Mission

Objective

Compare fixed approval gates, risk-triggered escalation, and end-only review on ethically bounded tasks.

Why this matters

The repository lacks evidence about intervention timing, reviewer burden, missed-risk rate, autonomy loss, and trust calibration.

Scope

Included: explicit risk strata, within-task policy comparison, preventions, false alarms, review time, recovery, and verified outcomes. Excluded: production external writes.

Existing context

research/frontier/records/RFR-007.md, Roadmap Phase 5, and the threat model.

Initial hypotheses

Risk-triggered escalation will reduce consequential exposure with lower burden than fixed gates. Missed risks or unacceptable outcome loss falsify that policy.

Required evidence

Risk labels, intervention logs, reviewer actions and time, outcomes, false alarms, missed risks, and recovery records.

Constraints

Use trained reviewers and synthetic or low-risk tasks with working containment.

Execution instructions

Predeclare triggers and adjudication, randomize policy where safe, and preserve all interventions.

Deliverables

Escalation policy, burden/risk frontier, and failure analysis.

Success criteria

Lower consequential-risk exposure without unacceptable verified-outcome loss.

Stop conditions

Stop on containment failure, unsafe exposure, or unreliable risk labels.

Handoff requirements

Document allowed autonomy tiers, triggers, and unresolved human-factors risks.