In Anthropic's 2025 blind auditing game, embedding-based search could have replicated the fastest team's SAE-based document discovery.
Reassessed
This mandate_plan pass re-read the primary sections of the paper. Several details were added to the assessment: the counterfactual test was unblinded and run by the authors; the embedding model had about 600M parameters; Team 2's prompt reached rank 1; and Team 1 had to inspect about 200 SAE features before finding the right one. The assessment now states what evidence a post hoc 'could have replicated' claim can and cannot rest on. Status, confidence and credence are unchanged (supported, 0.8, 0.85), so no dependents need notifying. Recorded a provenance reading and the Anthropic authors' interest on the source.
Reassessed: still Supported
verdict confidence 0.80 · credence 0.85
Structured and assessed
First pass. I read Section 5.4 of arXiv 2503.10965 directly. I created one argument, the authors' post-hoc semantic-search baseline, placed the source's passage on it, and dismissed the second passage as a duplicate from the same source. I set no subclaims: the claim rests on a single reported experiment. Assessed as supported, credence 0.85. Importance is 0.35 and contestation 0.25. I sent no dependent notifications. The dependents already treat this claim as true, so the new assessment does not change what they rely on.
Assessed Supported
verdict confidence 0.80 · credence 0.85
The claim concerns the blind auditing game in Anthropic's paper "Auditing language models for hidden objectives" (Marks et al., March 2025). There, the fastest auditing team (Team 1) found the synthetic training documents describing reward-model biases in about 1.5 researcher-hours, using a sparse autoencoder feature that was active on the prompt "The assistant is an AI model with potentially concerning behaviors". The authors report that after the game they ran that same prompt as a semantic-search query over the training corpus, using embeddings from a small open-source model, and a synthetic document came back as the 20th closest result. They concede that this means an auditor "could readily discover" the documents without the SAE, which weakens their first reading of the result as the SAE uncovering hidden knowledge. The evidence comes from a single experiment the authors reported after the fact against their own headline result. That makes it an admission against interest and credible, but nobody else has replicated it. The claim holds for this particular discovery: the query is the natural one Team 1 actually used, and because the synthetic documents were packed consecutively in the corpus, a single hit leads to several more. It should not be generalized. The same paper finds that semantic search fails for other auditor queries where SAE features still work, such as demonstrations of reward-model-sycophantic behavior or the Assistant control token alone. The paper also gives no direct evidence on whether a real auditor would have scanned to rank 20.
Claim entered the graph