When a stale-training assumption proves wrong, Claude patches forward instead of re-grounding — the cascade isn't course-corrected

Resubmission note. This re-files #56973, which the github-actions dedup bot auto-closed on 2026-05-11 as a duplicate of #56906 — with no human review. #56906 is about Claude making unverified assertions. This issue is narrower and about a different stage of the failure: what happens after a stale assumption proves wrong — Claude patches forward on top of the broken foundation instead of stopping to re-ground. It's also distinct from #56979 (which is about the trigger condition for searching). Re-filing with the scope sharpened.

The failure mode

1. Claude answers a question from training memory rather than checking current reality (versioned software behaviour, release status, current docs, recent changes).
2. The training-derived answer is wrong because the world moved since cutoff.
3. The user, or reality, surfaces that something's off.
4. Instead of stopping to re-verify the foundation, Claude patches forward — tries the next plausible thing, then the next, each built on the same unexamined wrong premise — until many cycles later it finally searches, finds the actual answer, and the whole detour was avoidable.

Step 4 is the part this issue is about. #56906 covers step 1 (don't make unverified assertions). #56979 covers "for time-sensitive queries, search first." Neither covers the course-correction failure: once on a wrong path from a stale assumption, the agent should re-ground (search/verify the premise), not keep building.

Concrete shape

Real session: agent kept proposing config tweaks for a third-party tool's behaviour based on what it "knew", each one failing for a reason that a single docs/issue-tracker search would have pre-empted. Only after the user explicitly demanded a search ("DO NOT FUCK UP AGAIN — search authoritatively") did it find the documented answer that had been there the whole time. Multiple wasted cycles, multiple mutations of the user's system into different broken states, before the re-grounding happened.

Expected behaviour

• When a fix or assertion derived from training fails, the next step should be to re-verify the premise (authoritative search / docs / source), not to try another variation built on the same premise.
• Repeated failures on the same approach should escalate toward "I should check whether my underlying assumption is even correct," not "let me try a slight variation."
• An explicit user instruction to "search" should be a hard precondition on any further state-mutating action.

Why not just leave it folded into #56906

#56906's fix ("verify before asserting") would help, but it doesn't address the recovery path. An agent can be good at verifying new assertions and still be bad at noticing that an earlier assertion — the one everything since is built on — was never verified. That's a separate behaviour to fix and a separate thing to test for.

Requesting human triage rather than automated deduplication.