Repository navigation
Research and AI Judge
Research creates frozen, reproducible comparisons. It reuses the normal session engine while keeping research configuration, target order, Viewer inputs and evaluation evidence locked.
A Research run can independently control:
- Viewer Notes: OFF or CURRENT;
- Field Guide: OFF or CURRENT;
- manual Research prompt or selected prompt history;
- protocol and target pool;
- blinded conditions and participant order.
For prompt-comparison studies, 2–4 versions may be selected from the recent history. Lock freezes the exact content and hashes used by the run. Resume uses those snapshots.
Research never creates new Field Guide or Viewer Notes versions.
The Judge operates on frozen evidence. It distinguishes:
- confirmed target correspondence;
- confirmed surrounding context;
- plausible but unverified context;
- contradicted or inaccurate information.
General plausibility alone does not earn a confirmed hit. Existing frozen scores are not silently recalculated when the rubric or software changes.
Viewer Notes Impact comparisons use one shared Field Guide across conditions so the Field Guide does not become a hidden second variable.
Research exports preserve configuration, hashes, frozen inputs, evaluation provenance and result data required for independent review. The visible transcript does not expose provider-private continuation state.