CCAR-P Judgment Traps

Symptom over root cause

Tell: an option adds a control to contain a problem. Fix: prefer the option that removes or scopes the problem (least privilege beats a well-guarded over-permission).

Buried constraint

Tell: three sentences of scenario, then a latency/compliance/cost line at the end. Fix: underline that constraint first — it flips the answer.

Escalate on confidence

Tell: 'route to a human when model confidence < X'. Fix: escalate on a rigid policy or calibrated threshold; 0.92 ≠ 92% accurate.

Fail open

Tell: a guardrail that lets the request through when the check errors. Fix: fail closed — an open control is worse than none.

One filter to rule them all

Tell: a single output filter 'covers' safety. Fix: screen input, screen output, and authorize tool calls — three points, three blind spots.

Subagent amnesia

Tell: 'analyze the findings' handed to a subagent that never got them. Fix: subagents don't inherit coordinator context; pass it, or keep it in one context.

Bigger context = better

Tell: 'just use the long-context window'. Fix: stuffing docs + big tool schemas causes context saturation and tail-content loss. Retrieve what's relevant.

Schema-valid = correct

Tell: 'the JSON validates, ship it'. Fix: schema-valid ≠ semantically correct; validate against the source.

auto forces the tool

Tell: 'tool_choice: auto guarantees the tool runs'. Fix: auto permits; only any or a named tool forces.

Just fine-tune it

Tell: 'fine-tune Claude on our data'. Fix: customer fine-tuning isn't offered — steer via prompting, context, tools, and retrieval.

Blind retry

Tell: 'retry on timeout'. Fix: a timeout after a non-idempotent write must not be blind-retried.

Reactive escalation

Tell: 'escalate when something looks wrong'. Fix: escalation triggers are pre-defined — a schedule or an agreed threshold, set in advance.