Minerval

← evals

Replays

A scorecard says how a graph came out. A replay shows how it was built: the sources landing one by one, the Extractor listing claims, the Matcher searching and deciding, each Steward’s tool calls and verdict, the Curator’s merges; for a two-arm episode, the same sources building two graphs side by side. Everything in a recording is derived from what the run wrote; nothing is narrated.

What a replay cannot show: anything the trace did not record. A run traced off yields events with no steps, and the player says so. The player refuses a schema version it does not know.