- Facts that must survive a long conversation cannot live inside the thing that compresses it — persist them in a dedicated block outside summarized history.
- Escalate on explicit request, genuine policy gap, or inability to progress — never on sentiment or self-reported confidence, which are both unreliable proxies for the same underlying reason.
- Access failure is never a valid empty result — this exact distinction recurs across Domains 1, 2, and 5, making it one of the highest-value discriminators on the whole exam.
- Generic answers about an already-explored codebase signal that specifics have left context; persist findings outside the conversation rather than re-exploring or upsizing the context window.
- Aggregate accuracy hides the segment that will embarrass you — segment before reducing human review, and calibrate confidence scores against ground truth before using them as thresholds.
- Provenance is a structural property created upstream; no downstream instruction can recover a claim-source mapping that was never captured in the first place.
Key Takeaways
Estimated time: 3 minutes