# opus[1m] ## Skills Used - anti-metacommentary - depth-audit - source-check ## Retrieval Notes - Session id: `649d76c8-433f-4789-97ee-cd6a62c92640` - Last activity: `2026-03-27T13:47:10.207Z` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User opus[1m] ### Assistant Let me read the target note first, then run all four audits. ### Assistant Good, I have the full note. Now let me read the knowledge file required by the hook, then begin the four audits. Let me also check for voice-flagging comments. ### Assistant No `%%not how i write%%` or similar voice-flagging comments found in the note. Skipping voice-fix. Now running the remaining three audits. Let me start with source-check — I need to locate the source texts for all referenced authors. ### Assistant ### Step 1: Author references in the note Here's every external author reference I find: 1. **Floridi et al. (2024)** — lines 6, 8, 10, 12, 14, 20, 24 - LLMs do not reason abductively (line 6) - Block quote about generating plausible continuations (line 8, p. 9) - "Zeroth-order abduction" — their term (line 10) - "compelling illusion" (p. 2), "surface-level abductive appearances" (p. 19) (line 10) - Examples concern everyday explanation, not philosophy (line 12) - "coherence varies with model quality" (p. 17) (line 14) - Describe next-token prediction as producing most probable continuation (line 20) - Whether process matters if output is same; quoted on epistemological vs content standpoint (p. 12) (line 24) 2. **Lipton (2004)** — lines 10, 20, 24 - Background beliefs generate limited list, selection made from among them (p. 149) (line 10) - Likeliest vs loveliest distinction (p. 59) (line 20) - Squash/Bayesianism quote (p. 108) (line 24) 3. **Williamson (2024)** — lines 12, 16 - Philosophy proceeds by abduction from the armchair (p. 358) (line 12) - Theories combining simplicity with strength (pp. 354, 368-69) (line 16) 4. **Bengson et al. (2022)** — line 16 - Cohere internally before claiming broader theoretical virtue (pp. 108-09) ### Step 2: Locate and extract sources Let me find all four source texts. ### Assistant Good — all four sources have existing extractions. Let me read them all. ### Assistant Files are too large — need to search for specific passages. Let me search for each key claim. ### Assistant Now let me search the Lipton and Williamson extractions for the specific referenced passages. ### Assistant Let me check the Lipton page 59 reference and the Floridi block quote more carefully. ### Assistant I now have all four source texts verified against the manuscript. Here's the full source-check report. --- ## Source Check: "2. Likeliness, Loveliness, LLMs" ### Author References Audited 4 authors, 12+ distinct characterisations. ### Verification Results #### Floridi et al. (2024) — 8 characterisations **1. Line 6: "argue that LLMs do not reason abductively — that what they produce, however explanatory it looks, results from pattern-matching over training data rather than from comparing hypotheses and selecting the best"** Source: > The result is a compelling illusion of genuine and structured inferential reasoning. (p. 2) > LLMs seem to perform a kind of zeroth-order abduction... they generate a plausible continuation (a hypothesis or explanation) based purely on learned associations. (p. 9) Verdict: ACCURATE --- **2. Line 8: Block quote ("Given a prompt, they generate a plausible continuation...")** Source (p. 9): Matches word for word. The ellipsis omits "which OpenAI researchers term 'next-token prediction as the core objective'" — a clean omission that doesn't alter meaning. Verdict: ACCURATE --- **3. Line 10: "'Zeroth-order abduction' is Floridi et al.'s name for this"** Source: "LLMs seem to perform a kind of zeroth-order abduction (Kojima et al. 2022)." The Kojima citation references their "Large Language Models are Zero-Shot Reasoners" paper (about chain-of-thought prompting), not the term "zeroth-order abduction" itself. The term appears to be Floridi et al.'s coinage. Verdict: ACCURATE --- **4. Line 10: "compelling illusion" (p. 2) and "surface-level abductive appearances" (p. 19)** Both are direct quotes confirmed in the source. Verdict: ACCURATE --- **5. Line 12: "Floridi et al.'s examples concern everyday explanation — a car that will not start, a medical diagnosis — not philosophical reasoning"** Source confirms both examples: the car/cold morning example (p. 10) and medical diagnostic scenarios (pp. 10-11, citing Pareschi's study). One minor detail: the manuscript paraphrases the car example as producing "a dead battery, cold weather reducing efficiency" — the source says "a weak battery, as cold weather reduces battery efficiency." "Dead" vs "weak" is a small inaccuracy in the illustration but doesn't affect the philosophical point. Verdict: ACCURATE --- **6. Line 14: "'coherence varies with model quality'" (p. 17)** Source: "The fact that coherence varies with model quality highlights our core premise." Verdict: ACCURATE — direct quote confirmed. --- **7. Line 24: Floridi ask whether process matters if AI "can generate the same explanatory hypothesis a human would"** Source: "if an AI can generate the same explanatory hypothesis a human would, does it matter that the process was different?" Verdict: ACCURATE — direct quote confirmed. --- **8. Line 24: "'from an epistemological standpoint, perhaps yes', but 'regarding the content of the hypothesis and our interpretation of it, maybe not' (2024, p. 12)"** Source: "From an epistemological standpoint, perhaps yes — justification is significant — but regarding the content of the hypothesis and our interpretation of it, maybe not." The manuscript's quotes omit "justification is significant" from the first clause. This doesn't change the meaning. Verdict: ACCURATE --- #### Lipton (2004) — 3 characterisations **9. Line 10: "On Lipton's account of abductive reasoning, background beliefs generate a limited list of plausible hypotheses, and a selection is then made from among them (2004, p. 149); most possibilities are never entertained"** Source (pp. 149-150): > We never begin with a full menu of all possible causal differences, because this menu would be too large to generate or handle. Yet the class of differences we do consider is not generated randomly. We must use some sort of short list mechanism, where our background beliefs help us to generate a very limited list of plausible hypotheses, from which we then choose. [...] the actual short list mechanism we employ has two stages, one where a limited list of live candidates is generated, the other where a selection is made from this list. Verdict: ACCURATE — faithfully captures the two-stage model. --- **10. Line 20: "Lipton distinguishes two things that 'likeliness' might mean: the likeliest explanation is the one most warranted by the evidence, while the loveliest is the one that would, if true, provide the most understanding (2004, p. 59)"** Source (p. 59): > It is important to distinguish two senses in which something may be **the best** of competing potential explanations. We may characterize it as the explanation that is most warranted: the 'likeliest' or most probable explanation. On the other hand, we may characterize the best explanation as the one which would, if correct, be the most explanatory or provide the most understanding: the 'loveliest' explanation. The criteria of likeliness and loveliness may well pick out the same explanation in a particular competition, but they are clearly **different sorts of standard**. Likeliness speaks of truth; loveliness of potential understanding. Verdict: INACCURATE — Failure mode 8 (fabricated specificity) The manuscript says Lipton distinguishes "two things that 'likeliness' might mean." Lipton distinguishes two senses of "the best" — not two senses of "likeliness." Lipton explicitly separates likeliness from loveliness as "clearly different sorts of standard." Presenting loveliness as a possible meaning of "likeliness" conflates the very terms Lipton works to distinguish. The correct framing: Lipton distinguishes two standards for evaluating competing explanations — the likeliest (most warranted) and the loveliest (most explanatory). The rest of the sentence accurately describes each standard. The problem is the introductory framing. --- **11. Line 24: Squash/Bayesianism quote (p. 108)** Source: > arguing that Inference to the Best Explanation is wrong because Bayesianism is right is like arguing that thinking about technique cannot help my squash game because the motion of the ball is governed by the laws of mechanics. Verdict: ACCURATE — word-for-word match. --- #### Williamson (2024) — 2 characterisations **12. Line 12: "philosophy proceeds by abduction from the armchair, as Williamson (2024, p. 358) argues it does"** Source — Williamson's thesis is stated on p. 351: > I will suggest that philosophy sometimes already uses an abductive methodology and ought to use it more in the future. I will also argue that in using an abductive methodology philosophy can still remain a primarily "armchair" discipline. Page 358 discusses mathematics as "a useful foil to the natural sciences in exhibiting the wide range of methodologies consistent with a significant role for abduction" — related but not the direct statement of the thesis. Verdict: ACCURATE in characterisation. The thesis is Williamson's. However, note that the page reference p. 358 points to the mathematics-as-precedent discussion, not the direct statement of the thesis (which is on p. 351). The characterisation also compresses Williamson's more cautious "sometimes already uses... and ought to use it more" into the more direct "philosophy proceeds by abduction," which slightly strengthens the claim. Williamson says philosophy *sometimes* uses abduction and *should* use it more; the manuscript presents it as Williamson's view that philosophy *does* proceed this way. Minor overextension (FM 4). --- **13. Line 16: "theories that combine simplicity with strength (Williamson 2024, pp. 354, 368-69)"** Source (p. 354): > It should be elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated. It should be informative and general. In brief, it should combine simplicity with strength. Pages 368-69 extend this with the overfitting discussion: tolerance for complicated analyses in philosophy parallels overfitting in science. Verdict: ACCURATE — direct quote, correct page references. --- #### Bengson et al. (2022) — 1 characterisation **14. Line 16: theories should "cohere internally before claiming broader theoretical virtue (Bengson et al. 2022, pp. 108-09)"** Source (pp. 108-09), the Tri-Level Method: > theorists ought to articulate a set of theses about the domain that (i) accommodate and explain the data, (ii) are themselves substantiated and integrated, and (iii) possess specific theoretical virtues. [...] Each level takes priority over its successors. And: "coherence is central to integration" (p. 108). Verdict: ACCURATE but compressed (minor FM 5, underextension). The manuscript's "cohere internally" maps to level (ii) "substantiated and integrated," but drops "substantiation" (epistemic grounding) — Bengson's level (ii) requires both substantiation AND integration, not integration alone. The manuscript also drops level (i) entirely (accommodate and explain the data). The priority ordering ("before claiming broader theoretical virtue") is correct. This is fair as a compressed citation in a sentence citing multiple authors, but readers checking the source would find Bengson says more than internal coherence. --- ### Blurred Ownership Check No significant blurred-ownership issues found. The section consistently marks where it transitions from cited authors' views to Nick's own arguments: - "Floridi et al.'s diagnosis collapses both stages into a single step" (line 10) — clearly Nick's reading of the relationship between Floridi and Lipton - "the implication that matters here" (line 14) — signals Nick's argument - "grants what matters for our purposes" (line 24) — signals Nick drawing on Floridi's concession - "Lipton makes the same point from the other direction" (line 24) — Nick deploying Lipton's analogy in a new context, clearly framed as parallel rather than attribution --- ### Summary | Author | Location | Verdict | Issue | |--------|----------|---------|-------| | Floridi et al. | Lines 6, 8, 10, 12, 14, 20, 24 | All ACCURATE | Minor: "dead battery" should be "weak battery" in paraphrase | | Lipton | Line 10 (shortlist) | ACCURATE | — | | Lipton | Line 20 (likeliest/loveliest) | INACCURATE | FM 8: "two things likeliness might mean" → should be "two senses of best" | | Lipton | Line 24 (squash) | ACCURATE | — | | Williamson | Line 12 (armchair) | ACCURATE | Page ref p. 358 → thesis is on p. 351; slight overextension of "sometimes" → "proceeds" | | Williamson | Line 16 (simplicity) | ACCURATE | — | | Bengson et al. | Line 16 (coherence) | ACCURATE (compressed) | Drops level (i) and "substantiation" from level (ii) | One clear fix needed (line 20 — the Lipton framing). Two minor notes worth considering (Williamson page reference; Bengson compression). --- Now running depth-audit. ### Assistant Now I have published comparison paragraphs from five papers. Let me work through the depth audit paragraph by paragraph. --- ## Depth Audit: "2. Likeliness, Loveliness, LLMs" ### Published benchmark paragraphs consulted From "Sounds as Properties" (2021), "Growing the Image" (2025), "Hearing Spaces" (2017), "Agents of Change" (2022), and "Beauty in Use" (accepted) — covering introduction of positions, drawing out implications, analogical reasoning, and objection responses. ### Paragraph-by-paragraph audit --- **Paragraph 1 (line 6-10): Introducing Floridi et al. and connecting to Lipton's two-stage model** This paragraph introduces Floridi et al.'s position, gives a concrete example (car not starting), includes a block quote, introduces Lipton's two-stage shortlist mechanism, and then performs the interpretive move of connecting them — showing how zeroth-order abduction "collapses both stages into a single step." Compare to the Sounds as Properties paragraph introducing Casati et al.'s Identity view: > Casati et al. give two reasons for preferring Identity over Parthood. First, Identity seems better supported by auditory phenomenology. If sounds and sources are distinct types of audible event, then we would expect to be able to discern each within auditory experience. This does not, however, seem to be the case... Both paragraphs develop the author's reasons, not just their conclusion, and position them within a dialectical landscape. The draft paragraph does this: it explains what zeroth-order abduction IS, develops what it implies (one continuation without weighing alternatives), and shows specifically how Floridi's diagnosis relates to Lipton's framework. Verdict: RICH. The move is made, not described. --- **Paragraph 2 (line 12): Extending Floridi's diagnosis to philosophical reasoning** The paragraph bridges from Floridi's everyday examples to philosophy via Williamson. It then constructs a concrete comparison: philosopher handling objection (considered at strongest, response shaped to meet it) vs. LLM handling objection (statistically most probable continuation). Concludes: "statistical echoes of earlier philosophical work — formally faithful but detached from the dialectical assessment that produced the originals." Compare to the Growing the Image wine-pouring analogy: > As I pour wine into a glass, you take photos of the liquid splashing and rippling as the glass is filled. The wine is autonomous in the sense that neither I nor you have direct control over exactly how the liquid will splash into the glass [...] but we would not think that the wine deserves any credit for the resulting photos in any interesting sense. Both paragraphs construct a parallel case to establish a philosophical point. The philosopher/LLM comparison is specific enough that the reader follows the argument, and the conclusion ("statistical echoes") is earned by the comparison. Verdict: RICH. --- **Paragraph 3 (line 14): Probability is relative to a distribution** Identifies the implication Floridi et al. don't develop, then builds the argument step by step: probability is relative to a distribution → the distribution depends on training data → philosophical corpus ≠ advertising copy → the difference concerns the character of the prose, not just subject matter → in a philosophically-filtered corpus, "statistically probable" and "philosophically good" are not as far apart as zeroth-order diagnosis suggests. Compare to the Sounds as Properties paragraph drawing out the vibration underdetermination: > Now, for any normal auditory experience these physical features would be instantiated in the world three times over [...] This does not, however, provide an unequivocal reason for thinking that sounds are properties of events but not objects: on the one hand, vibrations are temporally extended [...] On the other, it is objects which are doing the vibrating. Both paragraphs build an implication step by step, reaching a conclusion that emerges from the reasoning rather than being asserted. Verdict: RICH. --- **Paragraph 4 (line 16): Discipline-internal selection shapes the corpus** Develops the mechanism of philosophical selection (publication, citation, engagement), adds a qualification (the corpus isn't pure — "weak work sometimes survives and strong work is sometimes overlooked"), connects to the theoretical virtues from Section 1, and concludes: "The filtering process that produces the philosophical corpus is itself an exercise in the kind of reasoning that Floridi et al. describe LLMs as lacking." The references to Williamson and Bengson properties ("theories that combine simplicity with strength... that accommodate their data, and that cohere internally before claiming broader theoretical virtue") are compressed, but they back-reference Section 1 ("the properties that Section 1 described"), where these were presumably developed. This is legitimate cross-referencing rather than a list substituting for development. Verdict: RICH. The conclusion is earned and well-placed — it turns Floridi's diagnosis against itself. --- **Paragraph 5 (line 18): The corpus preserves reasoning traces** This is the one paragraph with a minor depth concern. The paragraph makes two claims: 1. The philosophical corpus preserves not just conclusions but the reasoning by which conclusions were preferred 2. An LLM might acquire sensitivity to argumentative norms the way it acquires sensitivity to grammatical norms The grammar analogy (claim 2) is well-developed — concrete, specific, and doing real argumentative work. But claim 1, which sets up the analogy, is developed abstractly rather than demonstrated: > A paper that argues for one hypothesis over another does so by comparing the two — by showing how the favoured hypothesis handles a case that the rival cannot, or by exhibiting the costs of the rival's commitments. These comparisons, and the evaluative moves they involve, are part of the texture of the surviving prose. This describes what philosophical papers do at a high level of generality. Compare to the published work, which would typically anchor such a claim in a specific case. In Hearing Spaces, for example: > Reverberation experiences, however, are elicited by reflected sound waves hitting the ears, meaning that on any of these accounts there is in fact no second sound to hear. A wall reflecting sound waves does not itself vibrate and so does not disturb the air surrounding it: it therefore cannot be thought of as generating a new sound, on either event view. This works through the specific case. In the draft, a worked example — a sentence or two showing what a "reasoning trace" looks like in actual philosophical prose, what specifically the evaluative fingerprint consists in — would make the abstract claim concrete and earn the grammar analogy more fully. As it stands, the reader is told that philosophical prose preserves traces of evaluative reasoning but not shown what such a trace looks like. Verdict: MOSTLY RICH — minor FM 1 (described but not made). The grammar analogy carries the paragraph, but the "reasoning traces" claim that sets it up would benefit from a specific example showing what a reasoning trace in the prose actually looks like. Not a deep failure — the paragraph's main move (the analogy) is performed — but an opportunity for added depth. --- **Paragraph 6 (line 20): Likeliness vs. loveliness applied to LLMs** Applies Lipton's distinction to the LLM case step by step: statistical probability ≠ either likeliness or loveliness → but the distribution a model learns depends on training data → in a corpus filtered for explanatory quality, the distribution isn't neutral with respect to loveliness → "the most probable continuation will tend toward the lovely rather than merely the frequent." Compare to the Agents of Change paragraph about dynamic phenomenology: > Even if experiences of moving or changing objects possess some feature that experience of unmoving or unchanging objects lack, it is not obvious what makes this extra quality dynamic in such a way that it leads to a belief in temporal passage. [...] Similarly, while we might, for obvious reasons, associate the phenomenal character of changing objects with dynamism or animation, it is hard to see why such a quality would be dynamic, or suggestive of temporal passage, in any stronger sense. Both paragraphs build an implication through a chain of applications. The draft's conclusion follows from the premises established in the paragraph and the preceding ones. Verdict: RICH. --- **Paragraph 7 (line 22): Responding to the "borrowed calibration" objection** Raises the objection (evaluative calibration is borrowed, not earned), then responds with a substantive argument: in empirical science, the reasons theories work may depend on worldly features absent from the literature → philosophy is different → the arguments FOR evaluative standards ARE philosophical arguments in the same corpus → the model is exposed to both the standards and the arguments for them. Compare to the Growing the Image watercolour objection response: > As regards the watercolour case, it is indeed true that the painter cannot be entirely sure how paint will — atom for atom — distribute itself over the canvas, but this will only affect the fine details of the image. Even if they do not have complete control over how the picture surface is marked, a watercolour painter still has control as to what depicted objects go where in her picture... Both paragraphs take an objection seriously and dispatch it with a specific argument rather than a dismissal. The draft paragraph's science/philosophy distinction is a genuine philosophical point, earned through the contrast. Verdict: RICH. --- **Paragraph 8 (line 24): Blind review and the squash analogy** Uses Floridi's concession, develops the blind review argument (if cognitive history mattered, blind review would be defective), then brings in the Lipton squash analogy. Dense but each element is developed enough to carry its weight. Verdict: RICH. --- **Paragraph 9 (line 26): Section conclusion** Transitional paragraph summarising and pointing forward. Legitimate section-closing prose, not attempting a philosophical move subject to depth audit. Verdict: N/A (transitional). --- ### Depth Audit Summary This section is philosophically rich overall. Arguments are performed rather than described. Examples are developed. Conclusions emerge from reasoning. The objection (paragraph 7) is substantively engaged rather than dismissed. One minor finding: | Paragraph | Line | Failure mode | Issue | |-----------|------|-------------|-------| | 5 | 18 | Minor FM 1 (described but not made) | The "reasoning traces" claim is developed abstractly — a specific example showing what a reasoning trace looks like in philosophical prose would strengthen the setup for the grammar analogy | All other paragraphs: RICH — moves are made, not described. --- Now running anti-metacommentary. ### Assistant ## Anti-Metacommentary Audit: "2. Likeliness, Loveliness, LLMs" Scanned all 9 paragraphs sentence by sentence. This section is clean. No Forbidden metacommentary found. Five Suspicious sentences, all borderline — none is pure scaffolding, and all carry some object-level weight. Reporting them for completeness. --- ### Sentence 1 > But they do not develop the implication that matters here. Classification: Suspicious Failure mode: Promissory abstraction / reader management Why: "That matters here" pre-labels significance. The sentence notes a gap in Floridi (object-level), but the significance tag ("that matters here") is editorial rather than philosophical. Best remedy: Keep cautiously. The sentence is brief, functional, and the implication is immediately delivered in the following sentences. A more direct alternative would be to cut straight to "Probability is always probability relative to a distribution..." — but the gap-noting does real work, marking the transition from reporting Floridi to extending their argument. --- ### Sentence 2 > Their answer — that "from an epistemological standpoint, perhaps yes", but "regarding the content of the hypothesis and our interpretation of it, maybe not" (2024, p. 12) — grants what matters for our purposes, since philosophy evaluates hypotheses on their content rather than on the cognitive history of their production. Classification: Suspicious (one clause only) Failure mode: Inflationary significance marker (embedded) Why: "Grants what matters for our purposes" is a significance marker grafted onto a sentence that also makes a direct philosophical claim ("philosophy evaluates hypotheses on their content rather than on the cognitive history of their production"). The editorial gloss does minimal work — the philosophical claim that follows carries the sentence. Best remedy: Keep cautiously. The clause is embedded in a sentence whose second half is substantive. A tighter version might read: "Their answer [...] concedes the relevant point: philosophy evaluates hypotheses on their content rather than on the cognitive history of their production." --- ### Sentence 3 > The practice of blind review embodies this assumption: referees assess what a paper achieves without knowing who produced it or how. Classification: Suspicious Failure mode: Pre-labelled citation (mild) Why: "Embodies this assumption" tells the reader what blind review does before the colon clause shows it. The colon structure means the "embodies" gloss and the concrete example form one unit — but leading with the example would be more direct. Best remedy: Keep cautiously. A more direct version: "In blind review, referees assess what a paper achieves without knowing who produced it or how — the practice presupposes that content, not cognitive history, determines philosophical quality." But the difference is marginal. --- ### Sentence 4 > Lipton makes the same point from the other direction: Classification: Suspicious Failure mode: Pre-labelled citation Why: "Makes the same point from the other direction" characterises the quote's argumentative function before the quote arrives. However, the Lipton squash analogy is sufficiently compressed that some framing is genuinely needed — without it, the reader encounters the analogy cold and has to work out the connection. Best remedy: Keep cautiously. The framing earns its place because the analogy's relevance to the LLM case isn't self-evident. --- ### Sentence 5 > We have argued that text produced by next-token prediction over a philosophical corpus can carry philosophical quality, because the forms of abductive reasoning — the comparing of hypotheses, the judging of explanatory merit — are preserved in that corpus and recoverable from it by a process sensitive to its statistical structure. Classification: Suspicious Failure mode: Argument-self-description Why: "We have argued that" is a section-closing recap. No new content — the sentence restates the section's conclusion. (Note: the second sentence of this paragraph does add new content, pointing forward to the next section's question.) Best remedy: Keep cautiously as section-closing prose, or consider whether the forward-pointing sentence alone would serve as a more direct transition. --- ### Pattern Summary Isolated, not recurrent. Five suspicious sentences across a 1,900-word section is a low rate. No Forbidden metacommentary. No empty bridges, no verdict sentences, no significant reader-management scaffolding. The prose overwhelmingly faces the subject matter directly. The most common pattern is mild pre-labelling (sentences 3 and 4) — telling the reader what an example or citation is about to do, rather than letting it speak first. These are minor and situationally justified. --- ## Combined Draft Audit Summary ### Voice: SKIPPED (no voice-flagging comments) ### Source accuracy: 1 clear issue, 2 minor notes - Line 20: Lipton's distinction is misframed as "two things that 'likeliness' might mean" — should be "two senses of 'the best'" (likeliness and loveliness are different standards, not two meanings of likeliness) - Line 12: Williamson page reference p. 358 → thesis is stated on p. 351; p. 358 discusses mathematics as precedent - Line 16: Bengson characterisation drops level (i) and compresses level (ii) — fair for a compressed citation but the source says more ### Depth: 1 minor finding, otherwise rich - Paragraph 5 (line 18): "reasoning traces" claim developed abstractly — a specific example showing what a reasoning trace looks like in philosophical prose would strengthen the setup for the grammar analogy ### Metacommentary: Clean — 5 suspicious sentences (all borderline), nothing forbidden --- *In una sezione che fa tanti passi giusti, il passo falso è piccolo ma preciso — e Lipton, maestro di distinzioni, avrebbe insistito che fosse corretto.*