SRIC admission repair: runtime 0.3.3
27 October 2026 · Komorebi P. · Beta / work in progress · AI-assisted self-review
The three failures reproduced in the earlier audit are repaired in a separate 0.3.3 source edition. The original 0.3.2 sources and counterexample record remain intact. The corrected edition returns to the research reading list as a scoped beta instrument. No real-model result or broader semantic claim is promoted.
Changes and evidence
| Obligation | Correction | Observed qualification |
|---|---|---|
| A1: empty replay logits | Validate every capture and scoring observation before consumption. Require nonempty finite logits, finite scalar losses, and compatible latent-trace matrices. | Empty capture/scoring logits and malformed loss/trace controls are rejected. |
| A2: NaN hidden by aggregation order | Validate individual observations and every replay error before taking the maximum. | A NaN is rejected at all 48 individual positions: 8 calibration and 40 audit tasks. NaN and positive/negative infinity in scored output are rejected. |
| A3: absent identity inferred to be real | Require explicit engineered or real provenance. Real identity needs pinned source/checkpoint/base-model digests matching a precommitted expected record. Check identity stability around adapter calls. | Missing, ambiguous, contradictory, malformed, mismatched, and changing identity controls are rejected before acceptance. Artificial metadata is used solely for these schema controls. |
| Positive and finite-negative controls | Preserve the scientific fixture configuration and finite replay threshold behavior. | Full fixture remains CAUSALLY_VALIDATED with real-checkpoint evidence false. A finite mismatch remains ADAPTER_INVALID. |
| Usable command-line path | Remove an unused hard-coded fixture-file read. Require a new 0.3.3 seal. | A fresh custom-path fixture/seal/run/receipt round trip passes. An old-edition seal is rejected. |
The implementation adds observation/provenance validation at the pipeline boundary, standalone causal admission, and semantic escalation. Invalid evidence raises AdapterValidationError before an accepted result is emitted. The real CLI forwards the already-checked sealed identity to the pipeline. A matching identity record establishes consistency with declared provenance. It cannot independently attest arbitrary third-party adapter behavior.
Outcome: implementation and finite assertions verified in the stated range. The bounded run completed successfully on Windows with Python 3.12.14 and NumPy 2.3.5. All 15 unittest methods passed, including the parameterized controls above. Raw test output, contract, and edition manifest record the tested scope and source hashes. The runner's source copy has no Git checkout. Exact imported-source identities and new edition bytes are recorded separately.
The package guide gives both the qualification command and the four-command synthetic example. The collection workflow also runs the new suite with pinned NumPy. Workflow execution is separate from the recorded Windows run and does not expand its scientific interpretation.
Remaining boundaries
The real checkpoint, Torch/Transformers integration, GPU execution, semantic-estimator correctness, bootstrap coverage, and held-out scientific claims were not qualified here. This edition is a source-checkout runtime. No wheel or installer is qualified. The original release's other tests and external-source parity checks were not all rerun. Broad real-model work remains in finalization gate F12.
Gate F15 is closed for the three reproduced admission defects and their explicit regression controls in this edition. Completion labels remain experiment/prototype and experimental evidence. Fresh protocol seals and a separately audited real-model campaign are required before making external empirical claims. The earlier counterexample harness continues to target 0.3.2 and is not the repaired edition's acceptance suite.
This is sequential AI-assisted self-review, not independent peer review. No Lean execution or whole-program correctness certification is claimed. The repair review record binds this report, the computation, and its inspected source revision.