v0.2.2 — Measurement Infrastructure
Date: 2026-08-29
Statistical rigor for published metrics: noise-floor CIs, permutation control, shared RLHF priors.
What's New
- Noise-floor baseline — Bootstrap CIs for convergence score, verdict rate, capitulation rate across 5 pairs
- Permutation control — LLM judge validated (87.4% match rate, 77.8 sigma above vocabulary floor)
- Shared RLHF priors — Causal mechanism documented: non-Mistral models rubber-stamp due to shared conversational defaults
Key Numbers
- Zero new LLM calls, zero new debates — all post-hoc on existing data
- 21 new deterministic tests
- $0.00 total cost
See CHANGELOG and Release Notes for full details.