# Suite27 release validation

Validated on 2026-10-02 UTC / 2026-10-03 Asia/Saigon. These checks verify software contracts and mechanisms, not real-data AP improvement.

## Fresh final checks

| Check | Observed result |
|---|---|
| `python3 -m pytest -q tests` on frozen source | **116 passed, 28 subtests passed**, 100.98 seconds |
| `python3 -m compileall -q .` | Exit 0 |
| `bash -n run_suite27.sh smoke_test.sh` | Exit 0 |
| ZIP extraction / CRC | Passed |
| `smoke_test.sh` from extracted ZIP, isolated temporary output | Exit 0; all six variants, preflight, logging, reload, resume, report collection |
| Independent read-only review | No unresolved Critical or Important findings; see `evidence/final_code_review.md` |
| Small actual path-reader/cost probe | Finite hard scores, 25 actual memory splits/60 splits, 13 memory nodes with comparable enabled ancestor context |
| Suite26 production evidence integrity | All 80 archived report hashes/bytes and all 10 available source hashes verified; aggregate reconciliation within 2e-12 |

`evidence/final_tests.log` contains the final full-suite result. `evidence/archive_smoke.log` contains the complete successful extracted-source run. An earlier smoke attempt was correctly rejected for changed report hashes; the final isolated-output rerun passed without relaxing any integrity guard. The final delivery archive contains the same frozen Python/shell sources checked by these tests, plus documentation, report tables and provenance.

## Tested environment

Python 3.12.14, PyTorch 2.14.1+cpu, NumPy 2.3.5, pandas 2.2.3, scikit-learn 1.8.0, SciPy 1.17.0, joblib 1.5.3, pytest 9.1.1, Linux x86_64. Production dependencies are bounded in `requirements.txt`; full tests additionally use `requirements-test.txt`. Your Suite26 GPU/Python environment has not been executed here. Every real run records its exact runtime and refuses replay under a different runtime signature.

## What the regressions verify

- Canonical row/key mapping, shared schema, target-grade removal, chronological uncovered-course reconstruction, train-only scaling/vocabularies, preserved finite/missing flags and preprocessing roundtrip.
- Same-student, strictly earlier candidate histories; integer history embargo; exclusion of query/current/future labels from evidence and cumulative faculty priors; explicit foreign-student permission only for the named shuffle intervention. External cached-feature provenance remains an assumption.
- Total validation-semester row support distinct from per-faculty support, same-term one-step origins, positive/negative support, exhaustive failed-origin audits and explicit partial-coverage selection.
- Hand-computed Newton split gain, logistic loss reduction, hard topology induction, a genuine memory split when current features are constant, support/leaf constraints and single-class native behavior.
- Actual raw-ancestor feedback, skipped faculty predicates, NumPy/Torch evidence parity, finite empty/single/constant histories, both outcome-value channels masked together and auxiliary heads absent from primary predictions.
- Main-loss retrieval gradients through the soft proposal, retrieval parameter change, stale pair rejection, mandatory complete regrowth and matched maximum proposal budget versus selected final score budget.
- All six variants, independent origin/seed fits, unequal eligible-cell weighting, deterministic simplicity tie breaks, target/later-label isolation, class-aware stock staged/refit/reload probabilities and recipe/forest schema consistency.
- Real all-six/two-seed end-to-end training across two synthetic folds, **24 per-seed checkpoint reloads**, keyed score agreement within CSV precision, correct whole-model seed averaging, completed-fold reuse and refusal of configuration/input/code/runtime/artifact changes.
- Future or contemporaneous development semesters rejected before operational-threshold application and before CLI fitting; fixed earlier-development policies, all-faculty confusions/budgets, zero-target support visibility and stable empty workload CSVs.
- Keyed reference alignment, duplicate/conflicting IDs/labels rejection, partial coverage, missing/one-class metric cells, fixed eligible student-cluster bootstrap cells, valid/invalid replicate counts, all-faculty table printing, combined stdout/stderr and failure exit preservation.
- Collection manifest bytes/SHA256 match each actual archived payload; weights and default row predictions are excluded. Resource tables include checkpoint sizes and label host RSS as a process-lifetime peak.

## Cost probe limits

The reproducible 20,000-row probe has 153 numeric inputs, 153 presence flags, 14 faculty indicators (320 prepared features), 32 history slots, an untrained frozen reader and four depth-four trees. Observed root-cache time 0.041 s, complete fit 1.556 s, hard prediction 0.189 s and lifetime host RSS peak 457.168 MiB. Exact values are in `evidence/synthetic_cost_probe.json`. It excludes input IO, preprocessing timing, reader pretraining, coadaptation, full development grids and real academic data. It cannot be scaled linearly into a trustworthy full-run estimate. There is no production GPU or maximum-size benchmark in this release.

## Scientific limits that remain

The original caches are not included or regenerated. Their raw-feature/course-metadata availability and grade-release assumptions require institutional verification. Hard histogram growth is greedy and approximate; root memory grids can miss node-local thresholds. The soft-subforest update is a proposal with no convergence or monotone improvement guarantee. A simpler model may win, and path strength/cycle zero remain valid outcomes. F6/F7 have already informed this research and remain historical comparative benchmarks. A real supported-origin run and prospective unseen origin are necessary to establish academic performance, generalization, cold-history robustness and practical cost.
