What ships
- Sanitized aggregate evidence for the Qwen3-0.6B LoRA SFT and historical preference-training study
- Dependency-light contamination audit, paired bootstrap, sign-flip, exact McNemar, Holm correction, and blind-review completion gates
- Synthetic, model-free reproduction path plus tested Python wheel and source distribution contracts
- Responsive GitHub Pages workbench for exploring the published evidence boundaries
Evidence boundary
This release does not redistribute raw customer-service conversations, derived training rows, frozen holdout text, raw model completions, blind-review packets or keys, model weights, adapters, or machine-local paths. The automatic 36-case holdout is a deterministic post-hoc rubric, not a human-preference or production-quality benchmark. Human review remains incomplete at 0 valid decisions.
Key result
The automatic holdout gave Base / SFT / preference-model means of 54.92 / 63.31 / 61.72. SFT - Base was +8.39 with an unadjusted 95% bootstrap CI of [+2.08, +14.78] and Holm-adjusted p=0.102. Preference-model - SFT was -1.58 with CI [-6.58, +3.89], so this release makes no DPO-advantage claim.
Verification: 22/22 public tests pass; GitHub CI covers Windows and Ubuntu on Python 3.10-3.12; Pages and package smoke checks pass.