{"path":"research/constellation-candidates.md","content":"# Constellation Candidates: Unmined Convictions and Principles\n\n**Date**: 2026-08-19 · **Status**: Mining deliverable, uncommitted, awaiting founder review.\n**Task**: Surface convictions/principles held as commitments in the corpus but NOT yet canonized in `CLAUDE.md` § The Convictions/Principles in One Breath, as candidates for a landing-page constellation of 5–8-word bold tiles.\n**Method**: Full read of the seven concept docs, ux-principles (P1–P23), the-missing-layer, and 14 research docs (self-similar-decomposition, lowering-the-cost, incentives-analysis, curiosity-as-growth-fuel, islands-of-coherence, the-load-bearing-unsaid, conviction-and-critique, weighing-scheme design, convergence-wager synthesis, structure-versus-scale, engagement-gradient-priors, dogfood-run-1, decomposition-axes) plus `deliberus/agent_readable.py`'s docstring. Each candidate passed three gates: held as a commitment (founder-ratified or repeatedly load-bearing), distinctive to Deliberus, and not on or inside the exclusion list.\n\nCandidates are grouped convictions-first, then principles. Where a candidate's ratification status is partial, it is flagged rather than hidden.\n\n---\n\n## Convictions (beliefs about the world, truth-apt)\n\n### 1. Trust the process, not the people\n\n**Position**: Every trusted proxy — journalist, expert, politician — can be corrupted; transparent, inspectable reasoning structure removes the need to trust the person at all. Trust migrates from people to process.\n**Home**: `civilizational-vision.md` § Trust Shifts from People to Process; `vision.md` Key Quotes — *\"Focus on trusting the process instead of the people.\"* (Fredrik, 2012, article outline — \"the earliest formulation of trustless collaboration\").\n**Distinctive**: Fourteen years old, in the founder's own words, load-bearing across vision, civilizational-vision, and the Lean/trustless-collaboration analogy (Tao's 25 strangers needed no mutual trust because the kernel verified the join). It is the project's answer to institutional trust collapse and is stated nowhere in the canonized lists.\n**Type**: Both — a belief about where trust can safely live, and a design policy.\n**Tile**: **Trust the process, not the people.**\n\n### 2. A new medium, not a better tool\n\n**Position**: Moving discourse granularity from paragraph-level debate to premise-level structured argumentation is a change of medium comparable to the advent of language — language enabled fine-grained communication between minds; this enables fine-grained *reasoning* between minds.\n**Home**: `vision.md` § A New Medium, Not Just a Better Tool — *\"could be likened to the advent of language itself\"* (2012 Leverage notes, called \"the strongest framing of the mission anywhere in the project\" in `civilizational-vision.md`).\n**Distinctive**: Reframes the project category entirely; no debate-tool competitor claims medium-status, and the corpus repeatedly leans on it (evolutionary-transition sections in two docs).\n**Type**: Conviction.\n**Tile**: **Fine-grained reasoning is a new medium.**\n\n### 3. Definitions are moves, not metadata\n\n**Position**: A primary source of real-world disagreement is the same word carrying different concepts (\"freedom\", \"barnfattigdom\"); therefore a definition is a first-class move *in* the deliberation, not annotation around it, and the \"oh — THAT's why we disagree\" recognition is among the most powerful moments the platform offers.\n**Home**: `vision.md` § Semantic Disambiguation as Root Cause; `ux-principles.md` P18 — *\"Definitional claims are themselves arguments — 'by freedom I mean X' is a MOVE in the deliberation, not metadata.\"*; `worldview-lenses.md` (the libertarian/social-democrat freedom case: 80% shared structure, one definitional premise).\n**Distinctive**: Every predecessor treats definitions as glossary; here they are nodes with strength, votes, and lifecycle states. Repeatedly load-bearing since the 2013 concept-tracking sketch.\n**Type**: Both — the empirical claim (disagreement is substantially semantic) plus the design policy.\n**Tile**: **Definitions are moves, not metadata.**\n\n### 4. Worldviews divide us, not our hearts\n\n**Position**: The human heart and conscience are highly concordant once you trace where positions come from; what differs across people is worldviews, not care. (With the founder's own trauma-not-DNA caveat pre-empting the psychopathy rebuttal.)\n**Home**: `convergence.md`, founder verbatim in Swedish (Apr 30, 2026, PauseAI Sweden): *\"det mesta som skiljer våra värderingar åt är våra världsbilder snarare än våra hjärtan.\"*\n**Distinctive**: This is the *heart-level* ground under the convergence wager, in the founder's own voice — warmer and prior to the technical wager, and absent from the canonized line (which carries only the bet-can-lose accounting).\n**Type**: Conviction.\n**Tile**: **Worldviews divide us, not our hearts.**\n\n### 5. Understanding is the infrastructure of love\n\n**Position**: The founding image (pre-2011): put on glasses you'd never choose and see that the person you've argued with for years reasons with integrity. \"I see you\" precedes and outranks \"I agree\"; *\"respect at scale is the closest thing to love that infrastructure can produce\"* — and the bridging signal is the computational trace of that moment. Tolerance-as-containment (\"I tolerate your ignorance\") is explicitly rejected.\n**Home**: `civilizational-vision.md` § Love Through Understanding; `worldview-lenses.md` § The \"Rose-Tinted Glasses\" UX.\n**Distinctive**: No epistemics platform claims love as an output; the corpus does, carefully, with a precise mechanism (bridging, Procaccia/Konya under live conflict) rather than sentiment.\n**Type**: Conviction.\n**Tile**: **Understanding is the infrastructure of love.**\n\n### 6. The real crux is rarely written\n\n**Position**: What is most load-bearing in an argument is least likely to appear in it — five independent literatures (enthymeme, presupposition, doxa, hinges, tacit knowledge) arrive at the same gradient, and the dogfood record confirms it: the crux of a real debate is stated on one side and implicit on the other (run 3F), or unwritten on both (run 6). An extraction system that reads only what is written is structurally biased against the most important content.\n**Home**: `the-load-bearing-unsaid.md` — *\"what is most load-bearing in an argument is least likely to appear in it\"*; `frontier-extraction-experiment.md` H-findings; `CLAUDE.md` ontology-frontier baseline.\n**Distinctive**: Turns implicit-premise recall into the pipeline's most important quality axis — a claim about arguments-in-the-wild that no argumentation platform has made operational.\n**Type**: Conviction.\n**Tile**: **The real crux is rarely written.**\n\n### 7. Bedrock is proven by moves tried\n\n**Position**: The only thing distinguishing genuine bedrock from arrested inquiry (dogma, thought-terminating cliché) is the record of decompositions attempted and found wanting — so a terminus verdict without an effort certificate is indistinguishable from a refusal to think. Corollary from the unsaid research: a pseudo-sacred value yields to decomposition, a genuine hinge does not, *and which one you have is discovered by trying*.\n**Home**: `weighing-scheme-and-terminus-classifier-design.md` (deep-descent worked answer: *\"The only thing distinguishing the two is the record of moves tried\"*); `depth.md` § The Truly Ineffable (terminus = repeatedly-tried-and-withdrawn, the Mathlib deprecation pattern); `the-load-bearing-unsaid.md`.\n**Status flag**: The moves-tried metadata is design-ratified in spirit but unbuilt; the surrounding sacredness-brake re-scope (\"consent-and-timing, never epistemic exemption\") is founder-opened deliberation, NOT ratified — noted here, not claimed.\n**Distinctive**: An epistemology of stopping-points no other system has: the end of inquiry is earned, evidenced, and reopenable.\n**Type**: Both.\n**Tile**: **Bedrock is proven by moves tried.**\n\n### 8. Vagueness is a defensive asset\n\n**Position**: LLMs solved the labor cost of structure; what now binds is exposure — every premise you state is a surface someone can attack, so staying vague is rational armor, and legibility asks for its surrender. The largest remaining cost is social: being *seen* understanding the other side reads as defection in a polarized setting.\n**Home**: `incentives-analysis.md` §1 — *\"the corpus has never named vagueness as a defensive asset whose surrender is what we are asking for\"*; `lowering-the-cost.md` §1 (the social term as \"untouched, and probably the largest\").\n**Distinctive**: A hard-eyed incentive claim that reframes the whole adoption problem post-LLM; nobody else in the space has moved past \"friction\" as the diagnosis.\n**Type**: Conviction.\n**Tile**: **Vagueness is a defensive asset.**\n\n### 9. Supports are selfish; attacks are gifts\n\n**Position**: Adding support to your own claim serves you; adding a strong attack on your own claim is pure altruism that hands an opponent a weapon — so contribution systematically over-collects supports, and any strength metric computed from both inherits an invisible inflation. Registered as a prediction about *human* contribution against the measured machine baseline (1.17:1).\n**Home**: `incentives-analysis.md` §4.2 — *\"strength is systematically inflated, and the inflation is invisible because it looks like well-supported reasoning\"*; corrected/calibrated in `lowering-the-cost.md` §6.4; run 5's zero attack edges as first datapoint.\n**Distinctive**: A falsifiable sociological law about argument graphs, with its own early-warning instrument (support-to-attack ratio published beside strength).\n**Type**: Conviction.\n**Tile**: **Supports are selfish; attacks are gifts.**\n\n### 10. Unstructured reasoning gets summarized by others\n\n**Position**: In an AI-mediated information environment, whoever's reasoning is structured gets represented by their own reasoning; whoever's is not gets represented by someone else's summary of it. Structured reasoning is distribution — the agent-readable surface is a channel, not a courtesy.\n**Home**: `the-missing-layer.md` § Honest limits; `incentives-analysis.md` L8 — *\"whoever's reasoning is structured gets represented in AI-mediated discourse, and whoever's is not gets summarized by someone else\"*; collected into `agents-as-a-consumer-class.md`.\n**Distinctive**: The self-interested, 2026-native case for contribution — \"did not exist three years ago\" — and the reason machine readers are first-class (per `agent_readable.py`: a legibility platform illegible to AI minds is a live contradiction).\n**Type**: Conviction.\n**Tile**: **Unstructured reasoning gets summarized by others.**\n\n### 11. The graph is the group\n\n**Position**: Solitary reasoning is biased by design (Mercier & Sperber — the founding academic premise), and myside bias is the one failure uncorrelated with intelligence or reflection; therefore the adversarial *structure* — diverse contributors attacking each other's premises under shared rules — must carry the load that individual open-mindedness cannot.\n**Home**: `convergence-wager-typed-residues-synthesis.md` §4 — *\"Solitary reasoning is biased; the graph is the group\"*; `academic-foundations` (Argumentation Theory of Reasoning) since Session 1.\n**Distinctive**: Inverts the rationalist self-improvement frame: the unit of epistemic virtue is the structure, not the person. Nowhere in the canonized lists despite being the founding social-epistemology premise.\n**Type**: Conviction.\n**Tile**: **The graph is the group.**\n\n### 12. Silence becomes legible against insistence\n\n**Position**: A classificatory premise (run 6: \"occupied\", 54 uses on one side, zero on the other) is invisible from inside a single text, because there it is not missing — it is assumed. One side's silence only becomes legible against the other side's insistence, so this class of load-bearing premise can only be found by comparing an opposed pair, with a counting operation and no model at all. Founder conclusion: such premises must *always* be made explicit by the pipeline.\n**Home**: `decomposition-axes.md` § The classificatory premise — *\"One side's silence is only legible against the other side's insistence\"* (measured 2026-08-14, founder-prompted).\n**Distinctive**: A genuinely new extraction epistemology: cross-source contrast as the detector for what neither author argues. Also the deep reason adversarial pairing is an ingest *rule*, not a habit.\n**Type**: Conviction (with a pipeline consequence).\n**Tile**: **Silence becomes legible against insistence.**\n\n### 13. The job is holding the unsettled\n\n**Position**: Every encyclopedia smooths disagreement away, every summary resolves it, every forum loses it — nothing in existence keeps a live open question addressable over years; *\"that is not a gap in a market; it is a gap in the species' tooling.\"* A well-stated open problem is one of the most productive artifacts a knowledge community can own (Hilbert's 23), and typed residues have different afterlives: the leaf is where reframing, analogy-transfer, scope-narrowing, and the analysis-to-attunement handoff get their exact coordinates.\n**Home**: `structure-versus-scale.md` § What to expect (*\"the most distinctive, and the likeliest to be the real thing\"*); `convergence.md` § What a residue is, and what the map is for (the four ways the map earns its keep; resolution operators beyond decomposition).\n**Distinctive**: Names the product's most distinctive *function* — not mapping agreement but preserving live disagreement in addressable form — which the canonized convergence line (compression-is-the-product) states only half of.\n**Type**: Both.\n**Tile**: **The job is holding the unsettled.**\n\n### 14. A reader who never writes is success\n\n**Position**: The unit of value is the commons, not the user: *\"a library is not justified by whether one borrower reads better\"* (founder, 2026-08-16, rejecting the uplift evaluation frame), and Romer's nonrivalry is the formal warrant — value scales with corpus coverage, not user count. So the steep engagement gradient is survivable, a never-contributing reader is a success case, and the fit measures are commons-shaped (coverage, reuse, reachability, falling descent cost, the residue map) — each still able to fail.\n**Home**: `lowering-the-cost.md` §5 + §8 boxed founder correction; `engagement-gradient-priors.md` (*\"a reader who never contributes is a SUCCESS — the library test\"*); `the-missing-layer.md` § Honest limits.\n**Distinctive**: A deliberate, founder-ratified rejection of the entire engagement-metrics paradigm AND of individual-uplift evaluation — while explicitly refusing the escape from falsifiability.\n**Type**: Both.\n**Tile**: **A reader who never writes is success.**\n\n### 15. Dissolving a distinction is not examining it\n\n**Position**: Decomposition has a named failure mode — *leveling dissolution* — where analysis erases a distinction rather than examining it (companies dissolved into \"fictitious entities\" dissolves families, nations, and rights with them), with a checkable test: *after the dissolution, does anything about what to actually do change?*\n**Home**: `nativism-thread-and-false-dissolution.md`; carried in `CLAUDE.md` Session 18/19 records; second wild instance measured.\n**Distinctive**: The project's own acid, given an internal safety rail — decomposition is the product, and here is precisely how it goes wrong. No other decomposition-forward system names its own pathology with a test.\n**Type**: Conviction (a claim about how analysis fails).\n**Tile**: **Dissolving a distinction isn't examining it.**\n\n---\n\n## Principles (policies, wisdom-apt)\n\n### 16. Incomplete arguments are invitations, not failures\n\n**Position**: The `sorry` model: an argument sketched with explicit, visible, transitive gaps is a *feature* — each gap is a contribution opportunity with a known shape, and the contribution barrier drops to zero while rigor is maintained underneath. Depth asymmetry across the graph IS information: where markers remain, the invitation stands.\n**Home**: `ux-principles.md` P4 (title verbatim: \"Incomplete Arguments Are Features, Not Bugs\"); `depth.md` (*\"the invitation stands\"* / depth-as-community-care); Tao's PFR project as the existence proof.\n**Distinctive**: The single move that dissolved the Structure-Adoption Paradox that killed twenty predecessors; distinct from the excluded no-gap-without-a-move (which governs *display*) and inversion principle (which governs *guidance*) — this is the ontological stance that incompleteness is first-class.\n**Type**: Both.\n**Tile**: **Incomplete arguments are invitations, not failures.**\n\n### 17. Absence is a state, not a score\n\n**Position**: A sentinel meaning *nobody has looked* must never enter a comparison that assumes a measurement; every display needs vocabulary for \"I have no reading here\", because a one-dimensional scale cannot abstain unless abstention is designed in. (115 of 116 claims showed confident badges with nothing answered, for five months.)\n**Home**: `ux-principles.md` P22, founder-named 2026-08-17, grounded in his own 2009–2013 four-state badge drawing; *\"structural sycophancy\"* verbatim.\n**Distinctive**: Founder-named, incident-forged, and the purest display-surface instance of the project's honesty stack — yet absent from the canonized lists.\n**Type**: Principle.\n**Tile**: **Absence is a state, not a score.**\n\n### 18. Agreement and quality never share a vote\n\n**Position**: \"Is this well-argued?\" and \"do I agree?\" are separable axes, and collapsing them into one upvote is the epistemic original sin of every social platform. The upper-right quadrant — well-reasoned but conclusion rejected — is *\"the richest territory in deliberation... where minds change\"*, and it is only detectable because the axes are kept apart (votes for stance, the CQ-derived badge for quality).\n**Home**: `ux-principles.md` P2 (\"Separate Quality from Stance — Always\") + Anti-Pattern 2; `bridging.md` § The \"I Can't Dismiss This\" Moment.\n**Distinctive**: The prerequisite for the bridging signal, held since the earliest sketches; the canonized lists carry bridging's threat-model cousin (flattening) but not this axis-separation commitment itself.\n**Type**: Principle.\n**Tile**: **Agreement and quality never share a vote.**\n\n### 19. Machine work regenerates; human work doesn't\n\n**Position**: Cleanup, re-extraction, and any destructive maintenance may delete only machine-provenance claims — human contributions survive re-extraction, orphaned rather than destroyed, because machine output is regenerable and human work is not.\n**Home**: `dogfood-run-1-friction-log.md` proposal 3 — *\"machine output is regenerable, human work is not\"* — shipped as provenance-scoped cleanup, with 12 past contributions migrated after the correction endpoints were found mis-stamping user work.\n**Distinctive**: A data-ethics rule specific to hybrid human/LLM graphs, discovered in production and enforced in code; generalizes far beyond this project but was born here.\n**Type**: Principle.\n**Tile**: **Machine work regenerates; human work doesn't.**\n\n### 20. Writing an argument searches the graph\n\n**Position**: Composing IS retrieval: as a user formulates a claim, the system surfaces existing claims, their strongest counters, and the current belief distribution — so duplication is prevented at the moment of creation rather than repaired by a dedup algorithm afterwards.\n**Home**: `ux-principles.md` P13 — *\"you don't need a dedup algorithm if the user never creates the duplicate\"*; Sketch 6 (2013): *\"accessible instantly when someone begins writing similar words.\"*\n**Distinctive**: A 2013-era founder design axiom that dissolves the deduplication problem at the UX layer; the flywheel's front door.\n**Type**: Principle.\n**Tile**: **Writing an argument searches the graph.**\n\n### 21. Keep the friction where meaning lives\n\n**Position**: Remove friction at structure-building (the LLM absorbs the structuring tax) and *preserve* it at meaning-commitment (disambiguation, CQ answering, decomposition) — the asymmetry is hormesis applied to collective reasoning, and it is the answer to interpassivity: the machine scaffolds, never replaces; embeddings suggest, humans decide meaning; lens-switching must produce discomfort, not comfortable tourism (*\"if switching lenses feels like browsing a museum, the system has failed\"*).\n**Home**: `ux-principles.md` P14 + hormesis frame; `depth.md` § The Hormesis Frame; `analysis-and-attunement.md` (*\"embeddings SUGGEST; the graph COMMITS... the discomfort of disambiguation is productive and should be facilitated, not eliminated\"*); `worldview-lenses.md` § Žižek's Warning.\n**Distinctive**: A precise answer to \"won't the LLM do the thinking for you?\" that no AI-tooling competitor has: friction is relocated, not removed, and the relocation map is principled.\n**Type**: Principle.\n**Tile**: **Keep the friction where meaning lives.**\n\n### 22. Decompose only where someone contests\n\n**Position**: \"Decompose as far as possible\" is refuted (molecular beats atomic; every verifier has its own atomicity optimum); what replaces it: decompose along an axis only when someone contests along it. A claim is never \"atomic\" or \"decomposed\" — it is undecomposed along enumerable axes, and an undecomposed axis is a marked gap, not a defect.\n**Home**: `decomposition-axes.md` §4 — the founder's own Session-9 *\"recursively decomposed (eventually, not instantly tho)\"* given evidence-backed teeth; carried into the ontology-frontier baseline.\n**Distinctive**: The mature form of the project's core mechanic — contest-driven rather than maximal decomposition — which the canonized nothing-permanently-atomic line gestures at but does not state as the operating rule with its literature warrant.\n**Type**: Principle.\n**Tile**: **Decompose only where someone contests.**\n\n### 23. Ask what honoring requires, never its price\n\n**Position**: What is taboo is scalar comparison with the profane, not structured moral reasoning: never present pricing prompts (\"how many X is one Y worth?\"); frame value-decomposition in specification language (\"what does honoring this value require here?\"), and keep analyzing-an-argument visibly distinct from endorsing-its-tradeoff. Enforced by a test (`test_no_scalar_pricing_language`).\n**Home**: `weighing-scheme-and-terminus-classifier-design.md` (three settled design constraints, founder-ratified with the shipped scheme); `convergence-wager-typed-residues-synthesis.md` §5 (the taboo-tradeoff escape route: *\"this constrains the voice of the decomposition UX, not its possibility\"*).\n**Distinctive**: A voice-level ethics of decomposition derived from the sacred-values literature, shipped and test-enforced — distinct from the (unratified) brake question it sits beside.\n**Type**: Principle.\n**Tile**: **Ask what honoring requires, never prices.**\n\n### 24. An instrument that cannot fail is decoration\n\n**Position**: Counter-instruments must retain the power to falsify the wager they guard; a classifier or metric that can only confirm is decoration. Applied everywhere: the residue map must be able to score the wager as *losing*, the fit measures must each be able to come out badly, and *\"two numbers that can genuinely fail are worth more than five that reassure.\"*\n**Home**: Standing rule quoted in `weighing-scheme-and-terminus-classifier-design.md` §2; `lowering-the-cost.md` §8.2 (the can-it-fail filter table); project threat model.\n**Distinctive**: The measurement-layer sibling of the excluded confession principle — confession is about *reporting* failure, this is about instruments retaining the *capacity* to produce it. Repeatedly load-bearing, never canonized on its own.\n**Type**: Principle.\n**Tile**: **An instrument that cannot fail is decoration.**\n\n### 25. Even the scoring rules are attackable\n\n**Position**: There is no non-circular ground for any epistemic norms — so the ranking rules themselves stand in the graph as challengeable claims. A platform whose own evaluation norms are decomposable, attackable nodes has a structural answer to \"who decided what counts as better reasoning?\" that no black-box scorer has; likewise a strength score is *\"always exactly the state of the arguments... anyone can change it by arguing back\"* — computed by published mathematics, never a model's opinion.\n**Home**: `convergence-wager-typed-residues-synthesis.md` §4 (*\"put the scoring norms in the graph as challengeable claims\"*); `the-missing-layer.md` § The claim; the shipped verdict-claim hybrid (`CLASSIFIES_TERMINUS`) as its first instance.\n**Distinctive**: Reflexivity as architecture: the system's authority is fully inside its own contest mechanism. Adjacent to but not stated by the excluded propose-never-assert (which governs machine *proposals*, not the standing of the *norms*).\n**Type**: Principle.\n**Tile**: **Even the scoring rules are attackable.**\n\n### 26. A steelman counts when they sign it\n\n**Position**: Ratification is the cheap trust primitive: \"they agree this is a fair statement of their view\" costs one click and buys more credibility than any methodological transparency. It converts unconsented extraction into consent, delivers the stop-misrepresenting-me benefit that makes contribution self-interested, and defeats selective legibility via the opponent-proposes/you-ratify design (\"advantage the confident, disadvantage the evasive\"). Cultural form: change what counts as winning — *\"I can state your view so well that you would sign it.\"*\n**Home**: `lowering-the-cost.md` §3 Trust + §3 Social; `incentives-analysis.md` §4.1, §5, §8 (*\"one mechanism, four returns\"*).\n**Status flag**: Repeatedly load-bearing across two major analyses and ranked top-3 buildable, but not yet shipped and not formally founder-ratified as a named principle.\n**Distinctive**: The ideological Turing test made into a one-click artifact; Community Notes is cited as the production precedent.\n**Type**: Principle.\n**Tile**: **A steelman counts when they sign.**\n\n### 27. Criticism binds on mechanism, never possibility\n\n**Position**: Governing method for how the project takes criticism: take it completely on mechanism, refuse it on possibility, never let it choose the objective — and check every reference class for whether its binding constraint still binds (*\"twenty argument-mapping platforms failed\"* cites a class whose constraint, input friction, has been removed). The anti-abuse tell is a count, not an introspection: if mechanism-layer fixes stop shipping while conviction stays high, the composition has died.\n**Home**: `conviction-and-critique.md` — status line: *\"Governing method.\"*\n**Distinctive**: A named epistemology of building-under-critique that the project actually runs (the audit-by-counting is documented working); its internal falsifier form — *not what would prove me wrong, but what would make me build differently* — is equally unmined.\n**Type**: Principle.\n**Tile**: **Criticism binds on mechanism, never possibility.**\n\n### 28. An idea lives through its dependent\n\n**Position**: Neither conviction nor criticism makes an idea load-bearing — an idea can be generated, survive scrutiny, be filed correctly, and still be inert (four measured instances, one violated three days after being written while citing its own lineage doc). The countermeasure: when an idea recurs, stop restating it and give it one name, one home, and one thing that depends on it — something that visibly breaks if it is false.\n**Home**: `conviction-and-critique.md` §4 (the uncollected idea); `structuring-gradient-lineage.md`; P20's postmortem note.\n**Distinctive**: A third failure mode of knowledge work that neither falsificationism nor conviction-culture names, discovered by measuring this corpus against itself.\n**Type**: Principle.\n**Tile**: **An idea lives through its dependent.**\n\n### 29. Offload everything except the confession\n\n**Position**: Offloading epistemic processing onto the substrate is the *purpose*, not the hazard — Deliberus is meant to become an ideal computational substrate absorbing as much of the idealized structuring as possible, with review always invited, never conscripted. The danger term is delegation *without self-report*: the badge was not wrong to compute on people's behalf, it was wrong to have no way of saying \"I haven't looked.\" Every fix should make delegation safer, not smaller.\n**Home**: `conviction-and-critique.md` § The delegation clarification, founder verbatim 2026-08-17: *\"Deliberus should ideally over time be made capable of offloading as much of the idealized processing and structuring and scaffolding as possible.\"*\n**Distinctive**: A founder-corrective conviction that cuts *against* the fashionable cognitive-atrophy worry — pro-delegation, with the confession channel as the sole non-negotiable — and it governs every future daemon/badge/synthesis decision.\n**Type**: Both.\n**Tile**: **Offload everything except the confession.**\n\n---\n\n## Two borderline candidates worth the founder's eye\n\nThese passed distinctiveness but sit lower on the commitment gate; included because the constellation may want one of them.\n\n### 30. A terminus carries a date\n\n**Position**: A classified terminus is fallible not only under the current move-set but under the current *evidence base* — an empirical residue dissolved in 2015 can reopen in 2030, premises age out, and strength computation is currently blind to all of it; the published residue fraction is a snapshot presented as a state.\n**Home**: `convergence.md` § A terminus is dated; `fractal-scales-and-temporal-frame.md` (the temporal rung, \"entirely unbuilt\"); `incentives-analysis.md` §4.5 (the same gap from the incentive side).\n**Why borderline**: Held as a named *gap* more than a ratified conviction; but it is repeatedly load-bearing and the tile is honest about the world, not the roadmap.\n**Type**: Conviction.\n**Tile**: **Even settled questions carry expiry dates.**\n\n### 31. Native vocabulary first; adapters never leak inward\n\n**Position**: Deliberus deliberately diverges from external standards (AIF) at exactly the points where its distinctive insights live — one node type, decomposition ≠ inference, the concept/sense layer, the maturity/terminus layer. Export adapters are a clean boundary add if interoperability is wanted; they are never a design constraint on the native representation and never leak back into the core model.\n**Home**: `deliberus/agent_readable.py` module docstring (a decision record living in code).\n**Why borderline**: Firmly decided but architectural in register; as a tile it may read too inside-baseball for a landing page.\n**Type**: Principle.\n**Tile**: **Diverge from standards where insight lives.**\n\n---\n\n## Rejected near-candidates\n\n- **\"The compression is the product even when residues never resolve\"** — hint-listed, verified, but it is *verbatim inside* the canonized convergence-wager One-Breath line. Excluded as already-canonized.\n- **Confusion-conflict dissolves; interest-conflict only clarifies** — hint-listed, verified in `vision.md`, but it is one of the two guards carried *inside* the canonized opacity-is-a-cost line. Already-canonized.\n- **One graph, staged maturity** — hint-listed, but canonized inside the capture-immediately-publish-deliberately line (\"one graph, staged maturity, never a separate draft universe\"). The Aug-18 membrane finding (\"share the code, compartmentalise the contents\") is a fresh deepening of it, not a new commitment.\n- **Consent-and-timing, never epistemic exemption** — hint-listed, but the corpus explicitly marks it *\"recorded as deliberation, NOT a decision\"* (brake re-scope, Jul 10). Folded as a status note under candidate 7 rather than tiled; tiling an unratified position would misstate the record.\n- **Voice on the node** — hint-listed; a `[new]` proposal in `lowering-the-cost.md`, \"the only concrete mechanism this corpus has ever proposed\" for the attunement pole — but a proposal, not a commitment. Watch it: if ratified it is an obvious tile.\n- **Mint-time dedup carve** — hint-listed and genuinely decided (system-minted text is identity-decidable by construction, below the claim-sameness fork), but implementation-registered rather than constellation-shaped; too technical for a tile.\n- **Reuse must match structurally, never by identity** — hint-listed; real (run 6 + assembly-theory doc), but it now lives adjacent to the canonized assembly-ceiling/shared-code line (\"claim identity is its binding constraint\") and reads as that line's operational corollary.\n- **No permanent bottom is also the fuel** (depth solves curiosity's transience) — lovely, but the corpus itself files it as *\"a coincidence rather than a design triumph\"*; not a commitment.\n- **Notifications are epistemic, not engagement-maximizing** (P9) and **time-on-site is not a success metric** — real commitments, but closest to generic anti-engagement positions a rationalist site would also state; failed the distinctiveness gate at tile granularity (the distinctive version of this stance is candidate 14's library test).\n- **Making the unsaid explicit is political** (Bourdieu: doxa→orthodoxy) — sharp and worth carrying, but a single-doc warning that folds under the standing legibility-under-power red-team theme rather than a standalone commitment.\n- **The graph is a map, not the product** (P1) — canonical, but an interaction-architecture stance whose landing-page work is already done by beginner-first (excluded) plus candidate 16; kept out to protect the constellation's conviction density.\n- **Double-crux is a social achievement the graph can invite, never decide** — true and crisp (I2), but a boundary-note on one instrument rather than a corpus-wide commitment; near propose-never-assert (excluded).\n- **Momentary engagement is worthless; recurrence is the signal** — founder-defined (2026-08-19) and important, but it is a measurement protocol for one test, not a platform conviction. May graduate if the pointing test becomes a standing instrument.\n- **Health-check/ops maxims** (real-shaped probes, never truncate the recovery artifact) — excellent engineering lore, not constellation material.\n"}