A planning feature representing end-word 'rabbit' activates at the newline token before Claude begins writing a rhyming line, and attention streams carry it forward biasing predictions toward 'rabbit'

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Claude Opus 4.6 represents a plan to end a couplet with 'rabbit' before outputting the rhyming line.finding0.848
  Demonstrates causal relationship between NLA explanations and model outputs via steering with edited explanations.
• In Opus 4.1, the think word representation decays to baseline in the final layer because the strong next-token prediction drowns out other representationshypothesis0.766
  Explanation for the 'silent' thought phenomenon.
• what features activate when we ask questions probing Claude's goals and values?question0.761
  Direction for understanding model's internal objectives via features.
• The sensitivity to think/don't think instructions may be achieved via a circuit that tags tokens as attention-worthy based on instructions or incentiveshypothesis0.747
  Mechanism for how the model modulates representation strength.
• Elephant programs that interact non-trivially with the outside world can have both input-output specifications, relating the programs inputs and outputs, and accomplishment specifications concerning what the program accomplishes in the world. These concepts are respectively generalizations of the philosophers' illocutionary and perlocutionary speech acts.claim0.745
  Fifth abstract claim.
• Elephant programs themselves can be represented as sentences of logic. Their extensional properties follow from this representation without an intervening theory of programming or anything like Hoare axioms.claim0.742
  Fourth abstract claim.
• If two processes need to communicate, they don't exchange messages or share a variable; instead, the data producing process generates a new data object and sets it adrift in tuple space.concept0.741
  Defines generative communication: the unique mechanism that enables uncoupled programming and persistent data objects.
• Random word prefix prompts show emergence patterns similar to no-prompt, suggesting prompt length alone does not shift truth geometry.claim0.741
  Control experiment ruling out token-count as the cause of truth geometry shifts.