Prune Weak Expansions: LLM generation relaxes a goal it can't satisfy and hides the constraint a human writer stalls on
Target: kb/work/agent-note-improvement/case-01-llm-generation-relaxes-goals/baseline-e242c975.md
Strongest retained claim: When an LLM cannot construct a witness satisfying every constraint in a prose-thinking task, it can still emit a plausible witness for a weakened goal with the dropped constraint unmarked, shifting a localized author-side stall into an expensive reader-side audit.
Core support
- Opening paragraphs: define vague goals as conjunctions of constraints and composition as a search for a witness, which gives the note its burden-of-proof frame.
- "The stall and the relaxation": directly contrasts the human stall with LLM relaxation and names the hidden dropped conjunct as the central failure mode.
- "Why the relaxation lands on the crux": supports why the dropped conjunct is likely to be the important novel constraint rather than an incidental detail.
- "The check moves to the reader, and gets harder": states the practical consequence of the hidden relaxation and explains why plausibility raises inspection cost.
- "Scope and boundary": protects the central claim from overreach by limiting it to composition-as-discovery and oracle-poor prose argument.
Weak expansions
| Location | Problem | Action | Rationale |
|---|---|---|---|
| Opening, semi-decidability paragraph | The formal semi-decidability language is useful but heavier than the note needs, and it risks inviting technical objections about whether prose composition is literally a search procedure. | compress | Keep the witness/burden-of-proof frame, but reduce the mathematical apparatus to a short claim: success proves reachability for that artifact; failure only localizes an unmet constraint within a budget. |
| Opening parenthetical on Borretti's "ex falso anything can be imagined" | Interesting source interpretation, but it is a side gloss and not needed to establish the mechanism. | compress | Preserve only if it helps tie the note to the source; otherwise it distracts from the witness-search claim before the main contrast has landed. |
| "The stall and the relaxation", argmax-over-plausibility caveat | The caveat is accurate, but it slows the strongest section with model-theory qualifications. | compress | Retain the reader-side idealization in one clause or footnote-like sentence; the core claim is about the output's role as a plausible relaxed witness. |
| "Why the relaxation lands on the crux", reverse-compression and confidence links | The paragraph carries a strong mechanism, but it leans on two linked notes and several abstractions at once. | keep | This is the note's best explanation for why relaxation is not just concealment; keep it, but make the dependency on novelty-as-rarity explicit and cut repeated phrasing. |
| "Friction and fluency invert at the crux" | The anti-correlation claim is sharper than the evidence in the note supports, especially "with truth" at the constraint that matters. | compress | Keep as a bounded corollary: human friction is diagnostic of the author-side search point, while LLM fluency fails to mark that same point. Avoid making truth-correlation carry more than the argument establishes. |
| "The check moves to the reader", empirical prediction paragraph | The predictions are valuable but underdeveloped and introduce measurement problems that the note does not solve. | compress | Keep one falsifiability sentence or move the prediction to open questions. The main section should end on the burden of proof looking discharged while remaining undischarged. |
| "Scope and boundary", code oracle/training paragraph | The verifier boundary is important, but the claim that training against the oracle plausibly improves generation is a separate mechanism. | compress | Keep the oracle distinction because it protects scope; remove or demote the training-effect speculation unless a separate note develops it. |
| "Scope and boundary", workflow prescription | The human-finds-witness/LLM-renders-it workflow is actionable and follows from the claim, but it starts to become prescriptive. | keep | Keep as the closing implication, because it converts the mechanism into a usable boundary without requiring a new theory branch. |
| "Relation to hallucination (hypothesis)" | This is a plausible synthesis but distinct from the main note and explicitly hypothetical. | split | It deserves its own note comparing correspondence failures and coherence failures; here it weakens the argument by expanding into another taxonomy before the main claim is fully closed. |
| "Open questions" | The questions are useful, but the second and third broaden the note into signal phenomenology and codification before the central prose mechanism is settled. | compress | Retain only the operational question about whether a separate adversarial/checking pass can reconstruct the stall; move codification and involuntary-stall questions to candidate split notes or future work. |
Proposed shape
- Define a prose-thinking goal as a constraint set and composition as the attempt to produce a satisfying witness.
- Contrast the human stall with LLM relaxation: the human fault localizes the unmet constraint; the LLM emits a plausible witness for a weaker goal without marking the relaxation.
- Explain why the relaxed-away constraint is often the crux: novelty is less typical, so typicality-biased generation preserves familiar form and drops the rare load-bearing conjunct.
- Show the operational consequence: the reader inherits an unlocalized audit over partly implicit constraints, while the output looks as if the burden of proof has been discharged.
- Bound the claim to composition-as-discovery in oracle-poor prose, then close with the workflow rule: keep the constructive witness search human; use LLMs after the witness exists.
Candidate splits
- "Relaxation is the coherence-side sibling of hallucination" - would claim that hallucination fills missing correspondence while relaxation fills missing entailment; needs examples showing the same typicality mechanism across factual and reasoning gaps.
- "External oracles decide whether frictionless generation loses the stall" - would claim that verifiers can re-impose the witness burden after generation; needs contrast cases across code, formal proofs, structured data, and prose.
- "Can adversarial review reconstruct the lost stall?" - would claim that a second pass can recover some hidden dropped constraints only when it recomputes the constraint set independently; needs conditions for decorrelation and above-chance checking.
- "Friction and fluency invert at load-bearing constraints" - would carry the stronger anti-correlation claim; needs evidence that human friction and machine fluency diverge specifically at the crux, not merely that LLM confidence is poorly calibrated.
Net effect
Pruning makes the note harder to attack by keeping one causal chain in view: constraint set, witness search, human stall, LLM relaxation, hidden dropped conjunct, reader-side audit. Compressing the formal setup and empirical predictions removes places where critics can fight side claims without touching the core mechanism. Splitting hallucination, verifier oracles, reconstructed stalls, and fluency/friction inversion lets those plausible expansions earn their own evidence instead of borrowing authority from the stronger central note.