Trusted editing and untrusted monitoring protocols substantially outperform simple baselines at preventing backdoors from subversive untrusted models.
Not yet assessed. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.
Decomposition
This claim has not been assessed yet. Attention goes where its expected value is highest and someone funds it; once this claim's assessment is funded and runs, it may well decompose into subclaims.
Provenance
Where this claim has been said, linked to its canonical form.
These protocols improve substantially on simple baselines.
Refers to trusted editing (GPT-4 writes code, GPT-3.5 rates suspiciousness and edits suspicious code) and untrusted monitoring (another GPT-4 instance checks for backdoors, with anti-collusion techniques), tested against red-team subversion strategies in a programming backdoor setting.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Created by extractor · Nov 2, 2026. Every judgment on this page is accompanied by a reasoning trace.