Redcell and a human analyst, assessed against rubric 1.0, registered 2026-09-05 07:13Z, before either arm ran.
Historical replay · BSC block 120,123,441 The human runs are the original observations. The agent arm was executed later against that exact immutable BSC state using historical (archive) RPC access.
2 agent repetitions, 2 manual. Manual repetitions use a stopwatch. Time and cost are measured. Quality is graded blind by a language model (scripts/ledger-grade.mjs, DeepSeek, each anonymised answer scored three times against the rubric hashed before either arm ran, per-criterion median taken). Scores can differ between terse and detailed answers, so the two manual repetitions can score far apart: a full write-up against a one-line summary of the same finding.
