862103,"rows":1171}},"open-jaw":{"action":{"jaw_mae_m":3.924383687387137e-05,"jaw_max_abs_error_m":0.0006450019979839865,"jaw_rmse_m":5.7757522848878867e-05,"joint_mae_rad":4.22157219926287e-06,"joint_max_abs_error_rad":0.0011964933888951815,"joint_rmse_rad":2.7132586884069046e-05,"rows":1800},"constant_velocity":{"jaw_mae_m":9.737168403075452e-07,"jaw_max_abs_error_m":2.721904949741505e-05,"jaw_rmse_m":2.7608542301903443e-06,"joint_mae_rad":0.00014437330891681652,"joint_max_abs_error_rad":0.011306049310197339,"joint_rmse_rad":0.00041121686413679565,"rows":1800},"persistence":{"jaw_mae_m":9.737168403075452e-07,"jaw_max_abs_error_m":2.721904949741505e-05,"jaw_rmse_m":2.7608542301903443e-06,"joint_mae_rad":0.0008515914412609879,"joint_max_abs_error_rad":0.008187976307765257,"joint_rmse_rad":0.0018364379646155953,"rows":1800},"state":{"jaw_mae_m":8.022591092982116e-05,"jaw_max_abs_error_m":0.0005888317439760529,"jaw_rmse_m":9.177126950567431e-05,"joint_mae_rad":5.267018412881058e-05,"joint_max_abs_error_rad":0.011975742413801106,"joint_rmse_rad":0.0002865409337301875,"rows":1800},"training_mean":{"jaw_mae_m":9.785815384331015e-07,"jaw_max_abs_error_m":2.7211471949042646e-05,"jaw_rmse_m":2.760864628988399e-06,"joint_mae_rad":0.0011056802379387453,"joint_max_abs_error_rad":0.007895030168901656,"joint_rmse_rad":0.0018133093855438717,"rows":1800}},"release-after-lift":{"action":{"jaw_mae_m":0.00018088801454194404,"jaw_max_abs_error_m":0.003424592601542177,"jaw_rmse_m":0.0004758322453667224,"joint_mae_rad":9.34078195729173e-06,"joint_max_abs_error_rad":0.0016227888852025432,"joint_rmse_rad":4.125450832758917e-05,"rows":1582},"constant_velocity":{"jaw_mae_m":0.00020617527355783652,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0007818902949658236,"joint_mae_rad":0.0003350354038600312,"joint_max_abs_error_rad":0.012038322434476285,"joint_rmse_rad":0.0006677470041561846,"rows":1582},"persistence":{"jaw_mae_m":0.00020617527355783652,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0007818902949658236,"joint_mae_rad":0.0019362902617289155,"joint_max_abs_error_rad":0.01410291592476276,"joint_rmse_rad":0.002951416269034785,"rows":1582},"state":{"jaw_mae_m":0.0002868281334153309,"jaw_max_abs_error_m":0.003551033682136962,"jaw_rmse_m":0.0007658901079131594,"joint_mae_rad":9.80688764966578e-05,"joint_max_abs_error_rad":0.01312721155491846,"joint_rmse_rad":0.0004892309023124577,"rows":1582},"training_mean":{"jaw_mae_m":0.00020617820619647667,"jaw_max_abs_error_m":0.0032000477835342823,"jaw_rmse_m":0.0007818902949056648,"joint_mae_rad":0.0020377291596198883,"joint_max_abs_error_rad":0.01411427793254794,"joint_rmse_rad":0.002893522836683613,"rows":1582}}},"per_group":{"reset-010":{"action":{"jaw_mae_m":0.00011273040329501912,"jaw_max_abs_error_m":0.0034194350485574417,"jaw_rmse_m":0.000339365582631867,"joint_mae_rad":6.99847089641837e-06,"joint_max_abs_error_rad":0.0011938234395374161,"joint_rmse_rad":3.3569323029914595e-05,"rows":2271},"constant_velocity":{"jaw_mae_m":0.0001082431766226375,"jaw_max_abs_error_m":0.003200039687521458,"jaw_rmse_m":0.0005651117046476424,"joint_mae_rad":0.00026311075000510456,"joint_max_abs_error_rad":0.011306049310197339,"joint_rmse_rad":0.0005716088302518483,"rows":2271},"persistence":{"jaw_mae_m":0.0001082431766226375,"jaw_max_abs_error_m":0.003200039687521458,"jaw_rmse_m":0.0005651117046476424,"joint_mae_rad":0.0015332785458832042,"joint_max_abs_error_rad":0.012164085090524801,"joint_rmse_rad":0.0025710590474533933,"rows":2271},"state":{"jaw_mae_m":0.00018847402736352624,"jaw_max_abs_error_m":0.003550527983412443,"jaw_rmse_m":0.0005525713519712771,"joint_mae_rad":7.874182656286702e-05,"joint_max_abs_error_rad":0.012587215429136569,"joint_rmse_rad":0.00040645093698676654,"rows":2271},"training_mean":{"jaw_mae_m":0.00010824690254826172,"jaw_max_abs_error_m":0.0032000472650698304,"jaw_rmse_m":0.0005651117046148318,"joint_mae_rad":0.0016754530356766746,"joint_max_abs_error_rad":0.01217544709830998,"joint_rmse_rad":0.002512234362159536,"rows":2271}},"reset-011":{"action":{"jaw_mae_m":0.00011623116201030318,"jaw_max_abs_error_m":0.003424592601542177,"jaw_rmse_m":0.00034382782370384493,"joint_mae_rad":7.4937639629861036e-06,"joint_max_abs_error_rad":0.0016227888852025432,"joint_rmse_rad":3.883780261196504e-05,"rows":2282},"constant_velocity":{"jaw_mae_m":0.00010779420500296016,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0005637261599113838,"joint_mae_rad":0.0002683395316290226,"joint_max_abs_error_rad":0.012038322434476285,"joint_rmse_rad":0.0005837376958847701,"rows":2282},"persistence":{"jaw_mae_m":0.00010779420500296016,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0005637261599113838,"joint_mae_rad":0.0015646173355810303,"joint_max_abs_error_rad":0.01410291592476276,"joint_rmse_rad":0.0026427673047063447,"rows":2282},"state":{"jaw_mae_m":0.00018946655148114266,"jaw_max_abs_error_m":0.003551033682136962,"jaw_rmse_m":0.0005511863770324415,"joint_mae_rad":8.064796154546672e-05,"joint_max_abs_error_rad":0.01312721155491846,"joint_rmse_rad":0.0004141923917898112,"rows":2282},"training_mean":{"jaw_mae_m":0.00010779791229237315,"jaw_max_abs_error_m":0.0032000477835342823,"jaw_rmse_m":0.0005637261598594247,"joint_mae_rad":0.0017048507854917946,"joint_max_abs_error_rad":0.01411427793254794,"joint_rmse_rad":0.002585643215634521,"rows":2282}}},"slices":{"jaw_changing":{"action":{"jaw_mae_m":0.0009577632932941206,"jaw_max_abs_error_m":0.003424592601542177,"jaw_rmse_m":0.001295276589798923,"joint_mae_rad":3.4749007324002226e-05,"joint_max_abs_error_rad":0.0016227888852025432,"joint_rmse_rad":0.00012846278081461826,"rows":296},"constant_velocity":{"jaw_mae_m":0.0016472618440214486,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0022136045505812504,"joint_mae_rad":0.00040525067899570397,"joint_max_abs_error_rad":0.012038322434476285,"joint_rmse_rad":0.0014260563330773357,"rows":296},"persistence":{"jaw_mae_m":0.0016472618440214486,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0022136045505812504,"joint_mae_rad":0.0009560483140301947,"joint_max_abs_error_rad":0.007619712491781283,"joint_rmse_rad":0.002009142786496577,"rows":296},"state":{"jaw_mae_m":0.0016072798792400203,"jaw_max_abs_error_m":0.003551033682136962,"jaw_rmse_m":0.002119157310962271,"joint_mae_rad":0.00037692177520477765,"joint_max_abs_error_rad":0.01312721155491846,"joint_rmse_rad":0.0014537915077703868,"rows":296},"training_mean":{"jaw_mae_m":0.001647262663215867,"jaw_max_abs_error_m":0.0032000477835342823,"jaw_rmse_m":0.0022136045457512906,"joint_mae_rad":0.0012175796905620937,"joint_max_abs_error_rad":0.007608350483996104,"joint_rmse_rad":0.0020275040988468928,"rows":296}},"moving":{"action":{"jaw_mae_m":0.00010094915561240337,"jaw_max_abs_error_m":0.003424592601542177,"jaw_rmse_m":0.0002974425654946643,"joint_mae_rad":8.769177983490002e-06,"joint_max_abs_error_rad":0.0016227888852025432,"joint_rmse_rad":4.181829172260462e-05,"rows":3398},"constant_velocity":{"jaw_mae_m":6.069852338639634e-05,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.00040739574647308464,"joint_mae_rad":0.0003559643679086865,"joint_max_abs_error_rad":0.012038322434476285,"joint_rmse_rad":0.0006687348308924464,"rows":3398},"persistence":{"jaw_mae_m":6.069852338639634e-05,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.00040739574647308464,"joint_mae_rad":0.002075066040840347,"joint_max_abs_error_rad":0.01410291592476276,"joint_rmse_rad":0.003017996596860414,"rows":3398},"state":{"jaw_mae_m":0.00014068737650487441,"jaw_max_abs_error_m":0.003551033682136962,"jaw_rmse_m":0.00040455175888809625,"joint_mae_rad":9.548645187797334e-05,"joint_max_abs_error_rad":0.01312721155491846,"joint_rmse_rad":0.0004738264525914143,"rows":3398},"training_mean":{"jaw_mae_m":6.070124427057511e-05,"jaw_max_abs_error_m":0.0032000477835342823,"jaw_rmse_m":0.00040739658162674,"joint_mae_rad":0.0021225234408844273,"joint_max_abs_error_rad":0.01411427793254794,"joint_rmse_rad":0.0029315318173961553,"rows":3398}}},"test":{"action":{"jaw_mae_m":0.00011448501155073586,"jaw_max_abs_error_m":0.003424592601542177,"jaw_rmse_m":0.00034160937951837255,"joint_mae_rad":7.2467157411158375e-06,"joint_max_abs_error_rad":0.0016227888852025432,"joint_rmse_rad":3.6305619407398296e-05,"rows":4553},"constant_velocity":{"jaw_mae_m":0.0001080181484574489,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0005644176837046016,"joint_mae_rad":0.00026573145715770303,"joint_max_abs_error_rad":0.012038322434476285,"joint_rmse_rad":0.0005777197451110524,"rows":4553},"persistence":{"jaw_mae_m":0.0001080181484574489,"jaw_max_abs_error_m":0.00320004020598591,"jaw_rmse_m":0.0005644176837046016,"joint_mae_rad":0.0015489857978248775,"joint_max_abs_error_rad":0.01410291592476276,"joint_rmse_rad":0.00260724633751098,"rows":4553},"state":{"jaw_mae_m":0.0001889714883862367,"jaw_max_abs_error_m":0.003551033682136962,"jaw_rmse_m":0.0005518776259183269,"joint_mae_rad":7.969719665517812e-05,"joint_max_abs_error_rad":0.01312721155491846,"joint_rmse_rad":0.0004103492721513921,"rows":4553},"training_mean":{"jaw_mae_m":0.00010802186504245507,"jaw_max_abs_error_m":0.0032000477835342823,"jaw_rmse_m":0.0005644176836622054,"joint_mae_rad":0.0016901874229110483,"joint_max_abs_error_rad":0.01411427793254794,"joint_rmse_rad":0.002549291711743848,"rows":4553}}},"model_artifacts":{"action":{"bytes":6438,"file":"action.model.json","sha256":"611880e00116a6c52d474fe2d08c446bc5475564e4f3d550dd2bb88bb2631045"},"state":{"bytes":5007,"file":"state.model.json","sha256":"0c1aca45dcf66816f587311cf413a298e9076c9f282cd3e4aa20f68510814fe8"}},"physical_robot_ready":false,"protocol":{"alpha_grid":[0.0001,0.01,1.0,100.0],"control_dt_seconds":0.02,"development_score":"joint_rmse_rad/0.01 + jaw_rmse_m/0.001","excluded_inputs":["case","phase","simulation configuration","future state","applied end-of-interval controls","object state"],"feature_modes":{"action":"before q/qdot/jaw + issued joint/jaw targets","state":"before q/qdot/jaw only"},"normalization":"training rows only; model core records means/scales","selection_tie_break":"first alpha in ascending fixed grid","slice_definitions":{"jaw_changing":"abs measured jaw delta > 0.00001 m","moving":"any measured abs joint delta > 0.0001 rad"},"target":"next actuator joint-position and jaw-gap deltas","test_used_for_selection":false,"training_weighting":"equal total weight per retained training episode"},"qualification":"unqualified research pilot; internal consistency only","schema_version":1,"scope":"simulation-only","selection":{"action":{"alpha":0.0001,"trials":[{"alpha":0.0001,"development_score":0.34359880923257363,"metrics":{"jaw_mae_m":0.00011408709850354386,"jaw_max_abs_error_m":0.0034112555979426496,"jaw_rmse_m":0.000340067832374854,"joint_mae_rad":7.1141354587507735e-06,"joint_max_abs_error_rad":0.0012005909356011786,"joint_rmse_rad":3.530976857719615e-05,"rows":4555}},{"alpha":0.01,"development_score":0.3488523445371724,"metrics":{"jaw_mae_m":0.00011030944787074467,"jaw_max_abs_error_m":0.003269044603257664,"jaw_rmse_m":0.000345160387948776,"joint_mae_rad":8.181816234241458e-06,"joint_max_abs_error_rad":0.0012538429794856495,"joint_rmse_rad":3.691956588396459e-05,"rows":4555}},{"alpha":1.0,"development_score":0.37446811496940613,"metrics":{"jaw_mae_m":0.0001131313932677029,"jaw_max_abs_error_m":0.0031182654411507395,"jaw_rmse_m":0.0003555391951989955,"joint_mae_rad":4.7180629138626286e-05,"joint_max_abs_error_rad":0.004918626541840294,"joint_rmse_rad":0.00018928919770410597,"rows":4555}},{"alpha":100.0,"development_score":0.40049833398550805,"metrics":{"jaw_mae_m":0.00011524902086540773,"jaw_max_abs_error_m":0.0028229383383668533,"jaw_rmse_m":0.0003606345536169485,"joint_mae_rad":8.85840301824777e-05,"joint_max_abs_error_rad":0.012589703024714939,"joint_rmse_rad":0.00039863780368559564,"rows":4555}}]},"state":{"alpha":0.0001,"trials":[{"alpha":0.0001,"development_score":0.5927615229709541,"metrics":{"jaw_mae_m":0.00018672971727035493,"jaw_max_abs_error_m":0.0035440017197499595,"jaw_rmse_m":0.0005518057723999874,"joint_mae_rad":8.045498027119069e-05,"joint_max_abs_error_rad":0.012901945226615082,"joint_rmse_rad":0.0004095575057096664,"rows":4555}},{"alpha":0.01,"development_score":0.5931159026759466,"metrics":{"jaw_mae_m":0.00018541009658040093,"jaw_max_abs_error_m":0.0035009574428947242,"jaw_rmse_m":0.000551959882687009,"joint_mae_rad":8.210262251786841e-05,"joint_max_abs_error_rad":0.012939254286617633,"joint_rmse_rad":0.0004115601998893769,"rows":4555}},{"alpha":1.0,"development_score":0.5939541313602504,"metrics":{"jaw_mae_m":0.00018440188004749438,"jaw_max_abs_error_m":0.003446910566375889,"jaw_rmse_m":0.0005527103992896407,"joint_mae_rad":8.404463578661915e-05,"joint_max_abs_error_rad":0.01294704832590305,"joint_rmse_rad":0.0004124373207060971,"rows":4555}},{"alpha":100.0,"development_score":0.5946172957204006,"metrics":{"jaw_mae_m":0.0001823020328579418,"jaw_max_abs_error_m":0.0033305991007897048,"jaw_rmse_m":0.0005528614861221444,"joint_mae_rad":8.993230981100541e-05,"joint_max_abs_error_rad":0.012868733472039574,"joint_rmse_rad":0.0004175580959825623,"rows":4555}}]}},"shoe_skill_ready":false,"source_corpus":{"original_manifest_sha256":"c529bf61405d3d8d46a6a96ebf56a871436a991a6e11abb9a264c9d880ff6624","original_tapes_included":false,"snapshot":{"bytes":57320,"file":"corpus-manifest.json","sha256":"c529bf61405d3d8d46a6a96ebf56a871436a991a6e11abb9a264c9d880ff6624"}},"stage":"complete","status":"completed","timings":{"feature_and_target_extraction_seconds":{"development":0.7096925000660121,"test":0.7305656664539129,"train":2.772606123238802},"fit_and_development_selection_seconds":8.799432625062764,"held_out_loading_started_elapsed_seconds":14.398998374817893,"inference":{"action":{"measured_calls":1000,"p50_microseconds":198.25,"p95_microseconds":206.7521,"scope":"compute-only predict_features; pre-extracted held-out input vectors; no sensing or transport","warmup_calls":100,"worst_microseconds":231.667},"state":{"measured_calls":1000,"p50_microseconds":144.833,"p95_microseconds":151.5503,"scope":"compute-only predict_features; pre-extracted held-out input vectors; no sensing or transport","warmup_calls":100,"worst_microseconds":184.917}},"pipeline_elapsed_seconds_before_final_receipt":19.292199834017083,"source_collection_seconds":41.12971108290367,"source_load_and_validation_seconds":{"development":0.44564912491478026,"test":0.44243291579186916,"train":1.6538948728702962}},"training_mean_delta":[0.0009610293150821221,0.00028681519387991386,1.1362007785179132e-05,1.1868488215456264e-10,-0.0002929461388636014,-0.0009610293636950312,-7.577548372403187e-09]}
diff --git a/benchmarks/actuator-pilot-20261011/source.json b/benchmarks/actuator-pilot-20261011/source.json
new file mode 100644
index 0000000..713fb08
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/source.json
@@ -0,0 +1,36 @@
+{
+ "kind": "dvidia.actuator-pilot-source-reference.v1",
+ "source_commit": "0d3f05f186423e18b15ed8362d772b7d74ec0207",
+ "repository": "https://github.com/Dvidia-Inference/dvidia-training",
+ "scope": "Post-run source snapshot; this supplements, and does not alter, frozen run receipts.",
+ "files": [
+ {
+ "file": "src/dvidia_training/transition_model.py",
+ "sha256": "4154267b52885113996d58ba996890eac13d674c4c18c48b7282d27273b6a96e"
+ },
+ {
+ "file": "src/dvidia_training/transition_pilot.py",
+ "sha256": "cafb5e31bd9a07f224339e1f636abccc5f8aecd1efce0ecb4413bbb79441217c"
+ },
+ {
+ "file": "src/dvidia_training/transition_cli.py",
+ "sha256": "b98fb1bc37ef4688f07d4c09b799bb5a668ad7fe38faa73eaed0c94d0cd0ddcf"
+ },
+ {
+ "file": "src/dvidia_training/arm_env.py",
+ "sha256": "e4af9b75364a568284b1e6d96469deacf498dd87a60b067fa5e310fe15c54c00"
+ },
+ {
+ "file": "src/dvidia_training/arm_policy.py",
+ "sha256": "d6d53d87703fb8608390092309285afb6a2c8fd3103a7abd77668be85469eab3"
+ },
+ {
+ "file": "src/dvidia_training/episode_tape.py",
+ "sha256": "abae9a4bd0080c7f638c5c82c050fb1364a20769998c054354b918233669f5d2"
+ },
+ {
+ "file": "pyproject.toml",
+ "sha256": "7b58e1ffe288b8b989b19d2ceb6807a7f18a2d870e8ca21d0db7808670bb212e"
+ }
+ ]
+}
diff --git a/benchmarks/actuator-pilot-20261011/state.model.json b/benchmarks/actuator-pilot-20261011/state.model.json
new file mode 100644
index 0000000..d4f0013
--- /dev/null
+++ b/benchmarks/actuator-pilot-20261011/state.model.json
@@ -0,0 +1 @@
+{"alpha":0.0001,"coefficients":[[-0.16476651686325575,0.005524465821249382,0.009778917298570814,5.454068308367904e-05,0.006895507473024272,-0.16376548535413757,-0.0023671697826554115,0.00022035077627096834,0.00030069157590808073,-8.07936728908365e-05,0.00021946539341671778,-0.002170401697911017,-5.641600826716365e-06],[0.24295356400456378,-0.016491546035597753,-0.029356673592170285,-1.6294142215859593e-05,-0.020556353321446338,0.24154037543539827,0.005265227483965713,0.001250019846892589,-0.0012669675243507387,1.890705296528738e-05,-0.0008941175698073853,0.002861393488113036,-1.4003880924183908e-05],[-0.34111049117413367,0.017136275750794128,0.030496616391164903,3.488534205298826e-05,0.021477158992771524,-0.339049634768761,-0.007795192947196962,0.0009300034413357292,0.005511171134371082,-0.00014342080972782673,0.0011475504251036584,-0.004403633488758934,3.1997326196973464e-05],[0.00041572080596648296,-2.310874555174632e-05,-3.964526204147814e-05,-1.7564888532700156e-06,-2.8286233517837672e-05,0.0004116237232897221,1.03390613716476e-05,-1.006249423041349e-06,-1.705182771265903e-06,2.143451702108474e-06,-1.1697648478515562e-06,6.072563759956098e-06,-3.650931411206051e-08],[0.24546952909905856,-0.029108699203966627,-0.052130860386361576,3.567289757475267e-05,-0.0366771862689887,0.24401830005372138,0.00566517815950886,-0.002064914971913653,-0.003637086706291626,0.00012988372313685272,0.0007618668250137123,0.003224412424656125,-6.802993983225428e-05],[0.24530001645456012,-0.008600592173428984,-0.015214068481103998,-9.12669266423266e-05,-0.010730598158762936,0.24378722696234584,0.005145060167539058,-0.0003477744642675069,-0.0004682722380182639,0.0001225922010016553,-0.00034321665664961026,0.004136290449853921,1.0232135721697584e-05],[0.029760310999706673,-0.018770114522380806,-0.03361153054663389,2.7855735044700755e-05,-0.023445563916419485,0.029454107339652313,0.0005802394233204858,-0.0002861293366299175,-0.00025955644036304934,4.432863831106114e-05,4.4988698728784796e-05,0.0003133069129748709,1.0141868386620094e-05]],"control_dt_seconds":0.02,"feature_contract":{"acquisition":"required signals contemporaneous with boundary; no delayed-sensor propagation","availability":"before decision plus issued command; no future or applied control","names":["before.joint_position[0]","before.joint_position[1]","before.joint_position[2]","before.joint_position[3]","before.joint_position[4]","before.joint_position[5]","before.joint_velocity[0]","before.joint_velocity[1]","before.joint_velocity[2]","before.joint_velocity[3]","before.joint_velocity[4]","before.joint_velocity[5]","before.gripper_width","command.issued.joint_targets[0]","command.issued.joint_targets[1]","command.issued.joint_targets[2]","command.issued.joint_targets[3]","command.issued.joint_targets[4]","command.issued.joint_targets[5]","command.issued.gripper_width"],"origin":"sensor-shaped privileged native simulation; hardware sensors unqualified","units":["rad","rad","rad","rad","rad","rad","rad/s","rad/s","rad/s","rad/s","rad/s","rad/s","m","rad","rad","rad","rad","rad","rad","m"]},"feature_mode":"state","intercept":[0.0009610293150821192,0.0002868151938799142,1.1362007785179062e-05,1.1868488215456403e-10,-0.00029294613886360224,-0.0009610293636950372,-7.577548372409316e-09],"is_controller":false,"kind":"dvidia.compact-action-transition-model","license":"MIT","model_sha256":"6f6a8d907a5516a547b7b805561d5f3d0bede236a3f9a25bec759a11e7319bd1","normalization":{"constant_features":[false,false,false,false,false,false,false,false,false,false,false,false,false],"means":[-0.0899905108472486,-1.1519101433125067,1.3351944186598976,2.0685882094698694e-06,-0.18334919247409012,0.08884156095454729,0.04256845907588669,0.01272885406120272,0.0012493047510298584,0.00011965114235991223,-0.012338512856367419,-0.038178765041471746,0.06767697721595069],"scales":[0.17860326146676422,0.1767063627527591,0.31658147968533146,7.945667011746564e-06,0.22167432531872702,0.1777062994467045,0.06463297032762196,0.08931124694281727,0.1797614962369242,0.00022410177231653518,0.15581823686916052,0.05904040544125759,0.018225447281794958],"scope":"fit-training-only"},"output_contract":{"acquisition":"required after signals contemporaneous with the after boundary","contact_and_task_outcomes":"not predicted","definition":"after minus before at exactly one 20 ms control boundary","names":["delta.joint_position[0]","delta.joint_position[1]","delta.joint_position[2]","delta.joint_position[3]","delta.joint_position[4]","delta.joint_position[5]","delta.gripper_width"],"units":["rad","rad","rad","rad","rad","rad","m"]},"physical_robot_ready":false,"predicts_task_success":false,"schema_version":1,"scope":"simulation-only","sensor_origin":"privileged-native-simulation","training":{"positive_weight_rows":18228,"rows":18228,"rows_sha256":"7bff06c153c5071b3b162eec08ea23eaeeee032daf7d1fb0861c6d9cfc674248","sample_weights_sha256":"7787fd55dcedecc0ac29b9030d915ca9036afaa6247c5484333ac4dd9bfc8596","weight_sum":18228.0}}
diff --git a/docs/episode-tape.md b/docs/episode-tape.md
index 877f53e..1fb16eb 100644
--- a/docs/episode-tape.md
+++ b/docs/episode-tape.md
@@ -109,8 +109,12 @@ authenticity, sensor accuracy or task mastery.
## Next consumers
-The first consumers should be a synchronized episode inspector and a frozen
-training/confirmation split, followed by a compact state-based outcome model.
+The [compact actuator pilot](laptop-skill-pilot.md) now consumes current joint
+position, velocity, jaw gap and issued commands from these tapes to predict the
+next boundary's joint/jaw deltas. It excludes future measurements, authored
+phases and final applied controls from inputs, and compares separate source
+groups with simple baselines. It does not predict grasp retention or task
+success. A synchronized inspector and compact outcome model remain next builds.
The current visual predictor and movement head do not consume these tapes.
A future learner must choose deployable inputs explicitly and hold privileged
object state and authored phase labels out of its deployment feature vector.
diff --git a/docs/laptop-skill-pilot.md b/docs/laptop-skill-pilot.md
new file mode 100644
index 0000000..e30a1ce
--- /dev/null
+++ b/docs/laptop-skill-pilot.md
@@ -0,0 +1,205 @@
+# A laptop-first Skillspace pilot
+
+DVIDIA's next useful step is a small learned component that predicts the measured
+effect of a robot command. A large language model does not need to run at every
+control boundary. The current transition pilot learns from native simulator
+episode tapes and runs locally on a CPU. It does not learn shoe organization
+from footage or control a physical robot.
+
+The proposed shoe pilot is separate: collect complete visual examples together
+with synchronized robot state and action evidence, then evaluate a learned policy
+within a declared robot, gripper and task envelope. A large number of short videos
+alone cannot supply the missing robot action measurements.
+
+## What the existing tools contribute
+
+| Component | Implemented contribution | Remaining gap |
+| --- | --- | --- |
+| Skillspace intake and review | Local footage manifests, source hashes, lineage grouping, observation overlays and editable annotations | Independently reviewed shoe data and synchronized robot actions |
+| Coverage planner | Capture missions, declared review/rights gates and unresolved conditions | Collected evidence; the planner does not inspect or approve videos |
+| Visual learner | A CPU PCA/ridge next-frame predictor from frozen source groups | Object geometry, grasping and action-conditioned task behavior |
+| Motion alpha | Seeded image tracking, image speed and short future-path extrapolation | Calibrated depth, contact dynamics and measured live-stream latency |
+| Episode tapes | Before/after state, requested/issued commands and native measurements at control boundaries | Equivalent calibrated hardware sensing and action alignment |
+| Native arm and capsules | Bounded rigid-box placement, simulation candidates and compatibility checks | A validated shoe task, learned shoe policy and physical qualification |
+| Reaction and motor memory | Separate native anticipation, braking, retention and fitted actuator experiments | Whole-arm hardware measurements and integration into a shoe policy |
+
+These pieces provide evidence and reusable interfaces. They are not already a
+single end-to-end footage-to-shoe-policy system. See [episode tapes](episode-tape.md),
+[motion observations](motion-observations.md) and the [execution map](roadmap.md).
+
+## Run the compact transition pilot
+
+Use the `feat/laptop-skill-pilot-v0.1` candidate source branch and
+[installation instructions](installation.md). It is not included in the frozen
+Observation Lab v0.1 release or hosted demo. Corpus generation requires the
+optional `arm` extra; fitting uses NumPy and saved-model inference uses Python's
+standard library. Python 3.12+ is required by the package.
+
+```sh
+.venv/bin/python -m pip install ".[arm]"
+.venv/bin/dvidia-transition corpus \
+ --output runs/transition-corpus --seed-start 311000
+.venv/bin/dvidia-transition run runs/transition-corpus/corpus.json \
+ --output runs/transition-pilot
+.venv/bin/dvidia-transition inspect runs/transition-pilot
+```
+
+Use fresh output directories. Open `runs/transition-pilot/report.html` locally;
+`report.json` retains the machine-readable results. On Windows PowerShell,
+replace `.venv/bin/COMMAND` with `.\.venv\Scripts\COMMAND.exe` and put continued
+commands on one line. The corpus command writes a directory containing
+`corpus.json` and its episode tapes.
+
+The default corpus has 12 seeded native scene groups, each recorded with nominal,
+open-jaw and release-after-lift commands. All three interventions for one scene
+remain in the same split. The 8/2/2 training/development/test grouping is an
+algorithm demonstration, not 12 independently captured physical examples.
+Related recording/session identities must also remain connected; changing an ID
+does not create new evidence.
+
+The learned ridge model has 20 inputs and seven outputs at the native 20 ms
+control interval:
+
+| Contract | Values |
+| --- | --- |
+| Current observation | Six joint positions, six joint velocities and measured jaw gap |
+| Issued action | Six joint targets and the requested jaw gap |
+| Prediction | Change in six joint positions and measured jaw gap at the next boundary |
+
+All present recordings use privileged native simulator measurements. The input
+contract resembles encoder/command signals a robot could expose; no physical
+encoder or gripper has been validated here. Missing, stale or delayed measurements
+must retain their status instead of becoming zero-valued readings. This
+fixed-boundary model requires contemporaneous required observations; it does not
+propagate aged sensor readings to the present.
+
+Authored phase, intervention name, scene identity, terminal outcome, future object
+state and the final applied servo controls are excluded from the prediction
+inputs. The final applied controls are recorded after the transition and are not
+known when the action is requested. Object/contact truth remains separate from
+the deployment feature vector.
+
+Training groups fit the normalization and model. Development groups choose the
+regularization; the selected model and evaluation protocol are frozen before
+test scoring. The report compares the action-conditioned candidate with
+training-mean delta, persistence, constant-velocity and state-only predictions.
+These comparisons ask whether commands add predictive value beyond current
+state. Low next-step error does not establish grasp retention, terminal task
+success or a shoe policy.
+
+Report model inference separately from loading and feature extraction, with
+sample counts, median, p95 and worst times. Camera capture, image processing,
+communication, physical command application and stopping behavior are outside
+this timing boundary. A fast numerical prediction is one component of a future
+closed control loop, not evidence that the full robot reacts at that speed.
+Cold saved-model loading is not measured; corpus-load timings cover episode
+validation rather than deployment startup.
+
+The [frozen benchmark](../benchmarks/actuator-pilot-20261011/README.md) retains
+the trained models and exact run receipt. On the measured laptop, the 6,438-byte
+action model took 0.207 ms at warm p95. Commands improved aggregate joint and
+gripper RMSE, but aggregate gripper mean and worst errors slightly regressed
+against persistence. The first matched opening transition was substantially
+overpredicted. Those results motivate a history-aware gripper model and fresh
+confirmation groups; they do not establish a usable shoe controller.
+
+## Hardware and offline operation
+
+| Operation | Current dependency requirement |
+| --- | --- |
+| Generate native episode tapes | CPU, Python, NumPy and MuJoCo through the `arm` extra |
+| Fit or evaluate the transition model | CPU, Python and NumPy; no GPU or model API |
+| Read the saved HTML report | A local browser; no hosted service |
+| Run footage observation/motion tools | Their separate OpenCV/FFmpeg and model requirements |
+| Operate a physical shoe robot | A compatible robot/gripper, calibrated sensing and a measured adapter; not implemented by this pilot |
+
+After dependencies are installed, the transition experiment needs no network,
+account, language model, renderer, video conversion or downloaded model weights.
+This does not make the observation pipeline's detector or camera processing
+part of its inference measurements. Minimum RAM, supported laptop throughput
+and end-to-end control deadlines must come from measured runs on specified
+machines; these dependency requirements are not certified minimum hardware.
+
+A future imitation policy may need more expensive training. ACT learns action
+chunks from robot demonstrations and is a candidate for that stage, rather than
+the transition predictor implemented here. A smaller student policy may then be
+evaluated for local execution. GPU training and CPU execution are separate
+questions. LeRobot's deployment documentation explicitly makes CPU use
+conditional on checkpoint support and measured control rate. Neither source
+establishes the speed or shoe competence of a DVIDIA policy.
+[ACT research](https://tonyzhaozh.github.io/aloha/),
+[LeRobot policy inference](https://huggingface.co/docs/lerobot/main/inference).
+
+## Prepare the first shoe Skillspace
+
+The [shoe capture plan](../examples/shoe-organizing-coverage.json) starts with
+one shoe moved from a marked pickup area into one marked slot. It contains no
+fabricated evidence. Its benchmark is blocked because a validated shoe task
+environment and compatible hardware adapter have not been supplied.
+
+The proposed final criteria are toe direction within 15 degrees of the slot,
+center within 20 mm of its reference and observed stability for one second.
+These are study proposals, not validated pass thresholds. Define the shoe
+reference point, toe direction, slot coordinates, uncertainty and observation
+procedure using calibrated measurements before freezing a benchmark.
+
+Start with one fixed table and calibrated camera, verified reachable regions and
+an observed grasp region compatible with the arm's payload and gripper. Use
+tucked or absent loose laces, one shoe and no clutter. Pair matching, arbitrary
+footwear and lace manipulation are later tasks with different requirements.
+
+An initial workload suggestion is 120 complete attempts across 12 disjoint
+physical shoe-pair/source groups, with eight groups for training, two for
+development and two reserved for confirmation. Ten attempts per group is a
+planning starting point, not a fixed minimum or known sufficient sample count.
+Choose separate collection partitions before recording. Preserve true recording,
+session and pair lineage; shared sessions can connect otherwise different pairs
+into one group. Confirmation evidence never fills the training collection quota.
+
+The per-condition group targets in the JSON are revisable workload estimates.
+A group can contribute different complete attempts to multiple conditions;
+summing cell quotas does not give a count of independent examples. Include
+changed starts, slot offsets, different pairs, genuine missed grasps and observed
+supervised recoveries. Retain operator assistance and all unsuccessful attempts.
+Do not induce an unsafe drop to populate a failure cell.
+
+Five- or six-second clips are useful only when they preserve the evidence needed
+for their labels. A complete attempt must show acquisition, transfer, release
+and the final outcome, even when that takes longer. Retain originals and stable
+source IDs before cutting or augmenting recordings.
+
+Inspect the proposed plan with the existing coverage validator:
+
+```sh
+.venv/bin/dvidia-skillspace-coverage inspect \
+ examples/shoe-organizing-coverage.json
+```
+
+`inspect` validates the plan and computes declared gaps. It does not collect
+footage, perform human review or run a shoe benchmark. See
+[coverage rules](skillspace-coverage.md) for complete attempts, review, rights
+and connected-source grouping.
+
+## From collection to an installable skill
+
+The intended path is:
+
+1. A Skillspace defines the task, operating envelope and capture missions.
+2. Complete footage receives review, rights checks and source grouping.
+3. Robot teleoperation pairs observations with synchronized measured state and
+ issued actions. Human hand motion alone is not a robot action label.
+4. A trained policy consumes only explicitly available deployment inputs;
+ development evidence selects its revision and diagnoses weak conditions.
+5. A validated task simulator evaluates fresh held-out scenarios and matched
+ controls, retaining every attempted episode and false completion.
+6. A supported robot/gripper adapter receives an artifact with its feature,
+ calibration and compatibility contracts. Physical trials establish the
+ measured deployment envelope before qualification.
+
+The transition pilot adds a small learned action-to-motion component to this
+path. The existing capsule format supports one simulation arm/rigid-box adapter;
+installing that capsule cannot install shoe competence on an arbitrary arm.
+The next shoe milestones are real synchronized capture, observable task labels,
+a learned closed-loop policy and a validated shoe benchmark. Data volume and
+model speed become useful evidence when those measurements improve held-out
+task completion without concealing failures.
diff --git a/docs/roadmap.md b/docs/roadmap.md
index 4244a08..69f570e 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -1,6 +1,6 @@
# Execution map: footage to installable skills
-Working map, updated 10 October 2026. This is the team's sequence of builds and evidence
+Working map, updated 11 October 2026. This is the team's sequence of builds and evidence
gates. A completed software component does not establish robot competence.
## The product loop
@@ -31,8 +31,8 @@ needs fresh confirmation scenarios; it cannot silently reuse a known test set.
| --- | --- | --- | --- |
| 1 | Collection planning — product + data | Local coverage planner, versioned briefs and declared source-group tallies implemented. | First skill's observable goal, operating envelope, actual recordings and reviewed capture lineage. |
| 2 | Contribution review — data + product | Proposed official-site connection; the local planner imports declarations only. | Immutable review receipts tied to original bytes, plan revision, intended-use permission and complete outcomes; source-group and near-duplicate audits. |
-| 3 | Episode evidence — simulation | Recorder preserves state, requests, applied commands, sensor age and authored phases at control boundaries. | Task-specific episode corpus, retained failure cases and synchronized inspector. Tapes are not consumed by a learner yet. |
-| 4 | Task and state models — learning | Existing small visual predictor and bounded movement head; outcome/phase supervisor proposed. | Choose deployable sensor inputs, compare with simple baselines and measure held-out boundary, outcome and false-completion errors. Keep authored labels and privileged state outside deployment inputs. |
+| 3 | Episode evidence — simulation | Recorder preserves state, requests, applied commands, sensor age and authored phases at control boundaries. A compact actuator learner now consumes before-state and issued-command evidence. | Shoe-specific corpus, retained failure cases and synchronized inspector. Sensor-shaped simulator inputs still need hardware counterparts. |
+| 4 | Task and state models — learning | Small visual predictor, bounded movement head and next-boundary actuator predictor; outcome/phase supervisor proposed. | Compare with matched baselines and measure held-out boundary, outcome and false-completion errors. Keep authored labels and future state outside deployment inputs; qualify sensed physical inputs separately. |
| 5 | Task simulation — simulation | Existing rigid-box placement adapter and frozen benchmark; other tasks need environments. | Validated assets/contact behavior, robot/gripper/sensor contract, reset distributions and independently observable task outcomes. |
| 6 | Qualification and sample efficiency — evaluation | Local benchmark runner and frozen sample-count evidence exist. | Learning curves over independent source groups, multiple collection orders, frozen candidates, nominal/stress/control strata, all attempted trials and measured cost. |
| 7 | Installable skill — runtime | Simulation candidate capsule and compatibility checks exist. | Bundle model, task, observation/action contract, adapter, data/protocol hashes, calibration requirements and artifact-bound qualification results. |
@@ -44,6 +44,14 @@ No new task policy is trained by adding these components.
## First incoming skills
+The [laptop skill pilot](laptop-skill-pilot.md) supplies a proposed shoe collection
+plan and a separate CPU actuator-learning experiment. It uses twelve native
+scene groups and three jaw-command interventions per group, with training,
+development and frozen test partitions. Its seven outputs predict joint/jaw
+changes at one 20 ms boundary; it is neither a shoe policy nor a contact/outcome
+model. Physical shoe sensing, shoe task physics and robot integration remain
+the next task-specific builds.
+
Start with one bounded task and one supported robot/gripper rather than declaring
a universal skill. For shoe organization, distinguish pairing, alignment and
placing on a shelf: each needs a goal, observable result and task environment.
diff --git a/examples/shoe-organizing-coverage.json b/examples/shoe-organizing-coverage.json
new file mode 100644
index 0000000..2126b82
--- /dev/null
+++ b/examples/shoe-organizing-coverage.json
@@ -0,0 +1,86 @@
+{
+ "format": "dvidia.skillspace-coverage-plan",
+ "schema_version": 1,
+ "plan_id": "shoe-slot-pilot",
+ "plan_revision": 1,
+ "skill": {
+ "name": "Place one shoe into a marked slot · proposed pilot",
+ "goal": "Move one shoe from a marked pickup area into one marked slot and observe stable release. Proposed study criteria are toe direction within 15 degrees of the slot direction, center within 20 mm of its reference, and stability for 1 second. These require calibrated observable measurements and have not been validated as task criteria.",
+ "operating_envelope": "Proposed physical capture plan: one shoe, fixed table and calibrated camera, documented reachable pickup/goal regions, and verified arm payload and gripper grasp compatibility. Start with tucked or absent loose laces, no clutter, and observable pickup/release. Pair matching, arbitrary footwear, autonomous physical operation and shoe simulation are outside this plan. Per-condition group quotas estimate collection workload; they are not sufficient training sample counts."
+ },
+ "cases": [
+ {
+ "id": "nominal-placement",
+ "label": "Complete marked-slot placement",
+ "brief": "Record a complete supervised robot attempt from the initial shoe state through grasp, transfer, release and observed final stability. Retain synchronized camera frames, measured robot state and issued commands. This five-group target is a revisable workload estimate, not a known training threshold.",
+ "outcome": "success",
+ "required_groups": 5,
+ "priority": 8
+ },
+ {
+ "id": "different-shoe-pair",
+ "label": "Different physical shoe pair",
+ "brief": "Use a different physical pair within the verified grasp/payload envelope; retain its stable pair identity for both left and right shoes and every related cut. Document the grasp region, material and geometry. This five-group target estimates workload and does not establish generalization.",
+ "outcome": "success",
+ "required_groups": 5,
+ "priority": 9
+ },
+ {
+ "id": "rotated-start",
+ "label": "Changed initial orientation",
+ "brief": "Vary the shoe's initial toe direction within the documented reachable and observable envelope. Preserve the complete corrective placement and final orientation measurement. The three-group target is a planning estimate; it does not enumerate every possible orientation.",
+ "outcome": "success",
+ "required_groups": 3,
+ "priority": 7
+ },
+ {
+ "id": "goal-offset",
+ "label": "Changed slot position",
+ "brief": "Move the marked slot to a documented offset within the verified workspace. Record its calibrated position and direction, the complete attempt and the final measured result. The three-group target estimates workload; do not exceed reach or collision-clearance limits to fill the cell.",
+ "outcome": "success",
+ "required_groups": 3,
+ "priority": 7
+ },
+ {
+ "id": "empty-or-missed-grasp",
+ "label": "Observed empty or missed grasp",
+ "brief": "Retain a genuine unsuccessful supervised attempt with the jaws and shoe visible. Distinguish a requested closure from measured contact and retention. Do not induce an unsafe drop or reach to create a negative example. The two-group target estimates failure-evidence workload.",
+ "outcome": "failure",
+ "required_groups": 2,
+ "priority": 10
+ },
+ {
+ "id": "supervised-recovery",
+ "label": "Correction after a failed grasp",
+ "brief": "Record the initial missed grasp, a supervised corrective attempt and its observed successful placement in one complete episode. Retain all issued commands and identify operator assistance. The two-group target is an estimate, not evidence that an autonomous recovery policy exists.",
+ "outcome": "recovery",
+ "required_groups": 2,
+ "priority": 9
+ },
+ {
+ "id": "changed-lighting",
+ "label": "Different observable lighting",
+ "brief": "Record a complete placement under a documented lighting change while the shoe, gripper and final slot remain measurable. Retain camera/calibration identity and measurement uncertainty. The three-group target estimates workload; it does not certify visual robustness.",
+ "outcome": "success",
+ "required_groups": 3,
+ "priority": 6
+ }
+ ],
+ "benchmark": {
+ "status": "blocked",
+ "environment": "A validated shoe-organizing task environment and compatible physical robot adapter have not been supplied.",
+ "reason": "The existing native arm adapter models a bounded upright rigid box, not shoe geometry, deformation, laces or calibrated hardware. This capture plan contains no footage, executed shoe benchmark, trained shoe policy or physical qualification.",
+ "success_predicate": "Proposed, unvalidated: a calibrated observation verifies grasp, transfer and release of one shoe; toe direction is within 15 degrees, center within 20 mm, and the final shoe remains stable for 1 second. Freeze measurement procedures and task criteria before independent confirmation.",
+ "failure_cases": [
+ "Empty grasp or unsuccessful pickup",
+ "Slip, drop or loss of retention",
+ "Final shoe position or toe direction outside the proposed tolerance",
+ "Release followed by unstable final placement",
+ "Missing, stale or uncalibrated object/robot measurements",
+ "Start, goal, payload or grasp region outside the verified hardware envelope",
+ "Unsupported shoe geometry or loose-lace interaction",
+ "False declaration of task completion"
+ ]
+ },
+ "evidence": []
+}
diff --git a/pyproject.toml b/pyproject.toml
index da66afb..aa710c8 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -39,6 +39,7 @@ dvidia-motor-memory = "dvidia_training.motor_benchmark:main"
dvidia-observe = "dvidia_training.observation_pipeline:main"
dvidia-observation-studio = "dvidia_training.observation_studio:main"
dvidia-motion = "dvidia_training.motion_cli:main"
+dvidia-transition = "dvidia_training.transition_cli:main"
[tool.setuptools.packages.find]
where = ["src"]
diff --git a/src/dvidia_training/transition_cli.py b/src/dvidia_training/transition_cli.py
new file mode 100644
index 0000000..48c6761
--- /dev/null
+++ b/src/dvidia_training/transition_cli.py
@@ -0,0 +1,56 @@
+"""Local compact actuator learning; no language service or physical arm commands."""
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+import sys
+
+
+def write_report_page(directory):
+ """Embed the exact retained receipt in a self-contained offline report."""
+ directory = Path(directory)
+ receipt = directory / 'report.json'
+ if receipt.is_symlink() or not receipt.is_file() or receipt.stat().st_size > 16 * 1024 * 1024:
+ raise ValueError('Use a bounded, regular pilot receipt.')
+ raw = receipt.read_bytes().decode('utf8')
+ document = json.loads(raw, parse_constant=lambda value: (_ for _ in ()).throw(ValueError('Nonfinite receipt.')))
+ if type(document) is not dict or document.get('kind') != 'dvidia.actuator-transition-pilot.v1':
+ raise ValueError('Unsupported actuator pilot receipt.')
+ embedded = json.dumps(raw, ensure_ascii=True).replace('<', '\\u003c').replace('>', '\\u003e').replace('&', '\\u0026')
+ template = Path(__file__).with_name('transition_pilot.html').read_text(encoding='utf8')
+ with (directory / 'report.html').open('x', encoding='utf8') as output:
+ output.write(template.replace('__DVIDIA_RECEIPT__', embedded))
+
+
+def main(argv=None):
+ parser = argparse.ArgumentParser(description=__doc__)
+ sub = parser.add_subparsers(dest='operation', required=True)
+ corpus = sub.add_parser('corpus', help='Record the authored 12-group native fixture; requires the arm extra.')
+ corpus.add_argument('--output', required=True)
+ corpus.add_argument('--seed-start', type=int, default=311000)
+ run = sub.add_parser('run', help='Fit compact models, freeze them, and score the separate test groups.')
+ run.add_argument('manifest')
+ run.add_argument('--output', required=True)
+ inspect = sub.add_parser('inspect', help='Check retained artifact consistency; this does not verify physical accuracy.')
+ inspect.add_argument('directory')
+ args = parser.parse_args(argv)
+ try:
+ from .transition_pilot import generate_corpus, inspect_pilot, run_pilot
+ if args.operation == 'corpus':
+ result = generate_corpus(args.output, seed_start=args.seed_start)
+ elif args.operation == 'run':
+ result = run_pilot(args.manifest, args.output)
+ write_report_page(args.output)
+ else:
+ result = inspect_pilot(args.directory)
+ print(json.dumps(result, sort_keys=True, allow_nan=False))
+ return 1 if result.get('status') in ('failed', 'error') else 0
+ except (ValueError, OSError, ImportError) as exc:
+ message = str(exc) if isinstance(exc, ValueError) else f'Local operation failed ({type(exc).__name__}).'
+ print(json.dumps({'status': 'error', 'error': message[:300]}), file=sys.stderr)
+ return 2
+
+
+if __name__ == '__main__':
+ raise SystemExit(main())
diff --git a/src/dvidia_training/transition_model.py b/src/dvidia_training/transition_model.py
new file mode 100644
index 0000000..7413182
--- /dev/null
+++ b/src/dvidia_training/transition_model.py
@@ -0,0 +1,405 @@
+"""Original compact, action-conditioned boundary predictor (MIT).
+
+This learns seven *20 ms state deltas*, not a task policy, visual representation,
+contact predictor or robot controller. The present training observations are
+privileged native simulation values in sensor-shaped fields. The data-only
+artifact and prediction functions need only Python's standard library; fitting
+imports NumPy locally. Internal hashes establish consistency, not authenticity.
+"""
+from __future__ import annotations
+
+from hashlib import sha256
+import json
+import math
+import os
+from pathlib import Path
+import re
+import stat
+
+from .episode_tape import JOINT_LIMITS, PROTOCOL, canonical, digest, validate_tape
+
+KIND = 'dvidia.compact-action-transition-model'
+CONTROL_DT_SECONDS = .02
+MAX_MODEL_BYTES = 256 * 1024
+MAX_TRAINING_ROWS = 200000
+MAX_NUMERIC_MAGNITUDE = 1e6
+FEATURE_NAMES = tuple([f'before.joint_position[{i}]' for i in range(6)] +
+ [f'before.joint_velocity[{i}]' for i in range(6)] +
+ ['before.gripper_width'] +
+ [f'command.issued.joint_targets[{i}]' for i in range(6)] +
+ ['command.issued.gripper_width'])
+OUTPUT_NAMES = tuple([f'delta.joint_position[{i}]' for i in range(6)] +
+ ['delta.gripper_width'])
+FEATURE_CONTRACT = {
+ 'names': list(FEATURE_NAMES),
+ 'units': ['rad']*6 + ['rad/s']*6 + ['m'] + ['rad']*6 + ['m'],
+ 'availability': 'before decision plus issued command; no future or applied control',
+ 'acquisition': 'required signals contemporaneous with boundary; no delayed-sensor propagation',
+ 'origin': 'sensor-shaped privileged native simulation; hardware sensors unqualified',
+}
+OUTPUT_CONTRACT = {
+ 'names': list(OUTPUT_NAMES), 'units': ['rad']*6 + ['m'],
+ 'definition': 'after minus before at exactly one 20 ms control boundary',
+ 'acquisition': 'required after signals contemporaneous with the after boundary',
+ 'contact_and_task_outcomes': 'not predicted',
+}
+_MODEL_KEYS = {'schema_version', 'kind', 'license', 'scope', 'physical_robot_ready',
+ 'is_controller', 'predicts_task_success', 'sensor_origin', 'feature_mode',
+ 'feature_contract', 'output_contract', 'control_dt_seconds', 'alpha',
+ 'normalization', 'coefficients', 'intercept', 'training', 'model_sha256'}
+_PROVENANCE_KEYS = {'episode_id', 'session_id', 'source_recording_id', 'seed', 'case',
+ 'index', 'before_time_seconds'}
+_REQUIRED_BEFORE = ('joint_position', 'joint_velocity', 'gripper_width')
+_REQUIRED_AFTER = ('joint_position', 'gripper_width')
+
+
+def _keys(value, expected, name):
+ if type(value) is not dict or set(value) != set(expected):
+ raise ValueError(f'{name}: unknown or missing fields.')
+
+
+def _number(value, name, *, minimum=None, maximum=MAX_NUMERIC_MAGNITUDE):
+ try:
+ valid = type(value) in (int, float) and math.isfinite(value) and abs(value) <= maximum
+ except OverflowError:
+ valid = False
+ if not valid or (minimum is not None and value < minimum):
+ raise ValueError(f'{name}: bounded finite numeric value required.')
+ return float(value)
+
+
+def _vector(value, length, name, **limits):
+ if type(value) not in (list, tuple) or len(value) != length:
+ raise ValueError(f'{name}: exactly {length} values required.')
+ return [_number(item, name, **limits) for item in value]
+
+
+def _sha(value, name):
+ if type(value) is not str or re.fullmatch(r'[0-9a-f]{64}', value) is None:
+ raise ValueError(f'{name}: lowercase SHA256 required.')
+
+
+def _sensor_values(snapshot, required):
+ """Use only explicitly fresh required inputs; extras never become features."""
+ _keys(snapshot, {'timestamp_seconds', 'signals'}, 'observation')
+ timestamp = _number(snapshot['timestamp_seconds'], 'observation timestamp', minimum=0)
+ signals = snapshot['signals']
+ if type(signals) is not dict:
+ raise ValueError('Observation signals must be a dictionary.')
+ values = []
+ for name in required:
+ sensor = signals.get(name)
+ if sensor is None:
+ raise ValueError(f'missing:{name}')
+ _keys(sensor, {'status', 'captured_at_seconds', 'age_seconds', 'value'}, name)
+ status = sensor['status']
+ if status != 'available':
+ if status in ('missing', 'stale'):
+ raise ValueError(f'{status}:{name}')
+ raise ValueError(f'unsupported_status:{name}')
+ captured = _number(sensor['captured_at_seconds'], name, minimum=0)
+ age = _number(sensor['age_seconds'], name, minimum=0)
+ if captured > timestamp+1e-9 or not math.isclose(timestamp-captured, age, rel_tol=0, abs_tol=1e-9):
+ raise ValueError(f'inconsistent_acquisition_time:{name}')
+ if age > PROTOCOL['sensor_age_limit_seconds']:
+ raise ValueError(f'stale:{name}')
+ if age > 1e-9 or not math.isclose(captured, timestamp, rel_tol=0, abs_tol=1e-9):
+ raise ValueError(f'noncontemporaneous:{name}')
+ if name == 'gripper_width':
+ values.append(_number(sensor['value'], name, minimum=0))
+ else:
+ values.extend(_vector(sensor['value'], 6, name))
+ return values
+
+
+def _issued_values(issued_command):
+ _keys(issued_command, {'joint_targets', 'gripper_width'}, 'issued command')
+ targets = _vector(issued_command['joint_targets'], 6, 'issued joint targets')
+ if any(not low <= target <= high for target, (low, high) in zip(targets, JOINT_LIMITS)):
+ raise ValueError('Issued target exceeds the authored arm limits.')
+ gap = _number(issued_command['gripper_width'], 'issued gripper width', minimum=0)
+ if gap > .08:
+ raise ValueError('Issued gap exceeds the authored jaw limits.')
+ return targets + [gap]
+
+
+def extract_rows(tape):
+ """Validate a tape, then expose causal features and explicitly exclude gaps.
+
+Provenance, phases, time, intervention/case labels, contacts, object state,
+future observations and applied servo controls are never learner inputs.
+"""
+ tape = validate_tape(tape)
+ provenance = tape['provenance']
+ rows, exclusions = [], []
+ configured_dt = tape['configuration']['control_dt']
+ for step in tape['steps']:
+ before, after = step['before'], step['after']
+ interval = after['timestamp_seconds']-before['timestamp_seconds']
+ reason = None
+ if not math.isclose(configured_dt, CONTROL_DT_SECONDS, rel_tol=0, abs_tol=1e-9):
+ reason = 'unsupported_control_interval'
+ elif not math.isclose(interval, CONTROL_DT_SECONDS, rel_tol=0, abs_tol=1e-9):
+ reason = 'incomplete_20ms_boundary'
+ try:
+ if reason is None:
+ state = _sensor_values(before, _REQUIRED_BEFORE)
+ next_state = _sensor_values(after, _REQUIRED_AFTER)
+ features = state + _issued_values(step['command']['issued'])
+ except ValueError as exc:
+ reason = f'required_signal_unavailable:{exc}'
+ if reason is not None:
+ exclusions.append({'index': step['index'], 'reason': reason,
+ 'episode_id': provenance['episode_id']})
+ continue
+ current_targets = state[:6]+[state[12]]
+ target = [later-now for later, now in zip(next_state, current_targets)]
+ _vector(target, 7, 'target delta')
+ rows.append({'features': features, 'target_delta': target,
+ 'episode_id': provenance['episode_id'], 'session_id': provenance['session_id'],
+ 'source_recording_id': provenance['source_recording_id'],
+ 'seed': provenance['seed'], 'case': provenance['case'],
+ 'index': step['index'], 'before_time_seconds': before['timestamp_seconds']})
+ return {'rows': rows, 'exclusions': exclusions, 'attempted_steps': len(tape['steps'])}
+
+
+def _training_rows(rows):
+ if type(rows) not in (list, tuple) or not 2 <= len(rows) <= MAX_TRAINING_ROWS:
+ raise ValueError('Training requires 2–200000 explicit rows.')
+ detached = []
+ for row in rows:
+ if type(row) is not dict or set(row) not in (
+ {'features', 'target_delta'}, {'features', 'target_delta'} | _PROVENANCE_KEYS):
+ raise ValueError('Training row must contain causal features, target deltas and optional complete provenance.')
+ item = {'features': _vector(row['features'], 20, 'features'),
+ 'target_delta': _vector(row['target_delta'], 7, 'target delta')}
+ if _PROVENANCE_KEYS <= set(row):
+ for name in ('episode_id', 'session_id', 'source_recording_id'):
+ value = row[name]
+ if type(value) is not str or re.fullmatch(r'[A-Za-z0-9][A-Za-z0-9_.:-]{0,159}', value) is None:
+ raise ValueError('Training provenance requires short source identifiers.')
+ for name in ('seed', 'index'):
+ if type(row[name]) is not int or not 0 <= row[name] <= 10**9:
+ raise ValueError('Training provenance requires bounded nonnegative integers.')
+ if row['case'] not in ('nominal', 'open-jaw', 'release-after-lift'):
+ raise ValueError('Unsupported simulation case.')
+ _number(row['before_time_seconds'], 'training timestamp', minimum=0)
+ item.update({name: row[name] for name in _PROVENANCE_KEYS})
+ detached.append(item)
+ return detached
+
+
+def fit_model(rows, alpha=.01, feature_mode='action', sample_weights=None):
+ """Fit weighted ridge from the supplied training partition only.
+
+The caller owns episode/session splits and must exclude development and test
+rows. Weighted centering makes the intercept unpenalized. Constant training
+features have scale one and zero coefficients. No NumPy import occurs at load
+or inference time.
+"""
+ rows = _training_rows(rows)
+ alpha = _number(alpha, 'ridge alpha', minimum=0)
+ if feature_mode not in ('action', 'state'):
+ raise ValueError('Use action or state feature mode.')
+ if sample_weights is None:
+ weights = [1.]*len(rows)
+ else:
+ weights = _vector(sample_weights, len(rows), 'sample weights', minimum=0)
+ if sum(weight > 0 for weight in weights) < 2:
+ raise ValueError('At least two positive-weight training rows are required.')
+ import numpy as np
+ active = 20 if feature_mode == 'action' else 13
+ features = np.asarray([row['features'][:active] for row in rows], dtype=np.float64)
+ targets = np.asarray([row['target_delta'] for row in rows], dtype=np.float64)
+ weight = np.asarray(weights, dtype=np.float64)
+ weight_sum = float(weight.sum())
+ means = np.sum(features*weight[:, None], axis=0)/weight_sum
+ variance = np.sum((features-means)**2*weight[:, None], axis=0)/weight_sum
+ constants = np.sqrt(variance) <= 1e-12
+ scales = np.where(constants, 1., np.sqrt(variance))
+ normalized = (features-means)/scales
+ normalized[:, constants] = 0.
+ target_mean = np.sum(targets*weight[:, None], axis=0)/weight_sum
+ weighted_x = normalized*np.sqrt(weight[:, None])
+ weighted_y = (targets-target_mean)*np.sqrt(weight[:, None])
+ matrix = np.vstack((weighted_x, np.sqrt(alpha)*np.eye(active)))
+ rhs = np.vstack((weighted_y, np.zeros((active, 7))))
+ fitted, _, _, _ = np.linalg.lstsq(matrix, rhs, rcond=None)
+ fitted[constants] = 0.
+ if not all(np.isfinite(value).all() for value in (means, scales, fitted, target_mean)):
+ raise ValueError('Ridge fitting must produce finite coefficients.')
+ model = {
+ 'schema_version': 1, 'kind': KIND, 'license': 'MIT', 'scope': 'simulation-only',
+ 'physical_robot_ready': False, 'is_controller': False, 'predicts_task_success': False,
+ 'sensor_origin': 'privileged-native-simulation', 'feature_mode': feature_mode,
+ 'feature_contract': FEATURE_CONTRACT, 'output_contract': OUTPUT_CONTRACT,
+ 'control_dt_seconds': CONTROL_DT_SECONDS, 'alpha': alpha,
+ 'normalization': {'means': means.tolist(), 'scales': scales.tolist(),
+ 'constant_features': constants.tolist(), 'scope': 'fit-training-only'},
+ 'coefficients': fitted.T.tolist(), 'intercept': target_mean.tolist(),
+ 'training': {'rows': len(rows), 'positive_weight_rows': sum(x > 0 for x in weights),
+ 'weight_sum': weight_sum, 'rows_sha256': digest(rows),
+ 'sample_weights_sha256': digest(weights)},
+ }
+ model['model_sha256'] = digest(model)
+ return validate_model(model)
+
+
+def _decode_model(document):
+ if type(document) is dict:
+ try:
+ raw = canonical(document)
+ except (TypeError, ValueError, OverflowError, RecursionError) as exc:
+ raise ValueError('Model must be finite JSON data.') from exc
+ elif isinstance(document, (str, bytes)):
+ try:
+ raw = document.encode('utf8') if isinstance(document, str) else document
+ except UnicodeError as exc:
+ raise ValueError('Model must be finite UTF-8 JSON.') from exc
+ else:
+ raise ValueError('Model must be a plain dictionary or UTF-8 JSON.')
+ if len(raw) > MAX_MODEL_BYTES:
+ raise ValueError('Model exceeds 256 KiB.')
+ def pairs(items):
+ result = {}
+ for key, value in items:
+ if key in result:
+ raise ValueError('Duplicate model field.')
+ result[key] = value
+ return result
+ def constant(value):
+ raise ValueError('Model must contain finite values.')
+ try:
+ return json.loads(raw, object_pairs_hook=pairs, parse_constant=constant)
+ except (UnicodeError, RecursionError, json.JSONDecodeError) as exc:
+ raise ValueError('Model must be finite UTF-8 JSON.') from exc
+
+
+def validate_model(document):
+ """Return detached bounded data, rejecting unknown fields and false scope."""
+ model = _decode_model(document)
+ _keys(model, _MODEL_KEYS, 'model')
+ expected = {'schema_version': 1, 'kind': KIND, 'license': 'MIT', 'scope': 'simulation-only',
+ 'physical_robot_ready': False, 'is_controller': False, 'predicts_task_success': False,
+ 'sensor_origin': 'privileged-native-simulation', 'control_dt_seconds': CONTROL_DT_SECONDS}
+ for key, value in expected.items():
+ if type(model[key]) is not type(value) or model[key] != value:
+ raise ValueError('Unsupported model kind, version, license, interval or qualification scope.')
+ if canonical(model['feature_contract']) != canonical(FEATURE_CONTRACT) or canonical(model['output_contract']) != canonical(OUTPUT_CONTRACT):
+ raise ValueError('Unsupported causal feature or output contract.')
+ if model['feature_mode'] not in ('action', 'state'):
+ raise ValueError('Unsupported feature mode.')
+ active = 20 if model['feature_mode'] == 'action' else 13
+ _number(model['alpha'], 'ridge alpha', minimum=0)
+ normalization = model['normalization']
+ _keys(normalization, {'means', 'scales', 'constant_features', 'scope'}, 'normalization')
+ if normalization['scope'] != 'fit-training-only':
+ raise ValueError('Normalization must be fitted on training rows only.')
+ _vector(normalization['means'], active, 'normalization means')
+ scales = _vector(normalization['scales'], active, 'normalization scales')
+ if any(value <= 1e-12 for value in scales):
+ raise ValueError('Normalization scales must exceed 1e-12.')
+ constants = normalization['constant_features']
+ if type(constants) is not list or len(constants) != active or any(type(value) is not bool for value in constants):
+ raise ValueError('Constant feature mask must contain explicit booleans.')
+ coefficients = model['coefficients']
+ if type(coefficients) is not list or len(coefficients) != 7:
+ raise ValueError('Coefficients require seven output rows.')
+ for row in coefficients:
+ _vector(row, active, 'coefficients', maximum=1e12)
+ if any(constant and (coefficient != 0 or scale != 1.)
+ for constant, coefficient, scale in zip(constants, row, scales)):
+ raise ValueError('Constant training features must have scale one and zero coefficients.')
+ _vector(model['intercept'], 7, 'intercept')
+ training = model['training']
+ _keys(training, {'rows', 'positive_weight_rows', 'weight_sum', 'rows_sha256', 'sample_weights_sha256'}, 'training')
+ if any(type(training[name]) is not int or not 2 <= training[name] <= MAX_TRAINING_ROWS
+ for name in ('rows', 'positive_weight_rows')) or training['positive_weight_rows'] > training['rows']:
+ raise ValueError('Training counts must be bounded and consistent.')
+ if _number(training['weight_sum'], 'weight sum', minimum=0, maximum=MAX_TRAINING_ROWS*MAX_NUMERIC_MAGNITUDE) <= 0:
+ raise ValueError('Positive training weight required.')
+ _sha(training['rows_sha256'], 'training hash')
+ _sha(training['sample_weights_sha256'], 'sample weights hash')
+ _sha(model['model_sha256'], 'model hash')
+ if model['model_sha256'] != digest({key: value for key, value in model.items() if key != 'model_sha256'}):
+ raise ValueError('Model hash mismatch.')
+ return model
+
+
+def _predict_validated(model, features):
+ active = 20 if model['feature_mode'] == 'action' else 13
+ normalization = model['normalization']
+ normalized = [0. if constant else (value-mean)/scale
+ for value, mean, scale, constant in zip(features[:active], normalization['means'],
+ normalization['scales'], normalization['constant_features'])]
+ try:
+ delta = [math.fsum([offset] + [coefficient*value for coefficient, value in zip(row, normalized)])
+ for row, offset in zip(model['coefficients'], model['intercept'])]
+ except OverflowError as exc:
+ raise ValueError('Prediction arithmetic must remain finite.') from exc
+ return _vector(delta, 7, 'predicted delta', maximum=1e12)
+
+
+def predict_features(model, features20):
+ """Strict standard-library inference; state ablation ignores command values."""
+ model = validate_model(model)
+ features = _vector(features20, 20, 'prediction features')
+ return _predict_validated(model, features)
+
+
+def predict(model, before_snapshot, issued_command):
+ """Predict a boundary delta, abstaining on absent/stale/invalid observations.
+
+This function emits no actuator command. It does not verify latency, the scene,
+hardware compatibility, stopping distance, task completion or contact dynamics.
+"""
+ model = validate_model(model)
+ try:
+ features = _sensor_values(before_snapshot, _REQUIRED_BEFORE) + _issued_values(issued_command)
+ delta = _predict_validated(model, features)
+ except (ValueError, OverflowError) as exc:
+ return {'status': 'abstained', 'reason': str(exc)}
+ return {'status': 'predicted', 'delta': delta, 'control_dt_seconds': CONTROL_DT_SECONDS}
+
+
+def write_model(path, model):
+ """Create a bounded JSON artifact without overwriting any existing path."""
+ model = validate_model(model)
+ raw = canonical(model)+b'\n'
+ if len(raw) > MAX_MODEL_BYTES:
+ raise ValueError('Model exceeds 256 KiB.')
+ with Path(path).open('xb') as output:
+ output.write(raw)
+ return {'path': str(path), 'bytes': len(raw), 'sha256': sha256(raw).hexdigest()}
+
+
+def load_model(path):
+ """Load bounded data; reject links and detect file swaps where IDs exist.
+
+ POSIX retains nonblocking/no-follow flags. Windows uses binary mode and the
+ same pre-open regular-file check; identity comparison is possible only when
+ both path and descriptor stat expose nonzero inode identifiers.
+ """
+ path = Path(path)
+ try:
+ before = path.lstat()
+ if not stat.S_ISREG(before.st_mode) or before.st_size > MAX_MODEL_BYTES:
+ raise ValueError('Model must be a regular non-symlink file at most 256 KiB.')
+ flags = os.O_RDONLY | getattr(os, 'O_NONBLOCK', 0) | getattr(os, 'O_NOFOLLOW', 0) | getattr(os, 'O_BINARY', 0)
+ descriptor = os.open(path, flags)
+ except (OSError, ValueError) as exc:
+ raise ValueError('Model cannot be opened as a regular non-symlink file at most 256 KiB.') from exc
+ try:
+ with os.fdopen(descriptor, 'rb') as source:
+ info = os.fstat(source.fileno())
+ if not stat.S_ISREG(info.st_mode) or info.st_size > MAX_MODEL_BYTES:
+ raise ValueError('Model must be a regular file at most 256 KiB.')
+ before_ino, after_ino = getattr(before, 'st_ino', None), getattr(info, 'st_ino', None)
+ before_dev, after_dev = getattr(before, 'st_dev', None), getattr(info, 'st_dev', None)
+ if before_ino not in (None, 0) and after_ino not in (None, 0) and before_dev is not None and after_dev is not None:
+ if (before_dev, before_ino) != (after_dev, after_ino):
+ raise ValueError('Model file identity changed during opening.')
+ raw = source.read(MAX_MODEL_BYTES+1)
+ except OSError as exc:
+ raise ValueError('Model regular-file read failed.') from exc
+ return validate_model(raw)
diff --git a/src/dvidia_training/transition_pilot.html b/src/dvidia_training/transition_pilot.html
new file mode 100644
index 0000000..9c95a86
--- /dev/null
+++ b/src/dvidia_training/transition_pilot.html
@@ -0,0 +1,62 @@
+
+
+
+
+
+DVIDIA · Actuator Lab
+
+
+
+DVIDIAACTUATOR LAB · LAPTOP PILOT
+
+A small learned movement predictor
+
What changes after a command?
+
This experiment learns the next joint positions and gripper opening from current measurements and a requested movement. It compares the learned result with simple predictors on separate simulation groups.
+Reading retained evidence…
+
+
—Saved action-model size
+
—Warm model computation, p95
+
—Scored test transitions
+
+
Prediction error
Lower error is better. These are next-boundary actuator measurements, not grasp success or shoe placement. Each transition normally spans 20 ms. Longer trajectories and closed-loop control have not been tested here. Transitions are correlated; the held-out evidence comes from two reset groups and six recordings.
Human footage supplies visual examples and task structure. Synchronized robot demonstrations supply movement supervision. This pilot connects episode tapes to a compact predictor; it does not train from shoe videos or send commands to a physical arm.
+
The first shoe task
Move one shoe from a marked pickup area into one marked slot, then verify its position, direction and stable release.
Start with a fixed table and camera, one shoe and tucked laces. The separate collection plan records missing conditions and complete failures. A validated shoe environment, sensed object state, robot adapter and physical trials are still required.
+
Evidence and limits
Current observations come from native simulator state. The model uses only current joint position, velocity, gripper gap and issued targets; authored phase, future state and final applied controls are excluded from its inputs.
Inspect groups, baselines and timing scope
+
+
+
+
+
diff --git a/src/dvidia_training/transition_pilot.py b/src/dvidia_training/transition_pilot.py
new file mode 100644
index 0000000..380af80
--- /dev/null
+++ b/src/dvidia_training/transition_pilot.py
@@ -0,0 +1,628 @@
+"""Group-held-out laptop pilot for the authored 20 ms actuator transition.
+
+This is an offline predictor experiment, not a shoe policy or physical dynamics
+qualification. A group contains three fresh, related simulation resets; groups,
+not individual transitions or sibling interventions, define the split.
+"""
+from __future__ import annotations
+
+from copy import deepcopy
+from hashlib import sha256
+import json
+import math
+import os
+from pathlib import Path
+import platform
+import re
+try:
+ import resource
+except ImportError: # Windows has no process-lifetime getrusage interface.
+ resource = None
+import sys
+from time import perf_counter, perf_counter_ns
+
+from .episode_tape import CASES, MAX_BYTES, SOURCE_FILES, canonical, digest, record_episode, validate_tape, write_tape
+
+CORPUS_KIND = 'dvidia.actuator-transition-corpus.v1'
+REPORT_KIND = 'dvidia.actuator-transition-pilot.v1'
+ALPHAS = (.0001, .01, 1., 100.)
+MAX_MANIFEST_BYTES = 2 * 1024 * 1024
+MAX_REPORT_BYTES = 16 * 1024 * 1024
+MAX_GROUPS = 60
+MAX_TOTAL_TAPE_BYTES = 512 * 1024 * 1024
+MODEL_NAMES = ('action', 'state', 'persistence', 'constant_velocity', 'training_mean')
+SPLITS = ('train', 'development', 'test')
+BASE_CONFIG = {
+ 'timestep': .001, 'control_dt': .02, 'horizon': 18., 'settle_time': .4,
+ 'object_position': [.42, -.04, .31], 'target_position': [.54, .10, .31],
+ 'object_half_size': [.02, .02, .02], 'object_mass': .04, 'object_friction': .8,
+ 'table_height': .29, 'position_tolerance': .025, 'dwell_seconds': .25,
+ 'scene_jitter': .006, 'joint_torque_limit': 40., 'jaw_force_limit': 15.,
+ 'joint_target_rate_limit': 3., 'jaw_width_rate_limit': .16, 'gravity': 9.81,
+}
+LIMITATIONS = [
+ 'Authored rigid-box simulator with privileged joint telemetry; no shoe footage or shoe physics.',
+ 'Predicts the next 20 ms joint-position and jaw-gap deltas, not object motion, contact force or a control policy.',
+ 'Held-out results are a small grouped research test, not independent physical robot qualification.',
+ 'Models use before-state and optionally issued actuator commands; phases, cases and simulator configuration are not inputs.',
+ 'Compute-only inference timings exclude sensing, transport, actuator response and operating-system deadline guarantees.',
+ 'Hashes verify internal consistency, not provenance authenticity or measured physical accuracy.',
+]
+
+
+def _json(path, limit):
+ path = Path(path)
+ if path.is_symlink() or not path.is_file() or path.stat().st_size > limit:
+ raise ValueError('A bounded regular JSON file is required.')
+ with path.open('rb') as stream:
+ raw = stream.read(limit+1)
+ if len(raw) > limit:
+ raise ValueError('JSON file exceeds its byte limit.')
+ def pairs(items):
+ result = {}
+ for key, value in items:
+ if key in result:
+ raise ValueError('Duplicate JSON field.')
+ result[key] = value
+ return result
+ def constant(value):
+ raise ValueError('Nonfinite JSON value.')
+ try:
+ document = json.loads(raw, object_pairs_hook=pairs, parse_constant=constant)
+ canonical(document)
+ except (UnicodeError, RecursionError, json.JSONDecodeError, TypeError, OverflowError) as exc:
+ raise ValueError('Finite UTF-8 JSON is required.') from exc
+ return document, raw
+
+
+def _write(path, document):
+ raw = canonical(document) + b'\n'
+ with Path(path).open('xb') as stream:
+ stream.write(raw)
+ return {'file': Path(path).name, 'sha256': sha256(raw).hexdigest(), 'bytes': len(raw)}
+
+
+def _fresh(output):
+ path = Path(output)
+ if path.exists() or path.is_symlink():
+ raise ValueError('Output already exists; use a fresh directory.')
+ path.mkdir(parents=True)
+ return path
+
+
+def _id(value):
+ return type(value) is str and re.fullmatch(r'[A-Za-z0-9][A-Za-z0-9_.:-]{0,159}', value) is not None
+
+
+def _hash(value):
+ return type(value) is str and re.fullmatch(r'[0-9a-f]{64}', value) is not None
+
+
+def _finite(value):
+ try:
+ return type(value) in (int, float) and math.isfinite(value)
+ except OverflowError:
+ return False
+
+
+def _sources():
+ return {name: sha256(Path(__file__).with_name(name).read_bytes()).hexdigest() for name in sorted(SOURCE_FILES)}
+
+
+def generate_corpus(output, seed_start=311000, groups=12):
+ """Collect all predeclared attempts; failed recordings remain in the manifest."""
+ if type(groups) is not int or not 6 <= groups <= MAX_GROUPS:
+ raise ValueError(f'Use 6–{MAX_GROUPS} reset groups.')
+ if type(seed_start) is not int or not 0 <= seed_start <= 2**31-groups:
+ raise ValueError('Seed start must be a bounded nonnegative integer.')
+ output = _fresh(output)
+ (output/'tapes').mkdir()
+ held = max(1, groups//6)
+ train_groups = groups-2*held
+ source_hashes = _sources()
+ manifest = {
+ 'kind': CORPUS_KIND, 'schema_version': 1, 'scope': 'simulation-only',
+ 'physical_robot_ready': False, 'shoe_skill_ready': False, 'status': 'collecting',
+ 'protocol': {
+ 'revision': 'grouped-native-actuator-20ms-v1', 'cases': list(CASES),
+ 'split_unit': 'predeclared reset group; all sibling interventions stay together',
+ 'variation': 'seeded scene jitter plus predeclared bounded group mass/friction; neither is a model feature',
+ 'split_counts': {'train': train_groups, 'development': held, 'test': held},
+ 'source_files_sha256': source_hashes,
+ },
+ 'groups': [],
+ }
+ for index in range(groups):
+ group_id = f'reset-{index:03d}'
+ split = 'train' if index < train_groups else 'development' if index < train_groups+held else 'test'
+ config = deepcopy(BASE_CONFIG)
+ # Deterministic and declared before simulation, not fitted to observed outcomes.
+ config['object_mass'] = .035 + .010*index/max(1, groups-1)
+ config['object_friction'] = .70 + .20*index/max(1, groups-1)
+ seed = seed_start+index
+ group = {'group_id': group_id, 'split': split, 'seed': seed,
+ 'session_id': f'session-{seed}', 'reset_lineage_id': f'native-reset-{seed}',
+ 'configuration': config, 'configuration_sha256': digest(config), 'attempts': []}
+ for case in CASES:
+ group['attempts'].append({'episode_id': f'{group_id}-{case}',
+ 'source_recording_id': f'recording-{seed}-{case}', 'case': case,
+ 'status': 'planned', 'path': None, 'sha256': None, 'bytes': None,
+ 'error': None, 'recording_elapsed_seconds': None, 'provenance': None, 'outcome': None})
+ manifest['groups'].append(group)
+ # This declaration is immutable evidence of the plan before any attempt starts.
+ _write(output/'collection-plan.json', manifest)
+ start = perf_counter()
+ for group in manifest['groups']:
+ for attempt in group['attempts']:
+ attempt_start = perf_counter()
+ try:
+ from .arm_env import ArmConfig
+ tape = record_episode(episode_id=attempt['episode_id'], source_recording_id=attempt['source_recording_id'],
+ session_id=group['session_id'], seed=group['seed'], case=attempt['case'],
+ config=ArmConfig(**group['configuration']))
+ path = output/'tapes'/f'{attempt["episode_id"]}.json'
+ receipt = write_tape(path, tape)
+ attempt.update(status='completed', path=path.relative_to(output).as_posix(),
+ sha256=receipt['sha256'], bytes=receipt['bytes'],
+ provenance=deepcopy(tape['provenance']), outcome=deepcopy(tape['outcome']))
+ except (OSError, ValueError, ImportError, RuntimeError) as exc:
+ attempt.update(status='failed', error=f'{type(exc).__name__}: {str(exc)[:240]}')
+ attempt['recording_elapsed_seconds'] = perf_counter()-attempt_start
+ failures = sum(a['status'] != 'completed' for g in manifest['groups'] for a in g['attempts'])
+ manifest['status'] = 'completed' if not failures else 'completed-with-failed-attempts'
+ manifest['collection_elapsed_seconds'] = perf_counter()-start
+ _write(output/'corpus.json', manifest)
+ return manifest
+
+
+def _manifest(path):
+ manifest, raw = _json(path, MAX_MANIFEST_BYTES)
+ expected = {'kind', 'schema_version', 'scope', 'physical_robot_ready', 'shoe_skill_ready',
+ 'status', 'protocol', 'groups', 'collection_elapsed_seconds'}
+ if type(manifest) is not dict or set(manifest) != expected or manifest['kind'] != CORPUS_KIND or type(manifest['schema_version']) is not int or manifest['schema_version'] != 1:
+ raise ValueError('Unsupported transition corpus manifest.')
+ if manifest['scope'] != 'simulation-only' or manifest['physical_robot_ready'] is not False or manifest['shoe_skill_ready'] is not False:
+ raise ValueError('Unsupported corpus qualification flags.')
+ if manifest['status'] not in ('completed', 'completed-with-failed-attempts') or not _finite(manifest['collection_elapsed_seconds']) or manifest['collection_elapsed_seconds'] < 0:
+ raise ValueError('Corpus must retain completed or failed recording attempts.')
+ groups = manifest['groups']
+ if type(groups) is not list or not 3 <= len(groups) <= MAX_GROUPS:
+ raise ValueError('Corpus requires bounded train/development/test reset groups.')
+ protocol = manifest['protocol']
+ if type(protocol) is not dict or set(protocol) != {'revision', 'cases', 'split_unit', 'variation', 'split_counts', 'source_files_sha256'} or protocol['revision'] != 'grouped-native-actuator-20ms-v1' or protocol['cases'] != list(CASES):
+ raise ValueError('Unsupported grouped corpus protocol.')
+ if any(type(protocol[k]) is not str or not 1 <= len(protocol[k]) <= 200 for k in ('split_unit', 'variation')):
+ raise ValueError('Bounded corpus protocol descriptions required.')
+ hashes = protocol['source_files_sha256']
+ if type(hashes) is not dict or set(hashes) != SOURCE_FILES or not all(_hash(v) for v in hashes.values()):
+ raise ValueError('Source hashes are required.')
+ counts = {split: 0 for split in SPLITS}
+ lineage = {name: {} for name in ('group', 'seed', 'session', 'lineage', 'reset', 'recording', 'episode', 'tape')}
+ total_bytes = 0
+ paths = set()
+ root = Path(path).resolve().parent
+ def unique(name, value, group_id):
+ previous = lineage[name].get(value)
+ if previous is not None and previous != group_id:
+ raise ValueError(f'Duplicate {name} lineage across reset groups; split independence is invalid.')
+ lineage[name][value] = group_id
+ for group in groups:
+ if type(group) is not dict or set(group) != {'group_id', 'split', 'seed', 'session_id', 'reset_lineage_id', 'configuration', 'configuration_sha256', 'attempts'}:
+ raise ValueError('Unsupported reset-group fields.')
+ group_id, split = group['group_id'], group['split']
+ if not all(_id(group[k]) for k in ('group_id', 'session_id', 'reset_lineage_id')) or split not in SPLITS or type(group['seed']) is not int or not 0 <= group['seed'] <= 2**31:
+ raise ValueError('Bounded reset group identifiers and seed are required.')
+ if group_id in lineage['group']:
+ raise ValueError('Duplicate reset group.')
+ unique('group', group_id, group_id)
+ config = group['configuration']
+ if type(config) is not dict or set(config) != set(BASE_CONFIG) or not _hash(group['configuration_sha256']) or digest(config) != group['configuration_sha256']:
+ raise ValueError('Predeclared configuration hash mismatch.')
+ for name, value in (('seed', group['seed']), ('session', group['session_id']), ('lineage', group['reset_lineage_id']),
+ ('reset', (group['seed'], group['configuration_sha256']))):
+ unique(name, value, group_id)
+ counts[split] += 1
+ attempts = group['attempts']
+ if type(attempts) is not list or len(attempts) != len(CASES) or {a.get('case') for a in attempts if type(a) is dict} != set(CASES):
+ raise ValueError('Every group must retain all three intervention attempts.')
+ for attempt in attempts:
+ if type(attempt) is not dict or set(attempt) != {'episode_id', 'source_recording_id', 'case', 'status', 'path', 'sha256', 'bytes', 'error', 'recording_elapsed_seconds', 'provenance', 'outcome'}:
+ raise ValueError('Unsupported recording-attempt fields.')
+ if not _id(attempt['episode_id']) or not _id(attempt['source_recording_id']) or attempt['status'] not in ('completed', 'failed') or not _finite(attempt['recording_elapsed_seconds']) or attempt['recording_elapsed_seconds'] < 0:
+ raise ValueError('Every attempt needs bounded identifiers, status and elapsed time.')
+ for name, key in (('recording', 'source_recording_id'), ('episode', 'episode_id')):
+ value = attempt[key]
+ if value in lineage[name]:
+ raise ValueError(f'Duplicate {name} identifier or alias.')
+ unique(name, value, group_id)
+ if attempt['status'] == 'failed':
+ if any(attempt[key] is not None for key in ('path', 'sha256', 'bytes', 'provenance', 'outcome')) or type(attempt['error']) is not str or not 1 <= len(attempt['error']) <= 300:
+ raise ValueError('Failed attempt must retain an error and no fabricated tape.')
+ continue
+ if attempt['error'] is not None or not _hash(attempt['sha256']) or type(attempt['bytes']) is not int or not 1 <= attempt['bytes'] <= MAX_BYTES:
+ raise ValueError('Completed recording needs a bounded byte hash.')
+ name = attempt['path']
+ if type(name) is not str or not re.fullmatch(r'tapes/[A-Za-z0-9][A-Za-z0-9_.:-]{0,159}\.json', name) or name in paths:
+ raise ValueError('Tape must have a unique, relative corpus path.')
+ paths.add(name)
+ tape_path = root/name
+ # No symlink at any component and no escaping the corpus root.
+ if any(p.is_symlink() for p in (root/'tapes', tape_path)) or not tape_path.is_file() or tape_path.resolve().parent != (root/'tapes').resolve():
+ raise ValueError('Tape path must remain a regular file inside the corpus.')
+ if tape_path.stat().st_size != attempt['bytes']:
+ raise ValueError('Tape byte count differs from manifest.')
+ total_bytes += attempt['bytes']
+ if total_bytes > MAX_TOTAL_TAPE_BYTES:
+ raise ValueError('Corpus exceeds the 512 MiB tape budget.')
+ provenance = attempt['provenance']
+ if type(provenance) is not dict or provenance.get('episode_id') != attempt['episode_id'] or provenance.get('source_recording_id') != attempt['source_recording_id'] or provenance.get('session_id') != group['session_id'] or provenance.get('seed') != group['seed'] or provenance.get('case') != attempt['case'] or provenance.get('configuration_sha256') != group['configuration_sha256'] or provenance.get('source_files_sha256') != hashes or not _hash(provenance.get('scene_sha256')):
+ raise ValueError('Recording provenance differs from its declared reset group.')
+ unique('tape', attempt['sha256'], group_id)
+ if type(protocol['split_counts']) is not dict or protocol['split_counts'] != counts or any(count == 0 for count in counts.values()):
+ raise ValueError('Split group counts must match predeclared groups.')
+ failed = sum(a['status'] == 'failed' for g in groups for a in g['attempts'])
+ if (manifest['status'] == 'completed') != (failed == 0):
+ raise ValueError('Corpus status disagrees with failed attempt denominator.')
+ return manifest, raw, root
+
+
+def _load_attempt(root, group, attempt):
+ path = root/attempt['path']
+ with path.open('rb') as stream:
+ raw = stream.read(MAX_BYTES+1)
+ if len(raw) != attempt['bytes'] or sha256(raw).hexdigest() != attempt['sha256']:
+ raise ValueError('Tape source hash differs from manifest.')
+ tape = validate_tape(raw)
+ if tape['provenance'] != attempt['provenance'] or tape['configuration'] != group['configuration'] or tape['outcome'] != attempt['outcome']:
+ raise ValueError('Tape content differs from declared corpus provenance or outcome.')
+ return tape
+
+
+def _metrics(rows, predictions):
+ if len(rows) != len(predictions):
+ raise ValueError('Predictions must retain every evaluation row.')
+ if not rows:
+ return {'rows': 0, 'joint_rmse_rad': None, 'jaw_rmse_m': None, 'joint_mae_rad': None,
+ 'jaw_mae_m': None, 'joint_max_abs_error_rad': None, 'jaw_max_abs_error_m': None}
+ errors = [[p-t for p, t in zip(prediction, row['target_delta'])] for row, prediction in zip(rows, predictions)]
+ if any(len(p) != 7 or not all(_finite(x) for x in p) for p in predictions):
+ raise ValueError('Finite seven-output predictions required.')
+ joint = [abs(x) for error in errors for x in error[:6]]
+ jaw = [abs(error[6]) for error in errors]
+ return {'rows': len(rows), 'joint_rmse_rad': math.sqrt(sum(x*x for x in joint)/len(joint)),
+ 'jaw_rmse_m': math.sqrt(sum(x*x for x in jaw)/len(jaw)),
+ 'joint_mae_rad': sum(joint)/len(joint), 'jaw_mae_m': sum(jaw)/len(jaw),
+ 'joint_max_abs_error_rad': max(joint), 'jaw_max_abs_error_m': max(jaw)}
+
+
+def _score(metric):
+ if not metric['rows']:
+ raise ValueError('Development rows are required for fixed alpha selection.')
+ return metric['joint_rmse_rad']/.01 + metric['jaw_rmse_m']/.001
+
+
+def _weights(rows):
+ counts = {}
+ for row in rows:
+ counts[row['episode_id']] = counts.get(row['episode_id'], 0)+1
+ # Each retained episode contributes the same total, independent of its length.
+ return [len(rows)/(len(counts)*counts[row['episode_id']]) for row in rows]
+
+
+def _predictions(name, model, rows, mean):
+ from .transition_model import predict_features
+ if name in ('action', 'state'):
+ return [predict_features(model, row['features']) for row in rows]
+ if name == 'constant_velocity':
+ return [[.02*x for x in row['features'][6:12]]+[0.] for row in rows]
+ if name == 'training_mean':
+ return [mean[:] for row in rows]
+ return [[0.]*7 for row in rows]
+
+
+def _quantile(values, q):
+ values = sorted(values)
+ coordinate = (len(values)-1)*q
+ low = int(coordinate)
+ high = min(low+1, len(values)-1)
+ return values[low]+(values[high]-values[low])*(coordinate-low)
+
+
+def _timing(model, rows):
+ from .transition_model import predict_features
+ for index in range(100):
+ predict_features(model, rows[index % len(rows)]['features'])
+ times = []
+ for index in range(1000):
+ features = rows[index % len(rows)]['features']
+ start = perf_counter_ns()
+ predict_features(model, features)
+ times.append(perf_counter_ns()-start)
+ return {'warmup_calls': 100, 'measured_calls': 1000, 'p50_microseconds': _quantile(times, .5)/1000,
+ 'p95_microseconds': _quantile(times, .95)/1000, 'worst_microseconds': max(times)/1000,
+ 'scope': 'compute-only predict_features; pre-extracted held-out input vectors; no sensing or transport'}
+
+
+def _matched_branch(rows_by_case, models, mean):
+ from .transition_model import predict_features
+ nominal = { (row['group_id'], row['index']): row for row in rows_by_case.get('nominal', []) }
+ for changed in rows_by_case.get('release-after-lift', []):
+ reference = nominal.get((changed['group_id'], changed['index']))
+ if reference is None or reference['features'][:13] != changed['features'][:13] or reference['features'][13:19] != changed['features'][13:19] or reference['features'][19] == changed['features'][19]:
+ continue
+ result = {'status': 'matched', 'group_id': changed['group_id'], 'index': changed['index'],
+ 'before_time_seconds': changed['before_time_seconds'], 'before_state_identical': True,
+ 'joint_command_identical': True, 'issued_jaw_gap_m': {'nominal': reference['features'][19], 'release': changed['features'][19]},
+ 'observed_delta': {'nominal': reference['target_delta'], 'release': changed['target_delta']},
+ 'prediction_delta': {}}
+ for name, model in models.items():
+ result['prediction_delta'][name] = {'nominal': predict_features(model, reference['features']), 'release': predict_features(model, changed['features'])}
+ return result
+ return {'status': 'unavailable', 'reason': 'No retained nominal/release transition has exactly identical before-state and joint command with a changed issued jaw command.'}
+
+
+def _hardware():
+ import numpy
+ rss = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss if resource is not None else None
+ multiplier = 1 if sys.platform == 'darwin' else 1024
+ return {'platform': platform.platform(), 'machine': platform.machine(), 'processor': platform.processor() or platform.machine(),
+ 'logical_cpu_count': os.cpu_count(), 'python': platform.python_version(), 'numpy': numpy.__version__,
+ 'gpu_used': False, 'process_lifetime_peak_rss_bytes': int(rss*multiplier) if rss is not None else None,
+ 'memory_scope': 'process-lifetime high-water mark; includes earlier work in this process; not isolated learner memory' if rss is not None else 'process-lifetime RSS unavailable on this platform'}
+
+
+def run_pilot(manifest_path, output):
+ """Fit on train, select on development, freeze artifacts, then parse test tapes."""
+ output = _fresh(output)
+ start = perf_counter()
+ report = {'kind': REPORT_KIND, 'schema_version': 1, 'scope': 'simulation-only',
+ 'physical_robot_ready': False, 'shoe_skill_ready': False, 'status': 'running',
+ 'attempts': [], 'counts': {}, 'metrics': {}, 'model_artifacts': {},
+ 'freeze': None, 'hardware': None, 'timings': {}, 'limitations': LIMITATIONS[:],
+ 'qualification': 'unqualified research pilot; internal consistency only', 'stage': 'manifest-validation'}
+ try:
+ from .transition_model import extract_rows, fit_model, load_model, predict_features, write_model
+ manifest, manifest_raw, corpus_root = _manifest(manifest_path)
+ report['corpus_manifest_sha256'] = sha256(manifest_raw).hexdigest()
+ with (output/'corpus-manifest.json').open('xb') as stream:
+ stream.write(manifest_raw)
+ manifest_receipt = {'file': 'corpus-manifest.json', 'sha256': sha256(manifest_raw).hexdigest(), 'bytes': len(manifest_raw)}
+ report['source_corpus'] = {'snapshot': manifest_receipt, 'original_tapes_included': False,
+ 'original_manifest_sha256': report['corpus_manifest_sha256']}
+ report['protocol'] = {'control_dt_seconds': .02, 'target': 'next actuator joint-position and jaw-gap deltas',
+ 'alpha_grid': list(ALPHAS), 'development_score': 'joint_rmse_rad/0.01 + jaw_rmse_m/0.001',
+ 'selection_tie_break': 'first alpha in ascending fixed grid', 'training_weighting': 'equal total weight per retained training episode',
+ 'normalization': 'training rows only; model core records means/scales', 'test_used_for_selection': False,
+ 'feature_modes': {'action': 'before q/qdot/jaw + issued joint/jaw targets', 'state': 'before q/qdot/jaw only'},
+ 'excluded_inputs': ['case', 'phase', 'simulation configuration', 'future state', 'applied end-of-interval controls', 'object state'],
+ 'slice_definitions': {'moving': 'any measured abs joint delta > 0.0001 rad', 'jaw_changing': 'abs measured jaw delta > 0.00001 m'}}
+ report['counts'] = {'groups': {split: sum(g['split'] == split for g in manifest['groups']) for split in SPLITS},
+ 'attempts': len(manifest['groups'])*len(CASES), 'completed_recordings': 0, 'failed_recordings': 0,
+ 'rows': {split: 0 for split in SPLITS}, 'excluded_steps': {split: 0 for split in SPLITS},
+ 'attempted_steps': {split: 0 for split in SPLITS}, 'simulation_successes': 0, 'simulation_unsuccessful': 0}
+ by_split = {split: [] for split in SPLITS}
+ load_seconds = {split: 0. for split in SPLITS}
+ extraction_seconds = {split: 0. for split in SPLITS}
+ receipt_by_episode = {}
+ for group in manifest['groups']:
+ for attempt in group['attempts']:
+ receipt = {'group_id': group['group_id'], 'split': group['split'], 'episode_id': attempt['episode_id'],
+ 'case': attempt['case'], 'recording_status': attempt['status'], 'source_sha256': attempt['sha256'],
+ 'source_bytes': attempt['bytes'], 'recording_error': attempt['error'],
+ 'simulation_outcome': deepcopy(attempt['outcome']), 'rows': 0, 'attempted_steps': 0, 'exclusions': [],
+ 'validation_status': 'not_checked' if attempt['status'] == 'completed' else 'failed_recording'}
+ report['attempts'].append(receipt)
+ receipt_by_episode[attempt['episode_id']] = receipt
+ report['counts']['completed_recordings' if attempt['status'] == 'completed' else 'failed_recordings'] += 1
+ if attempt['outcome'] is not None:
+ report['counts']['simulation_successes' if attempt['outcome']['result']['success'] else 'simulation_unsuccessful'] += 1
+ def consume(split):
+ for group in manifest['groups']:
+ if group['split'] != split:
+ continue
+ for attempt in group['attempts']:
+ receipt = receipt_by_episode[attempt['episode_id']]
+ if attempt['status'] == 'failed':
+ continue
+ load_start = perf_counter()
+ try:
+ tape = _load_attempt(corpus_root, group, attempt)
+ except (OSError, ValueError) as exc:
+ receipt.update(validation_status='error', validation_error=str(exc)[:200])
+ raise
+ receipt['validation_status'] = 'validated'
+ load_seconds[split] += perf_counter()-load_start
+ extraction_start = perf_counter()
+ extracted = extract_rows(tape)
+ extraction_seconds[split] += perf_counter()-extraction_start
+ for row in extracted['rows']:
+ row['group_id'] = group['group_id']
+ by_split[split].extend(extracted['rows'])
+ receipt.update(rows=len(extracted['rows']), attempted_steps=extracted['attempted_steps'], exclusions=extracted['exclusions'])
+ report['counts']['rows'][split] += len(extracted['rows'])
+ report['counts']['excluded_steps'][split] += len(extracted['exclusions'])
+ report['counts']['attempted_steps'][split] += extracted['attempted_steps']
+ report['stage'] = 'train-development-loading'
+ consume('train')
+ consume('development')
+ train, development = by_split['train'], by_split['development']
+ if not train or not development:
+ raise ValueError('Both training and development require retained valid 20 ms rows.')
+ weights = _weights(train)
+ mean = [sum(w*row['target_delta'][i] for w, row in zip(weights, train))/sum(weights) for i in range(7)]
+ report['training_mean_delta'] = mean
+ models, selections = {}, {}
+ fit_start = perf_counter()
+ report['stage'] = 'training-selection'
+ for mode in ('action', 'state'):
+ trials = []
+ candidates = []
+ for alpha in ALPHAS:
+ model = fit_model([{key: value for key, value in row.items() if key != 'group_id'} for row in train], alpha=alpha, feature_mode=mode, sample_weights=weights)
+ metric = _metrics(development, [predict_features(model, row['features']) for row in development])
+ score = _score(metric)
+ trials.append({'alpha': alpha, 'development_score': score, 'metrics': metric})
+ candidates.append(model)
+ best = min(range(len(trials)), key=lambda i: trials[i]['development_score'])
+ models[mode] = candidates[best]
+ selections[mode] = {'alpha': trials[best]['alpha'], 'trials': trials}
+ report['timings']['fit_and_development_selection_seconds'] = perf_counter()-fit_start
+ report['selection'] = selections
+ report['stage'] = 'freeze-before-test'
+ for mode, model in models.items():
+ path = output/f'{mode}.model.json'
+ write_model(path, model)
+ raw = path.read_bytes()
+ report['model_artifacts'][mode] = {'file': path.name, 'sha256': sha256(raw).hexdigest(), 'bytes': len(raw)}
+ freeze = {'kind': 'dvidia.actuator-transition-freeze.v1', 'schema_version': 1,
+ 'corpus_manifest_sha256': report['corpus_manifest_sha256'], 'models': deepcopy(report['model_artifacts']),
+ 'selection': deepcopy(selections), 'training_mean_delta': mean,
+ 'held_out_transition_targets_read': False, 'elapsed_seconds_at_freeze': perf_counter()-start,
+ 'training_sources': [a['sha256'] for g in manifest['groups'] if g['split'] == 'train' for a in g['attempts'] if a['status'] == 'completed'],
+ 'development_sources': [a['sha256'] for g in manifest['groups'] if g['split'] == 'development' for a in g['attempts'] if a['status'] == 'completed'],
+ 'test_sources': [a['sha256'] for g in manifest['groups'] if g['split'] == 'test' for a in g['attempts'] if a['status'] == 'completed']}
+ report['freeze'] = _write(output/'freeze.json', freeze)
+ # Only now may held-out transition targets be parsed. Manifest terminal outcomes were visible during preflight. Reload the frozen model files.
+ frozen = {mode: load_model(output/artifact['file']) for mode, artifact in report['model_artifacts'].items()}
+ report['stage'] = 'held-out-evaluation'
+ report['timings']['held_out_loading_started_elapsed_seconds'] = perf_counter()-start
+ consume('test')
+ test = by_split['test']
+ if not test:
+ raise ValueError('Held-out split contains no valid 20 ms transition rows.')
+ report['metrics']['test'] = {}
+ report['metrics']['development'] = {}
+ predictions = {}
+ for name in MODEL_NAMES:
+ model = frozen.get(name)
+ predictions[name] = _predictions(name, model, test, mean)
+ report['metrics']['test'][name] = _metrics(test, predictions[name])
+ report['metrics']['development'][name] = _metrics(development, _predictions(name, model, development, mean))
+ def grouped(key):
+ result = {}
+ for value in sorted({row[key] for row in test}):
+ indices = [i for i, row in enumerate(test) if row[key] == value]
+ result[value] = {name: _metrics([test[i] for i in indices], [predictions[name][i] for i in indices]) for name in MODEL_NAMES}
+ return result
+ report['metrics']['per_case'] = grouped('case')
+ report['metrics']['per_group'] = grouped('group_id')
+ report['metrics']['slices'] = {}
+ masks = {'moving': lambda row: any(abs(x) > .0001 for x in row['target_delta'][:6]),
+ 'jaw_changing': lambda row: abs(row['target_delta'][6]) > .00001}
+ for slice_name, predicate in masks.items():
+ indices = [i for i, row in enumerate(test) if predicate(row)]
+ report['metrics']['slices'][slice_name] = {name: _metrics([test[i] for i in indices], [predictions[name][i] for i in indices]) for name in MODEL_NAMES}
+ by_case = {case: [row for row in test if row['case'] == case] for case in CASES}
+ report['matched_branch'] = _matched_branch(by_case, frozen, mean)
+ report['timings']['inference'] = {mode: _timing(model, test) for mode, model in frozen.items()}
+ report['timings']['source_load_and_validation_seconds'] = load_seconds
+ report['timings']['feature_and_target_extraction_seconds'] = extraction_seconds
+ report['timings']['source_collection_seconds'] = manifest['collection_elapsed_seconds']
+ report['counts']['case_attempts'] = {case: sum(a['case'] == case for a in report['attempts']) for case in CASES}
+ report['counts']['case_rows'] = {case: len(by_case[case]) for case in CASES}
+ report['artifact_bytes'] = sum(a['bytes'] for a in report['model_artifacts'].values())
+ report['hardware'] = _hardware()
+ report['status'], report['stage'] = 'completed', 'complete'
+ except (OSError, ValueError, ImportError, RuntimeError) as exc:
+ report['status'] = 'error'
+ report['error'] = f'{type(exc).__name__}: {str(exc)[:300]}'
+ report['timings']['pipeline_elapsed_seconds_before_final_receipt'] = perf_counter()-start
+ report_receipt = _write(output/'report.json', report)
+ artifacts = {report_receipt['file']: report_receipt}
+ for name in ('corpus-manifest.json', 'freeze.json', 'action.model.json', 'state.model.json'):
+ path = output/name
+ if path.is_file():
+ raw = path.read_bytes()
+ artifacts[name] = {'file': name, 'sha256': sha256(raw).hexdigest(), 'bytes': len(raw)}
+ _write(output/'integrity.json', {'kind': 'dvidia.actuator-transition-integrity.v1', 'schema_version': 1,
+ 'artifacts': artifacts, 'authenticity_verified': False, 'accuracy_verified': False})
+ return report
+
+
+def _inspect_pilot(output_dir):
+ """Check local artifact consistency without upgrading it to accuracy evidence."""
+ from .transition_model import load_model
+ root = Path(output_dir)
+ if root.is_symlink() or not root.is_dir():
+ raise ValueError('Pilot output must be a regular directory.')
+ integrity, _ = _json(root/'integrity.json', MAX_MANIFEST_BYTES)
+ if type(integrity) is not dict or set(integrity) != {'kind', 'schema_version', 'artifacts', 'authenticity_verified', 'accuracy_verified'} or integrity['kind'] != 'dvidia.actuator-transition-integrity.v1' or integrity['schema_version'] != 1 or integrity['authenticity_verified'] is not False or integrity['accuracy_verified'] is not False:
+ raise ValueError('Unsupported pilot integrity receipt.')
+ allowed = {'report.json', 'corpus-manifest.json', 'freeze.json', 'action.model.json', 'state.model.json'}
+ artifacts = integrity['artifacts']
+ if type(artifacts) is not dict or not set(artifacts) <= allowed or 'report.json' not in artifacts:
+ raise ValueError('Unsupported pilot artifact paths.')
+ for name, artifact in artifacts.items():
+ if type(artifact) is not dict or set(artifact) != {'file', 'sha256', 'bytes'} or artifact['file'] != name or not _hash(artifact['sha256']) or type(artifact['bytes']) is not int or not 0 < artifact['bytes'] <= MAX_REPORT_BYTES:
+ raise ValueError('Bounded artifact receipts are required.')
+ _, raw = _json(root/name, MAX_REPORT_BYTES)
+ if len(raw) != artifact['bytes'] or sha256(raw).hexdigest() != artifact['sha256']:
+ raise ValueError('Pilot artifact hash mismatch.')
+ report, _ = _json(root/'report.json', MAX_REPORT_BYTES)
+ if type(report) is not dict or report.get('kind') != REPORT_KIND or report.get('schema_version') != 1 or report.get('physical_robot_ready') is not False or report.get('shoe_skill_ready') is not False or report.get('status') not in ('completed', 'error'):
+ raise ValueError('Unsupported pilot report.')
+ if report['status'] == 'completed':
+ if set(artifacts) != allowed or set(report.get('model_artifacts', {})) != {'action', 'state'}:
+ raise ValueError('Completed pilot must include both frozen models and corpus snapshot.')
+ freeze, _ = _json(root/'freeze.json', MAX_REPORT_BYTES)
+ if freeze.get('kind') != 'dvidia.actuator-transition-freeze.v1' or freeze.get('held_out_transition_targets_read') is not False or freeze.get('models') != report['model_artifacts'] or freeze.get('selection') != report.get('selection') or freeze.get('training_mean_delta') != report.get('training_mean_delta') or freeze.get('corpus_manifest_sha256') != report.get('corpus_manifest_sha256'):
+ raise ValueError('Frozen selection differs from the evaluation report.')
+ if report.get('freeze') != artifacts['freeze.json'] or report['source_corpus']['snapshot'] != artifacts['corpus-manifest.json'] or report['source_corpus']['original_tapes_included'] is not False:
+ raise ValueError('Pilot frozen or corpus receipt differs from retained artifacts.')
+ for mode in ('action', 'state'):
+ artifact = report['model_artifacts'][mode]
+ if artifact != artifacts[f'{mode}.model.json']:
+ raise ValueError('Model artifact differs from the freeze receipt.')
+ model = load_model(root/artifact['file'])
+ if model['feature_mode'] != mode or model['alpha'] != report['selection'][mode]['alpha'] or model['training']['rows'] != report['counts']['rows']['train']:
+ raise ValueError('Frozen model disagrees with selected mode or alpha.')
+ manifest, _ = _json(root/'corpus-manifest.json', MAX_MANIFEST_BYTES)
+ expected_groups = {split: sum(g['split'] == split for g in manifest['groups']) for split in SPLITS}
+ attempts = report['attempts']
+ if len(attempts) != len(manifest['groups'])*len(CASES) or report['counts']['groups'] != expected_groups:
+ raise ValueError('Pilot recording denominator differs from retained corpus.')
+ if report['corpus_manifest_sha256'] != artifacts['corpus-manifest.json']['sha256'] or report['source_corpus']['original_manifest_sha256'] != report['corpus_manifest_sha256']:
+ raise ValueError('Original corpus manifest hash differs from its retained byte snapshot.')
+ declared = {a['episode_id']: (g, a) for g in manifest['groups'] for a in g['attempts']}
+ if len(declared) != len(attempts) or {a['episode_id'] for a in attempts} != set(declared):
+ raise ValueError('Recording attempt identifiers differ from the corpus.')
+ for receipt in attempts:
+ group, source = declared[receipt['episode_id']]
+ expected = {'group_id': group['group_id'], 'split': group['split'], 'case': source['case'],
+ 'recording_status': source['status'], 'source_sha256': source['sha256'],
+ 'source_bytes': source['bytes'], 'recording_error': source['error'], 'simulation_outcome': source['outcome']}
+ if any(receipt.get(key) != value for key, value in expected.items()):
+ raise ValueError('Attempt receipt differs from the retained corpus declaration.')
+ if type(receipt.get('rows')) is not int or receipt['rows'] < 0 or type(receipt.get('attempted_steps')) is not int or receipt['attempted_steps'] < 0 or type(receipt.get('exclusions')) is not list or receipt['rows']+len(receipt['exclusions']) != receipt['attempted_steps']:
+ raise ValueError('Attempt rows and exclusions must cover every attempted step.')
+ expected_validation = 'validated' if source['status'] == 'completed' else 'failed_recording'
+ if receipt.get('validation_status') != expected_validation:
+ raise ValueError('Completed pilot contains an unchecked source recording.')
+ for split, key in (('train', 'training_sources'), ('development', 'development_sources'), ('test', 'test_sources')):
+ expected = [a['sha256'] for g in manifest['groups'] if g['split'] == split for a in g['attempts'] if a['status'] == 'completed']
+ if freeze.get(key) != expected:
+ raise ValueError('Freeze source hashes differ from the retained grouped corpus.')
+ completed = sum(a['recording_status'] == 'completed' for a in attempts)
+ if report['counts']['attempts'] != len(attempts) or report['counts']['completed_recordings'] != completed or report['counts']['failed_recordings'] != len(attempts)-completed:
+ raise ValueError('Failed recording denominators are inconsistent.')
+ for split in SPLITS:
+ subset = [a for a in attempts if a['split'] == split]
+ rows = sum(a['rows'] for a in subset)
+ if report['counts']['rows'][split] != rows or report['counts']['excluded_steps'][split] != sum(len(a['exclusions']) for a in subset) or report['counts']['attempted_steps'][split] != sum(a['attempted_steps'] for a in subset):
+ raise ValueError('Pilot row and exclusion counts are inconsistent.')
+ if split in ('test', 'development') and any(metric['rows'] != rows for metric in report['metrics'][split].values()):
+ raise ValueError('Evaluation metrics omit retained rows.')
+ if freeze['elapsed_seconds_at_freeze'] > report['timings']['held_out_loading_started_elapsed_seconds']:
+ raise ValueError('Held-out loading preceded the frozen selection.')
+ return {'kind': 'dvidia.actuator-transition-pilot-inspection.v1', 'status': report['status'],
+ 'integrity': 'verified', 'verification_scope': 'internal artifact consistency only; original source tapes not included',
+ 'authenticity_verified': False, 'accuracy_verified': False, 'physical_robot_ready': False,
+ 'shoe_skill_ready': False, 'counts': report.get('counts', {}), 'artifact_count': len(artifacts)}
+
+
+def inspect_pilot(output_dir):
+ """Return bounded internal-consistency evidence; malformed data fails cleanly."""
+ try:
+ return _inspect_pilot(output_dir)
+ except (KeyError, IndexError, TypeError, AttributeError, OverflowError) as exc:
+ raise ValueError('Malformed pilot artifact schema.') from exc
diff --git a/tests/test_transition_cli.py b/tests/test_transition_cli.py
new file mode 100644
index 0000000..b8f47ab
--- /dev/null
+++ b/tests/test_transition_cli.py
@@ -0,0 +1,73 @@
+"""Offline receipt embedding, entry-point isolation and retained evidence."""
+import json
+from pathlib import Path
+import re
+import subprocess
+import sys
+import tempfile
+import unittest
+
+from dvidia_training.transition_cli import write_report_page
+
+
+class TransitionReportTests(unittest.TestCase):
+ def test_page_embeds_exact_receipt_without_executing_label_markup(self):
+ with tempfile.TemporaryDirectory() as directory:
+ root = Path(directory)
+ document = {'kind': 'dvidia.actuator-transition-pilot.v1', 'status': 'completed',
+ 'limitations': [''],
+ 'metrics': {'test': {}}}
+ raw = json.dumps(document, indent=2) + '\n'
+ (root/'report.json').write_bytes(raw.encode('utf8'))
+ write_report_page(root)
+ html = (root/'report.html').read_text(encoding='utf8')
+ embedded = re.search(r'', html, re.S).group(1)
+ self.assertEqual(json.loads(embedded), raw)
+ self.assertNotIn('',html,re.S).group(1)
+ self.assertEqual(json.loads(embedded).encode('utf8'),(root/'report.json').read_bytes())
+
+ def test_wrong_kind_and_nonfinite_receipts_are_not_rendered(self):
+ with tempfile.TemporaryDirectory() as directory:
+ root = Path(directory)
+ for raw in ('{}', '{"kind":"dvidia.actuator-transition-pilot.v1","x":NaN}'):
+ (root/'report.json').write_bytes(raw.encode('utf8'))
+ with self.assertRaises(ValueError):
+ write_report_page(root)
+ self.assertFalse((root/'report.html').exists())
+
+ def test_help_does_not_import_training_or_native_runtime(self):
+ code = '''import builtins
+original = builtins.__import__
+def guarded(name, *args, **kwargs):
+ if name.split('.')[0] in ('numpy', 'mujoco'):
+ raise RuntimeError('Runtime imported during help')
+ return original(name, *args, **kwargs)
+builtins.__import__ = guarded
+from dvidia_training.transition_cli import main
+main(['--help'])
+'''
+ result = subprocess.run([sys.executable, '-c', code], cwd=Path(__file__).resolve().parents[1],
+ capture_output=True, text=True, timeout=10)
+ self.assertEqual(result.returncode, 0, result.stderr)
+ self.assertIn('corpus', result.stdout)
+
+
+if __name__ == '__main__':
+ unittest.main()
diff --git a/tests/test_transition_model.py b/tests/test_transition_model.py
new file mode 100644
index 0000000..2d6ba33
--- /dev/null
+++ b/tests/test_transition_model.py
@@ -0,0 +1,394 @@
+"""Causal boundaries, action sensitivity and bounded portable model artifacts."""
+from copy import deepcopy
+from hashlib import sha256
+import json
+import os
+from pathlib import Path
+import subprocess
+import sys
+from tempfile import TemporaryDirectory
+from types import SimpleNamespace
+from unittest.mock import patch
+import unittest
+
+from dvidia_training.episode_tape import canonical, digest, validate_tape
+from dvidia_training.transition_model import (
+ CONTROL_DT_SECONDS, FEATURE_NAMES, MAX_MODEL_BYTES, OUTPUT_NAMES,
+ extract_rows, fit_model, load_model, predict, predict_features,
+ validate_model, write_model,
+)
+from tests.test_episode_tape import fixture
+
+
+def learning_rows():
+ """Matched states with independently varied issued joint and jaw commands."""
+ rows = []
+ for joint in (-.5, 0., .5):
+ for gap in (0., .04, .08):
+ for velocity in (-.1, .1):
+ features = [0.]*6 + [velocity]*6 + [.04] + [joint]*6 + [gap]
+ target = [.012*joint-.003*velocity]*6 + [.1*(gap-.04)]
+ rows.append({'features': features, 'target_delta': target})
+ return rows
+
+
+def multi_step_fixture(count=7):
+ tape = fixture()
+ template = deepcopy(tape['steps'][0])
+ tape['configuration']['horizon'] = count*.02
+ tape['provenance']['configuration_sha256'] = digest(tape['configuration'])
+ before = deepcopy(tape['initial'])
+ tape['steps'] = []
+ for index in range(count):
+ step = deepcopy(template)
+ end = (index+1)*.02
+ after = deepcopy(tape['initial'])
+ after['timestamp_seconds'] = end
+ for sensor in after['signals'].values():
+ sensor['captured_at_seconds'] = end
+ step.update(index=index, before=deepcopy(before), after=after)
+ step['command']['issued_at_seconds'] = before['timestamp_seconds']
+ step['applied'].update(interval_start_seconds=before['timestamp_seconds'], interval_end_seconds=end)
+ if index < count-1:
+ step['result'].update(truncated=False, reason='running')
+ tape['steps'].append(step)
+ before = after
+ tape['outcome'].update(result=deepcopy(tape['steps'][-1]['result']),
+ simulation_seconds=count*.02, control_steps=count)
+ return validate_tape(tape)
+
+
+def rehash(model):
+ model['model_sha256'] = digest({key: value for key, value in model.items() if key != 'model_sha256'})
+ return model
+
+
+class TransitionExtractionTests(unittest.TestCase):
+ def test_exact_input_and_output_contract_excludes_privileged_scene_and_phase(self):
+ tape = fixture()
+ step = tape['steps'][0]
+ step['after']['signals']['joint_position']['value'][0] = .002
+ step['after']['signals']['gripper_width']['value'] = .075
+ result = extract_rows(tape)
+ self.assertEqual(result['attempted_steps'], 1)
+ self.assertEqual(result['exclusions'], [])
+ row = result['rows'][0]
+ self.assertEqual(row['features'], [0.]*12+[.08]+[0.]*6+[.08])
+ self.assertEqual(row['target_delta'][:6], [.002]+[0.]*5)
+ self.assertAlmostEqual(row['target_delta'][6], -.005)
+ self.assertEqual((len(FEATURE_NAMES), len(OUTPUT_NAMES)), (20, 7))
+ self.assertEqual(CONTROL_DT_SECONDS, .02)
+ self.assertFalse(any(token in name for name in FEATURE_NAMES
+ for token in ('phase', 'case', 'seed', 'time', 'object', 'contact', 'after', 'applied')))
+
+ def test_future_labels_object_state_and_applied_controls_do_not_change_inputs(self):
+ tape = fixture()
+ original = extract_rows(tape)['rows'][0]
+ changed = deepcopy(tape)
+ step = changed['steps'][0]
+ step['labels'].update(action_phase='recover_up', next_phase='verify')
+ step['after']['signals']['joint_position']['value'] = [.01]*6
+ step['after']['signals']['object_position']['value'] = [.45, .1, .33]
+ step['after']['signals']['grasp_contacts']['value'] = [1, 1]
+ step['applied']['native_position_controls'][0] = .01
+ step['applied']['ramped_joint_targets'][0] = .01
+ new = extract_rows(changed)['rows'][0]
+ self.assertEqual(new['features'], original['features'])
+ self.assertNotEqual(new['target_delta'], original['target_delta'])
+
+ def test_actual_issued_command_not_policy_intent_enters_input(self):
+ tape = fixture()
+ tape['provenance']['case'] = 'open-jaw'
+ tape['steps'][0]['command'].update(intervention='open-jaw')
+ tape['steps'][0]['command']['policy_requested']['gripper_width'] = 0.
+ row = extract_rows(tape)['rows'][0]
+ self.assertEqual(row['features'][19], .08)
+
+ def test_required_missing_signals_excluded_with_reason_without_zero_fill(self):
+ for boundary, name in [('before', 'joint_position'), ('before', 'joint_velocity'),
+ ('before', 'gripper_width'), ('after', 'joint_position'),
+ ('after', 'gripper_width')]:
+ tape = fixture()
+ missing = {'status': 'missing', 'captured_at_seconds': None, 'age_seconds': None, 'value': None}
+ tape['steps'][0][boundary]['signals'][name] = deepcopy(missing)
+ if boundary == 'before':
+ tape['initial']['signals'][name] = deepcopy(missing)
+ result = extract_rows(tape)
+ with self.subTest(boundary=boundary, name=name):
+ self.assertEqual(result['rows'], [])
+ self.assertEqual(len(result['exclusions']), 1)
+ self.assertIn(f'missing:{name}', result['exclusions'][0]['reason'])
+
+ def test_unrequired_missing_contacts_do_not_block_boundary(self):
+ tape = fixture()
+ for snapshot in (tape['initial'], tape['steps'][0]['before'], tape['steps'][0]['after']):
+ snapshot['signals']['grasp_contacts'] = {
+ 'status': 'missing', 'captured_at_seconds': None, 'age_seconds': None, 'value': None}
+ self.assertEqual(len(extract_rows(tape)['rows']), 1)
+
+ def test_stale_required_before_and_after_excluded(self):
+ tape = multi_step_fixture()
+ stale = {'status': 'stale', 'captured_at_seconds': 0., 'age_seconds': .12, 'value': [0.]*6}
+ tape['steps'][5]['after']['signals']['joint_position'] = deepcopy(stale)
+ tape['steps'][6]['before']['signals']['joint_position'] = deepcopy(stale)
+ result = extract_rows(tape)
+ self.assertEqual(result['attempted_steps'], 7)
+ self.assertEqual(len(result['rows']), 5)
+ self.assertEqual([item['index'] for item in result['exclusions']], [5, 6])
+ self.assertTrue(all('stale:joint_position' in item['reason'] for item in result['exclusions']))
+
+ def test_available_but_aged_required_before_and_after_are_excluded(self):
+ tape = multi_step_fixture()
+ # This signal is still available by the tape's 100 ms recording rule,
+ # but a 20 ms label must represent the actual boundary acquisition.
+ aged = {'status': 'available', 'captured_at_seconds': .019,
+ 'age_seconds': .001, 'value': [0.]*6}
+ tape['steps'][0]['after']['signals']['joint_position'] = deepcopy(aged)
+ tape['steps'][1]['before']['signals']['joint_position'] = deepcopy(aged)
+ result = extract_rows(tape)
+ self.assertEqual([item['index'] for item in result['exclusions']], [0, 1])
+ self.assertTrue(all('noncontemporaneous:joint_position' in item['reason']
+ for item in result['exclusions']))
+
+ def test_partial_terminal_boundary_is_excluded(self):
+ tape = fixture()
+ step = tape['steps'][0]
+ step['after']['timestamp_seconds'] = .001
+ for sensor in step['after']['signals'].values():
+ sensor['captured_at_seconds'] = .001
+ step['applied']['interval_end_seconds'] = .001
+ step['result'].update(terminated=True, truncated=False, success=True, reason='success')
+ tape['outcome'].update(result=deepcopy(step['result']), simulation_seconds=.001)
+ result = extract_rows(tape)
+ self.assertEqual(result['rows'], [])
+ self.assertEqual(result['exclusions'][0]['reason'], 'incomplete_20ms_boundary')
+
+ def test_tape_validation_precedes_extraction(self):
+ tape = fixture()
+ tape['physical_robot_ready'] = True
+ with self.assertRaises(ValueError):
+ extract_rows(tape)
+
+
+class CompactTransitionTests(unittest.TestCase):
+ def setUp(self):
+ self.rows = learning_rows()
+ self.model = fit_model(self.rows, alpha=1e-6)
+
+ def test_action_sensitive_predictions_and_state_ablation_on_matched_states(self):
+ passive = fit_model(self.rows, alpha=1e-6, feature_mode='state')
+ first = [0.]*12+[.04]+[-.5]*6+[0.]
+ second = [0.]*12+[.04]+[.5]*6+[.08]
+ low, high = (predict_features(self.model, features) for features in (first, second))
+ self.assertAlmostEqual(high[0]-low[0], .012, places=7)
+ self.assertAlmostEqual(high[6]-low[6], .008, places=7)
+ self.assertEqual(predict_features(passive, first), predict_features(passive, second))
+
+ def test_weighted_train_only_normalization_and_unpenalized_intercept(self):
+ rows = [{'features': [0.]*20, 'target_delta': [2.]*7},
+ {'features': [2.]*20, 'target_delta': [4.]*7},
+ {'features': [1000.]*20, 'target_delta': [1000.]*7}]
+ model = fit_model(rows, alpha=1e6, sample_weights=[1., 3., 0.])
+ self.assertEqual(model['normalization']['means'], [1.5]*20)
+ self.assertAlmostEqual(model['normalization']['scales'][0], 3**.5/2)
+ self.assertEqual(model['intercept'], [3.5]*7)
+ # Inference on a novel state cannot change fitted normalization.
+ baseline = deepcopy(model['normalization'])
+ predict_features(model, [9999.]*20)
+ self.assertEqual(model['normalization'], baseline)
+ self.assertEqual(model['training']['rows'], 3)
+ self.assertEqual(model['training']['positive_weight_rows'], 2)
+
+ def test_constant_features_have_zero_coefficients_and_safe_scale(self):
+ normalization = self.model['normalization']
+ self.assertTrue(normalization['constant_features'][0])
+ self.assertEqual(normalization['scales'][0], 1.)
+ self.assertTrue(all(row[0] == 0. for row in self.model['coefficients']))
+ changed = [row['features'][0]+1 for row in self.rows]
+ self.assertEqual(len(changed), 18)
+
+ def test_missing_and_stale_observations_abstain(self):
+ snapshot = deepcopy(fixture()['initial'])
+ command = {'joint_targets': [0.]*6, 'gripper_width': .04}
+ self.assertEqual(predict(self.model, snapshot, command)['status'], 'predicted')
+ missing = deepcopy(snapshot)
+ missing['signals']['joint_position'].update(status='missing', captured_at_seconds=None,
+ age_seconds=None, value=None)
+ result = predict(self.model, missing, command)
+ self.assertEqual(result['status'], 'abstained')
+ self.assertIn('missing:joint_position', result['reason'])
+ snapshot['timestamp_seconds'] = .2
+ for sensor in snapshot['signals'].values():
+ sensor.update(captured_at_seconds=.2)
+ snapshot['signals']['joint_velocity'].update(status='stale', captured_at_seconds=0., age_seconds=.2)
+ result = predict(self.model, snapshot, command)
+ self.assertEqual(result['status'], 'abstained')
+ self.assertIn('stale:joint_velocity', result['reason'])
+
+ def test_available_but_delayed_observation_abstains_without_age_propagation(self):
+ snapshot = deepcopy(fixture()['initial'])
+ snapshot['timestamp_seconds'] = .02
+ for sensor in snapshot['signals'].values():
+ sensor.update(captured_at_seconds=.02)
+ snapshot['signals']['joint_position'].update(captured_at_seconds=.019, age_seconds=.001)
+ result = predict(self.model, snapshot, {'joint_targets': [0.]*6, 'gripper_width': .04})
+ self.assertEqual(result['status'], 'abstained')
+ self.assertIn('noncontemporaneous:joint_position', result['reason'])
+
+ def test_direct_prediction_rejects_clock_contradictions_and_invalid_commands(self):
+ before = fixture()['initial']
+ before['signals']['joint_position']['captured_at_seconds'] = .01
+ self.assertEqual(predict(self.model, before, {'joint_targets': [0.]*6, 'gripper_width': .04})['status'], 'abstained')
+ before = fixture()['initial']
+ for command in ({'joint_targets': [0.]*5, 'gripper_width': .04},
+ {'joint_targets': [0.]*6, 'gripper_width': .2},
+ {'joint_targets': [0.]*6, 'gripper_width': .04, 'phase': 'approach'}):
+ self.assertEqual(predict(self.model, before, command)['status'], 'abstained')
+
+ def test_inference_load_and_extraction_do_not_import_numpy_or_mujoco(self):
+ with TemporaryDirectory() as directory:
+ model_path = Path(directory)/'model.json'
+ write_model(model_path, self.model)
+ code = '''import builtins
+original = builtins.__import__
+def guarded(name, *args, **kwargs):
+ if name.split('.')[0] in ('numpy', 'mujoco'):
+ raise RuntimeError('Optional numeric/native dependency imported')
+ return original(name, *args, **kwargs)
+builtins.__import__ = guarded
+from dvidia_training.transition_model import load_model,predict_features,extract_rows
+from tests.test_episode_tape import fixture
+model=load_model(__import__('sys').argv[1])
+assert len(predict_features(model,[0.]*20))==7
+assert len(extract_rows(fixture())['rows'])==1
+'''
+ result = subprocess.run([sys.executable, '-c', code, str(model_path)],
+ cwd=Path(__file__).resolve().parents[1], capture_output=True, text=True)
+ self.assertEqual(result.returncode, 0, result.stderr)
+
+ def test_serialization_roundtrip_hashes_detachment_and_no_overwrite(self):
+ with TemporaryDirectory() as directory:
+ path = Path(directory)/'model.json'
+ receipt = write_model(path, self.model)
+ self.assertLess(receipt['bytes'], MAX_MODEL_BYTES)
+ self.assertEqual(receipt['sha256'], sha256(path.read_bytes()).hexdigest())
+ restored = load_model(path)
+ self.assertEqual(predict_features(restored, self.rows[0]['features']),
+ predict_features(self.model, self.rows[0]['features']))
+ restored['intercept'][0] = 100.
+ self.assertNotEqual(restored['intercept'][0], self.model['intercept'][0])
+ with self.assertRaises(FileExistsError):
+ write_model(path, self.model)
+
+ def test_loading_without_posix_flags_still_checks_regular_file_and_symlink(self):
+ import dvidia_training.transition_model as transition_model
+ portable_os = SimpleNamespace(O_RDONLY=os.O_RDONLY, O_BINARY=getattr(os,'O_BINARY',0),
+ open=os.open, fdopen=os.fdopen, fstat=os.fstat)
+ with TemporaryDirectory() as directory:
+ path = Path(directory)/'model.json'
+ write_model(path,self.model)
+ with patch.object(transition_model,'os',portable_os):
+ self.assertEqual(load_model(path),self.model)
+ with self.assertRaises(ValueError):
+ load_model(Path(directory))
+ link = Path(directory)/'link.json'
+ try:
+ link.symlink_to(path)
+ except OSError:
+ pass # Some Windows installations do not grant symlink permission.
+ else:
+ with self.assertRaises(ValueError):
+ load_model(link)
+
+ def test_model_path_swap_is_detected_before_parsing_when_file_identity_exists(self):
+ import dvidia_training.transition_model as transition_model
+ with TemporaryDirectory() as directory:
+ path = Path(directory)/'model.json'
+ other = Path(directory)/'replacement.json'
+ write_model(path,self.model)
+ write_model(other,self.model)
+ before = path.lstat()
+ with path.open('rb') as stream:
+ opened = os.fstat(stream.fileno())
+ if not before.st_ino or not opened.st_ino:
+ self.skipTest('Path/descriptor identity is unavailable on this platform.')
+ original_open = os.open
+ def swapped_open(name,flags):
+ path.unlink()
+ other.rename(path)
+ return original_open(name,flags)
+ with patch.object(transition_model.os,'open',side_effect=swapped_open):
+ with self.assertRaisesRegex(ValueError,'identity changed'):
+ load_model(path)
+
+ def test_unknown_fields_false_scope_and_model_hash_tampering_refused(self):
+ for key, value in [('unexpected', 'executable'), ('physical_robot_ready', True),
+ ('is_controller', True), ('predicts_task_success', True),
+ ('scope', 'hardware'), ('control_dt_seconds', .04),
+ ('schema_version', True)]:
+ model = deepcopy(self.model)
+ model[key] = value
+ rehash(model)
+ with self.subTest(key=key), self.assertRaises(ValueError):
+ validate_model(model)
+ changed = deepcopy(self.model)
+ changed['intercept'][0] += .1
+ with self.assertRaisesRegex(ValueError, 'hash mismatch'):
+ validate_model(changed)
+
+ def test_nonfinite_bool_extreme_values_bad_shapes_and_weights_refused(self):
+ for value in (True, float('nan'), float('inf'), 10**1000, 1e7):
+ rows = deepcopy(self.rows)
+ rows[0]['features'][0] = value
+ with self.subTest(value=type(value).__name__), self.assertRaises(ValueError):
+ fit_model(rows)
+ features = [0.]*20
+ features[0] = value
+ with self.assertRaises(ValueError):
+ predict_features(self.model, features)
+ for weights in ([1.], [-1.]*18, [0.]*18, [True]*18):
+ with self.subTest(weights=weights), self.assertRaises(ValueError):
+ fit_model(self.rows, sample_weights=weights)
+ for features in ([0.]*19, [0.]*21, {'features': [0.]*20}):
+ with self.assertRaises(ValueError):
+ predict_features(self.model, features)
+
+ def test_artifact_nested_contracts_are_bounded_and_finite_even_with_recomputed_hash(self):
+ for path, value in [(('normalization', 'means'), [0.]*19),
+ (('normalization', 'scales'), [0.]*20),
+ (('normalization', 'constant_features'), [1]*20),
+ (('normalization', 'scope'), 'all-partitions'),
+ (('training', 'rows'), True),
+ (('training', 'rows_sha256'), 'x'*64),
+ (('training', 'extra'), 'unchecked'),
+ (('coefficients',), [[0.]*20]*6),
+ (('intercept',), [0.]*6)]:
+ model = deepcopy(self.model)
+ target = model
+ for key in path[:-1]:
+ target = target[key]
+ target[path[-1]] = value
+ rehash(model)
+ with self.subTest(path=path), self.assertRaises(ValueError):
+ validate_model(model)
+ model = deepcopy(self.model)
+ model['intercept'][0] = float('nan')
+ with self.assertRaises(ValueError):
+ validate_model(model)
+
+ def test_duplicate_nonfinite_oversized_and_symlink_artifacts_refused(self):
+ for raw in ('{"kind":"first","kind":"second"}', '{"coefficient":NaN}',
+ b'\xff', '\ud800', b' '* (MAX_MODEL_BYTES+1)):
+ with self.subTest(raw_type=type(raw).__name__), self.assertRaises(ValueError):
+ validate_model(raw)
+ with TemporaryDirectory() as directory:
+ path = Path(directory)/'model.json'
+ write_model(path, self.model)
+ link = Path(directory)/'link.json'
+ link.symlink_to(path)
+ with self.assertRaises(ValueError):
+ load_model(link)
+
+
+if __name__ == '__main__':
+ unittest.main()
diff --git a/tests/test_transition_pilot.py b/tests/test_transition_pilot.py
new file mode 100644
index 0000000..9d78cb1
--- /dev/null
+++ b/tests/test_transition_pilot.py
@@ -0,0 +1,318 @@
+"""Grouped split, frozen evaluation and failure denominator checks."""
+from copy import deepcopy
+from hashlib import sha256
+import json
+from pathlib import Path
+from tempfile import TemporaryDirectory
+import unittest
+from unittest.mock import patch
+
+from dvidia_training.episode_tape import CASES, canonical, digest, validate_tape, write_tape
+from dvidia_training.transition_pilot import (ALPHAS, CORPUS_KIND, _manifest, _metrics,
+ generate_corpus, inspect_pilot, run_pilot)
+from tests.test_episode_tape import fixture
+
+
+def authored_tape(group_index, case):
+ tape = fixture()
+ tape['configuration'].update(horizon=.06, object_mass=.04+.001*group_index, object_friction=.8+.01*group_index)
+ tape['provenance'].update(episode_id=f'reset-{group_index:03d}-{case}',
+ source_recording_id=f'recording-{group_index}-{case}', session_id=f'session-{group_index}',
+ seed=group_index+10, case=case, scene_sha256=digest({'group': group_index}),
+ configuration_sha256=digest(tape['configuration']))
+ template = deepcopy(tape['steps'][0])
+ before = tape['initial']
+ steps = []
+ for index in range(3):
+ step = deepcopy(template)
+ start, end = index*.02, (index+1)*.02
+ step.update(index=index, before=deepcopy(before))
+ # First nominal/release transition is identical, then only jaw command changes.
+ targets = [.02*(index+1)+.001*group_index]*6
+ requested = {'joint_targets': targets, 'gripper_width': .04}
+ intervention = 'open-jaw' if case == 'open-jaw' else 'release-after-lift' if case == 'release-after-lift' and index >= 1 else 'none'
+ issued = deepcopy(requested)
+ if intervention != 'none':
+ issued['gripper_width'] = .08
+ step['command'].update(issued_at_seconds=start, policy_requested=requested, issued=issued, intervention=intervention)
+ after = deepcopy(before)
+ after['timestamp_seconds'] = end
+ for sensor in after['signals'].values():
+ sensor.update(captured_at_seconds=end)
+ delta = [.005*(target-q) for target, q in zip(targets, before['signals']['joint_position']['value'])]
+ after['signals']['joint_position']['value'] = [q+d for q,d in zip(before['signals']['joint_position']['value'],delta)]
+ after['signals']['joint_velocity']['value'] = [d/.02 for d in delta]
+ old_gap = before['signals']['gripper_width']['value']
+ after['signals']['gripper_width']['value'] = old_gap+.1*(issued['gripper_width']-old_gap)
+ step['after'] = after
+ step['applied'].update(interval_start_seconds=start, interval_end_seconds=end, ramped_joint_targets=targets,
+ ramped_gripper_width=issued['gripper_width'],
+ native_position_controls=targets+[issued['gripper_width']/2]*2)
+ result = {'terminated':False,'truncated':index==2,'success':False,'valid':True,'reason':'horizon' if index==2 else 'running'}
+ step['result'] = result
+ steps.append(step)
+ before = after
+ tape['steps'] = steps
+ tape['outcome'].update(result=deepcopy(steps[-1]['result']), simulation_seconds=.06, control_steps=3)
+ return validate_tape(tape)
+
+
+def corpus(directory, failed=None):
+ directory.mkdir()
+ (directory/'tapes').mkdir()
+ groups = []
+ for group_index, split in enumerate(('train','development','test')):
+ example = authored_tape(group_index,'nominal')
+ group = {'group_id':f'reset-{group_index:03d}','split':split,'seed':group_index+10,
+ 'session_id':f'session-{group_index}','reset_lineage_id':f'lineage-{group_index}',
+ 'configuration':deepcopy(example['configuration']),
+ 'configuration_sha256':example['provenance']['configuration_sha256'],'attempts':[]}
+ for case in CASES:
+ tape = authored_tape(group_index,case)
+ attempt = {'episode_id':tape['provenance']['episode_id'], 'source_recording_id':tape['provenance']['source_recording_id'],
+ 'case':case,'status':'completed','path':f'tapes/{tape["provenance"]["episode_id"]}.json',
+ 'sha256':None,'bytes':None,'error':None,'recording_elapsed_seconds':.001,
+ 'provenance':deepcopy(tape['provenance']), 'outcome':deepcopy(tape['outcome'])}
+ if failed == (split,case):
+ attempt.update(status='failed',path=None,error='RuntimeError: deliberately failed attempt',provenance=None,outcome=None)
+ else:
+ receipt = write_tape(directory/attempt['path'],tape)
+ attempt.update(sha256=receipt['sha256'],bytes=receipt['bytes'])
+ group['attempts'].append(attempt)
+ groups.append(group)
+ manifest = {'kind':CORPUS_KIND,'schema_version':1,'scope':'simulation-only','physical_robot_ready':False,
+ 'shoe_skill_ready':False,'status':'completed-with-failed-attempts' if failed else 'completed',
+ 'protocol':{'revision':'grouped-native-actuator-20ms-v1','cases':list(CASES),
+ 'split_unit':'reset group','variation':'predeclared','split_counts':{'train':1,'development':1,'test':1},
+ 'source_files_sha256':groups[0]['attempts'][0]['provenance']['source_files_sha256']},
+ 'groups':groups,'collection_elapsed_seconds':.001}
+ path = directory/'manifest.json'
+ path.write_bytes(canonical(manifest)+b'\n')
+ return path
+
+
+def rewrite(path, manifest):
+ path.write_bytes(canonical(manifest)+b'\n')
+
+
+class TransitionPilotTests(unittest.TestCase):
+ def test_frozen_evaluation_artifacts_metrics_slices_and_matched_branch(self):
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ manifest = corpus(root/'corpus')
+ report = run_pilot(manifest,root/'pilot')
+ self.assertEqual(report['status'],'completed',report.get('error'))
+ self.assertEqual(report['counts']['attempts'],9)
+ self.assertEqual(report['counts']['rows'],{'train':9,'development':9,'test':9})
+ self.assertEqual(report['counts']['simulation_unsuccessful'],9)
+ self.assertFalse(report['physical_robot_ready'])
+ self.assertFalse(report['shoe_skill_ready'])
+ self.assertEqual(report['matched_branch']['status'],'matched')
+ self.assertEqual(report['matched_branch']['prediction_delta']['state']['nominal'],report['matched_branch']['prediction_delta']['state']['release'])
+ self.assertEqual(set(report['metrics']['test']),{'action','state','persistence','constant_velocity','training_mean'})
+ self.assertEqual([trial['alpha'] for trial in report['selection']['action']['trials']],list(ALPHAS))
+ self.assertEqual(report['timings']['inference']['action']['measured_calls'],1000)
+ self.assertEqual(inspect_pilot(root/'pilot')['integrity'],'verified')
+ with self.assertRaisesRegex(ValueError,'fresh'):
+ run_pilot(manifest,root/'pilot')
+
+ def test_held_out_tapes_are_not_parsed_before_both_models_and_freeze_exist(self):
+ import dvidia_training.transition_pilot as pilot
+ original = pilot._load_attempt
+ observed = []
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ manifest = corpus(root/'corpus')
+ def observed_load(corpus_root,group,attempt):
+ if group['split'] == 'test':
+ self.assertTrue((root/'pilot'/'freeze.json').is_file())
+ self.assertTrue((root/'pilot'/'action.model.json').is_file())
+ self.assertTrue((root/'pilot'/'state.model.json').is_file())
+ freeze = json.loads((root/'pilot'/'freeze.json').read_text())
+ self.assertFalse(freeze['held_out_transition_targets_read'])
+ observed.append(attempt['episode_id'])
+ return original(corpus_root,group,attempt)
+ with patch.object(pilot,'_load_attempt',observed_load):
+ report = run_pilot(manifest,root/'pilot')
+ self.assertEqual(report['status'],'completed',report.get('error'))
+ self.assertEqual(len(observed),3)
+
+ def test_failed_recordings_keep_denominators_and_exclusions_are_not_zero_fill(self):
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ manifest = corpus(root/'corpus',failed=('test','open-jaw'))
+ report = run_pilot(manifest,root/'pilot')
+ self.assertEqual(report['status'],'completed',report.get('error'))
+ self.assertEqual(report['counts']['attempts'],9)
+ self.assertEqual(report['counts']['failed_recordings'],1)
+ self.assertEqual(report['counts']['completed_recordings'],8)
+ self.assertEqual(report['counts']['rows']['test'],6)
+ self.assertEqual(report['counts']['case_attempts']['open-jaw'],3)
+ failed = [a for a in report['attempts'] if a['recording_status']=='failed']
+ self.assertEqual(len(failed),1)
+ self.assertEqual(failed[0]['rows'],0)
+ self.assertIsNone(failed[0]['source_sha256'])
+ self.assertEqual(inspect_pilot(root/'pilot')['integrity'],'verified')
+
+ def test_missing_after_measurement_is_excluded_with_reason_and_denominator(self):
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ path = corpus(root/'corpus')
+ manifest = json.loads(path.read_text())
+ attempt = manifest['groups'][2]['attempts'][0]
+ tape_path = root/'corpus'/attempt['path']
+ tape = json.loads(tape_path.read_text())
+ tape['steps'][-1]['after']['signals']['joint_position'] = {
+ 'status':'missing','captured_at_seconds':None,'age_seconds':None,'value':None}
+ tape_path.unlink()
+ receipt = write_tape(tape_path,tape)
+ attempt.update(sha256=receipt['sha256'],bytes=receipt['bytes'])
+ rewrite(path,manifest)
+ report = run_pilot(path,root/'pilot')
+ self.assertEqual(report['status'],'completed',report.get('error'))
+ self.assertEqual(report['counts']['rows']['test'],8)
+ self.assertEqual(report['counts']['attempted_steps']['test'],9)
+ self.assertEqual(report['counts']['excluded_steps']['test'],1)
+ self.assertIn('missing',report['attempts'][6]['exclusions'][0]['reason'])
+ self.assertEqual(inspect_pilot(root/'pilot')['integrity'],'verified')
+
+ def test_shared_settings_or_scene_assets_do_not_imply_duplicate_reset(self):
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ path = corpus(root/'corpus')
+ manifest = json.loads(path.read_text())
+ reference = manifest['groups'][0]
+ changed = manifest['groups'][2]
+ changed['configuration'] = deepcopy(reference['configuration'])
+ changed['configuration_sha256'] = reference['configuration_sha256']
+ for attempt in changed['attempts']:
+ attempt['provenance']['configuration_sha256'] = reference['configuration_sha256']
+ attempt['provenance']['scene_sha256'] = reference['attempts'][0]['provenance']['scene_sha256']
+ rewrite(path,manifest)
+ # The distinct declared seeds/sessions still define distinct groups.
+ _manifest(path)
+
+ def test_rehashed_report_cannot_hide_missing_attempts_or_differing_frozen_selection(self):
+ for field in ('attempts','selection'):
+ with self.subTest(field=field), TemporaryDirectory() as directory:
+ root = Path(directory)
+ path = corpus(root/'corpus')
+ report = run_pilot(path,root/'pilot')
+ self.assertEqual(report['status'],'completed',report.get('error'))
+ if field == 'attempts':
+ report['attempts'].pop()
+ else:
+ report['selection']['action']['alpha'] = 999.
+ raw = canonical(report)+b'\n'
+ (root/'pilot'/'report.json').write_bytes(raw)
+ integrity_path = root/'pilot'/'integrity.json'
+ integrity = json.loads(integrity_path.read_text())
+ integrity['artifacts']['report.json'].update(sha256=sha256(raw).hexdigest(),bytes=len(raw))
+ integrity_path.write_bytes(canonical(integrity)+b'\n')
+ with self.assertRaises(ValueError):
+ inspect_pilot(root/'pilot')
+
+ def test_raw_source_hash_mismatch_is_retained_as_error(self):
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ manifest = corpus(root/'corpus')
+ tape = next((root/'corpus'/'tapes').glob('*nominal.json'))
+ raw = tape.read_bytes()
+ tape.write_bytes(raw.replace(b'0.04',b'0.05',1))
+ report = run_pilot(manifest,root/'pilot')
+ self.assertEqual(report['status'],'error')
+ self.assertIn('source hash',report['error'])
+ self.assertTrue((root/'pilot'/'report.json').is_file())
+ self.assertEqual(len(report['attempts']),9)
+ self.assertEqual(inspect_pilot(root/'pilot')['status'],'error')
+
+ def test_seed_session_recording_and_hash_aliases_are_rejected(self):
+ for alias in ('seed','session_id','reset_lineage_id','source_recording_id','sha256'):
+ with self.subTest(alias=alias), TemporaryDirectory() as directory:
+ root = Path(directory)
+ path = corpus(root/'corpus')
+ manifest = json.loads(path.read_text())
+ a,b = manifest['groups'][0],manifest['groups'][2]
+ if alias == 'configuration':
+ b['configuration'] = deepcopy(a['configuration'])
+ b['configuration_sha256'] = a['configuration_sha256']
+ for attempt in b['attempts']:
+ attempt['provenance']['configuration_sha256'] = a['configuration_sha256']
+ elif alias == 'scene':
+ for attempt in b['attempts']:
+ attempt['provenance']['scene_sha256'] = a['attempts'][0]['provenance']['scene_sha256']
+ elif alias in ('source_recording_id','sha256'):
+ b['attempts'][0][alias] = a['attempts'][0][alias]
+ if alias == 'source_recording_id':
+ b['attempts'][0]['provenance'][alias] = a['attempts'][0][alias]
+ else:
+ b[alias] = a[alias]
+ if alias in ('seed','session_id'):
+ for attempt in b['attempts']:
+ attempt['provenance'][alias] = a[alias]
+ rewrite(path,manifest)
+ with self.assertRaisesRegex(ValueError,'Duplicate'):
+ _manifest(path)
+
+ def test_paths_symlinks_duplicate_json_nonfinite_and_corrupt_frozen_artifact_rejected(self):
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ path = corpus(root/'corpus')
+ original = path.read_bytes()
+ manifest = json.loads(original)
+ manifest['groups'][0]['attempts'][0]['path'] = '../outside.json'
+ rewrite(path,manifest)
+ with self.assertRaisesRegex(ValueError,'relative'):
+ _manifest(path)
+ path.write_bytes(b'{"kind":1,"kind":2}')
+ with self.assertRaisesRegex(ValueError,'Duplicate'):
+ _manifest(path)
+ path.write_bytes(b'{"kind":NaN}')
+ with self.assertRaisesRegex(ValueError,'Nonfinite'):
+ _manifest(path)
+ path.write_bytes(original)
+ tape_path = root/'corpus'/json.loads(original)['groups'][0]['attempts'][0]['path']
+ real_path = tape_path.with_suffix('.original')
+ tape_path.rename(real_path)
+ try:
+ tape_path.symlink_to(real_path)
+ except OSError:
+ real_path.rename(tape_path)
+ else:
+ with self.assertRaisesRegex(ValueError,'regular'):
+ _manifest(path)
+ tape_path.unlink()
+ real_path.rename(tape_path)
+ report = run_pilot(path,root/'pilot')
+ self.assertEqual(report['status'],'completed',report.get('error'))
+ frozen = root/'pilot'/'state.model.json'
+ frozen.write_bytes(frozen.read_bytes()+b' ')
+ with self.assertRaisesRegex(ValueError,'hash'):
+ inspect_pilot(root/'pilot')
+
+ def test_joint_and_jaw_metrics_keep_separate_units_and_baseline_errors(self):
+ metric = _metrics([{'target_delta':[1.,0.,0.,0.,0.,0.,.003]}],[[0.]*7])
+ self.assertAlmostEqual(metric['joint_rmse_rad'],(1/6)**.5)
+ self.assertAlmostEqual(metric['jaw_rmse_m'],.003)
+ self.assertEqual(metric['joint_max_abs_error_rad'],1.)
+ self.assertIsNone(_metrics([],[])['joint_rmse_rad'])
+
+ def test_generator_predeclares_plan_and_preserves_all_attempts_on_missing_runtime(self):
+ with TemporaryDirectory() as directory:
+ root = Path(directory)
+ def fail(**kwargs):
+ self.assertTrue((root/'corpus'/'collection-plan.json').is_file())
+ raise RuntimeError('deliberate native recorder failure')
+ with patch('dvidia_training.transition_pilot.record_episode',side_effect=fail):
+ manifest = generate_corpus(root/'corpus',groups=6)
+ self.assertEqual(manifest['protocol']['split_counts'],{'train':4,'development':1,'test':1})
+ self.assertEqual(manifest['status'],'completed-with-failed-attempts')
+ self.assertEqual(sum(len(g['attempts']) for g in manifest['groups']),18)
+ self.assertTrue(all(a['status']=='failed' for g in manifest['groups'] for a in g['attempts']))
+ _manifest(root/'corpus'/'corpus.json')
+ with self.assertRaisesRegex(ValueError,'fresh'):
+ generate_corpus(root/'corpus',groups=6)
+
+
+if __name__ == '__main__':
+ unittest.main()