{"path":"research/what-human-judgment-is-for.md","content":"# What Human Judgment Is Actually For\n\n**Date**: 2026-08-27\n**Occasion**: The founder, on being shown that ratification is an unbuilt half: *\"I don't really know if it makes sense to have human users vote/ratify… Let's not install meaningless buttons for users to press and judgments we hold our breath for them to make that don't actually change anything… when they do engage it should MATTER, it should be filled with MEANING, so they can FEEL their humanness rippling across the graph.\"* With three constraints: meet the residual-error taxonomy structurally; do not make ratification a permission-to-exist, because that is how numbers beat truth; and find where human holistic, attunement-shaped attention beats a model *under the best imaginable conditions*.\n**Status**: analysis and a proposal. Nothing built, nothing ruled.\n\n---\n\n## 1. The taxonomy answers the question, and the answer is not where the buttons are\n\nThe eight residual-error classes sort by *where the error lives*, and the triage already found that **the naked ones live outside the graph** — in what enters it, the shape it crystallized, and the human gate around it. Read that list asking *which of these can only a human fix*, and it collapses to two, plus one the corpus already ratified on other grounds:\n\n| Class | Can a machine fix it? | |\n|---|---|---|\n| **Coherent absence** (2) — what never became a step | **No, in principle.** A model cannot know what the corpus it was given is a *sample of*. The cross-source pass finds absences *within* supplied material; nothing inside the system can see that the material itself is unrepresentative. | **Human** |\n| **Frame lock-in** (7) — the reuse flywheel's shadow | **No, structurally.** Recognising a frame *as* a frame requires standing outside it, and reuse is the whole cost-curve bet, so the better the flywheel works the more expensive re-framing becomes. A coherence-optimising system will assimilate, not re-frame. | **Human** |\n| Same-hand bias (1) | Only by an independent generator — which a human is, *conditionally* (see §2). | Human, conditionally |\n| Identity scattering (3) | Largely machine, with a measured ceiling; the value-claim end wants a human bar. | Mixed |\n| House choices (4) | Needs **the operator**, not users. Fredrik's judgment, not a crowd's. | Operator |\n| Dead instruments (5) · Laundering (8) | Measurement and plumbing. | Machine |\n| Ratification decay (6) | Circular — this is the thing being designed. | — |\n\n**So the honest answer to \"what do we truly need human judgment for\" is: for the two things that live at the graph's boundary rather than inside it.** Everything inside is instrument work, and instruments do not need a crowd.\n\n## 2. Ratification is the *worst* available use of a human, on three independent grounds\n\nThis is the load-bearing negative result, and it did not come from taste.\n\n**One — it destroys the very independence it is recruited for.** Same-hand bias is the error class where a human is most obviously valuable: they are a different generator. But **a human anchored on a machine proposal is not an independent generator; they are a second sample of the machine's framing.** Ratification shows the person the machine's output and asks them to react to it. Whatever independence they had is spent before they answer.\n\n**Two — the complementarity literature says the passive mode is where teams lose.** Human-AI complementarity turns out to be *hard*: teams frequently fail to beat the better of human-alone and AI-alone ([arXiv:2605.04070](https://arxiv.org/abs/2605.04070); [arXiv:2510.26518](https://arxiv.org/abs/2510.26518)). Two conditions recur where it does work: **humans alone outperforming AI alone on that task**, and **humans actively interrogating AI recommendations rather than passively accepting them** — with clinical evidence that teams beat both sides only when practitioners are *trained to question* the AI. Ratification is the passive-acceptance mode by construction. It is the shape the evidence says produces no gain.\n\n**Three — it manufactures the one thing that makes brigading work.** The founder's worry about unprivileged groups is **not overblown**; it is the documented Wikipedia dynamic. The research anchor is an asymmetry rather than a conspiracy: *\"there are a minority of editors who are willing to put in a tremendous amount of energy to push a point of view… there aren't enough people willing to put in the time on the other side\"* (Bruckman, Georgia Tech), and the mechanism by which persistence wins is stated plainly in Wikipedia's own policy — *\"the first person to step out of the war in essence loses, in that the article's text will reflect the update made in the final round.\"*\n\n**Read that mechanism carefully: persistence wins because the object has ONE STATE.** An article is a single current text, so whoever edits last holds it. A graph where contributions accumulate **has no single state to fight over** — unless we build one. **Ratification-as-permission-to-exist is exactly the construction of a single state**, and it hands it to whoever has the most time.\n\n## 3. What humans are for: four acts, each additive, each mapping to a real gap\n\nThe design principle that follows — **narrowed 2026-08-27 after the first phrasing broke under\nquestioning, see §4d**: **contributions are added, never overwritten, and nothing needs permission\nto exist.** A contribution cannot be un-contributed by other people arriving in greater numbers,\nbecause there is no single current text for them to win. What this does *not* protect is weight:\nthe strength layer is itself an aggregate.\n\n### Act one — *\"Nobody here has said X.\"*\n\nThe absence judgment, against class 2. Someone who lives in a debate carries a felt sense of its shape: which arguments are conspicuously missing, which side is over-represented, which register nobody in this corpus writes in. **That is not retrieval; it is a comparison against a discourse the system has never seen.**\n\nIt is also, decisively, **a low-rung act**. It needs no ontology fluency, no decomposition skill, no understanding of the strength layer — which is how it satisfies the founder's two constraints at once, since the engagement gradient puts decomposers at a fraction of one percent while readers are the overwhelming majority. *\"You know this argument. What's missing?\"* is answerable by a reader.\n\n### Act two — *\"This whole region is framed wrong.\"*\n\nThe reframe judgment, against class 7, and the taxonomy notes it has **no mechanism at all** — supersession exists for edges, never for frames. This is the most obviously holistic act on the list: it is a judgment about the overview rather than about any claim, and it is the one an assimilating system is structurally worst at.\n\n### Act two, continued: the two gap-types, and the one the ratified rule cannot reach (2026-08-27)\n\n*Founder, on the disengagement analysis: \"the anxiety is a result of lack of clarity. It could also\nbe the fact that there is a gap that is not seen because it's inside the frame the graph already\nstands in, while an outsider could contribute a consideration that opens the frame.\" That is a\nsharper distinction than the analysis it corrects, and it exposes a limit in a rule this project\nhas already ratified.*\n\n**Gap type one — the visible gap.** An unanswered critical question, an unsupported value premise, a\nthin branch. The graph knows it is there, can display it, and can rank it: `worth_asking` computes\nwhich open question would move the conclusion most. The ratified rule applies cleanly — *no gap\ndisplayed without a move offered* — and PACE's appraisal comes out **closable**, which is the\ncuriosity branch.\n\n**Gap type two — the frame gap.** A consideration that is missing because it lies outside the frame\nthe graph was built inside. **The graph cannot display it, because displaying it would require\nseeing it** — which is coherent absence (class 2) and frame lock-in (class 7) experienced from the\ninside.\n\nSo the rule does not fail here; **it does not reach**. *No gap displayed without a move* governs\ngaps that get displayed, and this one never is. What a person meets instead is a felt sense that\nsomething about the framing is wrong, with no affordance anywhere for what they sense — and that\nappraises as **not closable**, which is precisely the inhibition branch.\n\n**The inversion that matters: here the anxiety is correct, and must not be designed away.** For a\nvisible gap, discomfort is a UX failure to be fixed by offering a move. For a frame gap, the\ndiscomfort *is the signal* — the person is registering something true that the map does not contain.\nSoothing it would be silencing a working alarm. **You cannot make a frame gap feel closable, and you\nshould not try. You give it an exit.**\n\n**What exists and what does not.** `REFRAMES` is a live edge type, so the graph can hold a reframe\n**once someone can state it as a claim**. There is nothing for the state *before* that — and the\nbefore-state is the whole problem, because a person who could already articulate the missing\nconsideration would simply add it. The taxonomy's note that act two has **no mechanism at all** is\ntherefore more specific than it looked: what is missing is not the reframe *record*, it is the\n**path from felt unease to a statable claim**.\n\n**Design direction, and the pattern already exists here.** Not a contentless \"something is off\"\nflag — that is a downvote without a reason, which this project's own rating discipline rejects, and\nit would be uninformative and brigadeable. The move is **articulation help**: the person marks a\nregion, and the system asks a small number of *frame-opening* questions — what is assumed here that\nyou would not assume · whose situation is absent · what would have to be true for this framing to be\nthe right one — converting felt unease into a claim that can then be attacked, supported and\nsupersede. That is the inversion principle (*the machine holds the user's hand through the descent*)\napplied to reframing instead of decomposition, and the **weighing descent is its working sibling**:\n14 questions that turn an unarticulated weighing into structure.\n\n**Two cautions carried from this corpus's own findings.** The question set is *itself a frame*, so it\nwill carry the distinctions of whatever register produced it and break on the next — it needs a\ndoes-not-fit value from day one. And it must not be **invented**: a trigger list authored from\nimagination catches imagined cases, which this project measured this month. Harvest it instead from\nthe frame-opening moves already recorded — the classificatory premise found in the legal pair, the\nleveling dissolution in the nativism thread, the unsurfaced weighing in the bloc-arithmetic cluster.\nThose are real instances of a frame being opened from outside, and they are the right training\nmaterial.\n\n**And the modality already chosen supplies what the artifact cannot.** A second person in a\nco-present session *is* the outside frame — which is why that session is the right first test rather\nthan the solo mode, and why it is the setting where act two could first be observed happening at\nall.\n\n### And the population is the horizon map (2026-08-27)\n\n*Founder: \"every new user on Deliberus is also a separate outside frame, a separate reality tunnel,\na parallax that is a horizon to explore towards… if we build it right and it builds traction,\nDeliberus will engage these frames with each other.\"*\n\n**The unification is exact, not figurative.** A **horizon** is this ontology's ruled name for *where\nshared mapping currently ends, moving as the corpus grows*. Another person's frame is precisely\nthat: a boundary of what has been mapped together, which moves when it is engaged. So the residue\nmap and the user population are **two views of one object** — the first counts horizons in the\ngraph, the second is where the graph's remaining horizons live. And it lands on the inversion\nalready recorded here: the parallax gap stops being the analytical pole's hard limit and becomes\n**the mechanism the absence judgment depends on**. Extending it from a pair of positions to a\npopulation is the natural next step rather than a new claim.\n\n**Three things qualify it, and the third is a design requirement rather than a caution.**\n\n**Co-presence is not engagement, and this is measured.** Run 6 put two opposed texts on one question\ninto one graph: **402 candidate pairs inside the debate, zero at the auto-link threshold**, and the\ncrux unwritten on both sides. Two frames in one corpus and the machinery could not connect them. So\n*\"Deliberus will engage these frames with each other\"* names **the work**, not a consequence of\nhaving them both present.\n\n**At scale, frames engaging each other is the thing that has already failed once.** Bail's month-long\nexposure to opposing views increased polarization. The difference the project bets on — reasons\nrather than flags — is real and unrun, and the honest form is that this is the bet, not the rebuttal.\n\n**And N users is not N frames.** If arrival self-selects on one disposition, a thousand users from\none reading community are **one frame sampled a thousand times** — which is the correlated-error\nfinding of the same week, one level up, and it means **frame diversity has to be selected for rather\nthan awaited**. Register diversity was promoted from preference to precondition for exactly this\nreason; the population-level version is the same rule about people.\n\n**What already exists for the work part.** The cross-source premise pass is the frame-engagement\nmachinery: it reads two opposed sources and proposes the premises **neither states** — the\nclassificatory premise, the default one side treats as needing no justification. It ran on\n2026-08-27: 11 premises across 4 adversarial pairs for $0.64, zero junk. It is **operator-gated and\nnot in the pipeline**, which is the concrete distance between the vision in this section and what a\nuser would meet today.\n\n### Act three — *\"I do not accept that as a statement of my position.\"*\n\n**Corrected 2026-08-27 on founder challenge, and the correction improves the act.** This was first written as *\"that is not what I meant\"* and called **incorrigible**, on the grounds that experience reports are observer-privileged. That over-read the ratified conviction. The ruling splits interior claims in two: **experience reports** are observer-privileged and decomposition never attacks the datum, while **self-interpretations** — first-person causal, counterfactual and satisfier claims — get **no elevated credence for being first-person**. A claim about *what I meant when I wrote that* is a report about a past mental state, which sits on the fallible side; the corpus's own citation of Mercier & Sperber on post-hoc rationalisation says so directly. People are routinely wrong about their own reasons.\n\n**What is genuinely unchallengeable is narrower, and it is not a report at all.** *\"I do not endorse that rendering as a statement of my position\"* is a **speech act**, not a description of history. It cannot be factually wrong, because it is not asserting a fact — it is withholding endorsement, in the present, by the only person with standing to give it. The archaeology of what was actually in someone's head stays open to analysis like anything else.\n\nThat distinction is not pedantry: it changes what the act is *for*. The value was never historical accuracy about an intention. It is that **the person gets to refuse a rendering and supply their own**, and the graph then holds both — the opponent's version and the repudiation, each attributed. Always-mint exists precisely so that comparison is possible, and this is the live test's designated primary evidence.\n\n### The naming collision this exposed, and it is a real one\n\n[incentives-analysis.md](incentives-analysis.md) §5 and §7 say **\"ship ratification\"** and count four returns from it, including an **opponent-proposes / you-ratify** design that defeats strategic silence. Read beside §2 above, that looks like a flat contradiction of this document. It is not — **it is one word doing two jobs**, and they need separating before the corpus contradicts itself:\n\n| | Who is judged | Who judges | Verdict |\n|---|---|---|---|\n| **Consent** (incentives-analysis's sense) | *your own* position, as rendered by someone else | **you**, the person whose position it is | **Ship it.** No majority semantics — the ruling belongs to one person, so numbers cannot win it, and it delivers the misrepresentation-protection benefit that makes contributing self-interested. This *is* act three. |\n| **Ratification** (this document's sense) | a **machine's** proposal about the world | whoever is checking | The passive-acceptance mode; keep it as an operator practice that the checker can **fail**, never as a public permission gate. |\n\n**And §5's mechanism survives intact under the better name.** If an opponent may render your premises and you must respond, strategic silence — the attack §5 identifies as *\"indistinguishable from good citizenship\"* and invisible to every existing instrument — stops paying. That is a genuine defense and it is additive: your refusal adds an object rather than deleting theirs.\n\n### Act four — *\"Here is what this costs, from where I stand.\"*\n\nStake testimony, and it is the answer to the founder's own worry rather than a casualty of it. The people most vulnerable to being out-voted are frequently the people whose testimony about *consequences* is the scarcest input a debate has. Being on the receiving end of a policy is knowledge no model holds and no amount of reading supplies. It belongs to the attunement pole, and it is additive by construction: one person's account of what a policy costs them is not erased by a thousand people who do not share it.\n\n## 4. What this makes of ratification, votes, and validity\n\n**Logical validity is not what humans are for, and the founder is right that his 2010 answer expired.** But the reason is narrower than \"LLMs got good\": most real arguments are not deductive at all, so validity was always a small slice. The scheme-plus-critical-questions architecture already handles the defeasible majority, and a machine handles the deductive minority. Neither needs a crowd.\n\n**Votes survive, demoted.** Agree/disagree feeds the discursive-dilemma instrument, which is an *instrument reading* rather than a gate. That is the right role and the right size: aggregation informs, and never decides.\n\n**Ratification survives in exactly one shape.** Not permission-to-exist, but **a check on machine proposals that the ratifier can fail** — i.e. what the seed-flawed-proposal probe measures. That is the interrogation mode the literature says works, and it is a small operator-and-close-collaborator practice, not a public button. It should never gate a human's own contribution, only a machine's.\n\n**And the corollary the corpus should carry**: an unratified *human* claim is not a lesser claim. The gray band exists to say *unexamined*, not *unpermitted*.\n\n## 4b. The same two classes, named from the ideology-critique side twenty years earlier\n\nThe founder's recollection is right and the fit is closer than a resemblance. Both nakedest classes are already in the corpus as philosophy, in [zizek-schmachtenberger-connections.md](zizek-schmachtenberger-connections.md), and **neither concept ever travelled out of that document** — a search returns zero uses of \"unknown knowns\" anywhere else in the corpus.\n\n**Frame lock-in is Žižek's *unknown knowns*.** His addition to Rumsfeld's matrix: *\"the disavowed beliefs, suppositions and obscene practices we pretend not to know about, although they form the background of our public values\"* — the frameworks that **structure perception while remaining invisible to the one perceiving**. That invisibility-from-inside is exactly why a coherence-optimising system cannot re-frame itself, arriving from ideology critique rather than from a reuse-economics argument.\n\n**Coherent absence is his reading of the Real, and the corpus already stated it as a build requirement.** From §1.2: *\"the most interesting nodes in a Deliberus argument graph would not be the claims that different worldviews contest, but the **gaps** — the topics, questions, or connections that a worldview simply cannot represent\"*, with the implication written out: *\"the system should be designed to detect and highlight not just disagreements between worldviews, but structural gaps within worldviews.\"* **That is the coherent-absence instrument, specified in 2026 and never built, and never recognised as the same object the error taxonomy names.**\n\n### And the new part, which runs the other way\n\n§1.3 treats the **parallax gap** — two linked perspectives with no neutral ground between them — as the hard *limit* on the analytical pole: decomposition is always done from a position, so there is no view from nowhere.\n\nRead against §1.1 that limit inverts into a mechanism. **You cannot see your own frame; someone standing in another one can.** So the irreducibility of perspective is not only the cost the convergence wager has to price — it is **the resource the absence judgment runs on**. This is why act one has to be performed by a *person other than the author*, and why no amount of prompting gets a single model there: a model asked to find what is missing searches the frame it is already inside.\n\nIt also answers Žižek's own challenge to the worldview filter, which the corpus recorded and left open: *can unknown knowns be surfaced through a technological interface, or does making them explicit transform them into something else?* His objection is that glasses you consciously put on are no longer ideology. **The answer this analysis supplies is that the interface is not the instrument — the other person is.** Deliberus's job is to hold the two frames in one addressable structure so the gap between them becomes visible as an artifact, which is a different claim from promising anyone transcendence of their own position.\n\nThe standing warning travels with it unchanged: **the point of wearing the other view is discomfort, not tourism**, and a system that makes it comfortable has performed the containment it meant to expose.\n\n## 4c. RATIFIED into canon 2026-08-27 (founder: *\"Go ahead\"*) — both lines are live in the One-Breath blocks\n\nThe founder's reaction to §4b: *\"Stuff like this deserves a permanent mention in our project claude.md I feel, yes? It is a kind of conviction or at least principle that feels crucial for the success of our endeavour.\"* Per the constellation precedent a canon add is ratified explicitly, which he then gave. **Both lines are now in `CLAUDE.md`** — the conviction beside *Opacity Is a Cost, Not a Mystery*, the principle in the principles block — carrying the positionality correction below rather than the first wording.\n\n**It is a conviction with a principle attached, and they belong in different blocks.** The claim about how minds work loses to evidence; the rule about how to build loses to regret.\n\n### The conviction, and the framing matters\n\nThe first version was stated as a limitation — *\"you cannot see your own frame; someone standing in another one can.\"* True, memorable, and it leads with a deficit. The same content stated as a capacity is both more accurate to what the project is for and closer to its own tagline:\n\n> **What is invisible from inside a position is visible from outside it — so minds make each other legible.** The frame you reason inside is the one thing you cannot examine, and someone positioned differently examines it for free. That makes **positionality** — the fact that everyone reasons from somewhere — a *resource* rather than only a cost.\n\n**Corrected on founder challenge 2026-08-27, and the challenge was right**: the draft first said this made *irreducibility* a resource, and asked whether a frame that an outsider can see is irreducible at all. It is not, and the two claims must be kept apart.\n\n- **Positionality** is the property that everyone reasons from somewhere. It has two consequences from one cause: your frame is invisible **from inside** (the cost) and visible **from outside** (the resource). That is what this conviction is about, and it is nearly undeniable.\n- **Irreducibility** is the further claim that some disagreements survive being made **fully explicit**. Nothing here touches it. Making a frame visible converts an unstated assumption into a stated claim — it becomes *arguable*, never thereby *agreed*.\n\n**Visibility is not reducibility**, and the corpus already carries that distinction twice. The founding conviction's own guard says legibility *dissolves* confusion-conflict and only *clarifies* interest-conflict. And a `structural` horizon is precisely a disagreement that survived full decomposition — which is what an irreducible parallax looks like **after** it has been made visible, not instead of. So the parallax claim is untouched by this inversion; it has only been made precise about *which* part is irreducible. The wager is about that part, and the horizon map exists to count it.\n\n**It is the companion to `Opacity Is a Cost, Not a Mystery`, and should sit beside it.** That conviction says why another mind holds its position is findable. This one says **who does the finding**: not a lone reasoner introspecting, and not a model auditing itself, but two positions making each other visible. Weaving minds together, stated epistemically rather than warmly.\n\n### The principle it implies\n\n> **Never ask a mind to audit its own frame.** The absence judgment comes from someone other than the author; a model asked what is missing searches the frame it is already inside; and a human anchored on a machine's proposal has already adopted that proposal's framing.\n\nThat is the line a future session needs *at the moment of designing a surface*, and it is what rules out ratification-as-the-human's-job in one sentence.\n\n### Why it is well-anchored rather than elegant\n\nThree independent supports, one of them measured in this corpus this week:\n\n- **Bourdieu's doxa** is what orthodoxy and heterodoxy *share* while they argue — invisible to both sides precisely because they are arguing inside it.\n- **Žižek's unknown knowns**, and his unanswered challenge to the worldview filter, which this resolves: the interface was never the instrument, the other person is.\n- **The cross-source premise run**, 2026-08-27: the occupation premise was recoverable only from the **pair**. Inside either text it was not missing — it was assumed. Within-source passes had been running for months and could not reach it by construction.\n\n**The caveat that must travel with it**: this does not say any outside view sees your frame. It says an outside view *can*, and specifically one positioned differently **on the thing in question**. A second observer who shares your frame is a second sample of it, which is the same-hand-bias failure wearing a human face.\n\n## 4c-ii. The bias literature, which supports two of these and narrows a third (2026-08-27)\n\nChecked against the primary sources rather than assumed, in\n[cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md).\n\n**Supports the no-expert-gate position.** Myside bias has **very little relation to intelligence**,\nin both naturalistic and within-subjects paradigms — Stanovich's *dysrationalia* names rational\nthinking as a capacity distinct from the one IQ tests measure. Gating contribution on expertise\ntherefore buys nothing on the axis that matters, which is independent support for *an act's value\nis not its rung*.\n\n**Supports the principle ratified above.** Pronin's bias blind spot is not cured by awareness, and\nthe follow-up work finds it **positively related to cognitive sophistication** while *not* mediated\nby actual susceptibility — the sophisticated are not less biased, only more confident they are not.\nThat is *never ask a mind to audit its own frame* with a literature under it, and it is also why\nthe correction to act three matters: a report about one's own past reasoning is the one testimony\nthat deserves no elevated credence, so only the **present refusal to endorse** can carry the act.\n\n**Narrows what may ever be built.** The measurement literature finds correlations between bias\nmeasures are low, **suggesting no general factor of susceptibility at all**, composite scores are\nunreliable, and test–retest across seven classic tasks ranged 0 to .82. So any future proposal to\nweight contributions by a contributor's demonstrated rationality would be scoring a trait too\nunreliable to exist. **This is an empirical kill rather than a values one**, which is the more\ndurable kind: a values argument invites a counter-values argument and this does not.\n\n## 4d. The additivity principle, and the hole it does not cover\n\n*Ratified into the One-Breath principles 2026-08-27, after a first phrasing was withdrawn. The\nreasoning is recorded here so the principle can be reviewed and refined rather than inherited.*\n\n### The phrasing that was withdrawn, and why\n\nThe first proposal was **\"humans contribute objects; machines compute over them.\"** The founder's\nobjection was that it is not self-explanatory — *what is an object, and what exactly does it rule\nout?* — and pressing on the first question broke it.\n\n\"Object\" was meant as *a node or edge: something with an identity, an author and a place, that\npersists and can be pointed at*, with the contrast class being a **verdict**. That separation does\nnot survive contact with the schema. A vote **is** stored as an edge. It is an object too. So the\nword was not carrying the distinction it appeared to carry.\n\nThe real distinction is one level in: **does the contribution mean something on its own, or only\nrelative to a tally?** An absent premise means what it means whether one person adds it or a\nthousand. A vote means nothing alone and everything in aggregate. That is what decides whether\nnumbers can win, and it is the property the principle is actually about.\n\n### The hole, which the first phrasing papered over\n\nStated as *no majority semantics*, the principle claims a protection the system does not deliver.\nThe additive design removes tally-semantics from the **contribution** layer — but **the strength\nlayer is itself an aggregate**. A thousand people each attaching a supporting edge does move a\nbadge. §4.2 of [incentives-analysis.md](incentives-analysis.md) already knew this from the other\ndirection: supports are self-serving and attacks altruistic, so QBAF strength inherits the\nasymmetry.\n\nSo brigading does not land on *existence*. It lands on *weight*. A principle claiming otherwise\nwould tell a future session the problem was solved, which is the dead-instrument failure applied\nto a rule rather than to code.\n\n### The narrowed form, and what it actually rules out\n\n> **Contributions are added, never overwritten — nothing needs permission to exist.**\n\nConcretely ruled out: a ratify-to-publish gate · any surface where disagreement hides or removes a\nclaim · a merge that destroys originals (already ruled independently in the sameness work) · a\n\"consensus version\" that replaces variants · deletion-by-disagreement in any form.\n\nThe mechanism it buys is precise and worth stating in one line: **there is no single current text,\nso the edit-war dynamic has nothing to run on.** Wikipedia's own policy names why persistence wins\nthere — *\"the first person to step out of the war in essence loses, in that the article's text will\nreflect the update made in the final round\"* — and that sentence only makes sense where an article\nhas one current text. A graph that accumulates rather than replaces removes the prize.\n\nExplicitly **not** claimed: that strength is un-brigadeable. That is a separate problem with\nseparate defences — provenance, attribution, sameness-merging (so a hundred paraphrases of one\npremise stop counting as a hundred), and the `necessary` support ceiling, which is the shipped\nexample of removing a payoff rather than detecting an abuse.\n\n### For later review\n\nThree things would sharpen or falsify this. **Whether the aggregate defences hold** — none has met\nan interested party, let alone a strategic one, so the whole strength-layer half is untested\nagainst the adversary class that matters. **Whether \"nothing needs permission\" survives at scale**\n— it is stated for the contribution layer, and a moderation surface for abuse is a different\nquestion that has not been asked here. And **whether existence-without-weight is honest to a\ncontributor**: a claim that exists but sits in the gray band forever may feel like permission\ndenied by other means, which is a UX question the principle does not answer.\n\n## 5. What would be built, if this is ruled\n\nEach is small, and each is a place to put a human rather than a button to press:\n\n1. **An absence prompt** on any debate cluster — *\"what argument is missing here?\"* — landing as a first-class claim with its own attribution, never as a flag on someone else's.\n2. **Frame as a representable object**, which the taxonomy already names as the cheapest first move and a decision rather than code: run the sorting test on \"decomposition frame\", and if two people can disagree about a frame while the system keeps working, it belongs in the graph — with frame-level supersession as edge-supersession's sibling.\n3. **An author channel** for *that is not what I meant*, wired to always-mint so the correction sits beside the rendering it corrects.\n4. **A stake field** on claims about policies with recipients — testimony, attributed, never weighed as evidence for the empirical question it accompanies.\n\n**The measurement that would tell us it worked** is not participation counts, which the engagement gradient says will be steep no matter what. It is whether contributions of these four kinds are ones **no machine pass had already produced** — independence, which is the whole reason to ask a person at all.\n\n## Sources\n\n- [arXiv:2605.04070](https://arxiv.org/abs/2605.04070) *Toward Human-AI Complementarity Across Diverse Tasks*; [arXiv:2510.26518](https://arxiv.org/abs/2510.26518) *Human-AI Complementarity: A Goal for Amplified Oversight* — complementarity is hard, gains are modest, and the conditions that predict it are humans-outperforming-alone and active interrogation rather than passive acceptance.\n- Bruckman (Georgia Tech) on the participation asymmetry in point-of-view pushing; Wikipedia's own edit-warring policy for the single-state mechanism (*\"the first person to step out of the war in essence loses\"*).\n\n## See also\n\n[the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) (the eight classes and the triage this rests on) · [engagement-gradient-priors.md](engagement-gradient-priors.md) (why a low-rung act matters more than a high-value one) · [parity-and-the-zone-interior.md](parity-and-the-zone-interior.md) (experience reports as observer-privileged) · [analysis-and-attunement.md](../analysis-and-attunement.md) (the pole act four belongs to) · [self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md) (always-mint, which act three depends on) · [steadying-an-outside-judge.md](steadying-an-outside-judge.md)\n"}