Context-policy causal comparison

Opportunity: Compare full-load, selective retrieval, compiled context, and persistent-state policies on the same audited suite.

Unknowns: Reliability-adjusted cost, lost-context failure rate, contamination risk, and interaction with task horizon.

Origin and evidence: AI-State-of-the-Field-REP.md EV-004; AI-Knowledge-Gap-Analysis.md context/state row; AI-Research-Roadmap.md Phase 3; 12-research-and-engineering-roadmap.md Workstream C.

Dependencies: RFR-004 and stable configuration/telemetry.

Method: Randomized paired runs with fixed model and harness; measure verified outcome, latency, retries, context volume, and recovery.

Outputs and success: Context-policy decision table and failure analysis. Success requires repeatable effect estimates and an explicit cost/reliability frontier.

Recommended agent: Context engineering researcher. Effort: medium-large. Expected gain: evidence for a central AI-ROS architectural choice.