Tool output formatters that quietly dropped search limits misled agents and their verifiers, and the audit prompt found the same bug in Hermes and OpenClaw
AspireAI community · arXiv (William Jeffries, Independent Researcher)
Agents and verifiers read the text a tool prints, not the tool's internal state. This paper documents what happens when that printed text misrepresents what the tool actually did, even when the computation was correct. The author hit it in a production RAG research system at the Pilsudski Institute in New York, where output formatters omitted search limits, result caps and identifier namespaces, and also caught an unlabeled estimate during code review before release.
Feeding agents the same tool results in two formatter versions, one with the omissions and one with all missing details stated, changed both what the agents claimed and what the verifiers accepted. Unsupported agent claims dropped substantially for three of the four failure types, and in one test verifiers went from accepting incorrect answers 90 times out of 90 to rejecting them 90 times out of 90. A control confirmed the verifiers could catch the errors once given the missing facts and asked to judge correctness.
The practical part: the audit prompt, written with prior knowledge of the author's own system, rediscovered the defects that were still live and surfaced twelve more that were uncatalogued. Pointed at OpenClaw and Hermes, it found warped formatters in both. The author opened pull requests for both, and the OpenClaw fix has been merged.
Caveat: single-author, exploratory work, and the audit prompt was written knowing the author's own system's defects.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.