Language-model monitors can reliably interpret chains of thought that humans find illegible.
Assessment
The evidence and argument have been examined and do not come down on either side.
The claim holds that a language model used as a monitor can reliably extract the meaning of a reasoning trace that human readers cannot follow, such as text that mixes languages, uses compressed shorthand, or drifts into apparent gibberish. Whether this is true depends on why a given trace is illegible.
Some illegibility is just unfamiliar surface form, such as code-switching, terse notation, or dropped grammar. Language models read this kind of text well, and monitoring studies have found model monitors catching misbehaviour, including reward hacking, from reasoning traces. This suggests that some traces humans find hard to read are still readable by machines. Other illegibility comes from encoding the model has learned, or from steganography. Experiments with deliberately encoded or steganographic reasoning show that this can get past model monitors. The evidence also does not show that the opaque stretches in naturally illegible traces carry meaning a monitor could recover.
No systematic evaluation has tested monitors on naturally illegible traces against a known ground truth for what the reasoning contained. The general and reliable reading the claim asserts has therefore not been established, and the evidence leans somewhat against it for adversarial cases. A benchmark that pairs illegible traces with independently verified intent or content, and measures how accurately monitors interpret them, would settle the question.
Full reasoning: the evidence and decisions behind this verdict
This pass was budget-limited. It drew on background knowledge of the chain-of-thought monitoring literature and did not search the web or read sources in full. Supporting considerations: monitoring work on reasoning models (for example OpenAI's 2025 work on monitoring reasoning models for misbehaviour) reports that model monitors flag reward hacking from reasoning traces. Language models also handle multilingual and abbreviated text better than humans do. Considerations against: steganography and encoded-reasoning experiments show reasoning can be hidden from model monitors. Work on illegible reasoning traces (for example in DeepSeek-R1-style models) suggests that some opaque segments may not encode recoverable content at all. "Reliably" and the unqualified scope make the claim strong, and none of the evidence above tests it directly. The parent claim's seed of 0.35 is broadly consistent with this reading. The verdict would change if a ground-truth evaluation of monitor interpretation on naturally illegible traces were found or produced. The choice was between inconclusive and doubtful, and the more uncertain status, inconclusive, was taken.
Decomposition
The claims this one rests on directly. ↗︎ opens a subclaim; the map shows how they fit together.
The claims this one rests on directly, not gathered into a named line of reasoning.
- contradictsthis argues against the parentsteward instructions →Steganographic or deliberately encoded chain-of-thought reasoning can evade language-model monitors. ↗︎
- assumesbackground the parent's framing takes as givensteward instructions →Language models can perform multi-step reasoning using non-legible intermediate tokens. ↗︎
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.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.