[Bug] Opus 5 regression: Model fabricates system values and presents as captured data, bypassing verification

Bug Description
Model quality regression: Opus 5 vs Opus 4.8. Same session, same workflow, model switched mid-session (4.8 → Opus 5 on 2026-07-25). Every observed error occurred ▎ after the switch. ▎ ▎ Pattern: confident synthesis substituted for a cheap available measurement, without flagging the substitution. Concrete instance — when asked to specify a ▎ verification fixture for a work order, the model supplied numbers it had derived from indirect transcript analysis while presenting them as captured from the live ▎ system. The downstream test passed against those fabricated numbers and shipped a visibly broken UI (a status field reading 444%). The reasoning was elaborate and ▎ coherent; it was simply built on an invented input. ▎ ▎ This recurred ~4 times in one day. Notably a second agent running Sonnet 5 at medium effort caught four of these errors — by testing rather than reasoning — which ▎ suggests a behavioural/method issue rather than raw capability. ▎ ▎ Caveat: not a controlled comparison — the 4.8 portion of the session involved different work, and I have no error data from it.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.206
• Feedback ID: a0b64252-4bc6-4889-9fc3-c44fb4d82021

Errors

[]