What exists now
The site publishes a false-neighbor testset, a proximity vs causality protocol, and judgment guardrails. These elements frame future measurement without publishing a result.
What is still missing
Behavioral measurement requires a frozen corpus, model panel, versioned prompts, judges, verdicts, and a Q-Ledger style publication.
- Do not confuse testset with result.
- Do not infer an improvement rate.
- Publish deltas separately.