# Suite34 validation — 21 October 2026

**Software validated on CPU. No real AQ experiment or CUDA training run was performed.**

## Version 1.0.1 patch verification — 22 October 2026

- Regression tests were run before changing production code and reproduced the
  missing bootstrap, missing shell export, missing metadata and unchecked overrides.
- Updated full suite: **49 tests run, 48 passed, one CUDA-only test skipped**,
  61.727 seconds. Exact output: `examples/verification_patch_tests.txt`.
- A separate full synthetic launcher again completed **112 cells** across seven
  folds, including A3, HGB and CacheReference, with full checkpoint replay.
- Repeating that unchanged command with `--resume` reused all 112 cells and
  reconstructed the complete report with no warnings. `examples/sample_report.txt`
  is this complete resumed 1.0.1 log; it supersedes the original example log.
- The CUDA workspace is set before PyTorch import, valid explicit overrides are
  preserved, and invalid/empty overrides fail. Tests inspect actual fresh-process
  import order and the shell child environment.
- The environment check now exercises CUDA deterministic forward/backward on the
  user's GPU. This check is marked `NOT_RUN_CPU` locally: no hardware claim is made.
- Independent focused review reported no Critical, Important or Minor findings;
  its runtime test rerun passed seven tests with one CUDA hardware skip.
- Package/model source hashes and the strict resume signature are retained.
  An existing 1.0.0 run needs a separate output directory under this version.

The evidence below describes the original 1.0.0 validation, retained for provenance.

## Evidence

- Full command: python -m unittest discover -s tests -v
- Result: **41 tests passed**, 48.351 seconds on the authoring runtime.
- Exact output: examples/verification_tests.txt.
- Separate complete launcher execution: two synthetic faculties × seven folds ×
  one seed × A0,A0M,A1,A2,P1,A3,HGB,CacheReference = **112 result cells**.
- **5,376 prediction rows**, each result cell retaining all 48 fixture targets.
- 84 neural models; selection and complete-history refit both exercised.
- Whole-target CPU checkpoint replay for every trained neural/HGB cell.
- A1/A2/P1 parameter equality and A0M within 5% confirmed in output artifacts.
- F6/F7 beta1/beta2 thresholds numerically identical for every faculty/arm/seed.
- Full launcher log: examples/sample_report.txt. These scores are synthetic
  software evidence only, not an estimate of AQ performance or study runtime.
- Syntax/import compilation and bash syntax validation passed.

## Tested contracts

Exact CF mask/finite filtering, equal-size shuffled sidecar rejection, missing
folds, feature/schema checks, future-history rejection, absolute semester
ordering, support-only fallback, all-missing/constant/categorical/OOV preprocessing,
full-path gradients, projection/auxiliary gradients, parameter matching, label
invariance, metric ties/undefined cases, deterministic budgets, checkpoint schema
guards and exact CLI replay.

Reporting tests cover macro cell weighting, full untruncated 1,500-row tables,
paired faculty-cluster uncertainty, complete/undefined/operational promotion
guards, retrospective decision isolation and frozen F7 threshold policies.

End-to-end tests cover every fold and neural arm, ordinary resume without
duplicates, source/configuration/artifact changes, required completion manifests,
uncommitted staging recovery, synthetic destination isolation, full restored
preprocessing/chronology audits, launcher failure logging and exit propagation.

The AMP overflow regression uses a real CPU GradScaler to verify skipped updates
and reduced persistent scale. It does not certify CUDA/FP16 kernel behavior.

## Independent review and fixes

Independent whole-package review found two Important defects, both corrected with
observed failing regression tests followed by the green full suite:

1. Synthetic cache overrides could bypass destination protection. Generation now
   rejects conflicting overrides and validates both actual destination trees
   before mutation; marked fixtures also validate their filenames/configuration.
2. Resumed logs lost preprocessing/chronology-selection events. These events now
   form hashed per-cell artifacts and are replayed in the final single log.

An empty completion manifest and interrupted staging directories were also
covered by failing regressions and corrected. No Critical issue was reported.
The full suite is the coordinator's verification, not an independent duplicate run.

One documented nonblocking limitation remains: malformed arguments/configuration
can fail before log-path setup and are reported in the terminal. Normal
environment/test/data/training/reporting failures are captured in the active log.

## Still requiring the server

Actual canonical joblib preflight and smoke execution, CUDA/AMP execution, resource
and runtime measurement, complete 14-faculty/3-seed university-data training, and
all scientific performance conclusions. The adapter was grounded in inspected
Suite32 source and original CSV headers; actual joblib bytes and the original
cache builder were unavailable locally.

Local versions are recorded in examples/validation_summary.json. Keep a PyTorch
build compatible with your server GPU; check_environment.py tests its NumPy bridge
and an available CUDA kernel before fitting.
