Mission

Objective

Compare full-history, recent-window, artifact-first, selective retrieval, structured-summary, and compiled-context policies on the same audited suite.

Why this matters

The repository does not know which context policy maximizes verified outcomes per total cost or how context loss and contamination interact with task horizon.

Scope

Included: randomized paired runs, stale/conflicting-state injection, reliability, latency, retries, context volume, recovery, and cost. Excluded: changing models or harnesses during the comparison.

Existing context

research/frontier/records/RFR-005.md and Roadmap Phase 3.

Initial hypotheses

Selective or compiled context will preserve reliability at lower total cost than full history. Context omissions or contamination may reverse that result.

Required evidence

Frozen suite/configuration, randomization, run records, policy payloads, verified outcomes, recovery traces, and cost data.

Constraints

Hold model, harness, task version, and grading fixed. Report task-policy interactions rather than only aggregate effects.

Execution instructions

Predeclare policies and paired-run analysis before execution.

Deliverables

Context-policy decision table, cost/reliability frontier, and failure analysis.

Success criteria

Repeatable effect estimates sufficient to choose or reject a default context policy.

Stop conditions

Stop if task drift, grader drift, or incomplete telemetry breaks causal comparability.

Handoff requirements

Record the selected policy, exceptions, and revalidation triggers.