Use README.md for the public entry point,
docs/README.md for task-oriented navigation, and
docs/PROJECT_STATUS.md for the concise current
state. This file governs repository work and preserves campaign constraints;
it is not the recommended first read for users.
The controlling whole-moonshot verdict is FULL_MOONSHOT_NOT_PROVEN.
Documentation, a passing unit test, or a new development commit cannot promote
that verdict. Keep R7, V1089, R97, and later additive experiments as separate
evidence lineages.
ABI owns acquisition and packaging of capabilities from foreign open-weight teachers: source qualification, probing, semantic labeling, quarantine, normalization, provenance, information accounting, minimization research, immutable artifacts, and fail-closed verification.
LayerCake owns runtime hosting, cake installation, transfer between compatible LayerCake hosts, routing, orchestration, and inference performance. Keep the repositories and evidence lineages separate. ABI may link to LayerCake and test its public host interface; it must not copy LayerCake implementation or represent LayerCake research as ABI evidence.
The user authorized R108 on 2026-11-14: one trained-task grounded-generation
qualification attempt with a preselected Qwen3-4B-Instruct-2507 control, then
the retained matched R99 LoRA/distillation and parent/full-model diagnostics
only if that control qualifies. Follow
experiments/trained_grounding_r108/PROTOCOL.md. Freeze the paired inputs and
all gates before inference. No training, model/prompt sweep, native host edits,
or broad-English certification. Failure ownership remains explicitly separated.
R108 ended incomplete at its registered control-process timeout: 13/96 initial rows, no replay or recipients. Keep the frozen complete verifier rejecting it. The additive interruption observer is evidence preservation, not qualification. Recorded GPU power/thermal throttling requires an operating-mode check before a separately authorized retry; do not override thermal protections, lower gates, extend the expired deadline, or overwrite the interrupted evidence.
The user authorized continued bounded progress after R106. R107 is a distinct
natural-text retention/intervention audit, not another control-qualification
sweep. Its preregistration is experiments/retention_intervention_r107/PROTOCOL.md.
It preserves all R103-R106 gates, uses zero training, and must not promote a
neutralized model or call WikiText likelihood a fluent-English certificate.
No automatic acquisition follow-on is authorized by a diagnostic result.
R107 completed all 18 executions with exact raw replay. Its primary parent interference contrast is +0.03224 passage BPB, CI [+0.02064, +0.04419], but neutralization worsens parent repetition. Do not promote a neutralized state. The fixed coherence route was untrained in R99: adapter distillation preserves its selected tensors and all 64 active rows exactly, unlike the historical shared-weight controls. This does not establish trained-route fluency or superiority over route-local LoRA. Preserve the negative generation evidence, all historical gates, and the separation from native host certification.
The controlling technical release is R7:
- public tag:
abi-final-validation-v2-repaired-r7-2026-08-30 - tagged commit:
3f82a9f4d67dda5c8ea13bd59b2d8f1bbd3dd128 - release: https://github.com/Yoder23/abi/releases/tag/abi-final-validation-v2-repaired-r7-2026-08-30
- archive SHA-256:
fc50f423986149b5d4670ec9e28698540f64be96034efa26e5704c4469921e88 - strict certificate:
results/abi_final_validation_v2/strict_validation_r7_tar_bound.json - blind review:
results/abi_final_validation_v2/blind_redteam_r7/
R7 passes bounded technical validation. Human review, different-hardware reproduction, and registered minimum-information certification are now open, not complete.
The independent audit preserved at
reviews/independent_2026-09-14/ returned
FULL_MOONSHOT_NOT_PROVEN; treat that as the controlling whole-moonshot
verdict. The later replay-surface repair is a repository/reproducibility fix,
not a scientific promotion. Never combine measurements from R7, V1089, and
R97 as though they describe one final artifact. Never mark Phase 5 clean
replay complete unless the six exact baseline tensors named in the review
manifest are restored or the phase is prospectively rerun and resealed.
The later V1089 compiler campaign is a separate local machine-evidence line, not a replacement public release. Phase 0 and Phase 1 are complete; Phase 2 is machine/handoff ready but has 0/21,000 human preferences; Phase 3 machine gates pass conditionally; Phases 4-7 have bounded machine certificates; and Phase 8 has a same-machine clean-export rehearsal only. Its exact B40 core passes the quality contract across 4,200 observations and three seeds, its chemistry/civics/Python packages pass the registered exclusion and composition gates, and the integrated Phase 7 product records 13.24x CPU and 13.51x GPU throughput ratios against registered comparators. Do not call this an unconditional campaign pass, independent reproduction, global minimum, deployed multi-source result, or universal ABI-over-LoRA/distillation result.
The additive R8 native-neural-transfer campaign is a Level 0 negative result, not a new release line. Its v10 package-only hypernetwork failed the public Pythia prerequisite: AFTER−BASE was +0.001953 across 1,024 paired rows (95% bootstrap CI -0.019531 to +0.024414). The held-out commitment was never revealed. R8 therefore does not establish native cross-model transfer and does not alter R7.
The additive R9 neural-ISA diagnostic is also negative and does not alter R7. Its stricter v2 capability-specific Pythia backend achieved 0.128472 training fit and 0.125 unseen-depth AFTER accuracy versus 0.073242 BASE; ZERO also reached 0.125. Exact live replay covered 8,192 evaluation and 288 training rows, and 5/5 hostile controls rejected tampering. The recipient-state GRU branch is closed; the capability-blind universal backend was not run.
The additive R10 runtime copy/paste campaign separates a bounded component pass from an overall negative result. Identical 2,040-byte source-extracted canonical-IR packages achieved 1.0 AFTER and RESTORED accuracy on all 61,440 Pythia, Qwen2, and T5 recipient rows, with exact removal and exact fresh live replay. The source model's native decoder achieved only 0.59375 to 0.65234375 on the newly sampled composed prompts, below the registered 0.99 source gate. R10 therefore proves only synthetic runtime-owned package paste/remove/restore; it does not prove lossless source behavior, LayerCake integration, native neural transfer, or teacher extraction, and it does not alter R7.
The additive R11 native-neural ABI construction campaign is a bounded synthetic pass and does not alter R7. Four held-out capabilities learned in an ABI-native teacher substrate were extracted as four 2,053-byte packages and executed through frozen DistilGPT2, Pythia, Qwen2, and T5 native output heads. Strict recomputation passed all 90,112 observations: teacher, recipient AFTER, and RESTORED outputs were exact; controls and removal were causal; recipient training was zero; source and recipient states were frozen. Seven hostile cases failed closed. A fresh live replay of the exact revision-001 package bytes reproduced both raw row files byte-for-byte. An independent acquisition replication was functionally exact but produced different float32 package hashes, so bit-reproducible acquisition is not claimed. R11 proves only the synthetic ABI-native construction mechanism; arbitrary open-weight teacher, English/domain, LayerCake, and LoRA/distillation claims remain open.
The additive R12-A public campaign proves one bounded conventional-teacher frontend construction and does not alter R7. Qwen2.5-0.5B was ordinarily LoRA-trained on a fresh synthetic capability. Its best native public result was 510/512 exact, below the registered 512/512 prerequisite. Its 24 atomic probes were exact, and a fixed zero-parameter extractor emitted an unchanged 2,053-byte R11 package that executed 512/512 unseen prompt-disjoint compositions exactly without the teacher. No held-out secret was created or revealed. Do not broaden this into exact teacher-relative transfer, pre-existing knowledge extraction, English/domain, LayerCake, or LoRA/distillation claims.
The additive R13-B held-out campaign passes bounded enumerable capability extraction locally and does not alter R7. Four committed finite capabilities reached 96/96 exact Qwen atomic probes; four 2,053-byte packages scored 2,048/2,048 against the oracle and 6,144/6,144 AFTER plus 6,144/6,144 RESTORED across frozen Pythia, Qwen2, and T5 recipients. Seven fresh-process replay files were byte-exact. A first hostile audit exposed a verifier defect and a second exposed an audit-harness defect; both are preserved, and the repaired audit passed 9/9 controls before the final seal. Native source/package agreement was only 372/2,048. R13-B therefore proves finite capability canonicalization, not behavioral transplantation, pre-existing knowledge, English/domain extraction, or LayerCake acceptance. Bulk replay assets remain unpublished, so it is not a new public release.
The additive R14 non-exhaustive campaign is a bounded negative and does not alter R7. It replaced complete atomic probing with 256 mixed, answer-free source observations per capability and evaluated 10,000 unseen-depth plus 1,000 order-counterfactual cases from a 4,642,668,576-case space. One of three fresh capabilities was recovered exactly; two latent selections were wrong. Three fresh source replay files were byte-exact and 7/7 hostile controls passed. R14 therefore does not certify reliable non-exhaustive extraction.
The additive R15A foreign-neural-state campaign is a bounded local pass and does not alter the R7 public release. A generic frontend frozen after 320 public Qwen learning events recovered 8/8 fresh held-out synthetic capabilities from anonymous 7,168-element effective output-weight deltas, with zero behavioral queries, answers, operation tables, or oracle calls at extraction. Eight 2,053-byte R11 packages were exact on 80,000 unseen and 8,000 counterfactual rows and executed at 1.0 AFTER/RESTORED across frozen Pythia, Qwen2, and T5 recipients. Full-state controls were at most 0.265, live replay regenerated 19,584 source rows and reproduced 135,168 recipient rows byte-for-byte, and 9/9 hostile mutations failed closed. R15A proves only bounded recovery of a deliberately learned synthetic capability from a before/after foreign neural delta. Pre-existing knowledge, English/domain extraction, semantic labeling, teacher-quality generation, after-only extraction, LayerCake ingestion, and LoRA/distillation superiority remain open.
The additive R15B pre-existing-representation campaign is a bounded local pass and does not alter the R7 public release. An unchanged Qwen2-7B-Instruct source with zero training produced 24/24 atomic anchors under four secret semantic-slot mappings. A physically isolated generic decoder recovered all four transitions from anonymous pre-answer residuals plus eight output-head rows, emitted four unchanged R11 packages, and scored 40,000/40,000 deep oracle rows. Strict and live verification passed across Pythia, Qwen2, and T5, and 10/10 hostile mutations failed closed. A fresh blind review passed with zero critical/high findings. Its live-tensor and source-identity findings were repaired by exact replay of all four representation bundles and SHA-256 inventory of all 10 files/15.24 GB in the pinned source snapshot. R15B proves only narrow pretrained arithmetic representation recovery with an externally registered ontology. Deep teacher behavior, English/domain discovery, autonomous labeling, LayerCake ingestion, minimality, and LoRA/distillation superiority remain open.
The additive R16 factual/semantic campaign is a bounded local pass and does not alter the R7 public release. On a preregistered hidden selection, an unchanged Qwen2-7B-Instruct source exposed 16/16 facts across registered chemistry and geography namespaces. A physically isolated compiler received answer-free questions, candidate strings, and source sequence scores and emitted two immutable structured packages totaling 1,222 bytes. Source, package, and target-only evaluation were 48/48 exact; other-domain and removed conditions abstained 48/48; a rotated-score control was 0/48. Live evidence replayed byte-exactly and the expanded verifier rejected 21/21 hostile mutations. A blind review passed with no Critical or High finding. This proves only bounded structured factual extraction and segregation. The registered candidate vocabularies and ontology prevent claims of autonomous discovery, fluent English transfer, arbitrary domains, LayerCake ingestion, minimality, or LoRA/distillation superiority. A clean preregistered replication bound to the repaired public protocol repeated every gate on a second selection; final blind review and public reconstruction remain open.
The additive R97 reasoning campaign is a bounded strictly verified local pass and does not alter the R7 public release. The frozen 530,050-parameter R96 bridge scored 1,399/1,400 on a separately sealed prospective catalog, versus 1,088 for the live pinned teacher, 60 for the unchanged LayerCake parent, and 17 for random. It passed all gates, strict raw recomputation, 12/12 hostile mutations, and byte-identical live replay. Preserve R90/R92/R93/R94/R95 as negative evidence. R97 proves only teacher-derived two-hop nonce reasoning transfer into LayerCake; broad English, arbitrary domains, global minimality, heterogeneous-host execution of this neural package, independent hardware, and universal ABI-over-LoRA/distillation claims remain open.
The four published immutable capability packages execute through the canonical capability runtime/conformance interface in the declared LayerCake v25, Qwen2.5-0.5B, and Pythia-160M environments. The proof includes:
- three capability-blind physical certification environments;
- 301,543 content-bound reachable-filesystem rows;
- 5,043 locked matrix rows;
- 3,072 live causality rows in 24 distinct condition processes;
- 2,100 live isolation rows;
- zero capability archives or success IDs in the certification inventories;
- strict recomputation consuming zero scientific status booleans;
- 19/19 fail-closed hostile controls before and after publication;
- exact public-manifest reconstruction in a clean environment; and
- a fresh blind Codex PASS, including 12/12 USTAR/V7 prefix controls.
Never broaden R7 into any of these unproven claims:
- arbitrary-model or universal host compatibility;
- transformer tensor or weight transplantation;
- teacher knowledge extraction, labeling completeness, or source diagnosis;
- teacher-quality English or specialist generation;
- superiority to LoRA, distillation, or fine-tuning;
- human-perceived quality;
- independent different-hardware reproduction;
- a globally minimal English substrate; or
- completion of the full ABI moonshot.
The R7 artifacts are capability-runtime/conformance packages. They are not evidence that ABI has already extracted a compact fluent English core from an arbitrary teacher.
- Preserve every failed release and negative result additively. R5 and R6 are historical failures, not files to rewrite or relabel.
- Treat stored
PASS,status, and gate fields as untrusted. Recompute claims from raw rows, exact counts, hashes, source inventories, and live failure records. - Fail closed on missing packages, raw rows, receipts, source files, hashes, or stale evidence.
- Do not overwrite immutable evidence paths. Use a new revision suffix.
- Do not substitute evidence from a different model, artifact, host, seed, or release lineage.
- Keep human ratings at
0/21,000until three independent raters actually complete and attest the frozen forms. - Keep independent hardware
PENDINGuntil a genuinely independent operator returns the signed different-hardware packet. - Keep minimum-information certification
PENDINGuntil the registered protocol is executed and verified.
- Prefer bounded, preregistered experiments over unbounded nearby sweeps.
- Use GPU for teacher extraction/training when appropriate; CPU-only is not an ABI ideology. Record hardware and one-time/per-host costs honestly.
- Keep generated caches, temporary reconstructions, model weights, and bulk experiment intermediates out of Git. Publish only required immutable assets and verification evidence.
- Run focused tests for changed code, then the full supported suite before a release-ready commit.
- Keep README, active mission, current status, technical claims, final results, review packet, and changelog consistent with the controlling release.
Active user authorization (2026-11-05): continue bounded evidence-led research
and publish versioned, clearly documented updates to both repositories. R98
stopped at its source prerequisite (206/240; 31 truncated responses), with no
student training. R98-C1 removed truncation with 31/31 exact token-prefix
matches but still failed (226/240; emails 22/30). Preserve both failures and
the C1 offline-tokenizer operational error. No further cap or prompt sweep.
The next bounded step is experiments/grounded_english_r99/PROTOCOL.md:
source-substring sentence selection to address measured email verbosity,
fresh prospectively frozen development prompts, then matched ABI-plan,
LoRA and distillation learners only if normalized-data qualification passes.
No threshold or optimizer sweep, no LayerCake source changes and no inherited
release claims. R99's normalized source passed 230/240, but scaled FP16 failed
after one update. The bounded numerical audit reproduced 50,688 non-finite
gradient elements at step two, while unit-scale FP16 and FP32 stayed finite.
Only the separately frozen experiments/grounded_english_r99_fp32/PROTOCOL.md
restart of all three arms is authorized next; it retains the original R99
deadline and every quality gate. No further numerical or optimizer sweep.
The restart has now completed and failed: ABI plan 10/48, distillation 20/48,
LoRA 27/48, unchanged parent 0/48; the gate was 44/48. Preserve all final
checkpoints and raw rows. Close the autoregressive-JSON-plan branch. A successor
must materially address measured grounding/fact-retention and plan-validity
failures, with a fresh preregistration and evaluation; do not automatically
scale this mechanism. No candidate is promoted. An automatic machine screen
cannot certify English fluency.
Continuation authorization (2026-11-05): before another learner, execute only
the additive experiments/grounded_english_r99/FIT_AUDIT.md diagnostic. It
compares autonomous training-example generation with teacher-forced response
and supplied-fact token prediction on the frozen checkpoints, with zero new
training or teacher queries. It cannot reopen the generated-plan branch or
promote an oracle-prefix score as autonomous quality. A successor still needs
a new representation, fresh prospective evaluation and matched controls.
R99-D1 completed: autonomous training-example passes were 7/24 ABI plan,
12/24 adapter distillation, 13/24 LoRA and 0/24 parent. The restricted learners'
correct-prefix fact-token accuracy was 97-99%, but visible-response accuracy
was only 76-80%. Preserve the diagnostic; do not describe this as an isolated
novel-identifier or LayerCake runtime failure. The next bounded intervention is
experiments/acquisition_capacity_r100/PROTOCOL.md: one conventional full-model
distillation capacity control, unchanged parent/data/seed/update count, matched
retained controls and new prospective development prompts. It cannot promote
an ABI product and authorizes no sweep or subsequent automatic run.
R100 completed and failed every registered screen: full-model control 20/48, LoRA 20/48, adapter distillation 14/48, ABI plan 6/48, parent 0/48; full-control training fit 17/24. The full optimizer intervention changed 86,969,088 parameter elements, but did not establish a strictly positive paired training gain over both response-only controls. Preserve all evidence. Do not repeat full-scope or generated-plan variants or infer data insufficiency/host incapacity as proven. The next proposal must prospectively distinguish acquisition data, optimization, parent suitability and autonomous realization, with matched controls and no historical gate changes. This control certifies no product or moonshot claim.
New user authorization (2026-11-05): run one bounded diagnostic to isolate the
acquisition failure, document and publish it. R101 is governed by
experiments/execution_continuation_r101/PROTOCOL.md: no training or teacher
queries; test full/padded/cached path agreement and correct-prefix continuation
on fixed full-model, LoRA, distillation and parent checkpoints. Oracle-assisted
outputs cannot count as autonomous English. Path disagreement is an integration
finding requiring localization, not automatically a native LayerCake defect.
No historical changes, new parameter sweep or automatic training follow-on.
R101 completed with measured execution-path agreement and exact replay of all 288 unassisted sequences. Its primary remaining-fact intervention was null: 24/34 before and after 50% oracle-prefix assistance, paired CI [-0.14706, 0.14706]. The full-model correct-prefix response-token accuracy was 91.44% on training examples versus 78.91% on development. Preserve the complete raw-logit and token evidence. Do not call joined oracle-assisted scores autonomous English or blame native LayerCake. Data coverage/generalization, optimization fit and parent suitability still require prospective discrimination. No automatic next run.
User continuation (2026-11-10): preregister one R102 entity-coverage versus
optimization-exposure factorial, using conventional LoRA and adapter distillation,
under experiments/coverage_exposure_r102/PROTOCOL.md. Four fixed trajectories,
two fixed endpoints; no full-model/JSON-plan variant or learning-rate search.
Entity renaming is conventional provenance-preserving augmentation, not ABI
novelty or broad semantic coverage. Freeze all synthetic targets and prospective
evaluation inputs before GPU work. Keep native LayerCake untouched. A local
Git freeze is required; hold external publication pending the user's answer to
whether this run should be published as before or kept local for review.
Publication clarification (2026-11-10): the user explicitly answered "Publish as before." The temporary privacy hold is lifted. Publish the sealed R102 tag, complete required evidence and an additive public-reconstruction receipt without moving the tag, modifying its historical local-only fields or promoting its negative scientific result. This publication step authorizes no new model run.
R102 preflight A1: the first joint sampler stopped before source/model execution.
Preserve its 4,096-proposal rejection trace and same-pool component analysis.
PREFLIGHT_A1.md prospectively registers an implementation repair before the
first successful input freeze: component token-profile matching in the same
bounded proposal stream. No scientific threshold, training budget, data factor
or evaluator is changed. The initial sampler and negative evidence stay intact.
R102 completed and strictly recomputed locally. Neither primary factor met its support criterion; all quality screens failed. Repeat-data LoRA was 51 -> 46/96 and adapters 32 -> 39/96; renamed-data controls were 14 -> 34 and 10 -> 24. The low-exposure repeated checkpoints exactly reproduce R99. Preserve all eight states and the failed preflight. The original-example panel is not in-sample for augmented data; do not mislabel it as augmented training-fit failure. Do not run more same-recipe alias/step-count variants. A successor requires a prospective recipient/interface positive control and grounded-content diagnosis, not an assumption that native LayerCake is broken or already English-qualified. No teacher-quality English, ABI novelty or product certificate is promoted.
New user authorization (2026-11-10): execute one bounded recipient/interface
diagnostic and publish its results. R103 is controlled by
experiments/recipient_interface_r103/PROTOCOL.md: first qualify the pinned
external English control on raw outputs, then compare raw/native versus the
historical Task/Response interface on four unchanged research checkpoints.
No training, new checkpoint, prompt sweep or native LayerCake source change.
Qualification failure stops comparisons; keep control, acquisition/interface
and host failures separate. Preserve every output and preregistered threshold.
R103 completed at the qualification stop: 41/48 raw passes versus 44/48, with email/format/tone each 4/6 versus 5/6. No recipient comparison or training ran. The contradiction proxy flags a future conditional in one row; preserve that limitation and every locked score, rather than silently rescoring the study. Even disregarding that flag does not clear qualification. A successor requires prospective evaluation-validity and control qualification work before recipient attribution; do not infer native LayerCake failure or launch a substitute run.
User continuation (2026-11-10): R104 may validate the evaluator, qualify one
pinned English control, then conditionally run the bounded recipient/interface
matrix and publish. Follow experiments/evaluator_control_r104/PROTOCOL.md.
Preserve R103; no historical rescoring. Keep 44/48 and 5/6 floors, zero collapse
and truncation. Finite author-specified scorer fixtures are not human fluency
validation. No new training, prompt/model sweep, native LayerCake change or
weight-editing backend is authorized. Keep prerequisite, acquisition/interface
and host defects separate. The weight-space research note is roadmap context,
not an experimental superiority result.
R104 completed at its prerequisite: 29/48, with summary 2/6, email/tone 0/6, clarification 4/6. All 155 finite author-specified fixtures passed, but two live approval-scope limitations remain; do not call this general semantic validation. Removing every approval flag yields at most 30/48, not a passing repaired score. The legacy screen's 44/48 also misses the clarification floor. Preserve all negative evidence and both scores. Recipients and training did not run; native LayerCake remains untouched. No prompt/model sweep or automatic successor.
User continuation (2026-11-11): prove prerequisites before claiming a breakthrough.
R105 is one frozen external-label evaluator falsification under
experiments/evaluator_scope_r105/PROTOCOL.md: conventional NLI, all 1,200 ScoNe
test rows plus 64 approval-scope cases, one live replay, zero training/teacher/
recipient calls. External gold is not new human ratings; overlap is disclosed.
Do not substitute evaluator success for an English-control qualification or
relax any existing gate. Preserve R104 and keep native LayerCake unchanged.
R105 completed and rejected the conventional evaluator: 779/1,200 external gold
and 62/64 approval cases. Its two single-scoped-negation families score 4/200 and
5/200; raw replay is byte-identical. This is evaluator failure, not acquisition
or native LayerCake failure. Use the additive evaluator_scope_r105.complete
verifier; the frozen stage verifier alone permits an absent replay directory.
Preserve that hostile-audit failure and the repaired full-study checks. No
post-hoc label inversion, model/threshold sweep or English-control promotion.
User continuation (2026-11-11): R106 is a separately preregistered diagnostic
under experiments/grammar_interface_r106/PROTOCOL.md. Use externally labeled
grammatical minimal pairs and exact conditional likelihood, not a replacement
NLI judge. Qualify only a narrow grammatical-discrimination control before the
conditional frozen-recipient/interface comparison. All target tokens are supplied;
this cannot qualify free generation, reopen R104's matrix, or certify an English
core. Preserve the original generation gates and negative evidence. No training,
model/route/prompt sweep or native LayerCake modification. Publish bounded results.
R106 completed at its narrow control stop: 178/200, including irregular verbs 37/50 against 45/50. Raw replay is byte-identical. Recipients and training did not run. Preserve the apparent upstream label reversal and judgment limitations without rescoring. Sentence likelihood is not autonomous fluency; no generation gate, ABI acquisition or native LayerCake claim is promoted. Do not run an easier subset, normalization/prompt/model sweep or automatic successor.
-
Execute the frozen human-rating packet with three independent raters.
-
Run the external reproduction packet on genuinely different hardware.
-
Execute registered minimum-information certification.
-
Treat R8, R9, and the overall R10 source-gate failure as preserved falsification results. Any successor native-transfer mechanism requires a new additive preregistration and a materially different canonical IR or recipient injection architecture. It must first fit a capability-specific public control and solve public capability-level realization before universal or held-out use. Do not run further final-state/embedding-state GRU, nearby width, learning-rate, or step-count variants.
-
Prove teacher-to-ABI acquisition, capability labeling/segregation, compact English extraction, and quality comparisons against distillation and LoRA only in separately registered lineages.
-
Treat R11 as the passed construction prerequisite and R12-A as a bounded public frontend construction pass with a failed exact source prerequisite. Do not run another nearby learning-rate, step-count, data-size, or LoRA sweep. A successor must materially change source acquisition or teacher formulation, clear exact public behavior first, then extract a capability already encoded in an independently trained open-weight teacher using the unchanged R11 package/host boundary and held-out teacher-relative behavior. Capability-specific recipient optimization and answer-table substitutes do not qualify.
-
Preserve R13-B as the passed local finite-table control. Publish its hash-addressed bulk replay assets before claiming public reproducibility. Keep its capability-extraction verdict separate from Track A teacher-output equality and do not rerun nearby enumerable-table variants.
-
Advance capability extraction only through a separately registered non-exhaustive capability where the probe budget cannot enumerate its full specification. Track A behavioral transplantation remains independently open.
-
R14 executed the output-probability version of that test and failed 2/3 latent recoveries. Preserve it. Do not run nearby query-budget, LoRA, schedule, or affine-family variants. A successor must materially change the frontend through preregistered controlled access to foreign weights, activations, residual representations, or training deltas while keeping R11 frozen. It must keep capability recovery separate from Track A equality.
-
Preserve R15A as the passed learned-weight-delta prerequisite. Its source capability was deliberately installed through complete atomic supervision; do not describe it as pretrained knowledge extraction. The next registered stage must test pre-existing, non-enumerable source capabilities with contamination-resistant labels and holdouts. Keep the R11 package and recipient boundary frozen and report Track A teacher behavior separately.
-
Preserve R15B as the passed pre-existing representation prerequisite. Its source prompts explicitly specify modular arithmetic, its generated reasoning exposes the answer before the captured final position, and its labels are externally registered. The next stage must remove these shortcuts by extracting a broader pre-existing capability whose held-out behavior and semantic segregation are independently scored. Do not call R15B English or arbitrary-domain extraction.
-
Preserve R16 public v1 as a source-interface failure: 15/16 open answers, 43/48 shuffled-choice selections, 13/16 consistent extracted facts, and 48/48 semantic labels. Do not delete failing chemistry facts or lower the exact gate. A public successor may replace choice-letter prediction with answer-free generation plus registered candidate-sequence scoring before any held-out commitment.
-
R16 public v2 passed 48/48 answer-free generations, 48/48 candidate selections, 16/16 cross-view facts, and 48/48 registered semantic labels. A held-out successor must freeze its extractor, package schema, physical isolation, factual universe, label ontology, exact gates, and evaluation paraphrases before committing or revealing a secret selection. It must keep structured factual extraction separate from English transfer and LoRA/distillation comparisons.
-
Preserve the R16 held-out pass as bounded structured factual extraction: 16/16 facts, 48/48 package/source agreement, exact target-only behavior, exact other-domain/removal abstention, 0/48 rotated-score control, exact live replay, and 21/21 hostile cases after additive assurance repair. Blind review passed the bounded claim with no Critical or High finding. Before clean replication bound to the corrected public protocol has now repeated every gate on a second selection (27 distinct facts across both runs). Final blind review of exact replication commit
26fb029passed with zero findings. Before promotion beyond local scope, require durable publication and reconstruction. A successor must test free-form linguistic realization or a non-enumerable specialist capability; do not rerun nearby fact-count, prompt, candidate-set, or ontology variants. -
Preserve R17 public interface v1 as a source prerequisite failure: 38/72 exact extraction outputs and 28/48 exact evaluation outputs. Compiler and LayerCake were not invoked. One v2 interface/evaluator repair is authorized because the raw failures isolate question mood, negative polarity, future tense, and an unnatural negative-question order. Do not change the 24 signatures, lexical split, compiler, controls, or gates. If v2 fails the 120/120 source gate, close this interface branch.
-
R17 v2 reached 67/72 extraction and 47/48 evaluation source exactness, so the registered 120/120 gate failed and the exact-template branch is closed. Preserve all rows. Do not run another prompt-wording, sample-count, or exact-template variant. A successor must score functional meaning, grammaticality, and teacher-relative quality separately and use a materially different package representation.
-
R18's frozen factorized representation passed its public prerequisite: 48/48 functional evaluation rows versus 47/48 for the source and 46/48 for independent modal consensus, with zero regressions, 0/48 control, 48/48 removal, strict verification, byte-exact live replay, and 12/12 hostile rejection. Its preregistered hidden replication then failed 46/48 with two source-correct regressions in present-positive plural questions: cross-number pooling selected singular
Doesinstead of pluralDo. Preserve the failure. Do not rerun that selection or tune another source prompt. A successor must materially change factorization and must earn a new hidden replication. Do not call R18 unrestricted English, autonomous extraction, LayerCake ingestion, minimality, or LoRA/distillation superiority. -
R19's same-number polarity-contrast mechanism passes both disclosed development sets at 96/96 total functional rows with zero regressions, 0/96 controls, 96/96 removal, strict verification, 12/12 hostile rejection, and byte-exact live replay. Preserve all operational/protocol failures that preceded the complete result. Its separately committed fresh lexical replication also passes at 48/48 package functional rows versus 43/48 teacher and 48/48 modal, with zero regressions, 0/48 control, 48/48 removal, 22/22 hostile rejection, and byte-exact live replay. This promotes only the registered supplied-slot 24-signature realization family. Do not run another nearby lexical, prompt, template, or signature variant. A successor must materially broaden the representation and free-form English behavior.
-
Preserve R20 as a public source/representation failure: 10/72 extraction and 19/120 evaluation outputs matched the registered strings, and the compiler could not identify six sufficiently supported programs. Compiler, package, and LayerCake paths were not invoked. Do not repair this by hand-coding the six oracle renderers, loosening the exact gate after seeing outputs, or tuning another teacher prompt. A successor must use a materially generative representation and separately score semantic content, fluency, adherence, hallucination, and teacher-relative quality against a matched sequence-distillation baseline.
-
Preserve R21 as a strictly verified public prerequisite pass and a strictly verified hidden replication failure. Public ABI-factorized packages scored 116/120 on all three seeds versus 96/120 teacher, 94-99 labeled monolith, and 71-80 raw sequence; fresh replay covered 1,080 GPU rows, 54 CPU rows, and 24/24 package removal, restoration, and targeted corruption controls. The committed hidden split scored 113/120 on every factor seed versus 107/120 teacher, but summary was 13/20 against the frozen 15/20 gate. Do not lower or rescore that gate, reuse its seed, or call R21 general English. Teacher self-labeling failed both registered interfaces (583/600 and 534/600); the 600/600 registered ABI label control is supervised ontology labeling, not autonomous discovery. A successor requires a materially broader semantic-plan representation and a proposition-level evaluator frozen before a new hidden selection.
-
Preserve R22 as a failed teacher-normalization prerequisite: 520/600 functional and 255/600 exact-field targets, despite 600/600 grounded and non-collapsed outputs. No student training or LayerCake execution occurred. Do not retry nearby rewrite prompts or reinterpret this as a LayerCake failure.
-
Preserve R23 as a strictly verified bounded semantic supplied-content pass. On a new seed committed after evaluator freeze, the exact 24 R21 package bytes scored 120/120 on all three factorized seeds versus 96/120 live teacher and both controls. Fresh replay regenerated 1,080 GPU and 54 CPU outputs, passed all 24 removal/restoration controls, and rejected 24/24 targeted tensor corruptions. Preserve the zero-row live-launch failure and its metadata-only repair. Do not call R23 unrestricted English, autonomous discovery or labeling, arbitrary-domain extraction, global minimality, or LoRA/distillation superiority. A successor must materially expand beyond supplied three-field behavior and retain prospective semantic evaluation.
-
Preserve R97 as the passed prospective two-hop nonce-reasoning endpoint: 1,399/1,400 candidate versus 1,088 teacher, 60 unchanged parent, and 17 random; 1,087/1,088 teacher-correct retention; zero collapse; 1,400/1,400 sparse execution; strict recomputation; 12/12 hostile rejection; and byte-identical live replay. Do not run a nearby label-length, wrapper, optimizer, seed, or reasoning-family variant. A local successor must materially broaden to fluent English or a non-enumerable specialist capability, use a separately sealed prospective evaluation, and include matched LoRA and sequence-distillation controls under explicit information and compute budgets. Human quality and independent hardware remain external gates and cannot be replaced by self-ratings or same-machine reruns.
Do not claim the ABI moonshot until those acquisition and external-quality gates are supported by their own evidence.