# Execution and review evidence — 2026-10-01

## Executed

**45 unittest cases passed**, with no test failures or warnings in the saved final test log. The full suite took 19.719 seconds in this environment. Evidence: `validation/unit_tests.txt`.

Tests cover canonical/structured row alignment, full history, prior-semester reconstruction, global chronology including uneven faculty calendars, train-only preprocessing/vocabulary, canonical IDs, unknown courses, grades/missingness, outcome-independent candidate selection, time-safe shuffling, keyed references and conflicting labels/metadata, paired coverage, dev-only F2 thresholds, IT confusion, faculty-fold macro AP and student-cluster bootstrap, complete pipeline/replay/resume, below-cap optimizer budget parity, and cross-fold feature intersection.

Model tests establish leaf mass conservation, seed-equivalent hard routing, adjacent-float32 split conversion, natural-count leaves after balanced split discovery, empty/partial padding and history permutation invariance, target-label independence, outcome-sensitive internal routes, target-independent/conditionally monotone readiness, and state roundtrip. **The main BCE alone** gives nonzero finite gradients to feature selectors, thresholds, leaves, memory keys/values, readiness association, target requirement, importance and gap weights.

A complete synthetic run executed two folds and two faculties through selection, fresh full-history refit, all three learned variants, normal/intervention inference, K sensitivity, keyed prediction output, metrics, operational thresholds, bootstrap and checkpoint reload. Replaying fold 1 reproduced 48 scores with maximum absolute difference **2.8357e-08** (CSV precision and batch grouping), below the 2e-07 verification tolerance. Tiny synthetic ranking results are not AQ evidence. Evidence: `validation/smoke_verification.json`, `validation/smoke_console.txt`, and `validation/smoke_results/`.

A separate **default-geometry one-batch CPU probe** used batch 128, 167 current features (136 raw + 17 course + 14 faculty), 64 trees of depth 6, history 32, hidden 32 and readiness 16. It executed forward/backward with finite probabilities and gradients. Leaf-mass error was at most 2.3842e-07. Parameter count was 1,166,896. Forward/backward took approximately 0.249/0.398 seconds, and peak process RSS was approximately 1,043 MiB for this probe. These measurements include this CPU process and test tensors; they are **not** GPU allocation, V100 throughput, or full-data host-memory measurements. Evidence: `validation/default_geometry_probe.json`.

Shell syntax (`bash -n run_suite26.sh`) and Python compilation (`python3 -m compileall -q .`) passed. The final archive is checked for ZIP integrity and contains source, tests, runner, requirements, design/review/docs and selected validation evidence; it contains no real student data or AQ-trained weights.

## Environment actually used

Python 3.12.14; PyTorch 2.14.1+cpu; NumPy 2.3.5; pandas 2.2.3; scikit-learn 1.8.0. CUDA was unavailable. See `validation/environment.json`. Requirements state supported minimum versions, not a claim that every version was executed. The original user log reports Python 3.13.5, PyTorch 2.7.1+cu118 and scikit-learn 1.7.1 on V100; that server environment was not accessible here.

## Independent review and corrections

Read-only code review confirmed the architecture's internal memory-conditioned split learning and found three Important issues: multi-semester fixed-origin feature provenance, lost legacy score aliases, and selection/refit budget mismatch below the cap. All three were corrected and regression-tested. The reviewer rechecked the corrections and confirmed the default protocol resolved those findings. The model gradient test was expanded to readiness parameters.

Additional verification caught an adjacent-float threshold conversion error; downward rounding now preserves sklearn routing on float32 inputs. A common schema is computed over every requested faculty/fold. Original-log review corrected the auxiliary head description to jointly trained, and verified that F7's original memory increment loses AP in **11 of 14** cells.

## Not established

- Real AQ training/inference, discrimination, calibration, superiority over tree-only/Committee/Council or any substantial performance boost.
- V100 CUDA execution, throughput, memory sufficiency or run duration.
- Full-scale host memory and cache-loading cost over all real rows.
- Timestamp-level grade availability or provenance of arbitrary precomputed canonical/covered features. These remain explicit cache assumptions.
- Latent readiness dimensions as identified skills, global monotonic risk, exact gradients through discrete feature choice, or an optimal/perfect framework.

The original Suite26 log is analyzed in `REVIEW_AND_RATIONALE.md`, and its printed complete summaries are extracted in `validation/original_result_summary.json`. That log is evidence about the original estimator, not this rewrite. F6/F7 informed this project-level redesign and must be described as reused historical comparison folds. An untouched later semester/cohort is needed to strengthen a final generalization claim.

Snapshot report paths under `validation/smoke_results/` record the local synthetic execution location. They are examples for inspection, not pretrained deployment paths; run the package on the existing AQ caches to produce the server's own outputs.
