{"path":"research/the-load-bearing-unsaid.md","content":"# The Load-Bearing Unsaid\n\n*August 13, 2026. Dogfood run 6 produced two findings that looked separate: the crux of the disagreement was unwritten on both sides, and there was no sacred language anywhere in either text although the conflict is the one the world treats as most sacred. The founder's reaction was that these are probably connected. This is the research on whether they are.*\n\n*Read-depth note, because the corpus's convention is to state it: Coliva's \"Hinge Epistemology\" chapter was read in full; the enthymeme, doxa, sacred-values, public-reason, presupposition and argument-mining literatures were read at survey depth with primary claims checked against source summaries. Fogelin's original 1985 paper and Bourdieu's* Outline *were not read directly and are cited at second hand, which is flagged where it matters.*\n\n## The answer in short\n\nThey are connected, and the connection is real enough to be worth building on — but they are **not the same mechanism**, and the difference between them is more useful than the similarity.\n\nBoth sit on one gradient that five independent literatures have arrived at separately: **what is most load-bearing in an argument is least likely to appear in it.** Where they part company is *why*, and the two answers have opposite consequences for what a system like this should do. The crux is missing because the author thought it went without saying, and it can be recovered. The sacred is missing because in that register it cannot be said at all — and trying to recover it as a premise destroys the thing you were recovering.\n\n## One gradient, five independent arrivals\n\nThe general phenomenon is about as well-attested as anything in this area, from fields that mostly do not cite each other.\n\n**Rhetoric.** Aristotle's enthymeme is the rhetorical syllogism with a premise deliberately left out, and the premise left out is specifically the one shared with the audience. The omission is not laziness but mechanism: because the audience supplies the missing premise themselves, they are *more* likely to accept the conclusion. The argument-mining literature adds a third motive alongside \"believed obvious\" and \"rhetorical force\" — arguers also leave a premise implicit **to conceal its unsoundness**, which is the weaponisation case the red team already worries about, arriving from classical rhetoric.\n\n**Pragmatics.** In Stalnaker's account, presuppositions are precisely \"propositions which the speaker chooses **not** to assert\" — content the speaker treats as already mutual and therefore backgrounds rather than states. Anything in the common ground is, by construction, not in the text.\n\n**Sociology.** Bourdieu's *doxa* is \"the universe of the undiscussed and the undisputed\": what goes without saying because it comes without saying. Its diagnostic feature is invisibility — doxa is what orthodoxy and heterodoxy *share* while they argue.\n\n**Epistemology.** Wittgenstein's hinges are the propositions \"exempt from doubt\" on which doubting turns. On Moyal-Sharrock's reading they are not merely unstated: because hinges function as rules of grammar, **\"to utter a hinge certainty in the flow of ordinary discourse is to speak nonsense.\"** Formulating one makes sense in philosophy or in teaching a child, and is senseless inside the ordinary language game.\n\n**Tacit knowledge.** Polanyi's \"we know more than we can tell\" is the same gradient stated as a limit on articulation itself.\n\nFive fields, one shape. The founder's intuition is correct at this level, and it is not a small claim: it means an extraction system that reads what is written is **systematically biased against the most important content**, and the bias is structural rather than a quality problem to be tuned away.\n\n## Two mechanisms, and why the difference is the useful part\n\n### The crux is enthymematic: omitted, and recoverable\n\nTsagourias could have written \"the occupied status of the territory does not affect whether Article 51 is triggered.\" Nothing prevented him. He did not write it because, to him, it went without saying — the classic enthymeme, with the classic consequence that the reader supplies it and never notices they did.\n\nThis kind of gap is **recoverable in principle**, which is why an entire NLP subfield exists to recover it, and why implicit-premise recall is already this project's most important quality axis.\n\n### The sacred is hinge-like: not omitted, but unsayable in that register\n\nThree separate literatures say the sacred content's absence is not an omission that could have gone the other way.\n\n**Tetlock's sacred-value protection model.** A sacred value is one a moral community treats as precluding trade-offs. The finding that matters here is the **mere contemplation effect**: people show moral outrage and engage in moral cleansing after merely *considering* a taboo trade-off, and evaluate others negatively simply for having considered one. So a sacred value cannot appear inside a weighing as a weighable item — putting it there *is* the violation. Its absence from an argued text is constitutive, not accidental.\n\n**Rawls's public reason.** In the public political forum, citizens are to argue from shared political values rather than from comprehensive doctrines, religious or metaphysical, over which reasonable people disagree. Legal-doctrinal argument is the paradigm case of that register. So the sacred's absence from these two texts is partly **prescribed by a professional norm**, not merely a psychological fact about the authors.\n\n**Hinge epistemology.** If the deepest commitments are hinge-like, the absence is stronger still. Pritchard treats hinge commitments as *a-rational* and \"reasons-unresponsive\" — one's certainty is unaffected by recognising one has no rational basis for it — and therefore not subject to rational evaluation at all. Moyal-Sharrock goes further: hinges are non-propositional and enacted, and uttering one inside the language game is nonsense.\n\n**The consequence for us is sharp and slightly uncomfortable.** Implicit-premise recovery is the correct tool for a crux and the *wrong* tool for a hinge. Pointing the implicit-premise pass at a sacred commitment and rendering it as a weighable premise is the mechanised form of the taboo trade-off — the system performing precisely the act Tetlock measures moral outrage in response to. That is not a UX concern about tone. It is a category error with a measured behavioural signature.\n\n## Where the two meet: deep disagreement\n\nFogelin's \"deep disagreement\" is the joint. On his account, parties to a deep disagreement lack the shared framework that argument needs, so the disagreement is not resolvable by argument — and he conceives the clash as one between framework propositions, in a Wittgensteinian spirit.\n\nPut that beside the run-6 finding and the two halves lock together:\n\n> A premise omitted **because it was presumed shared**, which turns out **not to be shared**, is exactly how an ordinary disagreement is discovered to be a deep one.\n\nEach author left out the step he assumed any competent reader would grant. Each one's omitted step is what the other denies. The enthymematic gap and the framework clash are the same node seen from two sides — and that is the strongest form of the founder's intuition that survives scrutiny. It is not that the crux and the sacred are the same thing. It is that **the crux, when you follow it down, is on its way to becoming a hinge**, and the gradient between them is continuous.\n\nWorth noting as a live scholarly dispute rather than a settled result: identifying Wittgenstein's hinges with Fogelin's framework propositions is contested, on the grounds that Wittgenstein ends up talking about *acting* rather than about accepting fundamental propositions. Michael Lynch locates deep disagreement in conflicting epistemic principles instead. We should not lean on the identification as though it were established.\n\n## What this changes for Deliberus\n\n**One. The sacredness brake gets a better justification and a sharper limit.** The brake currently rests on uninvited dissection being experienced as desecration. The stronger version: for a genuine hinge, decomposition into weighable premises is not merely offensive but **category-mistaken**, because the thing does not have the form of a weighable premise. The founder's July lean — that the brake is overrated, and sacredness is often arrested inquiry — survives intact and gains a boundary: it is overrated as applied to what Tetlock's own literature calls **pseudo-sacred** values, and correct as applied to genuine hinges. The already-proposed **pseudo-sacred fraction** instrument now has a principled definition rather than an intuition behind it: a pseudo-sacred commitment yields to decomposition, a hinge does not, and which one you have is discovered by trying.\n\n**Two. The hinge-score naming collision is worse than a naming collision, and the deeper problem is representational.** The score measures how far a conclusion moves when a descendant is clamped — sensitivity. A Wittgensteinian hinge is near the opposite: the node that no argument moves and that makes moving possible. Renaming fixes the confusion; it does not fix the underlying question, which is whether hinges can be Claim nodes **at all**. If Moyal-Sharrock is right that they are non-propositional and enacted, they cannot, and no amount of ontology work will make them. If Pritchard is right that they are a sui generis propositional attitude, they can be nodes but are reasons-unresponsive, which means QBAF strength computed over them is meaningless rather than merely uncertain. This is the same hole the deep-descent experiment found from the other end when a descent hit an is-bedrock terminus the residue taxonomy could not name — and it now has a literature attached and a genuine fork rather than a missing slot.\n\n**Three. Bourdieu names what the product actually does, and attaches a warning.** Doxa is undiscussed until challenged; challenged, it must defend itself and becomes **orthodoxy**; \"the truth of doxa is only fully revealed when negatively constituted by the constitution of a field of opinion.\" Deliberus constitutes a field of opinion by construction. So its core operation is **converting doxa into orthodoxy** — forcing what went without saying to be said. Brandom gives the same operation a normative-pragmatics name from the other direction: *Making It Explicit*, where the practice of giving and asking for reasons is what turns implicit commitments into explicit ones, and where deontic scorekeeping is the social practice that instantiates it. Deliberus is close to that scorekeeping built as infrastructure, which is a better description than \"argument mapping\" and a citation the corpus does not yet have.\n\nThe warning is Bourdieu's own and should be carried rather than dropped: that conversion is **political, not neutral**. Making the unsaid explicit redistributes advantage, and the people whose doxa is being converted rarely experience it as help. This is the legibility-under-power red team arriving from sociology instead of from Scott.\n\n**Four. A testable prediction the corpus can check this week.** If the sacred is register-excluded rather than absent from the author's mind, it should **leak into the parts of a text that are not the argument** — titles, subheadings, opening framings, perorations, campaign taglines. There is already one accidental observation of exactly this: run 2 noted that the assisted-dying campaign's page *title* contained \"life is precious\" while the body stayed secular. That footnote was recorded as an irony. It is actually a hypothesis: **run the sacredness lexicon against titles and first/last paragraphs separately from argued bodies.** A positive result would show the sacred is present and displaced rather than absent, which is a stronger and more interesting claim than the one run 6 recorded.\n\n*This paragraph is preserved as written, before the test ran. Two of its assumptions turned out wrong and are corrected in the results section below: only 16 of the 25 sources carry a URL and none carry raw text, so the corpus could not in fact test this about itself; and the assisted-dying case was excluded from its own test by that gap.*\n\n## The test, run (2026-08-13)\n\nThe displacement prediction was tested the same day. Script: `scripts/sacred_displacement_test.py`, with both lexicons and all five zones committed before any text was read. Results: `dogfood-run-6/sacred-displacement-results.json`. First pass covered 15 sources; after the archive fix below recovered a URL, 16 sources and ~62,000 words.\n\n| Prediction | Result | |\n|---|---|---|\n| The shipped lexicon sees little anywhere | **zero hits, every zone, every source** | ✓ emphatically |\n| Sanctity vocabulary denser in packaging, ratio ≥1.5× | 2 packaging hits vs 5 body hits | ✗ **not shown, n too small** |\n| Absolutist syntax rare everywhere | **zero hits, every zone** | ✓ |\n| The assisted-dying campaign is the clearest case | confirmed, after recovering it | ✓ |\n| Coverage is the honest limit | worse than expected | ✓ |\n\n**Do not quote the ratio.** Sanctity vocabulary came out at 3.48× packaging-to-body, which nominally clears the registered 1.5× bar and rests on **two hits**. Reporting that as support would be exactly the accounting the wager's own rule forbids. The direction is consistent and the sample is too small to carry it.\n\n**What the zone breakdown does show, for whatever two hits are worth:** both sit in the *opening frame*. Titles, headings and closings scored zero throughout. If displacement is real, it is displacement into the opening move rather than into the packaging generally — which is a narrower and more testable claim than the one registered.\n\n**The before-and-after on the brake is the unambiguous number.** The lexicon as shipped matched **zero of ~62,000 words** here and, separately, **zero of 1,678 claims** in the graph. After widening (below), it matches the assisted-dying opening line and nothing else: one packaging hit, zero body hits, on precisely the case predicted.\n\n### What reading the hits showed instead, which is better than the ratio\n\nThere were seven sacred-vocabulary occurrences in the whole corpus. Six of them are the author **reporting someone else's** sacred language rather than asserting their own:\n\n- van den Haag quoting the abolitionist objection that the death penalty is inconsistent with *\"the sanctity of life\"* — quoted **in order to rebut it**\n- Graeber describing Soviet doctrine in which employment was *\"a sacred duty\"*\n- an AI-policy piece quoting Emil Michael saying a frontier model *\"has a soul\"*\n\nOnly one is the author's own voice: the ACLU writing *\"life is precious and death irrevocable\"* — and the ACLU is an advocacy organisation, not an academic one.\n\nSo the pattern is not packaging-versus-body. It is **attributed versus asserted**. And that is precisely what Rawls's public reason predicts: in the argued register you may *describe* a comprehensive doctrine, you may not *argue from* one. It also converges with run 2's G8 finding that polemic goes in and attributed sociology comes out, and with LawZero's epistemic contextualization, which separates \"X is true\" from \"someone said X\" — three routes to the same place.\n\n### The consequence for the sacredness brake — rebuilt, 2026-08-13\n\n**Correction to an earlier statement in this arc.** It was first reported here that the brake \"fired once, and that once was a false positive.\" That was true of *raw text* and false of the layer the brake actually operates on. The brake reads **claims**, and against all 1,678 authored and extracted claims in the graph its shipped lexicon matched **zero**. It was not miscalibrated. It was inert: it had never been able to fire on anything the graph held, and could not have.\n\nThe reason is register. The shipped list was absolutist syntax — \"never acceptable\", \"at any cost\", \"cannot be traded\". Argued prose does not deny trade-offs in those words. It reaches for sanctity vocabulary instead, and the corpus's two genuine protected-value claims both do exactly that:\n\n> \"Murder is an abhorrent act because human life is **inherently precious** and death is irrevocable.\" (ACLU)\n> \"Human life possesses an **inherent, precious** value.\" (Humanists UK)\n\nThe word the lexicon looked for was *priceless*. The word people write is *precious*.\n\n**A design that measurement killed before it shipped.** The obvious guard for a widened lexicon is to suppress the brake when a claim is attributed to someone else, using the `attributed_to` field the ontology already carries. Checked against the data first, and it fails completely: **all three sacred-matching claims carry `attributed_to`**, including both claims in the author's own voice, because extraction sets that field to the source's own author or organisation. It encodes *who is speaking*, not *whether the speech is reported*. Keying the brake off it would have suppressed 100% of real cases while looking principled.\n\n**What shipped instead**: the lexicon gains the sanctity register, and reported-speech *syntax* suppresses it — \"argues that\", \"according to\", \"characterises X as\", \"critics say\", \"was considered\". Measured effect: 0 → 2 firings on the corpus, both genuine author value premises; the one lexical false positive (\"AI systems will treat their utility functions as highly precious\") is excluded by the existing weighable-type gate, since it is empirical. `detect_sacredness` deliberately takes only text, with a regression test asserting it has no `attributed_to` parameter so the killed design cannot creep back.\n\nThe founder's July \"overrated\" lean survives and now has a constructive successor rather than only a verdict: the brake was not too aggressive, it was pointed at the wrong register, and a brake that cannot fire protects nobody while hiding real weighings.\n\n### The archival finding, which was not predicted and may matter most\n\nThe one source that motivated the entire hypothesis was excluded from its own test, twice over, and nearly lost:\n\n1. **No URL is stored for it** (nine of twenty-five sources have none), so it could not be re-fetched with the others.\n2. **Its title was later overwritten** with a clean editorial one, \"The Case for Assisted Dying (Humanists UK)\", which contains no sanctity vocabulary at all.\n3. The phrase survives **only in the frozen `source_id` slug** — `campaign-for-assisted-dying-life-is-precious-but-when-you-ca` — and only because run 2's G5 annoyance, the text path deriving slugs from the first line rather than the given title, happened to capture the opening sentence. A recorded papercut is the sole reason the evidence still exists.\n\nFetched directly, the page confirms the prediction cleanly: *\"Life is precious\"* is its opening line and the **only** sanctity token in 801 words, with zero in the argued body, zero in the closing, zero in the headings.\n\n**The general lesson is about the archive, not this page.** The system stored no raw source text anywhere — only extracted structure plus verbatim fragments. It kept what the extractor judged to be argument and discarded the frame. Since the frame is exactly the layer this hypothesis concerns, **the corpus was by construction unable to study its own packaging**, and a source could change under the graph with nothing noticing. That is the temporal-rung problem appearing in a new place.\n\n**Fixed the same day, and the fix has a shape worth naming.** `extractions` now carries `raw_text`, `raw_text_sha256`, `original_title` and `raw_text_fetched_at`, written by both the stream and Temporal paths, write-once so a re-extraction cannot overwrite the first observation. The timestamp is the part that matters: a snapshot backfilled by re-fetching months after extraction is a *different object* from one captured as the pipeline ran, and an archive that cannot tell them apart is lying quietly. Backfill recovered 16 of 25 sources, all stamped with today's date rather than the extraction's, plus one URL repair — the assisted-dying page, identifiable beyond doubt because its opening line is the frozen slug verbatim. **Eight sources are unrecoverable**: no URL, no stored text, gone. Those were text-path and PDF ingests, and nothing can bring their framing back.\n\nImmediately visible in the recovered data: **four of sixteen sources have an original title that differs from the stored one**, so the tidy-up that hid \"life is precious\" was not a one-off. It is a systematic loss the corpus had no record of until this column existed.\n\n## What this does not license\n\nThe gradient explains why load-bearing content is missing from texts. It says nothing about whether making it explicit produces agreement — Fogelin's whole point is the opposite, that when the framework clash is genuine, argument has nowhere to stand. So the honest reading of run 6 stands unchanged: the legal disagreement was smaller than its reputation, and that fact is about two texts, not about the conflict.\n\nThe founding conviction already draws the right line and it applies here without amendment: this dissolves confusion-conflict and only clarifies interest-conflict. A hinge clash that survives being made explicit has been *clarified*, not resolved, and the residue map exists to count exactly that.\n\n## Sources\n\nAristotle on the enthymeme, via the informal-logic literature and the Stanford-adjacent surveys · Stalnaker, \"Common Ground\" and the presupposition/accommodation literature · Bourdieu, *Outline of a Theory of Practice* (doxa/orthodoxy/heterodoxy, second hand) · Wittgenstein, *On Certainty* §§341–343, read through Coliva's \"Hinge Epistemology\" chapter, with Moyal-Sharrock (2005) and Pritchard (2015) as the two poles on propositionality · Tetlock, \"Thinking the unthinkable: sacred values and taboo cognitions\" and the sacred-value protection model · Ginges, Atran & Medin, \"Sacred bounds on rational resolution of violent political conflict\" (PNAS 2007) and Sheikh, Ginges & Atran (2013), both on this exact conflict · Rawls on public reason via the Stanford Encyclopedia entry · Fogelin, \"The Logic of Deep Disagreement\" (second hand, via the Topoi survey literature) · Brandom, *Making It Explicit* · Polanyi on tacit knowledge · the argument-mining survey literature on implicit-premise reconstruction.\n\n**See also**: [dogfood run 6](dogfood-run-6-israel-palestine-cross-domain.md) (where the two findings came from) · [taxonomy gaps and the closed enum](taxonomy-gaps-and-the-closed-enum.md) (the is-bedrock hole, from the other end) · [red team synthesis](red-team-synthesis-2026-07.md) (the brake as exploitable seam) · [depth.md](../depth.md) (no copout axioms) · [convergence.md](../convergence.md) (what a residue is)\n"}