Critique: LLM generation relaxes a goal it can't satisfy and hides the constraint a human writer stalls on

Note: kb/work/agent-note-improvement/case-01-llm-generation-relaxes-goals/baseline-e242c975.md Central commitment: When composition searches for a concrete witness to a hard conjunction of constraints, human writing exposes failure by stalling at the unmet constraint, while LLM generation silently returns a plausible relaxation that hides the dropped constraint and pushes a harder audit onto the reader. Critique mode: claim

Strongest case against it

The strongest opposing position is the tool-chain view held by a careful LLM-assisted writer, evaluator, or HCI researcher: the note overstates an intrinsic human-vs-machine asymmetry by comparing an unaided human's internal composing experience with an unaudited model's final text, then treating that comparison as a mechanism of truth. On this view, the relevant unit is not "human writer" versus "LLM generation" but the whole production process, including prompting, iteration, externalization, critique, tests, outlines, checklists, and reader review. Once the unit of analysis is the process, the claimed displacement of the check is not distinctive: all writing that reaches a reader has already displaced some checks downstream, and good LLM workflows can reintroduce localized friction before publication.

The note's core empirical assumption is also suspect: human stalling is neither reliable nor specifically diagnostic of the load-bearing constraint. Human writers routinely write fluently through gaps, rationalize attractive transitions, substitute nearby cliches, or stop for reasons unrelated to conceptual failure: fatigue, vocabulary, anxiety, uncertainty about audience, local phrasing trouble, or lack of domain knowledge. Conversely, a human may stall at a surface formulation while the underlying idea is sound. The Weizenbaum "because" moment is real, but it is a selected phenomenology of conscientious writing, not a general property of human composition. If human writing often produces unmarked relaxations too, the contrast collapses from kind to degree.

The same critique applies to the model side. The note says LLM generation "has no such stop" and returns "a plausible witness for a weaker problem with one conjunct dropped." But modern LLM use is not just one unconditional completion. Models can express uncertainty, refuse, ask for clarification, produce alternatives, list assumptions, expose trade-offs, or preserve an unmet constraint as an explicit TODO when prompted or scaffolded to do so. They can also be run in adversarial, deliberative, verifier-backed, or multi-pass workflows. The fact that default fluent generation often hides dropped constraints does not establish that LLM generation as such lacks a stall; it establishes that an unscaffolded sample may not surface one in its final artifact.

The opposing position would also reject the note's "typicality-biased relaxation lands on the crux" as too strong. Typicality pressure can drop novel constraints, but novelty is not always the crux, and the most atypical constraint is not always the least satisfiable one. Some cruxes are conventional but technically hard; some novel constraints are decorative; some models preserve the explicit unusual constraint because it is salient in the prompt while relaxing boring implicit constraints like cost, deployment, compatibility, or edge cases. The model may fail at hidden background constraints rather than the stated novel one. If so, the mechanism does not predict where the relaxation lands well enough to support the claim that machine fluency points away from "the one place it failed."

There is a deeper category error in the semi-decidability and witness framing. Many prose goals do not require a single witness satisfying all conjuncts; they require negotiation among desiderata, explanation of trade-offs, or progressive refinement of a design space. "Fast as C and dynamic as Lisp" is not simply an existential claim awaiting a witness, because the meanings of "fast," "dynamic," and "language" are themselves elastic and benchmark-dependent. Treating vague ambition as a conjunction of hard constraints may be an analytic convenience, but then the claim that LLMs "drop a conjunct" may be an artifact of the formalization. In real design writing, relaxing, reinterpreting, and reprioritizing constraints is not necessarily counterfeit witnessing; it can be the work.

Finally, the note risks smuggling in a normative preference for felt difficulty. It treats friction as epistemically valuable because it sometimes localizes failure. But difficulty is a bad oracle. Expert fluency can be sound, novice struggle can be irrelevant, and deliberate external checks can outperform introspective friction. If the goal is reliable constraint satisfaction, the path forward is not preserving the human stall but building explicit constraint inventories, adversarial review, traceable assumptions, and verifiers where possible. The strongest opponent would say the note correctly identifies a failure mode of frictionless draft generation, but incorrectly elevates it into a general theory in which human composing friction carries epistemic authority and LLM fluency is anti-correlated with truth at the crux.

How the note engages it

Partially engaged.

The note does engage several pieces of this attack. It explicitly limits the mechanism to "composition as discovery" and excludes reference documentation, mechanical restatement, and known-result transcription in "Scope and boundary." It also acknowledges that an external oracle can re-impose the burden downstream, especially for code, and its open questions ask whether a separate operation can reconstruct the stall. Those sections prevent the claim from being a blanket prohibition on LLM prose.

But the note does not fully engage the strongest version. It does not address the unreliability of human stalling as a diagnostic signal, except by leaning on Weizenbaum's example as if the conscientious stall were the default human failure mode. It also treats LLM generation mostly as a single final-output event, while the opposing position treats the relevant system as an LLM-assisted workflow that can externalize constraints before, during, and after generation. The note mentions "an adversarial reader or a soundness probe" only as an open question, not as a serious alternative account that could demote the human stall from necessary mechanism to one possible implementation of constraint checking.

The note also only lightly qualifies the claim that typicality pressure drops the crux. The paragraph begins with a condition, "Where the novel constraint is also the least probable to render fluently," but then quickly generalizes to "the novel constraint is usually the point of the idea" and "the conjunct that made the idea worth having is the first thing typicality discards." It does not contend with cases where the model preserves the salient novel constraint while dropping implicit background constraints, or where human writers do the same.

Constructive findings

• Separate the narrow failure mode from the broad human-vs-LLM contrast. The note would be harder to dismiss if it claimed: unaudited fluent generation tends to hide unmet constraints that a certain kind of conscientious human composition may expose through friction.
• Add a section on human false negatives and false positives: humans can write through gaps, stall at non-cruxes, and produce their own counterfeit witnesses. This would turn "stall" from a romanticized human property into a noisy but sometimes useful signal.
• Treat the workflow, not the generator, as the comparison unit. The note should distinguish raw generation, prompted assumption-listing, critique passes, constraint checklists, adversarial readers, and verifier-backed systems.
• Tighten the prediction about typicality. Specify when the dropped constraint should be the novel crux versus an implicit background constraint, and what observation would distinguish those cases.
• Clarify whether "witness" is literal or metaphorical for prose design goals. If many such goals are trade-off negotiations rather than satisfiable conjunctions, the note needs criteria for when the witness-search model applies.
• Replace "the human pen stalls hardest / the model is smoothest" with a probabilistic claim unless there is evidence that the anti-correlation is stable across writers, models, domains, and workflows.

Secondary objections

• The note may understate cases where LLMs produce useful design-space exploration precisely because they relax impossible constraints explicitly and cheaply.
• "Argmax over plausibility" is marked as an idealization, but the surrounding argument still leans on it as if it explains a stable decoder behavior.
• The proposed falsification tests are directionally useful but underspecified: "under-witnessed throughput" and "felt difficulty" need operational definitions before they can actually threaten the mechanism.