{"path":"research/what-counts-as-evidence/corpus-map-research-docs.md","content":"# Evidence in the Deliberus corpus — what the research and design docs say\n\n**Scope**: the docs slice only (`docs/`, plus the ruling ledger, `TODO.md` and a few code\nlookups where a doc's claim needed checking). Every point cites its file and section. Line\nnumbers are given where they help locate a passage. Direct quotations are exact. Where I am\nunsure, I say so. **Snapshot date: 2026-09-26.** Three docs were edited today (commits `2f07509`,\n`5ab80b8`, `8f198d5`): `truth-graph-evidence-system.md`, ledger Arc P in\n`session26-open-rulings.md`, and `TODO.md`. Those edits reverse several older statements about\nevidence weighting, so where an older doc says something is \"live\" and a newer one says it is\n\"inert\", the newer one is the one to trust (see § 10).\n\nAll paths below are relative to `docs/research/` unless they start with\n`docs/` or name a code file.\n\n---\n\n## 1. What counts as evidence in Deliberus\n\n### 1a. There is no evidence type. Evidence is claims.\n\nThe founding position is that evidence is not a separate kind of object. It is claims whose\nsupport can itself be decomposed.\n\n- **The founder's open question** (Session 7, March 2026), verbatim in\n  `scheme-bounded-decomposition-and-evidence-as-subgraph.md` § The Founding Question (l.39):\n  *\"I don't want any copouts, I want people and AI agents surfing around in Deliberus to see\n  exactly what the best minds on our planet think in detail and what the best evidence shows -\n  evidence that also is decomposed into its constituent parts/arguments. This is similar to the\n  conviction I have that value premises are often decomposable and evidence is often too. What\n  actually constitutes un-decomposable, more atomic evidence is an open question to me,\n  scientifically, rationally, philosophically.\"*\n- **The answer, written 2026-08-16** (same doc, § What evidence currently *is*, l.11–25):\n  *\"There is no atomic evidence, only claims whose decomposition nobody has demanded yet.\"* The\n  doc presents this as the self-similar principle applied to evidence, which the founder had\n  already said it should be.\n- **The \"What Is Truly Atomic?\" table** (same doc, l.132–148). An observation decomposes into\n  working senses, a calibrated measurement and adequate conditions. \"Study found d=0.4\"\n  decomposes into sample size, methodology, test, confounders and replication. Arithmetic\n  decomposes into chosen axioms. A meta-analysis decomposes into its studies. \"I feel pain\" is\n  \"genuinely hard to decompose\". The conclusion: *\"The graph has no natural bottom. But it\n  doesn't NEED one.\"* The label is always *\"currently undecomposed,\" never \"bedrock.\"* The same\n  list appears in `docs/depth.md` l.93–96 (\"Empirical evidence\", \"Expert testimony\", \"Evidence as\n  subgraph\").\n- **Rescue of the foundation.** `evidence-division-and-the-foundation.md` § 2 (l.15): *\"there is no\n  primitive 'evidence' type — evidence is claims whose own support bottoms out empirically — so a\n  routing judgment is a link, links carry provenance, and links can be attacked.\"*\n- **Founder on the empirical horizon** (2026-09-14). A session had said that at an empirical\n  terminus *\"the map is done and evidence takes over\"*. The founder corrected it, verbatim\n  (`session33-the-horizon-vocabulary.md` l.11; ledger O2): *\"'terminus' does feel like a\n  misnomer for the factual case, since evidence breaks down into a subgraph, and all of the\n  subgraph claims can/should be dealt with in depth and are amenable to all sorts of various\n  analysis/breakdowns, right?\"* `the-canon-distillation-pass.md` item 11 grades this\n  \"demonstrated\": a session asserted the opposite on the record.\n\n### 1b. The word \"evidence\" names three objects\n\n`scheme-bounded-decomposition-and-evidence-as-subgraph.md` § What evidence currently *is* (l.13–19)\nsays the word names **three different objects**, and that this is the source of the confusion:\n\n1. **Properties on a claim**: `evidence_type`, `methodology_tier`, `effect_size`, CI bounds,\n   `sample_size`, `study_count`, `heterogeneity`, `source_span`. All extracted and shipped.\n2. **A claim linked by a support edge**: what `POST /claims/{id}/add-evidence` creates.\n3. **A decomposed subgraph from a document**: the evidence-as-subgraph vision, not built.\n\n*\"The first two are the same object at different granularities, and nobody had noticed.\"*\n\n### 1c. Every evidence property is an unminted claim\n\nSame doc, l.19: *\"`heterogeneity: 0.73` sits on a claim as a number nobody can argue with. Stated\nas a claim it reads 'the twelve studies disagree substantially with each other' — which is\nattackable, decomposable, and can carry evidence of its own. **Every evidence property is a claim\nthat has not been minted yet**.\"* The doc says this has the same structure as the reported-speech\nblind spot. It extends the idea to `sample_size` (\"a contestable claim about generalisability\")\nand to `methodology_tier` (\"a claim about what that design can establish\"). The consequence it\ndraws (l.25): the evidence-decomposition pass should first turn the already-extracted properties\ninto claims, and only then do the expensive document decomposition.\n\n**Refuted in part by measurement.** The expectation was that a minted heterogeneity claim with an\nattack edge would press on its parent \"with no new formula\". It was run in stress-suite wave two\nand **did not** (`synthetic-stress-suite-and-ontology-reflection.md` § Wave two, E1, l.69):\n*\"Minting converts a property into something contestable; weight still arrives only through\nscoring.\"* The scheme-bounded doc was corrected at source (l.23: *\"minting buys\ncontestability, not weight\"*). See § 10 for where the correction did not reach.\n\nThe same direction applies to methodology tiers. Ledger P4 (`session26-open-rulings.md` l.310)\ncalls the tier table *\"exactly the undecomposed 'standard for what counts as evidence' P2 wants\ndecomposed; the stated direction is to mint it as a claim.\"*\n\n### 1d. Evidence as subgraph\n\n- The vision, with a worked tree (`scheme-bounded-...md` § Evidence-as-Subgraph, l.316–343): a\n  meta-analysis supporting *\"UBI reduces poverty\"* resolves into extracted findings, each with CQs\n  (\"Sample representativeness? ← sorry\", \"Effect size heterogeneity? → 'I²=73%'\", \"Publication\n  bias?\"), inclusion criteria, and sorry markers for undecomposed study methodologies. The\n  principle is *\"A meta-analysis IS text. Run the pipeline on it\"* (l.343), and the recursion is\n  *\"bounded by the actual content of the evidence document.\"*\n- The founder, verbatim (l.41): *\"In accordance with the progressive complexification, uploading\n  a meta-analysis as a PDF to Deliberus as support of a claim should be allowed. But then that PDF\n  should be decomposed as far as possible - maximum granularity that is helpful to the rest of\n  humanity.\"*\n- A second statistical version appears in `truth-graph-evidence-system.md` § Evidence-as-subgraph\n  (l.425–442). A mother claim *\"Meta-analysis finds UBI reduces poverty by 14-27%\"* decomposes\n  into \"12 RCTs were included\", \"Pooled effect size d=0.21\", \"Studies used RCT methodology\" and\n  \"Poverty was measured by income thresholds\", each with CQs.\n- The stated reason a meta-analysis can be treated as a claim with children rather than an\n  attachment: it *\"IS a causal evidence claim\"* (`scheme-bounded-...md` § Causal cluster, l.177).\n- **Status**: the meta-analysis-to-studies decomposition is **not built**\n  (`scheme-bounded-...md` status banner, l.3).\n- **Hinge exception.** `philosophical-foundations-stress-test.md` l.237 (Pritchard's hinge\n  epistemology): *\"hinges are arational commitments that make evaluation possible and are not\n  improved by evidential support: any support edge pointing at a hinge misrepresents its epistemic\n  standing.\"* This is a named limit on \"evidence is claims\".\n\n### 1e. Evidential support as a formal notion\n\n`support-semantics-and-the-conjunction-problem.md` § 2 (l.29) lists **Oren & Norman's evidential\nsupport**: *\"An argument cannot stand at all unless supported by a chain reaching actual\nevidence.\"* § 3 (l.45) observes that this is *\"the sorry-marker intuition in formal dress, and it\nis a stricter reading than Deliberus currently implements — where an unsupported claim sits at an\nagnostic 0.5 rather than being unable to stand.\"* (The later Keynes split in § 4c below replaced\n\"agnostic 0.5\" with \"no reading\".)\n\n---\n\n## 2. Kinds of evidence the corpus distinguishes\n\nThe corpus has **several overlapping vocabularies** for kinds of evidence. Nothing reconciles\nthem.\n\n### 2a. The shipped `evidence_type` enum (code)\n\n`deliberus/extraction/models.py` l.20: `EvidenceTypeLit = Literal[\"citation\", \"quantitative\",\n\"anecdotal\", \"none\"]`. This feeds `compute_confidence` (l.265–286) as a bonus of\n+0.10 / +0.05 / 0 / −0.10 on a base set by `epistemic_status`: observed 0.85, causal_hypothesis\n0.55, prediction 0.40, counterfactual 0.45, feasibility 0.50, normative_assertion 0.70. The\nstress suite also writes `evidence_type='synthetic_example'`\n(`synthetic-stress-suite-...md` l.14), which is outside the enum. It is presumably written\ndirectly rather than through the extraction model; I did not verify this.\n\n### 2b. The four kinds from the April 5 branch (recovered today)\n\n`truth-graph-evidence-system.md` § Evidence And QEM Proposals, recovered note (l.421):\n**testimonial** (someone says X; source-based schemes and their CQs), **statistical** (study or\nmeta-analysis), **logical** (X follows from Y; relationship edges plus scheme classification),\n**observational** (I measured X; `epistemic_status: observed`). The Apr 8 rewrite dropped the\nlist. It was restored on 2026-09-26 and marked as a proposal (ledger P16 context).\n\n### 2c. The methodology tier ladder\n\n`METHODOLOGY_BASE_WEIGHT` in `models.py`: meta_analysis 0.70, systematic_review 0.68, rct 0.65,\nquasi_experimental 0.55, observational 0.45, case_study 0.38, anecdotal 0.32, expert_opinion 0.30,\ndefault 0.50 (`truth-graph-evidence-system.md` § Methodology-adjusted priors, l.459–475;\n`long-form-sources-and-meta-analysis-weighing.md` § 2 table).\n\n**Status (2026-09-26)**: *\"unratified and inert.\"* It was *\"invented on Apr 5 as a starting\nproposal and never ruled (ledger P4). They are in `models.py`, but the production badge path\n(`graph/badge_inputs.py`) carries no tier and nothing writes a tier onto an edge, so no badge a\nuser sees has ever been moved by them\"* (`truth-graph-...md` l.477).\n\nGaps the long-form doc names (§ 5, l.78–86):\n- *\"Mechanistic evidence has no tier at all.\"* The doc suggests *\"one tier per claim is the wrong\n  shape\"*, and that a claim may need *\"a tier per thing-it-is-claiming.\"*\n- *\"A flat top tier is too generous.\"* Hedges et al. 2025 found 57% of 1,000 meta-analyses had\n  dependent effect sizes, and 53% of those analysed them as independent. Medical meta-analyses\n  were often built on 3–5 studies.\n- *\"Publication bias, replication status, and funding conflicts have no representation\"*; each is\n  a natural CQ.\n\n### 2d. Scheme-based kinds (Walton)\n\nThe scheme itself carries the evidential kind. The source-based cluster (Expert Opinion, Position\nto Know, Witness Testimony, Popular Opinion, Popular Practice) decomposes *\"into credential\nchains\"*: *\"credibility of source transfers (defeasibly) to credibility of assertion\"*\n(`scheme-bounded-...md` l.156–166). The causal cluster (Cause to Effect, Effect to Cause,\nCorrelation to Cause, Sign) decomposes *\"into mechanistic chains + statistical evidence\"* and *\"is\nwhere meta-analysis decomposition lives\"* (l.168–177). The abductive cluster decomposes into\n*\"evidence quality + explanatory comparison + search thoroughness\"* (l.234–241). The CQ lists are\nin `walton-argument-schemes.md` § 2. Expert Opinion CQ6 is *\"Evidence: Is E's assertion based on\nevidence?\"* Position to Know has CQ1–3. Sign has correlation strength and alternative\nexplanations. Evidence-to-Hypothesis has three CQs.\n\nThe deductive/inductive/abductive table (`scheme-bounded-...md` l.266–281) says that for inductive\nschemes, \"premises\" means *\"Evidence + quality indicators (sample size, methodology,\nreplication)\"*. `bias` is itself a shipped Walton scheme (CLAUDE.md topic row on cognitive bias),\nso an accusation of bias is an argument with its own CQs, not a label.\n\n### 2e. Experience reports vs self-interpretations (the fallible-self-knowledge conviction)\n\n- **Founder, verbatim**, 2026-08-25 night (`parity-and-the-zone-interior.md` § Addendum —\n  fallible-self-knowledge caveat, l.55): *\"I believe firmly that humans may be wrong about the\n  causes of their own experiences. Not wrong about their fundamental experience, but about what\n  objective causes were behind it. They may be wrong about counterfactual what-if branches...\n  They may be wrong about what would best satisfy a need or desire they're sensing. They may in\n  short be wrong about correlates of internal direct experience and about the constitution and\n  conditionality of their subjective world. We need to structurally keep this in mind in all\n  future settings contexts and moments, in order to best assist in mapping what is yet beyond the\n  horizon.\"*\n- The distinction (l.57): first-person authority is *\"incorrigible about the experience itself\"*\n  and *\"fallible about everything surrounding it.\"* The surround is *\"ordinary empirical\n  territory: decomposable, evidence-attachable, challengeable like any claim, whoever asserts\n  them.\"*\n- Ontology consequence (l.63, \"buildable, parked\"): *experience reports* are observer-privileged\n  and decomposition *\"never attacks the datum.\"* *Self-interpretations* (causal, counterfactual,\n  satisfier) get *\"NO elevated credence for being first-person\"*, and are already typable via\n  `epistemic_status` causal_hypothesis/counterfactual.\n- Canon line, RATIFIED 2026-08-25 (*\"Sounds good\"*), l.65.\n- Empirical support cited (l.61, \"survey-grade; verify before citing outward\"): Nisbett & Wilson\n  1977, choice blindness (Johansson & Hall), affective forecasting / miswanting (Gilbert &\n  Wilson), misattribution of arousal (Dutton–Aron).\n- **Applied to act three** (`what-human-judgment-is-for.md` § Act three, l.164–168): the first\n  version, *\"that is not what I meant\"*, was called incorrigible. That was corrected 2026-08-27:\n  a claim about one's past intention is a self-interpretation and fallible. *\"What is genuinely\n  unchallengeable is narrower... 'I do not endorse that rendering as a statement of my position' is\n  a speech act, not a description of history.\"*\n- The bias literature supports this (§ 4c-ii, l.281–286). The bias blind spot is *\"positively\n  related to cognitive sophistication\"*, so *\"a report about one's own past reasoning is the one\n  testimony that deserves no elevated credence.\"*\n\n### 2f. Impact testimony / positional evidence\n\n- **Act four** (`what-human-judgment-is-for.md` § Act four, l.181–183): *\"Here is what this costs,\n  from where I stand.\"* It was renamed from *stake testimony* by founder ruling 2026-09-13,\n  verbatim: *\"Yes do rename to \"impact testimony\". Also dovetails with how people from Circling /\n  Authentic Relating talk about how they're \"impacted\" by communication, relating and by things\n  happening around them, or by their circumstances I guess, so feels coherent in that way as\n  well.\"* The doc: *\"Being on the receiving end of a policy is knowledge no model holds.\"* It is\n  *\"additive by construction.\"*\n- **The derivability test turned the reserved human residue from considerations into\n  evidence** (`derivability-of-missing-considerations.md`, verdict up front and § 3). *\"The\n  residue that no machine can reach is real but small, and it is mostly not considerations at all —\n  it is evidence (what a thing costs from where a particular person stands, a local fact nobody\n  wrote down).\"* § 3(b) l.148: *\"a consideration is a dimension that bears on the conclusion;\n  evidence is the ground for it. Machines derive dimensions; humans supply much of the ground.\"*\n  The tally (l.131): 47 considerations graded 0/1/13/22/9 at rungs 0–4, **0 at rung 5, and 4\n  pieces of evidence.** § 10 adds a caveat: the zero is model-graded and *\"biased toward zero by\n  construction.\"*\n- **Standpoint split** (§ 7, sharpening 4, l.282–294): *positional evidence* (\"evidential\n  advantage\", which *\"is shareable through testimony, story-telling and analogy\"*) is separate from\n  *positional question-setting* (a standpoint; zetetic deference).\n- **How testimony earns effect** (§ 9, l.403–439). It works *\"as evidence through the witness\n  testimony / position-to-know schemes\"*, whose CQs are *\"is the witness in a position to know? is\n  the testimony consistent with other witnesses? does the witness have a stake in deceiving?\"* The\n  saboteur's best move is fabricated testimony (*\"I was in a SiS home and it was fine.\"*). Four\n  defenses: position is itself a claim; testimony is *\"evidence for a premise, never a verdict\"*\n  (one edge, cannot cap or lift a whole); corroboration via the novelty dial; and *\"positional\n  contributions are verified positionally.\"*\n- **Positional tiers** (`cheap-to-add-slow-to-matter.md` § 3, NEW, unruled): Unverified,\n  Self-attested with record, Positionally corroborated. *\"Weight rises with tier; nothing is gated\n  on tier.\"* § 8 item 4 adds that verification fails for singleton positions, and proposes\n  **position grouping** and a **sole-witness label**.\n- **Open problem 5** (`docs/open-problems.md` l.41): testimony, long practice and lived experience\n  (grandma's *\"I only drive familiar roads in daylight\"*, a midwife's judgment) may be scored\n  \"unsupported\" *\"simply because its support doesn't come as steps.\"*\n- ⚠ See § 10 for a tension with `what-human-judgment-is-for.md` § 5 item 4.\n\n### 2g. Reported speech\n\n- **RATIFIED 2026-08-15: always mint the reported content** (`self-similar-decomposition-and-claim-ontology.md`\n  § RATIFIED, l.226–232). The founder chose this over link-only and mint-on-demand. Two hard\n  requirements follow. (i) The report→content link must *\"carry no strength, since the reporter's\n  accuracy and the content's truth are independent — disproving the content must never suggest the\n  quote was fabricated.\"* (ii) Every published count must distinguish asserted from minted.\n- **Both were satisfied only by accident** (§ audited 2026-08-16, l.234–242). REPORTS carried no\n  scheme, and `count_safe_summary` had no caller. The first was fixed 2026-09-17: the badge query\n  now names `REPORTS`, `CLASSIFIES_TERMINUS` and `ARGUES_REQUIREMENT` as excluded\n  (`weakest-link-arithmetic-...md` § 9d, l.615). `count_safe_summary` still had no caller as of\n  CLAUDE.md's last statement.\n- **The live minimum-wage example** (l.260): the proposition *a wage floor harms the workers it is\n  meant to help* is *\"reported by two authors who disagree about it and asserted by neither.\"*\n- Four failure modes: duplication, paraphrase drift, straw man and phantom asserter (l.252–262).\n  The straw man becomes *detectable* only because of always-mint.\n- **Attribution roles** (l.210–216): the `self`/`third_party` split fails because\n  `third_party` (73%) bundles *citing a supporting authority* (\"strengthens\") with *reporting an\n  opponent* (\"to attack\"). The proposal is `cited_support` beside `reported_opposition`. Measured\n  over 130 distinct values: 129 values mapped to 109 entities, one hallucinated string was caught,\n  and `attributed_to` was populated on 1,082 of 1,642 substantive claims.\n- **Fabricated consensus** (`fragile-checkers-...md` § 2b, l.67): in the graph, *\"three experts\n  disagree with you\"* must arrive as a claim (*\"most economists believe X\"*), which becomes a\n  checkable attribution claim. This is called *\"the anti-Asch instrument.\"*\n\n### 2h. Expert opinion and source credibility\n\nThe Dr Smith mother claim (`self-similar-...md` l.21–64) decomposes credibility into publications,\nappointment, \"publication-implies-expertise\" (definitional) and domain match (definitional). The\nfounder, Session 9, verbatim (l.46): *\"The freeform text in CQ polarity claims and the reasoning\nfield should be exposed to the acid of universal reason... There's always separate evidence in\ndifferent domains and contexts for a claim that generalizes over those domains, but then in such a\ncase the 'mother claim' (the more general one) should simply be modeled as being supported by (and\nreceive QEM/QBAF support transitively/programmatically) all the subclaims that it can and should\nbe decomposed into.\"* Also (l.64): *\"We must at some point in time enable and process such\ndecomposition without insisting that a muddled, too-general claim is atomic.\"*\n\n### 2i. Identity and endorsement as evidence-bearing factual claims\n\nHeld direction, founder 2026-08-31 (`the-ratification-record-and-the-triage-feed.md` l.220–228),\nverbatim: *\"This whole thing deserves some thought later. Claims about who endorses what should\nprobably be factual claims just like anything else. That way ppl won't fight over which username\nhas a tick/is the real Person Z... I don't wanna have to deal with that mess if I can help it...\"*\nSo *\"account @z is Person Z\"* is a claim with evidence. The platform issues no verification ticks.\n*\"Deliberately NOT designed further.\"*\n\n### 2j. Evidence in the FRI residue\n\n`empirical-deliberation-convergence-evidence.md` § (e) item 4 (l.88) lists, among the causes of\ndisagreement that resist deliberation, *\"different standards for what counts as evidence.\"* This\nis the empirical record's warning that the evidence standard is itself a place where people\ndisagree. Ledger P2 cites it against the founder's hope for person-independent probabilities.\n\n---\n\n## 3. How evidence enters, and provenance/attribution\n\n### 3a. Entry routes named in the docs\n\n1. **Extraction** (8-pass pipeline). Quantitative fields (`effect_size`, `sample_size`,\n   `study_count`, CI bounds, `heterogeneity`) are in the Pass-2b schema because `ClaimBase`\n   carries them. But *\"No prompt asks for them, no endpoint accepts them, no page shows them\"*.\n   The model fills one now and then: *\"1 of 57 in the minimum-wage fixture\"* (`TODO.md` l.24–27,\n   2026-09-26). An older doc called these fields \"Live\" and \"extracted\"; see § 10.\n2. **PDFs**, read natively via Gemini `inline_data` (`long-form-...md` § 2). Stored in Postgres as\n   bytes and never in git (§ 6).\n3. **`POST /claims/{id}/add-evidence`** (correction UX). *\"supports/attacks toggle pills + text\n   input; creates evidence claims with embeddings\"* (`docs/ux-principles.md` l.97). The edge\n   direction is `evidence → claim`, which is incoming.\n4. **`answer-cq`**: yes/no polarity pills on a critical question (same line).\n5. **Auto-connect.** It *discovers* evidence relations among existing claims rather than\n   generating them. Founder's pruning insight, verbatim (`scheme-bounded-...md` l.84): *\"If the\n   system would auto-decompose incoming new claims in a way that automatically prunes them based\n   on some criteria, like only finding or creating new premises that could one step removed be\n   allowing connections to other existing claims. Then perhaps it would be more computationally\n   feasible and helpful.\"* The doc inverts this into *\"Auto-CONNECT, not auto-decompose\"*:\n   *\"you're never generating premises, only finding real ones\"* (l.94–104). Since 2026-09-22 the\n   keeper's seven-key sameness check runs first (CLAUDE.md principles).\n6. **Agent-driven evidence loop** (proposal, Apr 5; `truth-graph-...md` l.501–509): find sorry\n   markers, search for papers, extract, auto-connect, update. Constraint: *\"agents scaffold, users\n   verify.\"* For agents as **readers**, the first tool set is *\"dispute-shaped and\n   fact-layer-first\"* (`agents-as-a-consumer-class.md` § 2a). It is read-only.\n7. **Long-form audio/video**: `/transcribe` ships. The missing piece is *\"only a fetch step\"*\n   (`long-form-...md` § 4, § 6).\n8. **Simulated lenses**: may ask on the tail's behalf, never testify (§ 3d below).\n\n### 3b. Provenance requirements\n\n- **Locator.** `source_span` gives the quote but *\"Neither [timestamp nor page] exists today\"*.\n  The locator is *\"the thing that lets a reader verify in eight seconds instead of ninety\n  minutes\"* (`long-form-...md` § 4).\n- **Segment level.** Long sources have an intermediate unit (chapter, lesson). The doc suggests\n  mapping segments onto argument groups before adding any node type (§ 4).\n- **A source is not the unit of evidential quality** (`long-form-...md` § 3): *\"a source is not\n  the unit of evidential quality. A claim is. Methodology belongs on the claim, and the\n  source-level tier is at best a prior.\"* The pipeline assigns `dominant_type` per extraction. The\n  truth-graph design said a new edge's base weight comes from *\"the methodology tier of the source\n  extraction.\"* Feeding the Physionic video in as one blob would give ten claims of very different\n  standing one strength. The doc names this *\"a new member\"* of the flattening family: *\"an\n  aggregate source flattens by averaging the evidential standing of its parts.\"*\n- **Lifecycle and provenance fields** (`truth-graph-...md` § Lifecycle And Provenance, l.160–167):\n  `origin`, `lifecycle` (raw/draft/candidate/canonical/superseded), `scope`, and lineage links.\n  `maturity_stage` and `retrieval_weight` shipped Apr 8 (l.187–190).\n- **Every published count labels its provenance** (CLAUDE.md principle; `cheap-to-add-...md` § 5):\n  asserted, minted, reported, simulated, witnessed-unverified.\n- **Bengio's epistemic contextualization** (`bengio-safety-from-honesty-and-deliberus.md` l.14–18),\n  Def 3.22: *\"A factual observation records a directly measured value, the output of an executed\n  computation, or a proven statement. In contrast, a communication act records that a named agent\n  uttered something...\"* Deliberus made *\"the same move... independently.\"* A roadmap item (l.56):\n  a cleaner split between *\"who asserted X\"* edges and *\"is X\"* strength.\n\n### 3c. Entry is judgment-free, not cost-free\n\n- `cheap-to-add-slow-to-matter.md` § 0 and § 8 item 1: *\"Cheap to add, slow to matter.\"*\n  *\"Distrust is spent on effect and on standing... never on the merits at entry.\"* This was\n  corrected 2026-09-05 after the founder asked *\"None at all on entry?\"*. Entry carries an account,\n  a rate limit, class rules and machine-side gates. What is forbidden is a judgment of **merits**\n  at the door.\n- **Contributions are added, never overwritten — nothing needs permission to exist**\n  (`what-human-judgment-is-for.md` § 4d, ratified 2026-08-27). ⚠ *\"This is about EXISTENCE, not\n  weight: the strength layer is an aggregate and is brigadeable (supports are self-serving, attacks\n  altruistic).\"*\n- Derivability § 9: *\"Existence is free; effect is earned.\"*\n\n### 3d. Simulation as a source (unruled)\n\n`derivability-of-missing-considerations.md` § 8 (l.321–388): *\"yes for questions, no for\ntestimony.\"* The eight constraints include: *\"Simulate to ask, never to testify\"*; a `simulated`\nprovenance kind; *\"Cap-only, never lift... carries zero evidence weight and can never raise a\nstrength\"*; and *\"Simulation is a gap detector, not a gap filler.\"* Net: *\"simulation is the\nlens-rotation pass with a muzzle — it may ask on the tail's behalf and may never answer for it.\"*\n\n### 3e. The graph converts persuasion into evidence objects\n\n`fragile-checkers-and-the-verification-bottleneck.md` § 2a (l.58–63): *\"Challenges to a verdict do\nnot go back to the judge.\"* `answer-cq` and `add-evidence` create attributed claim nodes, and the\nresponse is computed by deterministic QBAF. *\"The graph converts persuasion attempts into\nevidence-objects... Pressure has to write itself down.\"* Exposure item 3 (l.77): *\"a motivated\nstakeholder can still craft evidence text to game the scheme (and thus the edge's strength\nprior)\"*, and no instrument looks for this.\n\n---\n\n## 4. How evidence is weighed\n\n### 4a. QEM and CQ energy\n\n- **Selection of QEM** (`qbaf-gradual-semantics-research.md` § Why QEM Wins, l.60–69): gravity\n  toward the initial weight (*\"Unanswered CQs (0.5) stay at 0.5 until evidence arrives\"*), no\n  saturation, open-mindedness and symmetry. The doc also claims *\"Gravity = Bayesian prior\n  behavior... This IS the rational probabilistic behavior\"*, which is contradicted later (§ 10).\n- **The formula** (l.73–99): `energy = Σpos − Σneg` over CQ polarity strengths, and\n  `h(x)=x²/(1+x²)`.\n- **Three levels of theory** (`strength-layer-audit.md` § 1): QBAF is the structure, gradual\n  semantics is the class, QEM is one member. Potyka's version is a fixed point, while Deliberus\n  computes **a single pass** plus a recursive descent along decomposition. Cycles measured: *\"3\n  two-cycles and 1 three-cycle... against roughly 3,950 such edges, and zero cycles in\n  decomposition.\"*\n- **Unexamined edges contribute exactly zero** (`strength-layer-audit.md` § 2, l.61–65). An edge\n  whose own CQs are unanswered computes to 0.5, and claim-level contributions are offsets from\n  0.5, so *\"ten of them leave a claim at 0.500.\"* This is deliberate: *\"summing raw strengths would\n  let three unexamined supporters manufacture a strong claim out of silence.\"* A second gate: *\"a\n  user-created evidence edge carries no scheme until the correction pipeline's LLM pass assigns\n  one, and the badge only reads scheme-bearing edges.\"* The consequence: *\"the correction-UX\n  promise is a two-step loop presented as one: attach evidence, then answer the critical questions\n  it raises\"* (also `docs/ux-principles.md` l.99).\n- **Weighing is by scoring, not by attaching** (stress-suite E1). The *badge-boundary two-step\n  (attached ≠ scored)* applies to minted evidence properties too (`scheme-bounded-...md` l.23).\n- **E4** (`synthetic-stress-suite-...md` l.72): the first end-to-end weighing answer flipped\n  UNEXAMINED to GROUNDED. *\"a single answer registers as engagement before it registers as\n  direction.\"*\n\n### 4b. Polarity and direction: the severe defect, fixed\n\n`strength-layer-audit.md` § 2. *\"A supports B\" means A is evidence FOR B*, so a claim's evidence\nsits on its **incoming** edges. Defect one: `compute_claim_badge` never read the edge type, so\n*\"1,567 incoming ATTACKS edges were counted as though they were supports.\"* Defect two: the\ndirection preference was backwards (54 claims). Both were fixed 2026-08-17. Badges fell from\n**3,828 to 427**, and **3,401** now read NO DATA. The cause: *\"The defect lived in the layer\nnobody wrote a spec for.\"*\n\n### 4c. Weight versus direction (the Keynes split)\n\n- Commit `ff42b21` (2026-08-17, *\"feat: separate where the evidence points from how much of it\n  there is\"*) and `deliberus/evidential_weight.py`. Keynes 1921: *\"the weight of an argument rises\n  with relevant evidence even when the probability does not move.\"* Peirce: balance of reasons\n  versus amount of knowledge. Popper: corroboration is not probability. Dempster–Shafer: explicit\n  mass on \"don't know\". The invariant: *\"A reading with no weight has no direction, not a neutral\n  one.\"* *\"Weight is a count, deliberately, and it does not saturate.\"* Counted *\"on the inputs\n  rather than their sum, because evidence that cancels still carries weight and balance is a\n  finding.\"*\n- **P22, Absence Is a State, Not a Score** (`docs/ux-principles.md` l.372–386), founder-named\n  2026-08-17. Founder, verbatim: *\"Why the heck was this not the case to begin with? I'm not sure I\n  understand how the platform could default to this kind of structural sycophancy.\"*\n  (`strength-layer-audit.md` § 2b, l.87). The badge words trace to his own sketch: *\"all CQs\n  addressed (green), some addressed (blue/amber), none (orange), one failed (red)\"*. The rule:\n  *\"A sentinel value meaning nobody has looked must never be routed into a comparison that\n  assumes a measurement.\"*\n- **UNSETTLED** (sibling ruling 2026-08-23): *\"An examined claim whose evidence currently offsets...\n  reads UNSETTLED in a neutral shade\"*. *\"one decisively-answered question yields 0.524 and must\n  remain a visible lean.\"* The badge now has three honest empty states: NO DATA, UNEXAMINED and\n  UNSETTLED (`what-belongs-in-the-ontology.md` l.251).\n- **Keynes and the display default** (`cheap-to-add-...md` § 5): *\"Display default is unexamined,\n  never neutral-as-balanced.\"*\n- **Today's `/query` fix** (commit `5ab80b8`, 2026-09-26): after the split, an UNEXAMINED badge\n  carries strength None, and the context builder crashed comparing it to 0.5. Before that, *\"a NO\n  DATA claim carried the 0.5 prior and was counted as support, which is how most of the graph\n  reached the synthesis prompt.\"* Claims now fall into *borne out*, *cut against or split on*, and\n  *nobody has weighed yet*.\n\n### 4d. Methodology tiers in weighing\n\n- The design: *\"A meta-analysis starts with a stronger prior than an anecdote, but both can still\n  be challenged\"* (`truth-graph-...md` l.475). The claim is that the tier override is live in\n  `compute_edge_strength` and tested (`long-form-...md` § 2).\n- The audit adds *methodology damping* (`strength-layer-audit.md` § 2 end and § 5 item 6): *\"a\n  tier raises an edge's prior, and the claim now moves off its own prior by that edge's energy...\n  so methodology reaches a claim in damped form. Decide whether that is right.\"*\n- **But as of 2026-09-26 no production badge reads a tier** (ledger P4; `truth-graph-...md`\n  l.477). A TODO precondition on ruling P4 \"keep\" (`TODO.md` l.19–23): *\"If a methodology tier\n  ever reaches an edge with no answered critical questions, the badge reads UNSETTLED 0.511 at\n  weight 1, which breaks the rule... that an unexamined claim carries no number.\"*\n- `heterogeneity` and `study_count` **weigh nothing**: *\"A meta-analysis of twelve studies with\n  I² = 0.73 gets exactly the same 0.70 base weight as a meta-analysis of forty studies with I² =\n  0.05\"* (`long-form-...md` § 2). *\"Heterogeneity is contestation, and it is currently silent\"*\n  (§ 5).\n- **GRADE mapping** (`truth-graph-...md` l.665–679): study design → tier, risk of bias → CQs,\n  inconsistency → I², indirectness → relevance CQs, imprecision → CI. It was corrected today:\n  *\"The fields do; none of them weighs anything... no rating was ever built.\"* (Ledger P10.)\n- **Sample-size weighting** proposed at `log(N)/log(N_ref)` (ledger P16). The sample-size half is\n  unbuilt. The replication half is **superseded**: *\"five SUPPORTS edges from shared evidence are a\n  double-count, not five times the weight.\"*\n\n### 4e. Necessary vs corroborative, and min vs sum vs product\n\n- **The conjunction problem** (`strength-layer-audit.md` § 4). Dr Smith's parts are *jointly\n  required*, while three studies are *corroborative*. `DECOMPOSES_INTO` was *\"carrying two\n  different relations under one name.\"* The literature grounding is in\n  `support-semantics-and-the-conjunction-problem.md` § 2–3: Nouioua & Risch's necessary support\n  (the inverse closure principle), and *\"Corroborating studies are closer to deductive or\n  evidential support\"* (l.41).\n- **The splitting incentive** (`strength-layer-audit.md` § 3), with measured numbers for additive\n  parts: 3 at 0.70 → 0.516, 6 → 0.557, 12 → 0.671, and 3 at 0.90 → 0.557. *\"Six mediocre children\n  score exactly what three excellent ones do.\"*\n- **RULED 2026-08-31, the padding ruling** (*\"Go with option C, both\"*): untagged decomposition\n  children aggregate by their **weakest part**. 3 vs 12 children are now identical at 0.502\n  (`strength-layer-audit.md` top update; `weakest-link-arithmetic-...md` § 6a).\n- **Why min** (`weakest-link-arithmetic-...md` § 1b–1f):\n  - Min is the unique idempotent t-norm, so *\"cloning a conjunct is a no-op under min and under\n    nothing else.\"*\n  - It is the Fréchet–Hoeffding upper bound, *\"the correlation-agnostic upper bound, which is\n    exactly the right default posture for a system that cannot know the correlations\"*\n    (`evidence-division-...md` § 2b).\n  - For entailment-chained padding, *\"min is not a bound — it is Bayes's exact answer.\"*\n  - Product punishes honest depth: 3 parts at 0.70 → 0.343, 12 → 0.014. The legal conjunction\n    paradox is cited here (Cohen, Pardo, Clermont).\n  - Amgoud & Ben-Naim show there is no neutral arithmetic.\n- **The corroborative channel is Good's additive log-odds** (§ 1f): *\"for genuinely independent\n  evidence, log likelihood ratios add... Its known failure is naive Bayes's known failure — the\n  independence assumption.\"* This sides with the accrual school against Pollock (§ 1b).\n- **Where Bayes has no jurisdiction** (§ 1f): *\"likelihoods exist only where evidence can bear —\n  the empirical terminus class.\"* Also *\"QEM strengths are deliberately not calibrated\n  probabilities... no strength should be read as P(claim is true).\"*\n- **Worked trees** (§ 6). Minimum wage: old root 0.501, new 0.416; the failing scale premise is\n  transmitted 1:1. Deterrence: old 0.500, new 0.456. *\"a methodologically-attacked study reads\n  pristine under addition... and honestly weak under min\"*, and the naked value premise becomes\n  load-bearing.\n- **The up-transmission whisper** (§ 7). Weakness goes down at full strength, while strength goes\n  up only a little: *\"three parts each established at 0.9 lift their parent only to 0.507.\"* The\n  doc reads this as *\"it prices the coverage gap\"* and calls it currently *\"an accident of\n  inherited constants.\"* Proposal: **coverage-gated lifting**.\n- **The misfiling census** (CLAUDE.md, the no-answer-in-a-gate principle; `the-parts-review.md`\n  § 3). Of all 197 decomposition edges, 70 are `miscast_as_evidence` (35%), 81 are `does_not_fit`\n  and only 35 + 11 are genuine parts. *\"76% of every decomposition edge in the graph is not a\n  part-of relation.\"* A re-run agreed. Extraction files evidence as parts *\"20 times in 30\"*, and 0\n  of 84 auto_connect \"parts\" were genuine. Re-filing the 70 as supports is *\"a separate ruling — it\n  moves strength on every mother claim above them.\"* § 8 of the weakest-link doc: *\"a\n  corroborative-looking child is usually an evidence edge mislabelled as decomposition (three\n  studies are witnesses, not parts).\"*\n\n### 4f. The requirement cap (deflation mirror), built 2026-09-17\n\n- **Founder question** (§ 9d): *\"the irreplaceable human value and action also still definitely\n  makes them able to game and manipulate the system, right? Unless the similarity sensitivity and\n  auto merge can somehow defeat any human saboteurs.\"*\n- **Built** (founder: *\"Build the designed completion.\"*). `support_semantics.requirement_weight`:\n  an explicit tag gives 1.0, a confirmed part is presumed at 0.5, and an unconfirmed machine\n  proposal gets 0.0. An argued requirement claim (`POST /claims/{id}/parts/{part}/requirement`,\n  `ARGUES_REQUIREMENT`) makes the weight follow that claim's strength. Measured that day: *\"10 of\n  87 decomposed mothers carry a cap and 0 change strength.\"*\n- **Reception, 2026-09-19**, founder verbatim (§ 9e): *\"Well, ideally the RELEVANCE (which I guess\n  all this is about) is judged provisionally by the machine/LLM-passes or whatever and then humans\n  can challenge it but I'm wary of saboteurs/spammers adding bogus parts that aren't actually\n  relevant. I'm open to making the relevance claims explicit like you say, but I wonder how\n  foolproof / attack-proof this way of dealing with this issue is, we'll have to explore that\n  going forward I guess. But we will make do with what we've settled now on this issue.\"*\n- **Exposure table**: one bogus person-added part caps a claim at **0.75** if left alone, **0.55**\n  if argued down, and **0.50** if refuted. *\"That is not attack-proof. It needs a defender to show\n  up.\"*\n- **Derivability correction** (`derivability-...md` § 4, corrected 2026-09-03): *\"inside the space\"\n  does NOT mean the cap applies automatically.* The machine's derivable list is *\"roughly half\n  junk\"* (CQs-Gen 2025: 44% useful, and LLM classifiers labelled 75–85% useful). So the cap weight\n  is *relevance strength*, never mere existence.\n\n### 4g. Double counting\n\n- **The five cases** (`soft-canonical-clustering-and-reversible-merge-semantics.md` § Addendum\n  Aug 31, l.373–381):\n  1. one proposition, two nodes → claim merge;\n  2. one evidence node reaching one aggregation point by several paths → path-multiplicity +\n     fair-share (built, unadopted);\n  3. one relationship as parallel edges → edge supersession (live);\n  4. partial content overlap → Amgoud–David novelty weighting (proposed);\n  5. distinct claims sharing a source → source-as-node + provenance clusters (proposed).\n  \n  The criterion (l.391): *\"double-counting = more than one path from an evidence node to a single\n  aggregation point. Legitimate reuse — one subtree serving many unrelated claims — is not\n  double-counting at all; it is the flywheel.\"*\n- **The founder's one-step invariant** (l.323): *\"everything upstream/deeper will be the same.\"*\n  It divides the labour: merge handles case 1, and the aggregation layer handles cases 2–5\n  (l.385–389).\n- **Founder question** (l.368): *\"how hard can it be to let auto-merge prevent counting any\n  evidence — as in, any subgraph — twice?\"*\n- **The auto-merge hunch**, founder verbatim (`weakest-link-...md` § 5): *\"I have a hunch that part\n  of the protection against splitting subclaims thin could be based on steadily improving the\n  auto-merge of claims, such that we'd always see through the trivial attempts to split into\n  unwarranted subclaims? Investigate this. Maybe it's not enough, maybe it is.\"* The verdict (§\n  5d): not needed in the necessary channel (*\"the entailed part can never be the weakest\"*),\n  *\"right, and genuinely needed, where addition survives\"* (*\"two restatements of one study, three\n  papers laundering one dataset\"*), and insufficient alone for correlated distinct evidence.\n- **The open flank, correlated corroboration** (§ 9): *\"three genuinely distinct studies sharing\n  one dataset add three times in the corroborative channel.\"* This remains *\"the strength layer's\n  least-defended channel.\"*\n- **The reconciliation landscape** (§ 9b), across five fields: Amgoud et al. KR 2018 / Amgoud &\n  David AAAI 2021 similarity adjustment, Denœux's cautious rule, Jøsang cumulative vs averaging\n  fusion, robust variance estimation, and variety-of-evidence. The last carries a warning:\n  *\"replication can trump variety\"*, so *\"a shared source that is known should be a node, not a\n  discount.\"* The Choquet integral is a sixth. Proposed (unruled): *\"corroborative energy becomes\n  Σ (σᵢ − 0.5) × novelty(i | stronger co-contributors) × 0.3.\"*\n- **Stress-suite supersession double counts** (§ Wave two): three instances in one wave. Edge\n  supersession was ratified and shipped the same day. The bridge mother returned from 0.368 to\n  0.459.\n\n### 4h. Evidence routing and fair-share\n\n- **The alignment warning** (`decomposition-axes.md` § 2). Decomposition helps verification only\n  with *\"evidence aligned to each sub-claim\"*, and degrades it when *\"the same claim-level evidence\n  [is] repeated for every sub-claim\"* (arXiv:2602.10380). *\"Before the next decomposition axis\n  ships, the honest question is whether the architecture creates evidence alignment or merely\n  creates nodes.\"*\n- **§ 2b, routing.** *\"Aligned-versus-repeated is not about where evidence came from; it is about\n  whether the link between a piece of evidence and a sub-claim is specific.\"* Three requirements\n  follow: routing must be able to abstain; *\"Generalizability is a scope judgment, not a relevance\n  score\"* (47 papers are *\"highly relevant\"* to domain match and *\"transfer to it not at all\"*);\n  and *\"Generous routing inflates confidence\"*, since `subclaim_energy` has no correlation term.\n  The instruction: *\"route by linking, never by duplicating.\"*\n- **Founder correction, 2026-08-15** (`self-similar-...md` § The founder's correction to fair-share,\n  l.264–270): *\"she published 47 papers\"* is a bundle, not a unit. *\"Fair-share discounting was\n  correcting an artifact of under-decomposed evidence.\"* **E2 confirmed this numerically**: after\n  per-outcome routing, *\"every discount is 1.0\"* (stress suite, l.70).\n- **`deliberus/evidence_routing.py`**: built, tested, not adopted, because *\"it changes every\n  number in the graph\"* (`evidence-division-...md` § 4 item 3).\n- **Confirmation holism** (`evidence-division-...md` § 1): *\"'divide the evidence among the\n  subclaims' is not merely hard — it is metaphysically the wrong verb, because evidential bearing\n  is a property of claim-webs, not claim-atoms.\"* The answer: *\"Link, never divide\"*; abstention is\n  first-class; and *\"There is no THE division, and there does not need to be\"* (§ 3). The founder's\n  question, verbatim (header): *\"how to divide up the relevant evidence between different\n  subclaims under a mother-claim, and aggregate the evidence in a way that's epistemologically\n  honest and well-calibrated, while honoring the fuzzy boundaries between some semantic domains\n  that the different subclaims might be talking about — a central one that in my mind might\n  challenge the foundation of what I imagine Deliberus could become.\"*\n- The **join-semantics problem** (§ 2b): downward division and upward combination share one\n  silent assumption, *independence*.\n- The **falsifier** (§ 5): the foundation fails *\"if... mother-claim strengths remain dominated by\n  routing choices reasonable people make differently.\"* It was partially run as the slicing study\n  (`docs/open-problems.md` § 1): four filings over 87 mothers, *\"The verdict survived every filing\n  for 91% of them.\"*\n\n### 4i. Sameness as a precondition of weighing\n\nThis is ruled (CLAUDE.md principles; `soft-canonical-...md` l.320). The seven-key check runs\nbefore any relation. *same* → merge candidate, never an edge; *contains/overlaps* →\ndecomposition proposal plus classification; *opposes* → attack candidate. It matters for evidence\nbecause two renderings of one claim linked by `supports` is self-support. That is the\n14-false-links incident.\n\n### 4j. Never averaged: user probability and votes\n\n- Votes are demoted to telemetry. *\"Agree/disagree feeds the discursive-dilemma instrument...\n  aggregation informs, and never decides\"* (`what-human-judgment-is-for.md` § 4).\n  `docs/ux-principles.md` l.62: *\"The QBAF badge IS the quality signal — computed from CQ evidence,\n  not subjective votes.\"*\n- User probability estimates enter as a **decomposition invitation**, never averaged in (ledger\n  P6, open): *\"You estimate 30%. The graph suggests 70%. What do you know that the graph doesn't?\"*\n- *Bare judgments are telemetry* (founder-agreed 2026-08-31, `what-human-judgment-is-for.md` final\n  mirror).\n\n---\n\n## 5. Where evidence bottoms out\n\n- **The oracle gradient** (`fragile-checkers-...md` § 4b, l.107–109). Physics grants *\"ground truth\n  by experiment, compact law... and objective error metrics.\"* *\"A QBAF strength is not a physics\n  loss. It is an accounting of contributed reasoning, conditional on the corpus — 'given what\n  people have contributed,' never 'reality says.'\"* *\"The terminus ontology turns out to be a map\n  of the oracle gradient: an `empirical` terminus marks the leg of an argument that can, in\n  principle, bottom out in an AU-style verifier or a fact-layer check — the leg that reaches the\n  ground. The four evaluative termini (and `is_bedrock`) mark where no simulator will ever\n  arrive.\"*\n- **The fact layer ships first because that leg reaches the ground** (settled 2026-08-18;\n  `agents-as-a-consumer-class.md` § 2a–2b). Founder verbatim: *\"Enabling agentic fact checking is\n  at least as important as enabling agentic human-values-checking/alignment... And fact checking\n  is perhaps tractable sooner than we are ready to compile an axiomatic bedrock of shared human\n  values… I see both facts and values as indispensable to collaboratively refine and\n  consolidate.\"* The doc: *\"The reason facts ship first is leg length, not product difference.\"*\n  *\"The answer to a factual query is the contest, not the conclusion.\"*\n- **Two floors** (CLAUDE.md topic row, `decomposing-value.md`): a phenomenal floor and a normative\n  floor. The pain descent found an *is-bedrock* terminus the ought-biased residue types could not\n  name (`deep-descent-experiment.md`). `is_bedrock` was added and first fired live in the stress\n  suite.\n- **Horizon kinds are read, never set** (RULED 2026-09-15, *\"The second.\"*). The empirical kind's\n  structural signature is that *\"the subtree bottoms out in evidence edges (methodology-tagged\n  supports and attacks)\"* (`what-belongs-in-the-ontology.md` l.210). The is-bedrock signature: *\"an\n  is-claim with no evidence subtree\"* (l.214). *\"`empirical` and `definitional` are handovers, not\n  ends... they close the value question only\"* (l.204).\n- **No horizon is a dead end.** *\"Empirical → evidence\"* (`parity-...md` l.29). The empirical kind\n  *\"opens onto... evidence subtrees\"* (`what-belongs-...md` l.203).\n- **The permissive zone needs its evidential edges settled first.** Criterion 1, *\"The band's edges\n  are mapped\"*. The minimum-wage $15-vs-$14 case is *\"Not yet a zone. The band's edges (the\n  employment effects) are still empirically contested... It becomes a zone only after the\n  empirical horizon settles the edges\"* (`parity-...md` § Addendum 2026-09-14 table). Inside a\n  zone, *\"more evidence and more argument are the wrong moves\"*, and the right move is *impact\n  testimony and attributed weighting* (use 2, l.107).\n- **Hinges** are not improved by evidential support (§ 1d above).\n- **Depth has no bottom but a current bottom** (`docs/depth.md` l.98): *\"a classified terminus is a\n  fallible fixed point under the current decomposition moves.\"*\n- **Deliberus is not simulation for arguments** (`fragile-checkers-...md` l.107): it is *\"a thing\n  Deliberus can never honestly claim to be.\"*\n\n---\n\n## 6. Time and staleness of evidence\n\n- **Three states on three axes** (`supersession-and-bitemporal-lifecycles.md` § The relation):\n  (1) *the world changed*, which is valid time and was previously unbuilt; (2) *our knowledge\n  changed*, which is deliberately **not** lifecycle: *\"learning-better is an argument move — an\n  ATTACKS edge, a strength shift — kept visible and contestable\"*; (3) *the representation\n  sharpened*, which is supersession (W3C PROV `wasRevisionOf`). *\"stale, refuted, and superseded\n  are three different states, and a lifecycle that conflates them files the penguin under large\n  flightless duck.\"*\n- **Built 2026-08-22** (founder: *\"We should build it, unless you see a reason not to?\"*):\n  `recorded_at` backfilled onto all 4,768 claims (datable fraction 0.985); `valid_at`/`invalid_at`;\n  the staleness daemon `deliberus/staleness.py` (flag-tier, entropy-class, *\"per-kind decay\n  horizons as named provisional parameters, no-basis-never-flags\"*); ratification via `POST\n  /claims/{id}/validity`, which mints a challengeable verdict-claim ground; and the invalid-evidence\n  filter *\"at the same WHERE seams as the supersession filter.\"* The first loop: *\"a\n  three-year-old testimonial swept (one flag from 3,728 candidates, zero noise), ratified, publicly\n  declared LAPSED... and its support edge stopped weighing.\"*\n- **Decay horizons** (code, `deliberus/staleness.py` l.31–39, \"deliberately PROVISIONAL,\n  founder-tunable\"): meta_analysis 5y, quantitative 3y, citation 4y, anecdotal 2y. ⚠ *Unsure*:\n  `meta_analysis` is a methodology-tier value, not an `evidence_type` value, and the daemon looks\n  up `evidence_type`. That key may be unreachable. This is a code observation outside my slice.\n- **Founder's challenge, the same day, which stands**: *\"the DETECTOR is deterministic; its INPUTS\n  are not.\"* The evidence-kind label is an extraction-LLM judgment. At first, *content time*\n  (`valid_at`) *\"is currently populated by NOTHING\"*, so sweeps aged *transaction* time. The van\n  den Haag essay *\"entered the graph July 2026, so its citation-kind claims read as fresh until\n  2030.\"* *\"judgment was moved upstream and made ratifiable, not eliminated.\"*\n- **The machine-readable dating tier shipped that evening**: *\"3 of 31 sources machine-datable, 28\n  honest misses.\"* Van den Haag was dated **1986** by its copyright line, which also corrected the\n  operator's from-memory 1969. The LLM tier (`llm_content_dates`) is written and gated off.\n  *Event-bound validity* (*\"the sitting president supports…\"*) is a watcher design. *Timeless\n  claims* are exempt.\n- **Open** (`docs/open-problems.md` § 10, l.65): *\"the decay horizons are our guesses, not measured\n  facts about how fast each kind of evidence really rots; a settled verdict still never expires on\n  its own\"*, plus the maintenance problem.\n- **Temporal skew is accepted** (`the-residual-error-taxonomy.md` § Noted ninth, l.124): *\"strength\n  is age-blind by explicit ruling (staleness flags and ratified validity lapse are filtered out of\n  strength on purpose).\"* ⚠ See § 10 on how to read this.\n- **A staleness case turned up in the wild**: the EV-purchase piece's *\"depreciation figures are\n  dated\"* (`derivability-...md` § 2, *\"time-indexed\"*). The state-deficit answer is *\"time-indexed\n  by nature\"* (`supersession-...md` item 1).\n\n---\n\n## 7. Known defects, measured facts, and inert features\n\n### Measured numbers (with date and source)\n\n| Fact | Source |\n|---|---|\n| 1,567 incoming ATTACKS counted as supports; badges 3,828 → 427; 3,401 NO DATA; 54 read from the wrong side (2026-08-17) | `strength-layer-audit.md` § 2 |\n| Sample of 238 claims: 122 NO DATA, 116 with evidence, **115 of 116 at exactly 0.500**; no CQ polarity claim had ever received evidence | same doc, § 4b |\n| Additive padding: 3/6/12 children at 0.70 → 0.516/0.557/0.671; now all 0.502 under min | same, § 3; `weakest-link-...md` § 6a |\n| Support/attack graph: 3 two-cycles, 1 three-cycle over ~3,950 edges; decomposition acyclic | `strength-layer-audit.md` § 1 |\n| 197 decomposition edges: 70 miscast_as_evidence, 81 does_not_fit, 35 necessary, 11 corroborative (76% not part-of), replicated (⚠ corrected 2026-10-07: 84 of those 197 were the automatic linker's premise links, stored backwards, which could not have been judged parts; among the other 113, 59% are not parts: `auto-connect-upgrade.md` § 2c) | CLAUDE.md principles; `the-parts-review.md` |\n| Extraction filed evidence as parts 20 of 30; auto_connect 0 of 84 genuine parts; `split` 30 of 44 | `the-parts-review.md` § 3 |\n| Requirement cap live: 10 of 87 mothers capped, 0 changed (2026-09-17) | `weakest-link-...md` § 9d |\n| Slicing study: verdict survived four filings for 91% of 87 mothers | `docs/open-problems.md` § 1 |\n| Staleness: 1 flag from 3,728 candidates; 4,768 claims backfilled at 0.985 datable; 3 of 31 sources machine-datable | `supersession-...md` |\n| Reported speech: 65 of 1,642 substantive claims (4.0%) are reported-speech shaped, 2 decomposed; `attributed_to` on 1,082; `third_party` 73% | `self-similar-...md` § blind spot |\n| Implicit premises: 10 pipeline-produced across 107 arguments, the pass never ran on about 60% of the extracted corpus; hand extraction finds about 1.7× more per claim (n=2) | `self-similar-...md` § ATTEMPTED; `the-residual-error-taxonomy.md` § 2 |\n| Fact-layer assets (2026-08-18): 2,160 SUPPORTS, 1,794 ATTACKS, 3,515 scheme-bearing edges | `agents-as-a-consumer-class.md` § 2a |\n| Quantitative fields filled by accident: 1 of 57 claims in the minimum-wage fixture | `TODO.md` l.24 |\n| Tier on an unexamined edge would read UNSETTLED 0.511 at weight 1 | `TODO.md` l.19 |\n| Derivability tally: 47 considerations, rung 5 = 0 (model-graded), 4 pieces of evidence | `derivability-...md` § 3, § 10 |\n| CQ generation: 44% useful, LLM classifiers labelled 75–85% useful | `derivability-...md` § 7 |\n| Hinge: one leaf flip ≈ 2 percentage points at shipped weights | `qbaf-gradual-semantics-research.md` § Production |\n| AI-governance mapping: 20 claims typed by the author as neither observation nor claim were stored `observed`; `observed` gives +0.30 confidence over a hypothesis (0.85 vs 0.55); 24 of 116 corrected | CLAUDE.md topic row; `mapping-a-structured-position-on-ai-governance.md` |\n\n### Defects and inert features\n\n- **`heterogeneity` and `study_count` are inert** (`long-form-...md` § 2; stress suite, *\"visibly\n  weighing nothing — the documented inert fields, now demonstrable on demand\"*).\n- **Methodology tiers are inert in production** (ledger P4, 2026-09-26).\n- **Methodology is per source, not per claim** (`long-form-...md` § 3). It is a *flattening\n  mechanism*.\n- **No locator** on `source_span` (§ 4).\n- **Fair-share and evidence routing are built but not adopted** (`evidence-division-...md` § 2\n  item 3).\n- **`count_safe_summary` has no caller**, so blended counts are published (`the-residual-error-taxonomy.md`\n  § 8, *confidence laundering*). *\"aggregation is the operation that erases origin.\"*\n- **The bridging feed does not rank by the badge.** It averages `r.strength`, the classifier's\n  extraction-time confidence, or a flat 0.7 on correction-UX edges, *\"which no CQ answer ever\n  updates\"*. So *\"the badge and the bridging score are currently different evidence wearing one\n  name\"* (`docs/ux-principles.md` l.66; CLAUDE.md strength row).\n- **The recursive composition gap** (stress suite gap 1): *\"A bridge whose load-test calibration\n  is undercut renders as SUPPORTED.\"* Fixed the same day with `compute_badge_recursive`.\n- **The headline inherited the nearest edge's scheme** (gap 4): *\"STRONG ANALOGY\"* came from the\n  mouse study while the meta-analysis sat two hops away. Fixed.\n- **Note text lied**, saying *\"no evidence bears on this yet\"* while an attack stood (gap 3).\n  Fixed.\n- **Coherent absence** (`the-residual-error-taxonomy.md` § 2): *\"an absent premise accumulates no\n  attacks, so the recorded side's strength grows relative to the unrecorded side... a lean caused\n  by what got ingested is indistinguishable, in-graph, from a lean in the reasoning.\"* The defense\n  is *\"weakest of all\"*. *\"No instrument measures what the corpus is a sample of.\"* An\n  ingestion-balance count needing no LLM is proposed.\n- **Confidence laundering** (§ 8): *\"the aggregate inherits the record's authority while blending\n  imagination-tier inputs.\"*\n- **The critic trained on imagined rollouts** (CLAUDE.md topic row on QWM): *\"the strength layer\n  computes over machine-extracted edges with no ratification gate.\"*\n- **`confirmed=true` means \"a human wrote this\", not \"a human checked this\"**, and *\"there is no\n  ratify action\"* (`the-residual-error-taxonomy.md` § 6).\n- **The epistemic-status ladder cannot say \"this is open\"**: *\"all six epistemic statuses assert\n  something and NONE says this is open\"* (CLAUDE.md topic row, AI-governance mapping). Extending\n  the enum is a founder decision.\n- **QBAF strength is participation-dependent, not evidence-dependent**\n  (`bengio-safety-from-honesty-and-deliberus.md` assumption 5, l.70): *\"A claim can be strong\n  because nobody attacked it yet; P_n cannot. The badge is a map of a debate, not a probability\n  model of the world.\"*\n- **Supports are self-serving and attacks altruistic**, so QBAF inherits the asymmetry\n  (`incentives-analysis.md` via `what-human-judgment-is-for.md` § 4d).\n- **Sybil and source independence are undefended at every layer** (`cheap-to-add-...md` § 6, § 8\n  item 2). The single-source ceiling is only *\"a lone-error guard.\"*\n\n---\n\n## 8. Founder verbatim statements about evidence (collected)\n\n1. *\"What if we need even more atomic claims? Could the system include auto-finding premises for\n   each new claim submitted... But perhaps I need to let go of that rigorous analytical hope.\n   Perhaps that's a pipedream.\"* — Session 7, `scheme-bounded-...md` l.31.\n2. *\"Are they truly not premises? Are they not decomposable into subquestions about the expert's\n   experience, education etc?\"* — same doc l.35. This is the correction that made CQs premises.\n3. *\"I don't want any copouts... evidence that also is decomposed into its constituent\n   parts/arguments... What actually constitutes un-decomposable, more atomic evidence is an open\n   question to me, scientifically, rationally, philosophically.\"* — l.39.\n4. *\"...uploading a meta-analysis as a PDF... should be allowed. But then that PDF should be\n   decomposed as far as possible - maximum granularity that is helpful to the rest of humanity.\"*\n   — l.41.\n5. *\"If the system would auto-decompose incoming new claims in a way that automatically prunes\n   them... only finding or creating new premises that could one step removed be allowing\n   connections to other existing claims.\"* — l.84.\n6. *\"Any claim could be supported by different arguments, but each argument needed its premises in\n   order to be a logical argument. But of course there's also inductive arguments, right? And the\n   whole thing about what constitutes a good bayesian inductive argument and so on.\"* — l.264.\n7. Session 9, mother claim: *\"There's always separate evidence in different domains and contexts\n   for a claim that generalizes over those domains...\"* and *\"We must at some point in time enable\n   and process such decomposition without insisting that a muddled, too-general claim is atomic.\"*\n   — `self-similar-...md` l.46, l.64.\n8. Apr 5, ledger P2: *\"if there's inductive evidence for any claim, that's what determines the\n   probability assessment, we try to break down into smaller parts everything these probability\n   assessments are based on and crowdsource+AI-source the evidence and premises and CQ\n   questions/answers or whatever that this depends on. So that was in my mind not different for\n   different people ideally. But of course maybe it is irreducibly subjective\"* — OPEN.\n9. Apr 5, ledger P3: *\"I have a fear of letting fuzzy gooey subjectivity in to affect the\n   assessments in Deliberus without proper decomposition and deep analysis and evidence/cross\n   checking etc etc.\"* — OPEN.\n10. Apr 5, ledger P1 (RULED): *\"...be careful not to mistake my focus on any one aspect now for\n    tacit agreement about the other topics, ok? I want us to do a full conscious audit so I'm the\n    final arbiter...\"*\n11. Apr 5, P5 context: *\"I agree with your last paragraph, I share that intuition\"* (on QBAF\n    converging to probability, conditional on full evidence). Now SUPERSEDED.\n12. 2026-08-15, the registered prediction: *\"...asymptotically lowering the investment in terms of\n    defining the right ontology and filling up the graph with extractions and human- or\n    auto-populated implicit premises, evidence and shared fundamental knowledge/assumptions about\n    how the world works.\"* — `self-similar-...md` l.274.\n13. 2026-08-17: *\"The subtree energy and evidence computation is solvable, I'm sure. We just need\n    theoretical clarity.\"* — `support-semantics-...md` header.\n14. 2026-08-17: *\"Why the heck was this not the case to begin with?... structural sycophancy.\"* —\n    P22.\n15. 2026-08-17 (audit prompt): *\"I want and need more insight into all of this. The paper and\n    theory around it. How it relates to our graph and the architecture around it.\"* —\n    `strength-layer-audit.md` header.\n16. 2026-08-18: *\"Enabling agentic fact checking is at least as important as enabling agentic\n    human-values-checking/alignment... I see both facts and values as indispensable to\n    collaboratively refine and consolidate.\"* — `agents-as-a-consumer-class.md` § 2a.\n17. 2026-08-21: the evidence-division root question (quoted in § 4h). —\n    `evidence-division-...md` header.\n18. 2026-08-22, on the stress suite: *\"Create a bunch of synthetic examples and do recursive\n    decomposition on their subgraphs/trees of claims/evidence...\"* and *\"Fix all five gaps\"*;\n    temporal: *\"We should build it, unless you see a reason not to?\"*\n19. 2026-08-25: the fallible-self-knowledge caveat (§ 2e). The sacred-domain ruling: *\"It can and\n    should be possible to reason about everything...\"* (`parity-...md` l.47).\n20. 2026-08-26, cluster-carries-the-strength, from ~2008: *\"I've known this would be a weakness\n    since when I first was sketching all this out.\"* One-step invariant: *\"everything\n    upstream/deeper will be the same.\"* — `soft-canonical-...md` l.316, l.323.\n21. 2026-08-31, the auto-merge hunch (§ 4g), and the double-counting question (§ 4g). Padding\n    ruling: *\"Go with option C, both.\"* Review protocol: *\"...show me a few hundred\n    proposed-should-be-merged vs not claims and then ask me whether I agree or not...\"*;\n    *\"this is important for factual claims as well... We'll see in time what kind of\n    calibrations all of this needs\"* — `soft-canonical-...md` l.336–342.\n22. 2026-08-31, identity and endorsement claims as factual claims (§ 2i).\n23. 2026-09-02: *\"'since a genuinely missing consideration is by definition not derivable from the\n    sentence' — Now I'm feeling dubious about this.\"* — `derivability-...md` header. Deflation:\n    *\"...able to game and manipulate the system, right?\"* — `weakest-link-...md` § 9d.\n24. 2026-09-03: *\"So what's a conservative approach that still uses humans to fill up the tail\n    faithfully but has a healthy measure of distrust against both LLMs and humans?\"* and *\"I still\n    feel a tension re the saboteur etc. How can we know which human contributions belong or\n    don't?\"* — `cheap-to-add-...md` header; `derivability-...md` § 9. 2026-09-05: *\"None at all on\n    entry?\"*\n25. 2026-09-13, the impact-testimony rename (§ 2f). Housekeeping: *\"Any regular housekeeping should\n    be done that could be done with high confidence in safety/accuracy...\"*\n26. 2026-09-14: *\"'terminus' does feel like a misnomer for the factual case, since evidence breaks\n    down into a subgraph, and all of the subgraph claims can/should be dealt with in depth...\"*\n27. 2026-09-17: *\"Build the designed completion.\"* 2026-09-19: the requirement-cap reception (§ 4f).\n28. Apr 24/Jul 4–5 (via parity doc): *\"driving anywhere between 60 and 80 on a road can all be\n    defensible\"*; *\"a risk-function choice within a formally bounded permissible zone\"*.\n\n---\n\n## 9. Open questions and unruled proposals about evidence\n\n**In the ruling ledger, Arc P** (`session26-open-rulings.md` l.298–322):\n- P2: person-independent probability given the same evidence (his position, posed as a question).\n- P3: fuzzy subjectivity must not move assessments without decomposition (his fear, not yet a\n  rule).\n- P4: keep or mint the methodology-tier baseline (decision).\n- P6: probability estimate as a decomposition invitation.\n- P7: *\"Does Deliberus ever output a probability?\"* The options are never, only via the oracle\n  bridge, or as a failure detector.\n- P9: No Copout as a graph property versus response usability.\n- P10: a GRADE-like reading.\n- P14: thin graph, say so or answer from background knowledge (an unmarked Apr 5→Apr 8\n  reversal).\n- P16: sample-size weighting.\n- **O13**: attack-proofness of the requirement cap.\n\n**Strength and weighing** (TODO § weakest-link stack; founder-paced):\n- Conjunctive vs corroborative as an ontology addition (partly ruled by the default flip).\n- The three-way parts/evidence/miscast classifier with lazy review. **The parts review is built\n  but not deployed** (`the-parts-review.md`). Re-filing the 70 miscast edges is a separate ruling.\n- Coverage-gated lifting and the up-transmission whisper (`weakest-link-...md` § 7).\n- Latent derivable requirements: do they participate in the cap, or only in coverage?\n  (`derivability-...md` § 5, *\"Founder call\"*; partly resolved by relevance-weighting.)\n- A machine relevance proposal replacing the flat 0.5 presumption (§ 9e, unbuilt).\n- Similarity-adjusted corroboration (novelty dial), gated on the seven-key eval sitting (§ 9b).\n- Source-as-node + source-independence provenance for shared datasets, authors and citation\n  chains (§ 9, § 9b).\n- The fair-share adoption decision *\"with measured before/after strengths\"* and the\n  **correlation-exposure figure** beside the badge (`evidence-division-...md` § 3–4).\n- A general gluing check (sheaf) over decomposition subgraphs (§ 4 item 5).\n- Methodology damping (`strength-layer-audit.md` § 5 item 6).\n- Iteration to a fixed point (§ 5 item 4).\n- Holistic comparison of rival subtrees (the Allen & Pardo objection, `weakest-link-...md` § 4\n  obj. 3).\n- Reservation-value / Pandora's-box ranking of which requirement to open next (Terenin;\n  `derivability-...md` § 6).\n\n**Evidence representation**:\n- Methodology per claim rather than per source (`long-form-...md` § 7 item 2, *\"a founder\n  decision\"*); methodology on claim, edge or both (`truth-graph-...md` Open Q4, Phase 7).\n- A timestamp/page locator; segment structure; video ingestion; a mechanistic tier or *\"a tier per\n  thing-it-is-claiming\"*; *\"an aggregate source should have no independent tier at all and should\n  inherit only what its decomposition supports\"* (§ 5).\n- Publication bias, replication status and funding conflicts as CQs.\n- The evidence-decomposition pass (turn properties into claims first).\n- Evidence binding on edges (*\"direct replication, pooled estimate, methodological critique, or\n  weaker topical analogy\"*, `truth-graph-...md` l.492).\n- A `cited_support` vs `reported_opposition` role split.\n- A latent-variable split between *who asserted X* and *is X* (Bengio).\n- The epistemic-status \"open\" value (AI-governance mapping).\n\n**Human evidence**:\n- A **stake field** (`what-human-judgment-is-for.md` § 5 item 4).\n- Positional tiers, the single-source ceiling, the score-free nursery, position grouping and the\n  sole-witness label (`cheap-to-add-...md` § 2–4, § 8).\n- The `simulated` provenance kind and its eight constraints (`derivability-...md` § 8).\n- The test with people from affected positions, *\"the first honest measurement of the reserved\n  classes' size\"* (§ 10).\n- A `resolution` lifecycle marker for interior traversals that *\"must never propagate strength\"*\n  (`parity-...md` § Build implications).\n- Open problem 5, knowledge that does not come as steps (`docs/open-problems.md`).\n- Identity and endorsement claims (held).\n\n**Time**:\n- The LLM content-date tier (gated).\n- An event-bound validity watcher.\n- Calibrated decay horizons.\n- Terminus/verdict expiry.\n- Revisiting age-blind strength (`the-residual-error-taxonomy.md` § ninth).\n\n**Bayes and probability**:\n- The oracle bridge, *\"specced nowhere\"* (`TODO.md`).\n- An inductive-argument stack: BIRD, falsification tracking, calibration-weighted estimates\n  (`docs/inductive-reasoning-research.md` § 9). Largely overtaken; see § 10.\n\n**Checker independence**:\n- A two-family jury as behavioural confidence (`fragile-checkers-...md` § 3).\n- A QEM rationality-postulate audit (`bengio-...md` necessity § (b)).\n\n---\n\n## 10. Contradictions and tensions between docs\n\n1. **Methodology tiers: \"Live\" vs \"inert\".** `long-form-sources-and-meta-analysis-weighing.md` § 2\n   (Aug 13) marks `METHODOLOGY_BASE_WEIGHT` feeding QBAF base weight as **Live**, and its banner\n   echo sits in `scheme-bounded-...md` l.3. `strength-layer-audit.md` § 2 discusses methodology\n   *damping* as a live consequence. `agents-as-a-consumer-class.md` § 2a lists the tiers among\n   fact-layer assets. But `truth-graph-evidence-system.md` l.477 and ledger P4 (both 2026-09-26)\n   say *\"no production badge reads a tier\"* and nothing writes a tier onto an edge. My reading:\n   the tier is honoured inside `compute_edge_strength` when passed, but the production badge path\n   never passes one. The older docs are therefore stale. I have not verified this in code beyond\n   the ledger's claim.\n2. **Quantitative fields \"extracted\" vs \"arrive only by accident\".** `long-form-...md` § 2 and the\n   scheme-bounded banner say `study_count`, `heterogeneity` etc. *\"are all extracted, and reach the\n   synthesis prompt.\"* `TODO.md` l.24 (today) says no prompt asks for them and they appear by\n   accident (1 of 57).\n3. **Heterogeneity-as-claim \"needs no weighting change\".** The scheme-bounded doc was corrected\n   after E1 (*\"minting buys contestability, not weight\"*). But `long-form-...md` § 2 supersession\n   note and § 7 item 1 still say that minting with an attack edge means *\"the shipped strength\n   computation handles it with no new formula and no new weights\"* and that *\"heterogeneity should\n   move strength\"*. This is a single-surface correction that did not propagate.\n4. **Impact testimony: weighed as evidence or not?** `what-human-judgment-is-for.md` § 5 item 4\n   proposes *\"A stake field... testimony, attributed, never weighed as evidence for the empirical\n   question it accompanies.\"* `derivability-...md` § 3(b) calls it *evidence*. § 9 gives it effect\n   *\"as evidence through the witness testimony / position-to-know schemes\"* moving one premise by\n   one edge. `cheap-to-add-...md` § 3 has *\"Weight rises with tier.\"* One possible reconciliation:\n   it is not weighed for the *empirical* question, but is weighed for a premise about costs or\n   impacts. No doc states that reconciliation.\n5. **QEM as Bayesian.** `qbaf-gradual-semantics-research.md` (March) says *\"Gravity = Bayesian prior\n   behavior... This IS the rational probabilistic behavior.\"* `truth-graph-...md` (Apr) said QBAF\n   converges toward probability. Against both: `weakest-link-...md` § 1f (*\"deliberately not\n   calibrated probabilities\"*), Bengio assumption 5 (participation-dependence), and ledger P5\n   SUPERSEDED. The QBAF doc's *\"No saturation: many weak supports can collectively strengthen\"*\n   also sits uneasily with the padding ruling and double-count work, though those concern\n   decomposition children, not incoming supports.\n6. **The QBAF doc's formula vs shipped arithmetic.** It gives `energy = Σpos − Σneg` over raw\n   polarity strengths and says unanswered CQs give energy ≈ 0. The shipped claim level uses\n   *offsets from 0.5* (`strength-layer-audit.md` § 2). The two agree only when positive and\n   negative counts are equal. This is minor, and I am unsure whether edge-level CQ energy uses\n   offsets too.\n7. **The March inductive stack vs later rulings.** `docs/inductive-reasoning-research.md` § 9\n   proposes *\"Evidence quality scored by LLM\"*, *\"community probability estimate (aggregated from\n   user estimates, weighted by calibration score)\"*, and *\"Better-calibrated users' estimates\n   weighted more heavily.\"* Later docs reject each of these. P6 says never averaged. The bias\n   literature makes contributor-rationality scoring *\"an empirical kill\"*\n   (`what-human-judgment-is-for.md` § 4c-ii). The unrubricked `overall_quality_score` was removed\n   (ledger P8). The March doc has no supersession note.\n8. **Self-similar \"QBAF transitivity is the mechanism... No additional mechanism is needed\"** (Apr,\n   `self-similar-...md` § 4). The stress suite found the walk was one level deep until 2026-08-22\n   (gap 1). The April text is uncorrected in place.\n9. **Age-blind strength vs lapsed evidence stops weighing.** `the-residual-error-taxonomy.md` §\n   ninth: *\"strength is age-blind by explicit ruling (staleness flags and ratified validity lapse\n   are filtered out of strength on purpose).\"* But `supersession-...md` and\n   `docs/open-problems.md` § 10 say a ratified lapse makes the support edge *stop weighing*, and\n   CLAUDE.md says *\"filtered from strength.\"* Probably consistent in fact: unratified flags never\n   move strength, and ratified lapses remove the edge. The residual doc's phrasing is ambiguous and\n   could be read as saying lapses do not affect strength.\n10. **Thin graph: admit or answer?** The Apr 5 branch said admit. Apr 8 said background knowledge\n    with a caveat, never marked as a reversal. It is flagged in `truth-graph-...md` l.342 and open\n    as P14.\n11. **Decomposition gating: demand vs completeness.** The scheme-bounded doc once read the founder's\n    *\"maximum granularity that is helpful\"* as demand-bounded. It was corrected 2026-08-27\n    (l.25, `decomposition-axes.md` § 4): his criterion is separability. The correction is in\n    place; noted because older text elsewhere may still say \"contest-gated\".\n12. **Three vocabularies for evidence kinds**: `evidence_type` (4 values), the April four kinds,\n    and the 8-value methodology tier. None is mapped to another. The staleness daemon's\n    `meta_analysis` key suggests they were mixed at least once (unsure, code).\n13. **The derivability \"0 corpus-absent\".** It was stated flat in the Sept 2–3 surfaces and\n    softened on 2026-09-05 (§ 10). The doc reports the softening as propagated. Any surface still\n    quoting the flat zero is stale.\n14. **\"Every edge carries evidence\" vs parts miscast as evidence.** The mother-claim pattern\n    assumes decomposition children are parts. The census says 35% of `DECOMPOSES_INTO` edges are\n    evidence relations filed as parts, so the necessary (min) default now applies to many edges\n    that are really corroborative evidence. The re-filing is unruled.\n\n---\n\n*Read in full: scheme-bounded-decomposition-and-evidence-as-subgraph, evidence-division-and-the-foundation,\nlong-form-sources-and-meta-analysis-weighing, strength-layer-audit, weakest-link-arithmetic-and-the-merge-hunch,\nsynthetic-stress-suite-and-ontology-reflection, derivability-of-missing-considerations,\nwhat-human-judgment-is-for, parity-and-the-zone-interior, decomposition-axes,\nsupersession-and-bitemporal-lifecycles, truth-graph-evidence-system, bengio-safety-from-honesty-and-deliberus,\nforecasting-room-at-the-top-and-deliberus. Read in relevant sections: self-similar-decomposition-and-claim-ontology\n(most of it), fragile-checkers, the-residual-error-taxonomy (§ 1–2), science-reasoning-infrastructure,\nscience-lineage-philosophy (Toulmin, Walton, QEM, residues, no-bedrock), docs/inductive-reasoning-research\n(§ 2–5, § 9), qbaf-gradual-semantics-research, walton-argument-schemes (CQ tables, QBAF integration,\ncoverage), empirical-deliberation-convergence-evidence (framing, § e, implications), ledger Arc P,\nsoft-canonical-clustering (Aug 26/31 addenda), cheap-to-add-slow-to-matter, the-parts-review,\nsupport-semantics-and-the-conjunction-problem, what-belongs-in-the-ontology § 6c, agents-as-a-consumer-class\n§ 2a, the-ratification-record § identity, docs/ux-principles (P4/P22), docs/depth.md, docs/open-problems.md,\nTODO.md Arc P loose ends, and evidential_weight.py / staleness.py / models.py docstrings and constants.*\n"}