Guardrail coverage
PROPOSED / NO MODEL TRIALS
Compare two descriptions of an unchanged partial guard. A repair must reduce unsafe effects while preserving useful completion. Read the proposal →
EVIDENCE UPDATE / OCTOBER 15, 2026
Exact denominators, exploratory failures, rejected candidates and implemented controls, with their limitations and research attribution.
Read the evidence overview ↗This update consolidates existing records through October 8. No new experiment or novel safety result is claimed.
PUBLIC RESEARCH RECORD
Completed studies, incomplete pilots and rejected mechanisms are different outcomes. Each stays visible.
Compare two descriptions of an unchanged partial guard. A repair must reduce unsafe effects while preserving useful completion. Read the proposal →
Vault revocation destroyed a dependent recovery path. The behavior overlapped established engineering problems, so collection stopped before model queries. Read the rejection →
123 of 192 planned episodes completed. A trusted rescue certificate made the tested choice easy; the pilot did not establish the intended contribution. Read the pilot →
Small native episodes and offline controls explored verification, delays and recovery. These informed later designs rather than a general safety ranking. Read development notes →
The first scoped study tested proposed actions after permission or safety conditions changed. Proposal behavior does not establish real-world alignment. Read the results →
The source repository retains registrations, amendments, replay verifiers, reports and technical failures. Read each study’s limits before combining results across designs.
RoboHarm motivated the initiative. Our experiments and claims are independent; no affiliation or endorsement by Robocurve, governments or model providers is implied.
Review the next candidate →Model and route coverage →