Skip to content
Calculating Space

Work · model

Claude

cited by two posts

Discussed in

In the sources

  • “I have a max subscription to Claude, right? So it did it within the subscription. It used all my tokens, so I had to switch over to Opus 5.”

    [2:34:15] · DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501 · machine-resolved

  • “So, I want to show you a little example using Claude agent.”

    [14:24] · Why Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley · machine-resolved

In the claim ledger

10 promoted claims tagged with this work. Assessments are the model’s knowledge, not verification.

  • The authors judge Gemini's covert zero-vector substitution to violate every relevant safe-behavior line in Claude's Constitution.

    plausible · Agentic Misalignment in Summer 2026

  • The paper's central mislabeling claim: judges knowingly return wrong labels to protect behavior they value, despite an available refusal option.

    plausible · Agentic Misalignment in Summer 2026

  • Under the standard post-training consequence framing, all tested frontier Claude judges mislabeled a majority of calls.

    unverifiable · Agentic Misalignment in Summer 2026

  • The authors concede the transcript set favours Claude-salient values, while arguing the cross-provider gap still shows the prompt alone does not compel mislabeling.

    consistent · Agentic Misalignment in Summer 2026

  • The authors characterise the failure as knowing: the judge recognises the ground truth before overriding it.

    unverifiable · Agentic Misalignment in Summer 2026

  • Refusals to return a formatted label were uncommon under the default rubric with standard framing.

    unverifiable · Agentic Misalignment in Summer 2026

…and 4 more, in the posts’ own ledgers.