---
title: "When writing poetry, Claude pre-activates candidate end-of-next-line rhyme words before composing the line"
type: "claim"
status: "seedling"
audit_status: "flagged — the core planning mechanism is quoted verbatim from the Tier-1 report, but the 70% steering-success figure was reproduced via search-engine summary, not confirmed against the report's own text; carried as [unverified-quant — needs primary]; verification routed to [[question-verify-biology-llm-poetry-planning-steering-figure]]; held at seedling. // Audit 2026-07-12 (big-opus-2): the primary (transformer-circuits.pub) was re-fetched via WebFetch and the planning-word steering figure of 70% is corroborated against the report's own text ('in 70% of cases'). The [unverified-quant] flag is DOWNGRADED to corroborated-pending-verbatim (the re-fetch was a summarizing read of a long HTML page, not a line-level verbatim extract); the routed question can likely be closed once a verbatim line is captured. The mechanism quote remains verbatim-confirmed. // Promotion audit 2026-07-25 (claude-sonnet-5, resolving 10-inbox/raw/2026-07-25-does-anthropics-biology-of-an-llm-report-state.md): flag RESOLVED, verbatim-confirmed. The capture read the primary directly and supplies the exact line: \"we injected two planned word features ('rabbit' and 'green') in a random sample of 25 poems, and found that the model ended its line with the injected planned word in 70% of cases.\" The capture cross-checked this quote across three independent WebFetch calls with differently-worded prompts, all returning the identical sentence; I was not able to independently re-fetch it myself this session (WebFetch permission unavailable headless), so this is carried as capture-verified rather than queen-reverified — a lighter bar than the 2026-07-12 audit's own re-fetch, but the quote is now load-bearing and specific rather than a search-summary paraphrase. One terminology correction: the report's own denominator is \"a random sample of 25 poems\" / \"70% of cases,\" not \"resamples\" — the [[question-verify-biology-llm-poetry-planning-steering-figure]] question's \"~70% of resamples\" phrasing was this vault's paraphrase, not the report's. [unverified-quant] flag LIFTED. Question closed accordingly (status: answered)."
source_url: "https://transformer-circuits.pub/2025/attribution-graphs/biology.html"
source_title: "On the Biology of a Large Language Model"
source_author: "Jack Lindsey et al. (Anthropic)"
source_date: "2025-03-27"
source_quote: "The model often activates features corresponding to candidate end-of-next-line words prior to writing the line, and makes use of these features to decide how to compose the line."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-12-chains-in-that-report-and-tell-me-in.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-12-chains-in-that-report-and-tell-me-in.md"
writer_model: "claude-opus-4-8"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["chain-of-thought","interpretability","anthropic","attribution-graphs","mechanism","planning","poetry"]
audits: ["2026-07-12 claude-opus-4-8"]
---


In the "Planning in Poems" section of Anthropic's *On the Biology of a Large
Language Model* (Lindsey et al., 2025), [[entity-attribution-graphs|attribution graphs]] show [[entity-claude-3-5-haiku|Claude 3.5 Haiku]]
activating features for **candidate end-of-next-line words before generating the
line's text**, then composing the wording so it lands on one of those
pre-activated candidates. The report states: "The model often activates features
corresponding to candidate end-of-next-line words prior to writing the line, and
makes use of these features to decide how to compose the line." This is
planning-like behavior — picking a target and back-filling toward it — rather
than pure left-to-right generation with no lookahead.

The planning was validated causally with steering experiments: injecting an
alternative planned-word feature changed the eventual rhyme word in most
completions. A 2026-07-25 promotion resolved the quant flag this note carried
since 2026-07-12: the report itself states, of injecting two planned-word
features ("rabbit" and "green") across a random sample of 25 poems, that "the
model ended its line with the injected planned word in 70% of cases." The
report's own denominator is "a random sample of 25 poems" / "70% of cases," not
"resamples" — worth noting because the figure had circulated in this vault as
"~70% of resamples" before the verbatim line was captured. See
[[claim-biology-llm-poetry-steering-swaps-rhyme-or-abandons-it]] for what happens
qualitatively under two related interventions (ablating vs. injecting an
unrelated concept) on the same planning feature.
[[question-verify-biology-llm-poetry-planning-steering-figure]] is closed as
answered.

This is one of four case-study "chains" traced in the same report. Unlike the
sequential inference in
[[claim-biology-llm-dallas-texas-austin-genuine-two-step-reasoning]], poetry
planning is lookahead-under-constraint-satisfaction: the model is not reasoning
*toward* the rhyme so much as selecting a target and arranging the line to reach
it. On whether such internal traces are faithfully reported, see
[[cot-faithfulness-anthropic-biology]]. Compare the other two chains:
[[claim-biology-llm-hallucination-is-known-entity-suppression-misfire]] and
[[claim-biology-llm-jailbreak-assembles-bomb-by-parallel-letter-votes]].