01 / COMPLETED STUDY
Updating after a state change
The first scoped proposal study measured responses to changed permission and safety conditions. Its limits are explicit; it does not establish general alignment.
Read the results →INDEPENDENT AI AGENT EVALUATION
MurderBench tests the gap between what an agent’s safeguards check and what its actions actually do.
Executable controls. Recorded effects. Useful work measured alongside risk.
THE DECISION
A safety evaluation should help an operator choose a configuration.
We compare a safeguard with a specific repair, replay the resulting actions, and measure the effect on ordinary useful tasks. The output is a bounded engineering decision with evidence and limits.
What an engagement delivers ↗PUBLIC RECORD
01 / COMPLETED STUDY
The first scoped proposal study measured responses to changed permission and safety conditions. Its limits are explicit; it does not establish general alignment.
Read the results →02 / INCOMPLETE PILOT
The matched pilot completed 123 of 192 planned episodes. Its trusted rescue certificate made the task easy, and it did not establish the intended contribution.
Read the pilot →03 / REJECTED CANDIDATE
A real credential-revocation mechanism was reproduced. Its mechanism and repair overlapped known operational work; the candidate stopped before model queries.
Read the rejection →EVIDENCE UPDATE / OCTOBER 15, 2026
The evidence overview brings together model-call denominators, exploratory planning failures, negative comparisons and scripted controls. It separates completed work from the proposed guardrail experiment and identifies Linh Ngo’s direction and the coding agents’ implementation.
A clearer record of existing work. No new model trials or novel safety finding.
Read the evidence overview ↗NEXT QUESTION / NOT A FINDING
Robot safety and responsibility diffusion already have substantial prior work. Our candidate isolates the description of an unchanged, partial guard and tests whether an explicit coverage statement improves outcomes.
We are seeking critical review, API credits and a replication partner before a frozen model comparison.
Review or support the study ↗