# Claim limits and known defects

Human recipe adoption is distinct from every generated item being correct. Original misses remain visible in preserved certificates and additive human-gate records.

- Refusion: 48 supported/2 flags among50, explicitly bound to ruling-E <=2/50 precedent. D5 boundary narration .030 versus <=.02 remains a miss. Refusion te3l is its own input-bound embedding, reused byte-for-byte from extract36; not fused_v2 matrix reuse.
- AUDIO-v3: adopted changed-row20; deletion can leave dangling clauses, and real sounds at a clip boundary may be removed. Changed-row audit is not corpus-wide accuracy.
- Captioner: hallucinated buzz/tone and boundary classes persist. On the separate fixed9-window serving-cost probe:5 empty/whitespace outputs and one literal `io`. This is a probe defect, not a corpus defect-rate estimate.
- Text-assisted Omni3B cost bar >=2x cheaper than Captioner30B+B1 failed arithmetically (B1/candidate0.744x at batch1,1.132x at batch8). Valid-output comparative cost claims remain held due to the defective Captioner baseline. Timing tables are sums of measured warm resident-stage work, not observed full-stack wall latency.
- Omni3B trained66,080 examples versus adopted4B180,000. No audio-specific claim at this size. Audited supplied-text inference is distinct from sensor-only inference and does not transfer QA/human passes to it.
- Adopted270M router retains its G2 zero-v2 diagnostic miss. Scores are model likelihood diagnostics, not relevance probabilities. Original negative separability finding remains; adoption does not erase machine-gate misses.
- Fused_v2 original support3/50 missed<=2/50 but was adopted under ruling E. Marlin has noisy endpoint/temporal classes and one unviewed item in its recorded audit. DVD has quoted-text/OCR hallucination classes. Speech tags preserve unsupported-language/null reasons and listener uncertainty. See each layer certificate/addendum/HUMAN-GATE.
- Candidate banks are machine-filtered, not human-audited; F5 anchor-order mismatch blocks a headline benchmark claim. Frozen v1 retrieval2195/QA2042 instruments are unchanged.