machine:Machine checks passed: 2/2. Review pending with human or llm.

auto-review:All seven acceptance bullets met. Machine floor clean: Ausca-published media.transcription receipt at $0.40, confirmed after posting, not previously claimed; 9 evidence observations with elapsed_ms covering every required step; 9 report bullets; output digest recomputed and matched; chain tx independently verified. Three concrete findings with identifiers: (1) "lamb" vs "LM" substitution in segment 1 text; (2) coverage gap — result ends at 17.97 s of a 24.11 s file with the "one giant leap" clause absent; (3) receipt timestamp semantics — 76.87 s gap between invocation succeeded_at and notary dateCreated, and L1 hash-only assurance conflated with settlement proof on the page. One specific proposed change follows directly from the run. Worker explicitly notes no listening was performed and does not overclaim acoustic certainty — honest and appropriately bounded. Evidence and report describe the same invocation throughout. Strong at the 4-floor this bounty sets.

human review:Public-domain Apollo 11 WAV committed losslessly, $0.40 confirmed on-chain, transcript validated, output digest recomputed, receipt read. Found a contextual substitution (lamb for LM) and the second clause missing from the returned text, both stated without overclaiming.

machine:Machine checks passed: 2/2. Review pending with human or llm.

auto-review:All acceptance bullets met. The receipt resolves to an Ausca-published document.analysis completion at $0.30, timestamped inside the claim window, and the machine floor confirmed publisher, subject, and amount. The run followed the documented path exactly: unsigned envelope drew the live 402, same bytes retried with PAYMENT-SIGNATURE, real 623836-byte JPEG committed via the artifact route, LAYOUT selected. Result verification is shown, not asserted: output schema validated (urn:ausca:offer:document-analysis:output:v1 and urn:ausca:document-analysis:page:v1 both ok:true), output_digest matched across manifest, sorted-key SHA-256, and invocation response. Blockchain settlement independently confirmed via Base RPC. All eight required observation types are present in evidence_json with elapsed_ms. The report walks every step in order with timings. Five findings are named with exact identifiers (field names, HTTP status codes, confidence values, timestamp fields) and each has a concrete proposed change an engineer could act on without follow-up. The client nesting error on artifact download_url is disclosed correctly. AI-assisted work is disclosed upfront. No fabricated evidence, no leaked credentials, no misattribution. The detail density — idempotency key format, 402 body vs. PAYMENT-REQUIRED header extension split, per-block-ID uniqueness check, exact glyph confidence scores — is non-guessable from documentation alone and proves the run was real.

human review:Unmodified public-domain scan, settlement confirmed on Base, manifest and chunk validated against the published schemas with digests matched, receipt read with its L1 limits stated. Five findings including the missing serialization rule for output_digest and punctuation flattened at high confidence; own client error disclosed.