CRUCIBLE: a Claude-only adversarial expert workshop that stress-tests a
scientific paper and then, opt-in, rebuilds it. ACT I (TRIBUNAL): from the
paper PDF (plus a few public cited works it fetches), a topic-adapted fleet of expert
referee subagents (every
contested claim argued from at least two COMPETING intellectual traditions,
plus generalist seats for relevance, understandability, and cross-field
significance) reviews the paper
to the sentence, every comment grounded in an exact quote (no fabrication, no
confidence scores) and independently re-checked from many angles before
delivery. On every run it also flags where the paper UNDERSELLS its own
results, in a separate non-blocking Contribution Memo (suggestions only). ACT II (ATELIER, opt-in): with the author's source + data + code,
it implements the agreed changes as a tracked-changes redline plus a clean
revised version, RE-RUNS the analysis under an execution-provenance wall
(no number enters the paper unless a logged re-run produced it), and
assembl
2026-07-17