# Run log

2026-09-25: Implementation authorized with defaults documented in contract.md.
Initial implementation pass underway; corrective rounds used: 0 of 3.
Environment smoke check passed; model-specific checks remain.

Corrective round 1: Long-input oracle uncovered tokenizers 0.23.2 overflow
early truncation. Python/Rust parity alone had hidden shared data loss. Full
training was held. Replaced library overflow with explicit complete-token
windows, rebuilding token caches and checking complete token coverage. Initial
chunks-stride-42.pt is obsolete and must not be used.

Round 1 verified: manual cache 99,258 windows /24,285,133 tokens exactly
accounts for full content, specials, and overlap. Independent native/Python
boundary oracles and final-stride probe parity pass. No full training or final
test predictions occurred before this repair.

2026-09-25 19:53 UTC: Persistent local pipeline launched in tmux datafog-pii.
Official HF two-epoch recipe, measured batch16 float32 MPS, stride42. Diagnostic
learning and optimizer/RNG resume completed; MPS resume max parameter difference
9.31e-10, identical next-update loss and diagnostic outputs. Initial full-run
estimate 6,924 seconds plus overhead is below user-approved24-hour allowance.
Heartbeat complete-datafog-pii-prototype scheduled every30 minutes to continue
review/recovery and final handoff, with no routine progress notifications.

2026-09-25 22:08 UTC: Training completed in 8,102 seconds. Epoch2 selected by
lower validation token loss (0.00188162 versus epoch1 0.00357060).

Corrective round 2: Final export verification failed before calibration/test.
The ONNX export wrapper defaulted to training mode; export's restoration of the
wrapper mode recursively re-enabled dropout on the reference model. Independent
reload in evaluation mode showed all probe argmax values identical and max
absolute logit difference 1.15e-5. Fixed wrapper.eval() and assert evaluation
mode remains set. No weights changed and no final test was exposed.

2026-09-26 00:19 UTC: Pipeline completed all stages. Final review independently
recounted 90,018 test documents: TP242,853 / FP2,424 / FN5,886. All frozen
inference/evaluation input checksums and 16 packaged-file checksums verified.
Packaged native executable processed Unicode email and clean text with no
Python on PATH. Final report reviewed. Synthetic fixture recall and email
regex advantage explicitly retained; no tuning on exposed test results.
Prototype contract complete with documented representability/generalization
limitations. Two corrective rounds used. Follow-up heartbeat paused.