# Version1.1.1 hotfix verification

The raw-semester startup fix has its own regression/integration tests and CLI audit log. `semester_fix_unit_tests.log` is the current full-suite result. The504-cell training/replay/resume files below are retained evidence from version1.1.0, not a new training run of this loader-only patch. See `docs/SEMESTER_ZERO_FIX.md`.

# Suite35 PACE1.1 verification

All numerical results in this directory are **synthetic software verification**, not measured AQ performance.

## Final evidence

| Check | Result |
| --- | --- |
| Full unit suite |164 tests passed |
| Native input schemas | Actual predecessor canonical joblib,17-feature sidecar and DDTBHP contracts exercised |
| Full native-schema training |504 cells:36 variants ×2 faculties ×7 folds ×1 seed |
| Architectures/objectives | MLP,GRU,causal Transformer; grade,course,joint SSL |
| Primary arms | H0,H1,N0,N1,N2,N3,H2,HR |
| Saved-probability AP |504 recomputations matched reported AP |
| Full-pipeline checkpoint replay |504 cells passed |
| Automatic resume |504 cells replayed; zero retraining starts |
| Checkpoint/source integrity | Original inputs unchanged; checkpoint hashes and mtimes unchanged on resume |
| Development freeze | F1–F5 only, then F6–F7 retrospective |
| Consolidated log | Captured output matches the complete log byte for byte, including resume |
| Final raw reconciliation probes | Five unequal/missing-ID and consumed-evidence cases correctly rejected |

The training fixture uses dimension8,one layer,one SSL/neural epoch,one seed29 and2 HGB iterations to cover every execution path quickly. These are verification budgets; the real no-argument launcher uses the full research defaults. CPU PyTorch2.5.1 was used. Actual AQ cache/database bytes, full-scale runtime/memory and CUDA execution were unavailable.

`native_schema_summary.json` records the exact command, final modeling-source digest and checks. `source_sha256.json` records packaged code/configuration hashes. `native_execution_and_resume.log` is the complete single log. Detailed synthetic results/report, source audit, protocol freeze and probability-validated paired contrasts are included. No synthetic scores are presented as evidence of research improvement.

## Corrections covered

Tests cover finite CF coverage/permutation alignment, full versus covered support, exact grade identities, ambiguity and conflict rejection, numeric coalescence, missing-credit recovery/provenance, target-only identity joins, strict earlier-semester prefixes, automatic categorical CSV fallback, population-standard-deviation parity, empty histories, future exclusion, unique-prefix tensor storage, dense/indexed training parity, keyed chronology shuffling, checkpoint corruption guards, archival population counts and log snapshot replacement.

The native chronology intentionally inherits the earlier suites' previous-semester-grade availability assumption. Internal ordinal timestamp carriers are not observed publication dates. Passing these checks cannot certify upstream feature provenance or grade-revision timing.

## Reproduce

```bash
python3 -m unittest discover -s tests -v
python3 verification/run_native_smoke.py --directory /tmp/suite35_native_proof
```

Use a new directory for the second command. It creates only synthetic inputs, trains all504 cells, validates probabilities/checkpoints, resumes automatically and verifies complete log capture and immutable sources. For the actual AQ study use `./run_suite35.sh` from the package root.
