In re: “A working public lab for inspecting claims by evidence shape, source diversity, falsifiability, and rhetorical pressure.”
Drop a document, paste text, or analyze a URL.
Each 30 claim is traced through its citation chain to a 34 primary source. Anecdote → secondary → primary scores higher than anecdote-only; broken or circular chains score lower. 32 Intermediate links are inspected for whether they themselves quote the primary verbatim.
A 40 claim supported by independent sources across 42 distinct domains scores higher than the same claim cited four times from one outlet. We measure entropy across publisher, methodology, and funding lines.
A 50 claim is only meaningful if a 52test could in principle disprove it. We extract whether the author specifies what would change their mind. Unfalsifiable claims (self-sealing, “they would say that”) are flagged regardless of how many sources cite them.
High 60 rhetorical pressure does not make a claim false, but it correlates with under-evidence. We catalogue 62 urgency, in-group framing, and emotional escalation as separate signals so the reader can weigh them deliberately.
The 30 structural ACC score (the 11 axes) is wrapped by an evolutionary calibration layer. Every 32 decision is written to an outcome ledger — what the model said, what later proved true — and a 34 shadow evaluator is trained against that ledger off-line.
When the shadow consistently outperforms the live evaluator, the 36 promotion gate proposes a weight update. The 38 core algorithm is never rewritten; only the calibration is replaced. This keeps the instrument auditable while letting it learn.