Minerval

← evals

Background

Ground rules

The rules every eval is held to. They come from the master plan, issue #334.

The rules

  1. No answer key for contested questions. The judge’s standard is the The public rules every agent is given as the first part of its instructions.constitution, never its own opinion.
  2. Two places reality can check the graph: predictions, which resolve slowly, and agreement between models, which is fast but relative. Neither becomes a truth score for the rest.
  3. Expert consensus is never the referent. Distance from consensus is reported, never used as a gate.
  4. One run is one sample. Groups of about three, and a change must clear the spread (comparing runs).
  5. Every invariance is tested benign and adversarial. The adversarial case shows the invariance is not empty. Not yet met.
  6. Numbers inform; they never decide. No eval score changes a verdict or a claim’s importance (judgment over mechanism).
  7. Stability plus calibration is the argument. A graph that is stable under changes that should not matter, and calibrated where it can be checked, has earned some trust where it cannot be.
  8. Judges are reviewed before they are trusted. No judge number feeds a decision until a person has read its verdicts (judge review).

Detail

The plan

The full plan, with the nine suites it lays out and their reasoning: issue #334. Standing per suite: not measured yet.