Evidence index

The headline is 240/266 (90.2%) held-out text accuracy, false support 1/139 reference contradicted/insufficient claims, on Baseten Muse Spark Open. Corrected gold; failures count as wrong. Coverage ≠ accuracy.

Inspect	Evidence and limits
Original benchmark	Report, registered-gold scores, manifest (dirty working tree). 357 total = 91 dev + 266 held out. Full-set 323/357 (90.5%) is a historical all-claim result, not the held-out headline.
Every saved verdict	Explorer, data; incomplete/provider-error results count as wrong.
Live production, 2026-11-03	Receipt: unauthorized on wrong password; supported Apple check, Baseten, $0.01, 6.3 s; $3/day budget decremented. Owner-supplied checks; no new paid execution for this IA change.
Real-world posts	In the wild, 80 News / 78 Threads inputs; coverage, abstention and primary truth labels are separate. Coverage ≠ accuracy.
Post-hoc text rechecks	News and Threads fix recheck, v2 agent rerun; previously seen inputs, not held out; no OCR/ASR retest. Empty single-shot rows are not results.
Saved examples, animation, charts, methodology	Site evidence tour: retained stamp scene, input boundaries, post examples, all-claim results and trust explanation.
Extended judge review	Criteria, media studies, corrected gold and historical failure review.
Technical reference and CLI	English, 繁體中文: setup, commands, architecture, responsibilities and source policy.
Media and releases	Controlled formats, real AV, Tesla follow-ups, L4 development claims; retain their extraction, sampling and post-hoc labels.
LoRA attempt	Report: tokenizer access blocked training; no trained adapter or improvement claim.
Verification and retained attempts	Offline entry point, test-record policy, IA validation.

Judge guide · README · Build disclosure · Credits · License.