{"path":"research/the-harness-conviction.md","content":"# The Harness Conviction\n\n**Date**: 2026-08-30 · **Type**: founder conviction (verbatim, standing) + analysis + the analogy\nworkshop. **Entered The Convictions in One Breath the same day** — proposed under the\nwrong-by-default filter at the demonstrated grade (a session asserted the delegation-is-hazard\nopposite twice), ratified by the founder with one amendment, verbatim: *\"Cool - but with the two\nhuman-reserved classes named in the line, perhaps?\"* — so the canon line names coherent absence and\nframe lock-in explicitly. Every mapping and proposal below is individually marked ruled or unruled; the conviction\nitself is the founder's and needs no ruling.\n\n---\n\n## 0. Plain language first\n\nThe question this document holds: **what makes machine reasoning good, and what is Deliberus's role\nin that?** The founder's answer: the harness. A language model's quality profile — sycophancy,\nwiggling, bias, hallucination, forgotten premises at one end; grounded, unfazed, well-read reasoning\nat the other — is not fixed by the model alone. It is a function of the tools, records and\ninstruments the model reasons *through*. Deliberus's graph, ontology and instrument suite are that\nharness, and the claim is architecture-independent: it does not wait for a post-transformer\nbreakthrough, because the remedy operates outside the model.\n\n---\n\n## 1. The conviction, verbatim\n\nThree messages, 2026-08-30, prompted by a session glossing the interpassivity warning\n(`ux-principles.md` anti-pattern 9 / P14) in the delegation-is-the-hazard frame:\n\n> \"If I outsource food shopping and meal planning for a conference of 1000 ppl to a master chef with\n> a calculator and a spreadsheet, then it'll turn out way better than if I task it to an unaided\n> beginner chef. The quality of the LLM matters, the harness matters and the tools and structure of\n> the ontology graph and instruments etc that Deliberus exposes and wraps to the LLM as it uses\n> Deliberus as a tool/MCP or whatever, all of this matters deeply.\n>\n> It makes a real difference to whether we can expect sycophancy, wiggling, bias, hallucinations,\n> sloppy analogies and forgotten premises and just general AI slop... vs stringent, well-balanced,\n> wise, well-read, steely-eyed, epistemically unfazed, always-grounded, near-perfect reasoners based\n> on the best philosophy and science and the most accurate priors, evidence, inferences and\n> analogies that humanity can muster. Deliberus can ultimately get us closer to this, I'm convinced\n> of it.\"\n\n> \"Today's LLMs work a lot like biological brains. They inherited their flaws in some ways. They're\n> very messy and need help. Some of it will doubtlessly be about the internal architecture of\n> AIs/LLMs - perhaps we need something other than transformers eventually etc. But I believe\n> something like Deliberus can soon make a transformative difference even without such leaps.\"\n\nAnd the follow-up objection that decomposed one of the adjectives (§ 6):\n\n> \"'Well-balanced' well that one is tricky because there's often not just two perspectives on an\n> issue/topic... There could be like ten different ones, or one for each sentient being who ever\n> interacted/related with some part of reality that's being discussed.\"\n\n---\n\n## 2. Where it came from: the delegation correction's second statement\n\nThe gloss that prompted this quoted Žižek (\"systems that do cognitive work FOR users rob them...\")\nand Stanford's unguided-LLM finding as if delegation itself were the hazard — the exact frame the\n**delegation clarification** ([conviction-and-critique.md](conviction-and-critique.md) § The\ndelegation clarification) had already rejected and marked *do not regress*: offloading is the\npurpose; the hazard is offloading **without self-report**. The regression is logged there as\nevidence about the correction's stickiness, and P14's body — which still leads with the rejected\nframe — is where the next regression will come from until it is rewritten (proposed, unruled; the\nreframe consequences are recorded in `ux-principles.md` § Naming note).\n\n---\n\n## 3. What this adds to the March statement\n\nThe March 2026 conviction ([agents-as-a-consumer-class.md](agents-as-a-consumer-class.md) § 2) said\nagents can **check their conclusions** through the graph. This says more: **the whole quality\nprofile of machine reasoning is a function of harness and grounding, and Deliberus is that\nharness.** The scaffolded party is no longer only the human. The same structure that holds a\nperson's hand through a descent (the Vygotsky sense in `depth.md`) holds a model's: addressable\npremises, typed relations, provenance, strength computed deterministically rather than vibed, known\nmoves visible as known. And the second message makes the claim **architecture-independent**.\n\n---\n\n## 4. The law underneath: remove the constraint, don't repair the reasoner\n\nThe founder's chef varies the wrong variable — master versus beginner quietly argues *buy a better\nmodel*, while the thesis holds the reasoner fixed and varies the equipment (§ 8 replaces it). But\nthe analogy's core survives the diagnosis: **the calculator does not make the chef wise; it deletes\nan error class.** A harness is a stack of deleted error classes, and \"steel\" is slop minus the\ndeleted classes.\n\nThis is the cognitive-bias codex's own conclusion transferred to machines\n([cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md)):\nfor humans, changing the environment beats training the mind; for models, harness interventions\nbeat prompt exhortation — and for the same reason, which the founder's second message names. The\nbiological-brains parallel is not an analogy but a **common cause**: models trained on human output\ninherit human failure statistics, and approval-training selects for sycophancy the way social life\nselects for conformity. Wright's own market-selection thesis\n([wright-god-test-noosphere-and-deliberus.md](wright-god-test-noosphere-and-deliberus.md)) says the\nmarket *demands* approval-currying agents — slop is what selection produces, so the counterweight\ncannot be the model improving itself. It must be external and measured.\n\nAnd the precedent for *transformative without changing the substrate* is the success prior's whole\nlineage ([fractal-priors-for-success.md](fractal-priors-for-success.md)): writing, double-entry,\nstatistics, peer review — every prior externalization of reasoning was transformative **on unchanged\nwetware**. Nobody upgraded the brain; they built the notation. The same move is available for\ntransformers now.\n\n---\n\n## 5. The adjective list decomposes into the instrument inventory\n\nEach word in the founder's list names a failure class with a mechanism already built or named —\nwhich is what makes the conviction a design programme rather than a hope:\n\n| His word | Failure class | Mechanism | State |\n|---|---|---|---|\n| \"epistemically unfazed\" / no wiggling | verdict flips under conversational pressure (25–71% measured across frontier judges) | verdict-as-record: challenges become claims + typed edges, response computed by deterministic QBAF | **live** — verified against code 2026-08-25, silence-condition caveat owned |\n| no sycophancy toward \"everyone agrees\" | fabricated consensus (strongest measured persuasion tactic) | always-mint + `attributed_to` turns the claim into a lookup against an attributed record | substrate live (1,082 claims); judge-steadying reads designed 2026-08-27 |\n| \"always-grounded\" | ungrounded assertion / hallucination | addressability + the synthesis citation gate (`cited_claim_ids`, hard validation) | **live** 2026-08-13 |\n| no forgotten premises | models lose their own world model mid-generation (Tower of Hanoi; recoverable by injection) | the graph as persistent external structure; implicit-premise + cross-source premise passes | persistence live; cross-source pass built, propose-only |\n| no sloppy analogies | analogical stretch passing as inference | scheme classification + the analogy scheme's own critical questions | **live** — 558 analogy-cluster edges (413 example / 112 analogy / 33 precedent) of 3,666 scheme-bearing, measured 2026-08-30 |\n| \"well-balanced\" | corpus lean read as reasoning lean | per-cluster ingestion-balance instrument; provenance labels on every count | named, **unbuilt** — and see § 6: the countable form is only one of three meanings |\n| \"best priors humanity can muster\" | evidence weighted by vibe | methodology base weights; meta-analysis decomposition | live but partly **inert** — `heterogeneity` and `study_count` weigh nothing (measured) |\n\n---\n\n## 6. \"Well-balanced\" decomposed — the founder's own objection sharpens the weakest row\n\nThree meanings that must not be conflated:\n\n- **(a) Symmetric ingestion over an enumerable partition** — the countable instrument, valid only\n  where sides exist (a pro/con pair). Crude by design, entropy-class. This is all the table row\n  above refers to.\n- **(b) No perspective silently absent** — the founder's case: ten perspectives, or one per sentient\n  being. Uncountable in principle. This is **coherent absence arriving in balance clothing**, the\n  human-reserved error class, and the act that answers it (*nobody here has said X*) is low-rung.\n  The design answer is attribution plus visible-lean-as-invitation (the founder's 2026-08-21\n  ruling), never averaging.\n- **(c) A balanced verdict** — refused outright: the middle is not automatically wise (the DeepMind\n  steering result), so the graph shows the lean and treats it as an invitation.\n\nThe design never averages perspectives into balance; it holds them attributed and makes the gaps\nvisible. The objection strengthens the ceiling argument (§ 10) rather than weakening the row.\n\n---\n\n## 7. The check that must be able to fail\n\nThe check is **the human confirm-click on machine proposals**. An exam where every answer is marked\ncorrect is not an exam. Current state, measured this week: the codebase has five mechanisms that\n*propose* (terminus classifier, stance conflicts, staleness daemon, cross-source premises, claim\ndecomposition) and **no action that ratifies** — `confirmed=true` is the default on human-authored\nedges, so it means *a human wrote this*, never *a human checked this*\n([the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) class 6). And the measured human\nbehaviour is worse: people who **detect** a machine error follow its advice anyway, 80.7% of the\ntime (the cognitive-commons detect-but-defer result). \"Able to fail\" concretely: *no* must be a\nreal, cheap, sometimes-taken path — verified by occasionally seeding a known-flawed proposal and\nmeasuring the catch rate. A catch rate near zero means the gate is decoration.\n\nThe coupling with the conviction: **the better the harness makes the machine, the more load-bearing\nthe failable human check becomes** — good output is exactly what unstructured confirmation\nrubber-stamps. This is the surviving core of the anti-pattern-9 rewrite (Feynman: you must be able\nto tell having-thought from having-agreed).\n\n---\n\n## 8. The analogy workshop\n\n### 8.1 Why the chef fails as a vehicle\n\nIt varies the **reasoner** (master vs beginner), so it accidentally argues *buy a better model*. The\nthesis needs the reasoner held fixed and the **equipment** varied. All three candidates below do\nthat.\n\n### 8.2 The instrument-rated pilot — pedagogy repaired\n\n*Founder: \"I have no clue what the graveyard spiral is so maybe not the best pedagogical analogy\" —\ncorrect; the term is pilot-insider jargon and is dropped.* The self-decoding version: **pilots who\nfly into cloud without instruments lose control of the plane in about three minutes** (the classic\n\"178 seconds\" studies) — because the inner ear cannot tell a steady turn from level flight, so the\npilot feels level while turning toward the ground. The panel does not make the pilot smarter; it\n**deletes that error class**. And the instrument *rating* is trained trust of the record over the\nfeeling — which is the anti-wiggle discipline exactly: when in-context pressure says flip, the\nrecord says stand. Yields the phrase **instrument-rated reasoner** for what the harness makes of a\nmodel. Limits: no adversary in the picture, and physics supplies an oracle that values lack.\n\n### 8.3 The navigator's shared chart, in depth\n\nA nautical chart is a **cumulative, shared, correctable record of a territory**. Its properties map\none to one:\n\n- **Soundings are claims with provenance and a date.** Modern charts print a per-region reliability\n  grade (zones of confidence) and the survey date — provenance labelling on every count, in\n  production, on paper.\n- **Unsurveyed water is left blank and labelled** — never filled with invented depths. That is\n  *conservative relevance beats honest thinness* as cartographic practice: the chart confesses its\n  gaps rather than hallucinating coverage.\n- **Identity is free.** Coordinates decide whether two soundings are \"the same place,\" which is why\n  the chart's reuse flywheel ran frictionlessly for centuries — the condition\n  [the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) names, satisfied by geometry. Claims\n  lack a coordinate system; identity is judged. This is the chart's deepest disanalogy.\n- **Corrections are pushed, not just published.** Notices to Mariners go out weekly, and ships'\n  masters are *obligated* to report new dangers (SOLAS). The graph has the correction machinery\n  (supersession, the correction pipeline) but **no push channel to those who relied on a\n  now-corrected claim** — the load-bearing version needs a reliance record, which trades against\n  privacy. Named as a gap, unruled.\n- **The seabed shifts.** Charts decay physically (silting, moving sandbars), which is why the\n  correction cycle exists — the temporal rung's staleness problem, solved institutionally.\n- **Charts were strategic secrets.** Iberian and Dutch chart-secrecy is the historical form of\n  selective legibility: publishing your waters selectively. The chart is not adversary-naive.\n- **The check is external.** A wrong sounding is discovered by grounding — reality is the oracle,\n  and the chart has no *internal* consistency check. Two contradictory depths at one coordinate is\n  simply an error; the chart cannot hold disagreement as content.\n\n### 8.4 The ledger with audit, in depth\n\nDouble-entry bookkeeping records every transaction twice, so **the books must balance — an\nimbalance proves an error exists** (not where, but that). Its properties map to different parts of\nthe system:\n\n- **Imposed coherence, made computable.** \"The books must balance\" is the architecture bet five\n  centuries early: a single-entry list of what happened needs an editor to *notice* an\n  inconsistency; double-entry makes inconsistency **compute** as an imbalance — exactly the\n  wiki-versus-graph distinction\n  ([world-models-and-the-ontology-revival.md](world-models-and-the-ontology-revival.md)).\n- **Balance catches inconsistency, never omission.** A wholly omitted transaction leaves the books\n  balanced — coherent absence, in accounting clothing. Which is why **audit** exists: vouching\n  entries against external documents, and *confirmations* — writing to the counterparty and asking\n  directly. A confirmation letter is the consent channel: the one party who cannot be outvoted about\n  their own position, asked directly.\n- **Reconciliation is always-mint.** In commerce, my payable is your receivable; honest books\n  mirror, and reconciling my records against your statement finds the divergence. That is the\n  straw-manning detector in ledger language: one side's rendering of the other, compared against\n  what the other actually asserts.\n- **Segregation of duties is the same-hand answer.** No single person both authorizes and records a\n  transaction; no single model family should both generate and check (the two-family jury).\n- **Materiality is risk-weighted attention.** Audits do not check everything; effort goes where\n  misstatement risk is highest — the same shape as hinge-ordered questioning and `worth_asking`.\n- **Balanced fraud exists.** Fictitious revenue with fictitious receivables balances perfectly; the\n  ledger's history includes the full strategy-class arms race (internal control → external audit →\n  forensic accounting), and **auditor capture** (Enron/Andersen) is Porter's dark rider measured:\n  the checking institution captured by the checked. The ledger is the analogy that has met every\n  adversary class.\n- **The enabling invariant does not transfer.** Double-entry works because exchange conserves value\n  — the accounting identity forces the balance. Reasoning has no conserved quantity, so \"the books\n  must balance\" transfers as a *shape* (make inconsistency compute), never as arithmetic.\n\n### 8.5 Compare and contrast\n\n| Axis | The shared chart | The ledger with audit |\n|---|---|---|\n| What it records | a shared territory — world-facts | one's own commitments and transactions |\n| Identity | **free** (coordinates) | cheap (date + parties + amount) |\n| The check | external — reality grounds you; re-survey | **internal** (the balance) + institutional (the audit) |\n| What it catches | position error, hazards | inconsistency — **not omission** (audit covers absence) |\n| Cumulative across reasoners | **yes — the whole point** | classically no: each firm its own books; sharedness only at interfaces (reconciliation, confirmations) |\n| Adversary history | mostly entropy (seabed shifts) + chart secrecy | **full spectrum** including strategy-class fraud and auditor capture |\n| Oracle | the territory (empirical, unavailable for values) | conservation of value (formal, unavailable for reasoning) |\n| Decay handling | correction cycles, pushed to users | scheduled restatement (depreciation, impairment) |\n\n**The synthesis, and it is the point**: the chart solves *many minds, one record* without a native\ncheck; the ledger solves *a record that checks itself* without being shared. **Deliberus needs both\nhalves — the chart's sociality with the ledger's checkability.** And the encouraging historical\nfact: each institution grew the half it lacked. Charts grew checkability (reliability grades,\nsurvey dates, mandatory correction reporting); ledgers grew sharedness at every interface\n(reconciliation, confirmations, consolidated audit). Two artifacts converging on the same target\nfrom opposite ends — a shared, self-checking record — which is itself a small two-instance fractal\nprior for the thing Deliberus is.\n\n### 8.5b The identity disanalogy, founder-corrected (2026-08-30): claims DO have coordinates\n\n§ 8.3 said *claims lack a coordinate system; identity is judged*. The founder pushed back, verbatim:\n\n> \"Even though in Deliberus' case the relation between territory/reality and map/graph is much less\n> obviously objectively mappable from any kind of geometry, I think there are regularities and\n> dimensions to track. At the very least we can try to use vector embeddings for all claims and\n> contested-concepts... and any claims tagged with any contested-concept (which should be\n> auto-tagged, is that done?) ofc have a unique combination of meaning-vector-coordinates. The\n> territory and geometry mapped itself is created and grown and will continue to evolve based on\n> investments in and population of the Deliberus graph/DB. So it's more complex than Google Maps,\n> even though features on a map are based on intersubjective agreement as well, and maybe sometimes\n> editing wars occur there as well, not just on wikipedia.\"\n\nHe is right, and the corrected statement is sharper: **claims have coordinates; what the chart has\nand the graph lacks is a metric where distance IS the decision.** Nearby soundings are the same\nplace; nearby embeddings are only the same vocabulary (the measured cosine ceiling — same-topic\nyes, same-kind-of-argument no, and a low cosine is weak evidence of unrelatedness). Three coordinate\nlayers, in rising interpretability:\n\n1. **Raw embeddings** — continuous, cheap, axes unnamed; the discovery layer, never authority on\n   meaning (the standing embeddings-tension ruling). Coverage measured 2026-08-30: 3,582 of 4,852\n   claims carry embeddings.\n2. **Concept–sense tags** — the founder's proposal, and the interesting layer: a claim's location is\n   *which contested concepts it uses, in which senses* — discrete, **named, challengeable axes**,\n   unlike embedding dimensions. Partially built: the extraction pipeline auto-tags at extraction\n   time with the sense recorded on the edge (`USES_CONCEPT`, `r.sense`), measured 465 of 1,741\n   substantive claims tagged. **No corpus-wide backfill exists, and human-minted claims from the\n   correction UX are never tagged** — a backfill is a committable-tier daemon job by the daemon\n   taxonomy (same class as embeddings and attribution normalisation). **RULED 2026-08-30** —\n   founder: *\"Concept tagging should be corpus-wide backfilled and also any new claim regardless\n   where from should be auto-tagged... we shouldn't implement it this session, but soon.\"* Spec in\n   `TODO.md` § RULED 2026-08-30.\n3. **Structural fields** — scheme, claim kind, terminus, attribution: the mathlib direction already\n   ruled (the more of identity is structural, the more decidable matching becomes).\n\nTwo further corrections his message carries. **The territory here is constitutive** — mapping\ncreates addressable objects, so the coordinate system itself grows with the graph, unlike geography.\nAnd **even geography's \"free identity\" frays exactly where stakes appear**: OpenStreetMap has edit\nwars over contested borders and names, and Google Maps shows different national borders to\ndifferent countries. So the sharpened law, which improves § 8.3 and the reuse-flywheel condition:\n**identity is cheap where nobody has a stake; stakes make identity judged — in cartography too.**\nThe chart's centuries of frictionless reuse ran on waters nobody disputed; the flywheel condition\nis not *geometry versus meaning* but *unstaked versus staked*.\n\n### 8.5c Naming RULED same day: motivated stakeholder (founder: *\"'Motivated stakeholder' it is.\"*)\n\nThe adversary-taxonomy's middle class — an agent with a stake who does not model the detector — is\ncalled *motivated-stakeholder-class*, and the calibration sentence reads *no instrument here has ever met an\nmotivated stakeholder*. The founder: the term is not strong enough. Candidates on the table, with the\ntwo Claude-runnable headline tests applied: **motivated stakeholder** (collision-AGREES — the\ncorpus already uses *motivated* for exactly this class: motivated stopping, motivated cognition;\nimplies agent + stake without implying scheming, which stays strategy-class); **skin-in-the-game**\n(strongest plain-language register; better as the first-use gloss than as a class name);\n**invested incumbent** (rejected — *incumbent* wrongly excludes a newcomer with a stake; the class\nis about having a stake, not holding a position). Rename ripple measured: 17 mentions across 15\ndocs plus the CLAUDE.md threat-model quotes. Ruled the same day and the\ncorpus-wide rename executed; the ruling record lives at the taxonomy's definition site\n([fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md) § Three classes).\n\n### 8.6 Which analogy for which register (unruled — founder picks)\n\nThe **pilot** for the individual slop-versus-steel sentence, with the 178-seconds pedagogy. The\n**chart** for the commons and civilizational register. The **ledger** for the honesty, audit and\nfunder register — it is already the success prior's own lineage, and it is the one that has met an\nmotivated stakeholder.\n\n---\n\n## 9. The meta-closure: this comparison is itself the object under discussion\n\nThe chart–ledger comparison in § 8.5 is a two-instance structural analogy carrying a mechanism claim\n(*same bottleneck: many reasoners, one record, errors must surface*). Its critical questions are\naskable — the counter-analogy question becomes *is there a shared record that stayed uncheckable and\nthrived?* (open) — and it is representable today as an analogy-expressing mother claim with\nper-instance subclaims, per the 2026-08-30 mapping in\n[convictions-across-scales.md](convictions-across-scales.md): **the four-check transfer test is a\nregister-specific enrichment of the shipped analogy scheme's critical questions**, so a fractal\nprior stored in the graph gets a badge that reads out *how well-evidenced is this prior*. The\nfounder's question that opened this (*\"Is there an analogy scheme??\"*) has a verified yes: shipped,\nwith `example` and `precedent` as cluster siblings, 558 live edges.\n\n---\n\n## 10. Bounds and the ceiling\n\nFour bounds keep the conviction falsifiable, all already named in the corpus: **same-hand bias**\n(the checkers run on the model class they audit; cross-family jury buildable); **the harness is\nitself LLM passes** (steering enters upstream of the record, and the instruments measure collapse,\nnot amplification); **correlated errors** (stacking same-family models is one detector sampled\ntwice); and **\"near-perfect\" cashes out per layer** — oracle-true only where the terminus reaches\nground; on the evaluative layer the achievable maximum is *fully legible*, never *correct*.\n\nOne proposal, unruled: the listings-board worry (many agents reading one graph correlate by\nconstruction) has a legibility answer — **provenance makes the correlation inspectable**, so an\nauditor can tell correlated-because-same-slop from correlated-because-same-checked-record.\n\n**The ceiling composes with the human-role analysis.** At harness-limit the machine still cannot\nknow **what its corpus is a sample of** (coherent absence) or **stand outside the frame the graph\ncrystallized** (frame lock-in) — precisely the two error classes\n[what-human-judgment-is-for.md](what-human-judgment-is-for.md) reserves for humans. Harness the\nmachine to steel, and the residue of the human role is exactly what no harness reaches.\n\n---\n\n## 11. Sequencing consequence\n\nThe conviction raises the priority of the read-only MCP surface and does not touch its gate: the\nstrength layer must be honest first (the two accidentally-satisfied blockers — `count_safe_summary`\nuncalled, `REPORTS` non-propagating by accident), and the four cheap audit closures from the\npre-launch audit are the discharge path: wire the honest counter, add positive controls, publish the\ncorpus-composition sentence, build the two-family jury. The nearest executable step of\n\"near-perfect reasoners\" is wiring one function and adding positive controls.\n\n---\n\n## Cross-references\n\n[agents-as-a-consumer-class.md](agents-as-a-consumer-class.md) (§ 12 stub; the March statement, the\nMCP shape, the gate) · [conviction-and-critique.md](conviction-and-critique.md) (the delegation\nclarification and the logged regression) · [convictions-across-scales.md](convictions-across-scales.md)\n(the four-check test as analogy-scheme CQ enrichment; fractal priors as mother claims) ·\n[cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md)\n(remove the constraint, don't repair the reasoner) ·\n[fractal-priors-for-success.md](fractal-priors-for-success.md) (the unchanged-wetware lineage) ·\n[the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) (ratification decay; the audit) ·\n[fragile-checkers-and-the-verification-bottleneck.md](fragile-checkers-and-the-verification-bottleneck.md)\n(judge wiggle; verdict-as-record) · [the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md)\n(identity must be cheap to decide — the chart satisfies it by geometry) ·\n[what-human-judgment-is-for.md](what-human-judgment-is-for.md) (the two human-reserved classes) ·\n[wright-god-test-noosphere-and-deliberus.md](wright-god-test-noosphere-and-deliberus.md) (market\nselection for approval-currying) · [world-models-and-the-ontology-revival.md](world-models-and-the-ontology-revival.md)\n(imposed coherence; the wiki test) · [kulturkapital-and-the-unscalable.md](kulturkapital-and-the-unscalable.md)\n(the same partition at organisational scale) · [reflection-and-the-graph.md](reflection-and-the-graph.md)\n(Echo as the sycophant; Narcissus as same-hand bias) · [ux-principles.md](../ux-principles.md) (anti-pattern 9, the\nnaming note, P14)\n"}