reinstator

Provenance evidence for model distillation. Every number below comes from a pre-registered experiment; each page links to its criterion and raw files.

Pages

• Status, 2026-09-19 (overnight autonomous run) (docs/STATUS-2026-09-19.md)
• Distillation Provenance Research Plan (docs/PLAN.md)
• Phase 0 result (experiments/00_reproduce/RESULT.md)
• Phase 1 pilot: controlled students with known teachers (experiments/01_testbed/RESULT.md)
• H1: how much does the reference choice matter? (experiments/02_methods/h1_reference_relaxation/RESULT.md)
• H2: mixture estimation instead of single-teacher attribution (experiments/02_methods/h2_multi_teacher/RESULT.md)
• H3: idiolect features to break the near-relative tie (experiments/02_methods/h3_idiolect/RESULT.md)
• H4 first cut: probe and generation budgets (experiments/02_methods/h4_probe_budget/RESULT.md)
• H5 first cut: an anytime-valid stopping rule for probing (experiments/02_methods/h5_guarantees/RESULT.md)
• R2: false-positive sweep (experiments/02_methods/r2_false_positives/RESULT.md)
• R3: three-tier evidence report (experiments/03_integration/RESULT.md)
• Evidence report: S_human (experiments/03_integration/reports/S_human.md)
• Evidence report: S_qwen3 (experiments/03_integration/reports/S_qwen3.md)
• Evidence report: S_r1 (experiments/03_integration/reports/S_r1.md)