Weight: 15%. About 9 of the 60 questions.
| Topic | What it tests |
|---|---|
| 5.1 Conversation context | Case-facts blocks, trimming tool output, "lost in the middle", statelessness |
| 5.2 Escalation & ambiguity | When to escalate, when to clarify, when to proceed |
| 5.3 Error propagation | Structured errors, partial results, coverage notes, avoiding retry storms |
| 5.4 Large codebase context | Scratchpad files, subagent delegation, /compact, /clear, /context |
| 5.5 Human review & calibration | Stratified sampling, calibrated thresholds, per-segment accuracy |
| 5.6 Provenance & uncertainty | Claim-to-source mapping, conflicts, dates, confidence |
Mental model
- Context is a budget. Keep what's critical exact and persistent, and keep everything else compact.
- Failures and uncertainty should be visible and structured. Don't hide them and don't let them crash the system.
- Aggregate numbers hide problems. Measure per segment.
Updated 2026-10-28