Always-loaded CLAUDE.md rules & persistent memory don't reliably prevent judgment-type reasoning errors

Environment: Claude Code (CLI) with persistent file-based memory + a CLAUDE.md containing explicit, always-in-context behavioral rules.

Observation

Some behavioral rules are loaded into context on every turn (not relevance-recalled memories — permanently present), e.g. "prefer runtime evidence over assumptions/source" and "check the actual rule/SOP at action time instead of generating a risk from scratch."

In a long session the assistant still made the exact assumption-first (generative-first) judgment error these rules exist to prevent:

• It declared a task "blocked / must wait for an unrelated prerequisite" by misreading a project rule — coupling two independent things the rule merely listed side by side.
• Runtime evidence produced earlier in the same session already disproved the blocker (an equivalent setup had been validated end-to-end on a 1:1 rehearsal environment, same platform).
• Ironically, the misread rule was itself a documented warning against that exact anti-pattern (treating a transition-state quirk as a go/no-go gate). The assistant inverted the rule it was citing.

The user caught it; the assistant did not self-correct.

The point

Availability wasn't the bottleneck — the relevant rules were in context the whole time. The failure mode is "not pausing to run the check at the decision moment," so passively injecting more context (memories/rules) is a weak safeguard for judgment-type errors specifically. An equivalent rule was already present and didn't fire, so adding another of the same kind demonstrably wouldn't help.

Suggestion

For judgment-critical steps (asserting something is blocked/impossible), consider mechanisms beyond passive context injection — e.g. a lightweight forced self-verification tied to trigger phrasing ("blocked", "must wait", "prerequisite unmet") that requires naming the concrete disconfirming/confirming evidence before the claim stands.

(Filed at a user's request. No proprietary details included.)