Status: completed 2026-11-10; control qualification failed at 41/48.
The registered recipient/interface comparison was not executed.
FULL_MOONSHOT_NOT_PROVEN remains controlling.
R102 did not qualify any English core or support its two registered primary factors. R103 tests one measured interface difference: the R76 acquisition used raw prompts followed by a newline, whereas R98-R102 used a Task/Response wrapper. There is no training, checkpoint selection, optimizer change or new LayerCake runtime implementation.
Read the locked protocol.
First qualify the pinned Qwen2-7B-Instruct control on 48 new supplied-content inputs using native chat and raw autonomous outputs. Require at least 44/48, at least 5/6 per task, zero collapse and zero truncation. Qualification failure ends this run before recipient comparisons; no substitute model or retry.
If that prerequisite passes, compare both formats on a separate 48-input set:
| Checkpoint | Native interface | Benchmark interface |
|---|---|---|
| Qwen positive control | System/user chat template | Plain Task/Response completion |
| R76 parent | Raw prompt plus newline | Plain Task/Response completion |
| R100 full-model control | Raw prompt plus newline | Plain Task/Response completion |
| R99 LoRA | Raw prompt plus newline | Plain Task/Response completion |
| R99 adapter distillation | Raw prompt plus newline | Plain Task/Response completion |
All recipients retain their exact weights, tokenizer and capability route. Qwen is an external task-feasibility control, not a size-matched transformer baseline, transferred English core or teacher resident inside LayerCake. Differences in scale, precision and prior acquisition preclude architecture-only causal claims. Within-checkpoint input-format comparisons are paired.
The two primary contrasts are native-minus-benchmark for the parent and the full-model control, with 10,000-resample paired 97.5% intervals. Other checkpoints are descriptive controls. A recipient must still pass every unchanged mechanical quality condition and teacher-relative noninferiority; no screen alone promotes a fluent product core, broad English or CPU performance.
The new inputs reuse known supplied-content task templates, with new identifiers, day assignments and note order. They are not unseen semantic tasks. Every arm receives the same explicit 120-word request, fixed before qualification to align with the existing bounded-response screen. No source-span normalization, answer prefix, output repair, retrieval or teacher scoring is used.
All input sets, scientific code, tests, source inventory and checkpoint dependencies are frozen and committed before GPU work. Raw tokens/text, exact prompt rendering, route traces and resource receipts are retained. The bound is 90 minutes, at most 144 control and 384 recipient generations, zero optimizer steps and 1 GiB new evidence. Historical failures and unrelated running work are preserved.
Failure to qualify the control is a control/setup finding. If the control passes but the recipients fail, the limitation belongs to those acquisition/checkpoint/ interface combinations until a separate test localizes a native LayerCake defect. This diagnostic does not certify or disprove the underlying host architecture.
python -m experiments.recipient_interface_r103.run verify
python -m pytest -q tests/test_recipient_interface_r103.py tests/test_r103_completed_verifier.pyThe verifier is for completed evidence, not permission to overwrite a run.
Software preflight: 20 focused tests passed, and the full review passed 313 tests with four declared skips outside the Windows sandbox. The initial sandbox run's two temporary-file rename permission failures remain in the operational record and separate failed JUnit receipt. No model-result claim follows from these checks.
The pinned native-chat control produced all 48 raw responses. Strict verification
returned CONTROL_QUALIFICATION_FAILED: 41/48 passed versus the locked 44/48
minimum, and three tasks missed their 5/6 floors. There was zero collapse and
zero truncation. No recipient model was evaluated, no optimizer step ran, and no
checkpoint or native LayerCake source changed.
| Task | Raw passes / 6 |
|---|---|
| Summarization | 5 |
| Rewriting | 6 |
| Email drafting | 4 |
| Conversation | 6 |
| Format control | 4 |
| Tone control | 4 |
| Clarification | 6 |
| Abstention | 6 |
The seven failed rows include three required-literal omissions, two bullet-line format failures, one overlength response, and one flagged approval contradiction. The identifier error overlaps one omission. These counts are mechanical proxy outputs, not independent human judgments of English fluency.
The contradiction regex flags as soon as the details are confirmed in
r103-qualification-020-tone_control. That is a future condition, not an assertion
that approval already exists. This is an observed limitation of the literal
proxy, preserved in an additive regression test. The response may have other
unmeasured adherence issues; no new semantic score is assigned to it.
The frozen evaluator and 41/48 result remain unchanged. Even disregarding that one flag cannot reach 44/48, and email/format task floors still fail. This post-run sensitivity observation is not a repaired or promoted quality result. Do not claim that this control cannot speak fluent English, that recipient interfaces were compared, or that ABI transfer or native LayerCake failed.
The next bounded proposal should validate the evaluator's semantic boundaries and establish a genuinely qualified control before resuming recipient attribution. No new prompt, source substitution, model run or gate change is authorized by R103.
- Preregistration:
2c62b74; successful freeze commit:aac5f08. - 725-input freeze:
709b4c6a5da6802a53ea53d64041d60db0786b984aeceb4cc2d3425c8df8b31a. - Decision:
3420c88c6e608b32801bdc144031ac011abaf085af6f6c3861efa4afda1b31bc. - Raw responses:
d05ac3f2bd344c87acbd397f631ebb102f30823087d289a620504207e75658bd.
The external control used 48 requests, 8,690 input tokens and 4,190 generated tokens (including terminal tokens); unique rendered prompts contain 39,285 UTF-8 bytes and outputs 17,895 bytes. Generation time was 248.01 seconds; the complete control process recorded 415.51 seconds wall and 408.64 CPU-process seconds, including identity checks and loading. These are diagnostic costs, not CPU inference benchmarks or acquisition-efficiency claims.
The 7,615,616,512-parameter control used local GPU FP16, with peak CUDA allocation 15,330,695,680 bytes and sampled peak RSS 2,328,981,504 bytes. Sampled RSS is not an OS high-water guarantee. Its parameters were not copied into any recipient. The four inherited 87,710,428-parameter research checkpoints remain bound inputs, not newly acquired or English-qualified cores. Their previous acquisition costs are not erased. There are no new checkpoints, imported logits/hidden activations, trained bridges, or training tokens from this diagnostic.
The supported review passed 297 default tests with four declared skips plus 16/16 exact historical Phase 7 tests (313 total). Separately, 12 additive completed-evidence checks passed in 453.81 seconds: the actual failure recomputed, ten hostile mutations failed closed, and the observed temporal-regex limitation was reproduced. Tampered copies were used; the original study files did not change. The hostile-test receipt is separate from the original preregistered 20-test scientific surface.
Documentation checks, focused lint and source/wheel builds passed. Package API
version remains 0.4.0a1. Software success does not qualify the control or fill
the unexecuted recipient matrix. Public reconstruction is reported separately.
The complete review manifest
inventories 560 repository assets and all four required inherited LFS checkpoint
paths, with external host bindings recorded separately. SHA-256:
6a9cf5a7c46d02ea48331828cfd643209624b72b8c9215ebb6ce68c1d96dddc7.
Unlike an inference replication, raw evidence recomputation needs only the five
listed control tokenizer/config files, not the source's full model weights.
The research prerelease
is abi-research-r103-2026-11-10, source commit
cc6b59cd8b0b72730a0ae847015c42a9c7a22ab5, annotated tag object
db1d4b999e6351b28e552115603516a3d3f1c1fc. No earlier tag was moved.
Use sibling public checkouts named abi_release and layercake_release, with
the host at 865a2b9e59e6221caa1185f9b64479d2077953f3. Windows users should
enable Git's core.longpaths only in their review clones. Install the Python
dependencies named in the manifest and obtain the five pinned public Qwen
tokenizer/config files. The R102 cache setup
uses the same source revision and avoids Windows symlink privileges.
Download r103_manifest.json and review_r103_public.py from the release outside
the checkouts. Require manifest SHA-256
6a9cf5a7c46d02ea48331828cfd643209624b72b8c9215ebb6ce68c1d96dddc7
and helper SHA-256
11f309d9749bf2099e2084a5aa36b6886c1c7d50aa2397e481130c9455820fb8.
Hydrate exactly the manifest's four LFS paths, then run the helper:
$r103Manifest = Get-Content r103_manifest.json -Raw | ConvertFrom-Json
$r103PayloadPaths = $r103Manifest.lfs_paths -join ','
git -C abi_release lfs pull --include=$r103PayloadPaths --exclude=
python review_r103_public.py --abi-root abi_release --abi-commit cc6b59cd8b0b72730a0ae847015c42a9c7a22ab5 --manifest-sha256 6a9cf5a7c46d02ea48331828cfd643209624b72b8c9215ebb6ce68c1d96dddc7 --output public_r103_reconstruction.jsonSet HF_HOME to the public tokenizer cache before running this command. Existing
receipts are never overwritten. This recomputes raw evidence and runs 32 focused
checks; it does not run a teacher, evaluate a recipient, or train anything.
Public-only sibling checkouts matched all 560 inventoried assets, including the
four inherited LFS paths, and all 725 frozen bindings. The unchanged verifier
recomputed CONTROL_QUALIFICATION_FAILED with the identical decision hash;
32 focused tests passed. Both source checkouts remained clean. GitHub verification
passed on Python 3.10, 3.11 and 3.12 for the sealed source revision.
The public reconstruction receipt
has SHA-256
e224006dfe1a19c0eab10bfb185d41502d5ca181b1bb256a934cf3b1f3f22306
and verdict BOUNDED_CONTROL_QUALIFICATION_FAILURE_RECOMPUTED.
This reused the public-only R102 checkout directories, their already downloaded public LFS objects and public tokenizer cache, the same physical machine and installed Python dependencies. The R103 tag and helper/manifest were fetched from GitHub; no development artifacts were copied into those checkouts. This was not a brand-new clone, live generation replay, independent hardware or human quality review. It does not qualify the control or establish recipient quality. The receipt is additive; the scientific tag and frozen scores are unchanged.