# Bash SFT note 03 evidence packet

Sanitized public evidence for the 30 August 2026 Session 3 research note.

## Files

- `summary.json` - headline numbers, harvest mix, claim boundary
- `samples.json` - eyeball pass/fail completions from last40 wd=0.1
- `four-regime.json` - mix-composition arms A-D
- `locked-eval.json` - GFR / harvest / last40 locked execution aggregates
- `last40.json` - last-window recipes including the Kimi K3 continue
- `fair-baselines.json` - matched-prompt kept, stock-base, and shell-specialist comparison
- `report.md` - compact companion narrative
- `SHA256SUMS` - hashes for every packet file except itself

The packet excludes training JSONL, private traces, prompts, full completion
dumps, fixtures, local or remote paths, credentials, and weight tensors.

## Claim boundary

Arm A won a different corpus (imitation val loss). The kept product checkpoint
is last40 wd=0.1 at 92/98. Hardfam upsample and the Kimi K3 continue did not
beat it. Under the separate matched-prompt comparison, the kept checkpoint
passed 90/98, the exact stock base passed 8/98, and NL2Shell 0.8B passed 4/98.