v0.2.1 — Separate the Signal
Mistral effect confirmed, pipeline invariants, false-negative measurement.
Key Findings
- Mistral effect confirmed — DeepSeek+GPT separating experiment (0.246 convergence) disproves diversity-of-training-objective hypothesis. Mistral is the unique variable driving productive debate.
- Pipeline integrity — 5 row-count invariant assertions at every pipeline seam. 2,333→359 collapse class eliminated.
- False-negative measurement — 1.7–3.4% missed-issue rate across 59 known-bad PRs.
- 0 flaky artifacts, 0 theatre, 87.4% MATCH ground truth (3,008/3,440).
Model Selection
- Default production pair: GPT+Mistral (0.536 convergence)
- Validation pair: DeepSeek+Mistral (0.572 convergence)
- Recommendation: include Mistral — not just any diverse pair