{"path":"research/what-counts-as-evidence/corpus-map-founder-voice.md","content":"# Evidence in Deliberus: the founder's voice, rulings, convictions, history and open loops\n\n*Research slice for the evidence project. Written 2026-09-26 from a cloud session, read-only over\nthe repository. `.private/` was not searched (not needed for this slice; unreachable claims\nabout it are not made). Every quotation below is copied from the file cited; where a doc records only\na paraphrase, that is said. Status tags: **RULING** (he settled it, words on record) · **POSITION**\n(his stated view, not put as a standing rule) · **QUESTION** (he asked, left open) · **MUSING**\n(thinking aloud, hedged by him) · **CLAUDE** (assistant proposal, never ruled; included only where it\nframes his words).*\n\n---\n\n## 0. The shape of it in eight lines\n\n1. From 2011–13 evidence was an **entity**: a link from a Source to a Proposition, crowd-verified with Found/Missing votes and a `foundIndex`, with separate Trustworthy/Untrustworthy votes on sources.\n2. In March 2026 he rejected the black-box attachment: **evidence decomposes like value premises do**, a meta-analysis PDF \"should be decomposed as far as possible\", and \"what actually constitutes un-decomposable, more atomic evidence is an open question to me\".\n3. April 5, 2026 (the mobile truth-graph session) produced his two load-bearing epistemic positions: **the same evidence should yield the same probability assessment** once everything it rests on is decomposed, and a **fear of \"fuzzy gooey subjectivity\"** moving assessments without decomposition. Both are OPEN in ledger Arc P, and none of that night's machinery (the methodology tiers) was ever ruled.\n4. August 2026 turned evidence from properties into claims (*\"there is no atomic evidence, only claims whose decomposition nobody has demanded yet\"*, Claude's answer to his open question), split **direction from weight** (Keynes), fixed the badge that had counted 1,567 attacks as supports, and got his correction that **fact-checking is at least as important as values-checking and ships first**.\n5. His **evidence-division** worry (2026-08-21) is the one he said \"might challenge the foundation\"; the answer is holism managed by linking, abstention, fair-share discounting and a confession figure, none of it ruled as a whole.\n6. Late August–September moved the reserved human contribution to **positional evidence**: impact testimony (named by him 2026-09-13), lived experience, *\"positional contributions are verified positionally\"*, and his fallible-self-knowledge caveat (incorrigible about the experience, fallible about its causes).\n7. The horizon-vocabulary arc (2026-09-14→16) produced his correction that the empirical case is not an end: *\"evidence breaks down into a subgraph\"*. Terminus is now a horizon kind, read, never set.\n8. The biggest open decisions: P4 (the methodology ladder: keep, mint as claims, or drop), P7 (does Deliberus ever output a probability), P2/P3 (his same-evidence-same-probability position and the subjectivity fear), fair-share adoption, the 70 `miscast_as_evidence` edges, per-claim vs per-source methodology, the oracle bridge, and the evidence-decomposition pass itself.\n\n---\n\n## 1. Every founder verbatim statement on evidence, in date order\n\n### 2011–2013 (primary-source era)\n\n**1a. \"Logic is universal.\"** — Jan 2012, chat with Curtis SerVaas, answering whether validators should be divided by topic (climate vs evolution experts). `docs/research/chat-logs-analysis.md` § 3 (source `Curtis-jan22.txt`, archive not in repo). **POSITION** (early). Relevance: the system validates logical structure rather than domain claims directly; no topic-expert validators.\n\n**1b. The inductive \"hard problem\", 2012.** Quoted in `docs/technical-direction.md` § The Inductive Reasoning Problem as *\"Deductive arguments are simpler than inductive arguments — dealing with probability estimates for all claims can get hairy.\"* `chat-logs-analysis.md` § 3 renders the same passage as two \"icky issues\" he identified: (1) how to integrate statistics, evidence and inductive arguments, (2) how to weigh pros vs cons. **MUSING**. ⚠ The \"icky issues\" list is a paraphrase in the analysis doc; only the quoted sentence is presented as his words, and the raw log is not in the repo to verify.\n\n**1c. \"Focus on trusting the process instead of the people.\"** — 2012 article outline, `docs/vision.md` § Key Quotes; `simplenote-archive-analysis.md` § 5. **POSITION**. Earliest statement that reliability should come from process rather than from who the source is — the ancestor of \"no expert gate\".\n\n**1d. NOT the founder: \"I see it as a spreadsheet for knowledge rather than an expert system.\"** — `vision.md` line 42 attributes it to **Tom Boniecki** (Jan 2011, the Ivan Phillips thread), and so does `simplenote-archive-analysis.md` § 3, which goes on: *\"By adjusting weights and confidences you will be able to come to different conclusions but you will have to state those confidences explicitly…\"*. It is listed here only so it is not mistaken for his. His own 2012 line beside it in `vision.md` is about combining markets with \"the discerning force of logic, the common ground of science\".\n\n**1e. \"I'm an EA but also anticapitalist in the long run, but in order to figure out (and transition into) something better than capitalism, we need to build a collaborative sensemaking platform, Deliberus or something like it, a way for humanity to mindmeld with itself, to deeply and thoroughly process all knowledge, all beliefs and the evidence supporting it, and so distill our values to their essence, at least represented in language.\"** — `vision.md` § Key Quotes (undated in the doc; ~2025 messenger era by context). **POSITION**. Evidence named alongside beliefs as the thing processed.\n\n**1f. Sketch 7 (2013): \"Politiskt bevis/data\"** (political evidence/data, \"evidence sources with scoring\") and \"axiom 100% (grundläggande)\". `docs/sketches.md` § Sketch 7. Handwritten; the only evidence word in the sketches. The \"axioms established by users\" later inverted into No Copout Axioms (`history.md` § Consensus software board).\n\n**1g. Sketch 2's four VALID states**, including **\"none addressed (orange)\"** — the founder's words from the 2009–2013 argument-bundle sketch, per `TODO.md` line ~1858 and `strength-layer-audit.md` § 2b. The spec compressed four states to three numeric bands and the \"none addressed\" state vanished; its return in August 2026 as the gray UNEXAMINED badge is the Keynes split.\n\n### March 2026 (Session 7, Mar 29–31; Session 9, Apr 1)\n\n**1h. The foundational evidence conviction** — Session 7, `scheme-bounded-decomposition-and-evidence-as-subgraph.md` § The Founding Question (doc dated 2026-03-29):\n\n> \"I don't want any copouts, I want people and AI agents surfing around in Deliberus to see exactly what the best minds on our planet think in detail and what the best evidence shows - evidence that also is decomposed into its constituent parts/arguments. This is similar to the conviction I have that value premises are often decomposable and evidence is often too. What actually constitutes un-decomposable, more atomic evidence is an open question to me, scientifically, rationally, philosophically.\"\n\n**POSITION + QUESTION.** The single most-cited founder sentence on evidence. The last clause is his explicit open question.\n\n**1i. The meta-analysis PDF** — same doc, same session:\n\n> \"In accordance with the progressive complexification, uploading a meta-analysis as a PDF to Deliberus as support of a claim should be allowed. But then that PDF should be decomposed as far as possible - maximum granularity that is helpful to the rest of humanity.\"\n\n**POSITION** (treated by the corpus as settled design direction; never formally \"RULED\"). Note the bound: \"maximum granularity *that is helpful*\". ⚠ The corpus once read this as \"demand-bounded\" and attributed contest-gating to him; that was **corrected 2026-08-27** — his criterion is completeness/separability, contest-gating is only a budget heuristic (`scheme-bounded…` § What evidence currently is, Note 2026-08-27; `decomposition-axes.md` § Correction).\n\n**1j. Credibility CQs are premises, not modifiers** — same doc:\n\n> \"Are they truly not premises? Are they not decomposable into subquestions about the expert's experience, education etc?\"\n\n**QUESTION that functioned as a correction** — the doc records \"This was wrong. Fredrik's pushback was correct\" about treating critical questions as strength modifiers. Origin of the Dr. Smith credentials decomposition (publications, affiliation, domain match).\n\n**1k. Inductive vs deductive** — same doc § on argument types:\n\n> \"Any claim could be supported by different arguments, but each argument needed its premises in order to be a logical argument. But of course there's also inductive arguments, right? And the whole thing about what constitutes a good bayesian inductive argument and so on.\"\n\n**QUESTION/MUSING**. Led to the per-argument-type table (\"premises\" means evidence + quality indicators for inductive arguments).\n\n**1l. Auto-decomposition \"pipedream\" and the pruning insight** — same doc:\n\n> \"What if we need even more atomic claims? Could the system include auto-finding premises for each new claim submitted, so that all claims would be the conclusion of a few logical premises? That's the way I (or my left hemisphere at least) always imagined it. But perhaps I need to let go of that rigorous analytical hope. Perhaps that's a pipedream.\"\n\n> \"If the system would auto-decompose incoming new claims in a way that automatically prunes them based on some criteria, like only finding or creating new premises that could one step removed be allowing connections to other existing claims. Then perhaps it would be more computationally feasible and helpful.\"\n\n**MUSING → design direction** (auto-*connect* rather than auto-generate: \"never generating premises, only finding real ones\"). This governs evidence discovery too.\n\n> ⚠ **Corrected 2026-10-07, against this map's own rule that every quotation is his.** The words in\n> the parenthesis are Claude's, from the same document four paragraphs below his. His sentence says\n> *\"finding or creating new premises\"*. No ruling made *finding only* the direction, and it does not\n> govern evidence discovery by his word. [writing-out-the-unwritten.md](../writing-out-the-unwritten.md) § 3.\n\n**1m. Agents check reasoning, not facts (March framing)** — 2026-03-31, `agents-as-a-consumer-class.md` § 2 (from `session8-landing-page-and-civilizational-vision.md` § 8):\n\n> \"Agents checking against human-mapped structures isn't the point. The point is that they can check their reasoning through a Lean-like system just like Lean does today with agents, except the bedrock consists of human values stemming from our shared light of consciousness as a species, and I'm convinced our values are actually more atomic and similar among humans than what we often imagine.\"\n\n**POSITION**, later **amended by him** (see 1v).\n\n**1n. Mother claim / evidence across domains** — Session 9, 2026-04-01, `self-similar-decomposition-and-claim-ontology.md` § The Mother Claim Pattern:\n\n> \"The freeform text in CQ polarity claims and the reasoning field should be exposed to the acid of universal reason. That is, it should be recursively decomposed (eventually, not instantly tho) until we aren't muddling properly separable aspects of reality that can and should be judged separately. There's always separate evidence in different domains and contexts for a claim that generalizes over those domains, but then in such a case the 'mother claim' (the more general one) should simply be modeled as being supported by (and receive QEM/QBAF support transitively/programmatically) all the subclaims that it can and should be decomposed into.\"\n\n**RULING** — the doc's status is \"Decision made — implemented in code (ClaimBase hierarchy)\"; the Self-Similar Decomposition Principle is \"(April 2026 — DECIDED)\" in CLAUDE.md. The evidence-bearing clause: *separate evidence in different domains*, strength flowing up programmatically. This is the parent of every later evidence-routing question.\n\n**1o. Option B voting (2026-03-31)** — `docs/ux-principles.md` P2 \"Implementation decision (Mar 31, 2026 — Option B)\": humans vote agree/disagree only; quality comes from CQ evidence, not votes. **RULING** (recorded as a decision; his exact words are not quoted there — *paraphrase, his words not recorded*). He re-states the underlying skepticism in 1ee.\n\n### April 5, 2026 — the truth-graph mobile session (recovered 2026-09-26, ledger Arc P)\n\n**1p. The audit directive** — ledger P1, verbatim:\n\n> \"I wanna delve into each and every topic we've brought up in this whole session consciously and make explicit decisions, so be careful not to mistake my focus on any one aspect now for tacit agreement about the other topics, ok? I want us to do a full conscious audit so I'm the final arbiter of what direction we head in or what we actually build etc.\"\n\n**RULING** (procedural). Consequence: nothing from April 5 — including the evidence taxonomy (testimonial/statistical/logical/observational), methodology tiers, sample-size weighting, calibrated-confidence composite — was ever agreed. `truth-graph-evidence-system.md` headings were corrected from \"Decisions\" to \"Proposals\" on 2026-09-26.\n\n**1q. Same evidence, same probability** — ledger P2, verbatim:\n\n> \"if there's inductive evidence for any claim, that's what determines the probability assessment, we try to break down into smaller parts everything these probability assessments are based on and crowdsource+AI-source the evidence and premises and CQ questions/answers or whatever that this depends on. So that was in my mind not different for different people ideally. But of course maybe it is irreducibly subjective\"\n\n**POSITION, asked as a question — OPEN.** Ledger note: later corpus vindicates the method and challenges the uniqueness (rational permissivism, Buchak's risk band, the FRI collaboration's unmoved medians).\n\n**1r. The subjectivity fear** — ledger P3, verbatim:\n\n> \"I have a fear of letting fuzzy gooey subjectivity in to affect the assessments in Deliberus without proper decomposition and deep analysis and evidence/cross checking etc etc.\"\n\n**POSITION — OPEN**, \"not yet put as a standing rule\". Currently served by the Keynes split (UNEXAMINED carries no number) and *bare judgments are telemetry*.\n\n**1s. QBAF → probability intuition** — ledger P5: *\"I agree with your last paragraph, I share that intuition\"*, conditional on full evidence. The claim he agreed with (strength converges to probability as the graph deepens) is **SUPERSEDED** three times over; the surviving route is the unbuilt oracle bridge.\n\n**1t. Concept-sense auto-merge** — ledger P12, verbatim: *\"the approach I recently staked out for concept senses automatically merging for users later to consciously split them if deemed necessary. I reasoned that this would entail less friction and less wasted/manual time/effort for humans.\"* **POSITION, OPEN** (adjacent: identity scattering splits evidence across copies).\n\n### April–August 2026\n\n**1u. Registered prediction on cost** — 2026-08-15, `self-similar-decomposition-and-claim-ontology.md` § Registered prediction:\n\n> \"This is a living knowledge substrate which our interactions with will over time increase in what we get from them in terms of instructive/constructive value while asymptotically lowering the investment in terms of defining the right ontology and filling up the graph with extractions and human- or auto-populated implicit premises, evidence and shared fundamental knowledge/assumptions about how the world works. The costs here will likely go down over time while the usefulness will go up. This is my prediction. …\"\n\n**POSITION (registered prediction, unmeasured** — cost tracking was never implemented).\n\n**1v. The 47-papers correction** — 2026-08-15, `self-similar-decomposition…` § The founder's correction to fair-share discounting; `TODO.md` \"Decompose the evidence, not only the claims — founder correction 2026-08-15\". Recorded as: the 47 papers \"have contents, quality, relevance, and standing of their own; a serious treatment decomposes them, and then admits every other relevant paper and every indirectly related subtree in the graph.\" ⚠ **Paraphrase — his words not recorded**; the commit `970346b` describes it as the \"second correction\". **Treated as a correction he made and Claude accepted** (functionally a RULING on direction: evidence itself must decompose; a missing evidence-decomposition pass).\n\n**1w. Long-form sources** — 2026-08-13, `long-form-sources-and-meta-analysis-weighing.md` header: he asked that content like a two-hour review video be analyzable *\"both as a whole after submitting, and as its constituent parts\"*, whether meta-analysis weighing had been discussed, and *\"this whole video could be seen as a meta analysis I guess?\"* **QUESTION/MUSING.** Produced the finding that methodology is assessed per source and should be per claim.\n\n**1x. Strength theory** — 2026-08-17, `strength-layer-audit.md` line 5: *\"I want and need more insight into all of this. The paper and theory around it. How it relates to our graph and the architecture around it.\"* And `support-semantics-and-the-conjunction-problem.md`: *\"The subtree energy and evidence computation is solvable, I'm sure. We just need theoretical clarity.\"* **POSITION (confidence).** The Keynes direction/weight split followed the same day; commit `ff42b21` records he \"asked for a philosophically stringent fix rather than a guard\" (*paraphrase*).\n\n**1y. Fact-checking is at least as important — the August correction** — 2026-08-18, `agents-as-a-consumer-class.md` § 2a; commit `0c2b121`:\n\n> \"Enabling agentic fact checking is at least as important as enabling agentic human-values-checking/alignment. Both can play an extremely important role in the AI ecosystem and as infrastructure. And fact checking is perhaps tractable sooner than we are ready to compile an axiomatic bedrock of shared human values… I see both facts and values as indispensable to collaboratively refine and consolidate.\"\n\n**RULING in effect** — `TODO.md` marks \"Founder decision: the public framing — reasoning-verification or fact-checking\" as **ANSWERED 2026-08-18: both, and facts first.** The sub-claim \"fact layer ships first\" is his tractability ordering. Also the context quote (2026-08-17) that opened the arc: *\"…as an MCP/service for other agents to check their facts/conclusions or get recommendations or advice or whatever (this idea should be explored at depth if it hasnt already! had that idea many months ago!)\"*.\n\n**1z. Convictions may have evidence** — 2026-08-18, `convictions-across-scales.md` line 5:\n\n> \"Any and all convictions across the Deliberus corpus need not be without evidence: there might be a strong expectation that I hold for it based on analogies to similar patterns in the world. …\"\n\n**POSITION** (analogical evidence as legitimate support for a conviction; later: the four-check transfer test is the analogy scheme's CQ enrichment).\n\n**1aa. Offloading is the purpose** — session 21 (≈2026-08-17), `conviction-and-critique.md` § The delegation clarification:\n\n> \"Part of the purpose of Deliberus is to help people by all means mentioned earlier to offload large parts of the work and cost of traversing opacity and stringently following whatever virtues seem most epistemologically virtuous. They SHOULD of COURSE be INVITED to dive in and review and reflect and certify or fix by hand any and all small or large mistakes or inconsistencies or plot holes or missing implicit premises or logic jumps or WHATEVER, BUT Deliberus should ideally over time be made capable of offloading as much of the idealized processing and structuring and scaffolding as possible.\"\n\n**POSITION (corrected a Claude frame).** Governs machine evidence-work: delegation is fine; delegation *without self-report* is the hazard.\n\n**1bb. The evidence-division root question** — 2026-08-21, `evidence-division-and-the-foundation.md` header:\n\n> \"how to divide up the relevant evidence between different subclaims under a mother-claim, and aggregate the evidence in a way that's epistemologically honest and well-calibrated, while honoring the fuzzy boundaries between some semantic domains that the different subclaims might be talking about — a central one that in my mind might challenge the foundation of what I imagine Deliberus could become.\"\n\n**QUESTION (root, open)**. He also connected it to assembly theory's boundary problem (§ 2c, \"added 2026-08-21 on founder connection\" — paraphrase).\n\n**1cc. Synthetic stress suite** — 2026-08-22, `synthetic-stress-suite-and-ontology-reflection.md` line 3: *\"Create a bunch of synthetic examples and do recursive decomposition on their subgraphs/trees of claims/evidence on all of them, feed it all into the live argument graph and reflect on the ontology afterwards.\"* **Instruction.** Its wave two refuted the expectation that minting heterogeneity as a claim would add weight (minting buys contestability, not weight).\n\n**1dd. Staleness challenge** — 2026-08-22, `supersession-and-bitemporal-lifecycles.md`: \"the DETECTOR is deterministic; its INPUTS are not\" is recorded as \"the founder's challenge\" (*paraphrase*); plus *\"We should build it, unless you see a reason not to?\"* The challenge located the judgment inside `evidence_type` (an LLM label) and unpopulated content time (van den Haag 1969 reading as fresh).\n\n**1ee. Live-test research feeding** — 2026-08-25, `kartpaus-auto-enrichment-design.md` § The founder's ask:\n\n> \"In a live test for an election debate... I want Deliberus to automatically recursively decompose all claims in each mapping-break. It should find as much research of high quality as possible and feed everything into the graph.\"\n\n> \"All unstated premises need to be clarified and scaffolded. Ideally whenever the weighing questions are applicable they should be auto opened AND tentatively responded to in the graph as faithfully as possible according to our reading of the source/sources... automatic recursive decomposition and deepening exploration and relevant research/googling should fire/engage, at the very least... directly after a live test/recording that's transcribed and diarized and ingested... and I'm leaning towards that this should also occur (to some degree/cap at least) in normal use.\"\n\n**RULING** for the live test (direction RULED; post-session enrichment RULED MANDATORY for the live test); **lean** for normal use. By hand during election week per 2026-09-09 ruling.\n\n**1ff. Fallible self-knowledge (conviction-grade)** — 2026-08-25 night, `parity-and-the-zone-interior.md` § Addendum:\n\n> \"I believe firmly that humans may be wrong about the causes of their own experiences. Not wrong about their fundamental experience, but about what objective causes were behind it. They may be wrong about counterfactual what-if branches … They may be wrong about what would best satisfy a need or desire they're sensing. They may in short be wrong about correlates of internal direct experience and about the constitution and conditionality of their subjective world. We need to structurally keep this in mind in all future settings contexts and moments, in order to best assist in mapping what is yet beyond the horizon.\"\n\n**RULING-grade POSITION** (\"conviction-grade\"; ledger C4 RULED: self-interpretations get no elevated credence, only present refusal-to-endorse is incorrigible). The key evidential law for testimony.\n\n**1gg. No stop signs; proportional attunement** — 2026-08-25, same doc, sacred-domain addendum: *\"It can and should be possible to reason about everything, but we should be even more risk averse and epistemically humble in the personal/spiritual domain … But even this should not be artificially protected against analysis ultimately. Which of course, again, necessitates an equally deep attunement … I don't see any fundamental need to put up stop signs …\"* **RULING (ratified in substance 2026-08-25).**\n\n**1hh. The interior of the permissive zone is mappable** — 2026-08-24, same doc: *\"The internal process modeling should be clear which reasons and values weigh more for whom, but still be used to inform other people's internal/spiritual/values-reflections …\"* **RULING.** Attribution as the load-bearing requirement for personal-domain evidence.\n\n**1ii. Sustained exposure** — ledger I1, RULED (his): *\"it only tracks attitude change presumably over a quite short amount of time. I would be curious to see what happens when a human is presented with overwhelming reasoning and evidence going against their more or less sacred values.\"*\n\n**1jj. The meta layer is the gate** — 2026-08-27, `vision.md` § Why the meta layer is the gate; ledger M1 RULED:\n\n> \"I NEED to get all the philosophy, epistemology, ontology, process, method, evidence, logic, psychology, all the meta stuff right before we launch this, or it'll fall like a house of cards in a puff of wind.\"\n\n**RULING** (governing sequencing). \"Evidence\" is named explicitly among the things that must be right before launch.\n\n**1kk. Ratification — meaning, not meaningless buttons** — 2026-08-27, `what-human-judgment-is-for.md` Occasion: *\"I don't really know if it makes sense to have human users vote/ratify… Let's not install meaningless buttons … when they do engage it should MATTER, it should be filled with MEANING, so they can FEEL their humanness rippling across the graph.\"* **POSITION**.\n\n**1ll. Bias findings are design knowledge** — ledger F9 RULED: *\"Do you mean to tell me that It would not be beneficial to take into account the nature of any of the biases that have been studied?\"* Result: taxonomy-as-schema no; findings-as-design-knowledge yes; arrangement-as-evidence no.\n\n**1mm. Voting skepticism and the triage feed** — 2026-08-30, `the-ratification-record-and-the-triage-feed.md` § 3 and § 3b:\n\n> \"a human surfing around Deliberus should be subtly encouraged (but not conscripted/required) to evaluate different kinds of discernments made by LLMs in whatever circumstances we can't be 105% confident in the reliability of the targeted prompting of the LLM-step... but ideally humans would PRIMARILY be nudged (perhaps a feed could be created) to sift through suspected candidates for human-reserved classes of error …\"\n\n> \"I think in the past I've expressed serious skepticism about humans getting to vote on anything where they could lie based on motivated cognition or manipulation/corruption etc, so they should only be able to vote on whether they agree/disagree with a claim... but the triage-feed and the actions/judgments that it would invite from humans seem to be something else? …\"\n\nand 2026-08-31: *\"doesn't the skeptic-refusal (3b) of voting/ratification apply to this remaining seam as well, unfortunately?\"*; *\"To me the most fundamental question here is, should there even be a voting/ratification beyond the stance (agree/disagree) vote on a claim. …\"*. **RULING**: *\"Agreed on everything\"* (the demotion: bare judgments are telemetry) and *\"Anyhow yes I agree, blessing this whole proposal you gave now, go ahead.\"* (no ratify act; contributions and telemetry only). Evidence consequence: a human can move an evidential assessment only by contributing something that argues for itself.\n\n**1nn. The harness** — 2026-08-30, `the-harness-conviction.md`: *\"… It makes a real difference to whether we can expect sycophancy, wiggling, bias, hallucinations, sloppy analogies and forgotten premises … vs stringent, well-balanced, wise, well-read, steely-eyed, epistemically unfazed, always-grounded, near-perfect reasoners based on the best philosophy and science and the most accurate priors, evidence, inferences and analogies that humanity can muster. Deliberus can ultimately get us closer to this, I'm convinced of it.\"* **RULING** (canon, ratified 2026-08-30 with his amendment naming the two reserved classes).\n\n### Aug 31 – Sep 5 (derivability arc)\n\n**1oo. \"Don't we need humans for this?\"** — `session29-…` § 1: *\"This confuses me. I always thought the system somehow inherently had 'dealt with' this. That the badge/evidence score of a mother claim was maybe somehow implicitly assuming that all parts had been brought to bear. But we probably do need to 'check' this … Don't we need humans for this? Isn't this the reserved-for-humans responsibility of ensuring we're not in any frame lock-in? …\"* **QUESTION** → Popper's asymmetry in arithmetic (a raised, never total, lift; unruled).\n\n**1pp. The saboteur** — same doc § 2: *\"the irreplaceable human value and action also still definitely makes them able to game and manipulate the system, right? Unless the similarity sensitivity and auto merge can somehow defeat any human saboteurs.\"* **QUESTION.**\n\n**1qq. Doubting the non-derivability claim** — *\"'since a genuinely missing consideration is by definition not derivable from the sentence' — Now I'm feeling dubious about this. Reason about ten different random real example arguments found online. …\"* **QUESTION → correction** (sentence withdrawn corpus-wide; the human residue re-described as positional evidence).\n\n**1rr. Lived experience** — same doc line 97: *\"Positional? What does that mean?\"* → *\"Ah- the lived experience, otherwise known as\"* → *\"Very important. Hmm. Pondering this. To some extent we could backfill millions of lived experiences from online data. But yeah\"*. **MUSING** (parked with three catches, `TODO.md` § D).\n\n**1ss. Simulated tail** — 2026-09-03: *\"Can we not prompt models to simulate the perspectives of the tail?\"* **QUESTION** (Claude's answer: yes for questions, never for testimony; unruled).\n\n**1tt. Canon tweak + belonging** — 2026-09-03: *\"I guess you can go ahead with the canon tweak for now. But. I still feel a tension re the saboteur etc. How can we know which human contributions belong or don't?\"* **RULING (provisional)**: *An act's value is not its rung* re-pointed at the two positional acts (impact testimony; question-setting). The belonging question stays open.\n\n**1uu. Noise, not fakery** — *\"Sybil attack? What's that again? What I'm imagining is drowning in irrelevant content/text. Which can definitely destabilize and confuse and derail.\"* **POSITION (worry).**\n\n**1vv. The conservative approach** — 2026-09-03: *\"So what's a conservative approach that still uses humans to fill up the tail faithfully but has a healthy measure of distrust against both LLMs and humans?\"* **QUESTION** → `cheap-to-add-slow-to-matter.md` (single-source ceiling; positional tiers for testimony; unruled). Then 2026-09-05: *\"None at all on entry? … Critically examine our assumptions and rank them from weakest to strongest, according to the evidence at hand …\"* → the ceiling demoted to a lone-error guard; position grouping + sole-witness label proposed.\n\n### September 2026\n\n**1ww. Frame gaps get a contribution path** — 2026-09-09, ledger K; `what-human-judgment-is-for.md`: *\"a frame gap also deserves attention and to be exposed as a closable gap … When you are given the possibility to close gaps or otherwise contribute, the anxiety is short-lived and causes a productive contribution.\"* **RULING.**\n\n**1xx. Impact testimony, named** — 2026-09-13, `what-human-judgment-is-for.md` § Act four; commit `37f251c`:\n\n> \"Yes do rename to \"impact testimony\". Also dovetails with how people from Circling / Authentic Relating talk about how they're \"impacted\" by communication, relating and by things happening around them, or by their circumstances I guess, so feels coherent in that way as well.\"\n\n**RULING** (terminology). The act itself (\"Here is what this costs, from where I stand\") is ledger C3, **OPEN** as one of four acts.\n\n**1yy. Confident housekeeping** — 2026-09-13 (`soft-canonical…` Aug 26 addendum): *\"Any regular housekeeping should be done that could be done with high confidence in safety/accuracy and could spare human users from having to later clean up and organize the graph manually etc …\"* **RULING** (canon). Also *\"Confirmed the proposed ordering, thx\"* (sameness before any relation; the keeper's prompt says *\"OVERLAPS is judged by EVIDENCE, not by topic\"*).\n\n**1zz. Cluster carries the strength (~2008)** — `soft-canonical…` Aug 26: *\"I've known this would be a weakness since when I first was sketching all this out.\"* **POSITION** — provenance: identity scattering splitting evidence across copies was his worry from the start.\n\n**1aaa. Evidence is a subgraph, not a terminus** — 2026-09-14/15, ledger O2; `session33-the-horizon-vocabulary.md`; `what-belongs-in-the-ontology.md` § 6c:\n\n> \"'terminus' does feel like a misnomer for the factual case, since evidence breaks down into a subgraph, and all of the subgraph claims can/should be dealt with in depth and are amenable to all sorts of various analysis/breakdowns, right?\"\n\n**RULING** (O2 RULED 2026-09-16: *\"ah, yes, agreed, and choose \\\"definitional\\\"\"*, covering terminus → horizon kind, verdict → reading). Corrected Claude's statement that \"the map's work ends there and evidence takes over\".\n\n**1bbb. Motivated evidence in definitional fights** — `session33`: *\"… they will unfortunately often rationalize whatever defense and evidence they can to support their cause. But this is motivated reasoning in debate/politics and it is only one facet of the human condition, there is also honesty and curiosity and depth and empathy etc. …\"* **MUSING.**\n\n**1ccc. Read, never set; strength decides the zone** — 2026-09-15, ledger O4: *\"I'd rather let the graph score of how well-supported any and all 'options' claims are determine what shows up as a 'permissive zone'.\"* and RULED *\"The second.\"* with the caveat *\"But it could be premature, if the refreshed re-read subtree maybe isn't final … We cannot foresee magically all future moves carried out on the subtree.\"* **RULING.** Horizon kinds (incl. empirical: \"the subtree bottoms out in evidence edges\") are computed readings.\n\n**1ddd. Don't over-emphasize the guess** — 2026-09-19, ledger O12: *\"we shouldn't maybe put too much emphasis on what the current \\\"guess\\\" is… guess it's a question of what ends up in the UI/UX/graph-summaries etc, idk\"*. **QUESTION/OPEN.**\n\n**1eee. Relevance first, machine-proposed** — 2026-09-19, ledger O13: *\"ideally the RELEVANCE… is judged provisionally by the machine/LLM-passes or whatever and then humans can challenge it\"*; and *\"we will make do with what we've settled now\"*. **POSITION (direction), OPEN** — bears on whether a part (or an evidence item) presses on its whole.\n\n**1fff. The group premise, evidence both ways** — 2026-09-09, `the-group-premise-evidence.md`: *\"I'd also like us to explore and research evidence both for and against this group premise, start with reading up in the corpus, then do a bit more online research.\"* **Instruction** (method: corpus first, then outside evidence, both sides).\n\n**1ggg. Verbatim-preservation habit** — 2026-08-29, CLAUDE.md: *\"Can you make it a habit to preserve my verbatim phrasings when you note rulings?\"* **RULING (procedural)** — why this report distinguishes his words from paraphrase.\n\n**Not the founder, but load-bearing evidence testimony in the corpus**: Daniel Lindqvist on his hand-mapped file (2026-09-22): *\"Det låter som att jag menar att jag vet saker som enbart är antaganden\"* — the map had stored hypotheses as `observed` (0.85 vs 0.55 confidence). The ontology gap (no epistemic status can say *this is open*) is **the founder's call, not taken** (`mapping-a-structured-position-on-ai-governance.md`; SESSION-HISTORY Session 33).\n\n---\n\n## 2. Convictions and principles that constrain how evidence may be handled\n\nEach is canon (CLAUDE.md One-Breath blocks) unless marked. What it forbids or requires for evidence is stated after the dash.\n\n1. **Nothing is permanently atomic** (canon; Self-Similar Decomposition DECIDED Apr 2026). — Evidence is claims; every evidence property (`heterogeneity`, `sample_size`, `methodology_tier`) is an unminted claim. No evidence type may be treated as bedrock. Floor is **separability**, not contest.\n2. **Nothing is permanently terminal / read, never set** (RULED 2026-09-15). — An *empirical* horizon is a handover to an evidence subtree, not an end; the kind is a dated, conservative reading over the arrangement.\n3. **Contributions are added, never overwritten — existence, not weight** (ratified narrowed 2026-08-27). — No evidence is deleted by disagreement; conflicting evidence coexists; but the strength aggregate *is* brigadeable and must be defended separately (provenance, attribution, sameness-merging, the `necessary` ceiling — none tested against a motivated stakeholder, ledger D9).\n4. **Propose, never assert**, with the **confident-housekeeping** carve (RULED 2026-09-13). — Machine judgments on contested ground (methodology tier assignment, scheme, terminus, relevance of a part) are proposals; machine-written unconfirmed links failing a confident re-check are superseded automatically, old link kept visible.\n5. **The machine may act on the map; it may only propose about the territory** (`soft-canonical…` Aug 26 governing principle). — Merging duplicate evidence nodes is map-work; judging what evidence shows about the world is territory.\n6. **Bare judgments are telemetry; there is no ratify act** (agreed/blessed 2026-08-31). — No vote or click can raise an evidential assessment; only a contribution that argues for itself can. Assent = examined-no-objection telemetry. Humans vote stance only (P2 Option B, 2026-03-31).\n7. **Fallible self-knowledge** (1ff; conviction-grade). — First-person testimony is incorrigible about the felt datum, fallible about causes, counterfactuals, satisfiers. Self-interpretations get no elevated credence (C4 RULED); only the present refusal to endorse is incorrigible.\n8. **Analysis may go anywhere, and owes attunement in proportion to depth** (RATIFIED 2026-08-25). — Evidence-work may enter the personal domain; premature verdicts are the only prohibition; machine auto-research into claims about someone's own body/spirit waits on consent (kartpaus doc's personal-domain carve).\n9. **An act's value is not its rung** (re-pointed provisionally 2026-09-03). — Impact testimony and question-setting (positional evidence) are the highest-value human contributions and must never sit behind expert affordances.\n10. **Positional contributions are verified positionally** (Claude, 2026-09-03, answer to his belonging question; unruled as a rule). — Only people sharing a position can check testimony from it; testimony enters via `witness_testimony` / `position_to_know` schemes and their CQs.\n11. **Every published count labels its provenance** (canon). — Asserted vs minted vs reported; evidence counts included.\n12. **If a system cannot confess failure, it will report success** (canon). — Evidence instruments need confession channels (UNEXAMINED as no-number, `undecided` routing, unrouted fraction reported). Candidate: confession must pay (unruled).\n13. **A guard that never fires looks exactly like a guard that works** (canon). — The methodology ladder is the live example: in `models.py`, read by no production badge.\n14. **The Keynes split** (shipped `ff42b21`, 2026-08-17; UNSETTLED band founder-RULED 2026-08-23). — Direction and weight are separate; a reading with no weight has no direction; weight is a count and does not saturate. Attached ≠ scored.\n15. **Checking that keeps pace with persuasion, not faster decisions** (canon). — Evidence verification must scale at persuasion's rate; throughput is the persuader's mechanism.\n16. **Machine reasoning's quality is a function of harness and grounding** (RATIFIED 2026-08-30). — The ceiling is coherent absence and frame lock-in; evidence the corpus silently omits is invisible to the machine.\n17. **Never ask a mind to audit its own frame** (canon); same-hand bias (residual taxonomy). — Evidence checkers must not share the generator's model family where avoidable; the two-family jury is the proposed escalation.\n18. **The oracle gradient** (`fragile-checkers…` § 4b; \"settled 2026-08-18\" for the ordering). — The empirical leg can bottom out in a checkable fact; evaluative kinds never will. Fact layer ships first *because that leg reaches the ground*.\n19. **Deliberus maps reasoning, it does not endorse conclusions; synthesizes only inspectably** (canon). — No verdict-returning \"is X true\" endpoint; the answer to a factual query is the contest (incoming attacks, methodology-weighted evidence, unanswered CQs, horizon kind).\n20. **Sameness is decided before any relation** (RULED 2026-09-13; built 2026-09-22). — Two renderings of one finding must not become support for each other; *overlaps* is judged by evidence not topic.\n21. **A no-answer belongs in a gate** (RULED 2026-09-22). — Relevant here because 70 of 197 decomposition edges are **evidence miscast as parts** (`miscast_as_evidence`); re-filing is an open ruling.\n22. **Every new pipeline pass answers \"does this flatten contestation?\"** (canon). — An aggregate source tier is the **third flattening mechanism** (averaging the parts' standing).\n23. **Know which adversary class you face; prefer payoff removal to detection** (canon). — Fabricated testimony is the saboteur's best remaining move; defense is infrastructural (sybil) and positional, not epistemic.\n24. **Every claim is a link / no CQ without both endpoints** (canon). — Evidence edges and the questions on them are always addressable.\n25. **The meta layer is the gate** (RULING, 1jj). — Evidence handling is explicitly one of the things that must be right before launch.\n\nPre-2026 ancestors: *\"Focus on trusting the process instead of the people\"* (2012) and *\"Logic is universal\"* (2012) → no expert gate (E6, unruled row but \"your instinct is the literature's position\"); no contributor-rationality scoring (E7).\n\n---\n\n## 3. Historical evolution of the evidence concept, 2011 → Sep 2026\n\n**2010–2012, conversations.** Evidence enters as the inductive \"hard problem\" (1b): statistics, evidence and induction are hard to integrate; probability for all claims \"can get hairy\". Logic is universal (1a); users vote on the \"logical strength\" of links (Sasha, Nov 2010). Ben Hyink flags that the system would work best \"with established knowledge, scientific methodology, and empirically testable theories\" and poorly across radically different priors (`simplenote-archive-analysis.md`) — the fact/value boundary in embryo. 2012 Simplenote: *\"Queues of needy args/claims\"* — evidence-gap detection as feed design (`simplenote…` line 227).\n\n**Jan 2013, the \"Consensus software\" board (FB-01).** \"Limit content to atomistic statements\", per-statement probability estimates, missing/implicit-premise identification, \"axioms established by users\" (`history.md`). Sketch 7: \"Politiskt bevis/data\" with scoring; \"axiom 100%\".\n\n**2013, the Meteor prototype.** Evidence is its own collection: `evidences` link an element to a `source` (optionally a `reference` anchor, e.g. a Wikipedia section), with `evidence_votes` (Found/Missing) producing `foundIndex`; `sources` carry Trustworthy/Untrustworthy votes (`object-model.md` § Evidence; `meteor-prototype-analysis.md` § 6). Two real debates: 3 evidence entries of 33 elements (9% coverage), each 1–2 Found votes. Evidence here is **crowd-verified existence of a citation** (\"is the claimed support actually in the source?\") plus **source trust by vote** — both bare judgments, which the 2026 corpus later rules out as telemetry.\n\n**Mar 29–31, 2026 (Session 7).** The scheme-bounded turn: critical questions are premises; evidence-as-subgraph (a meta-analysis is text, run the pipeline on it); \"What is truly atomic? Almost nothing\" table (observation, study result, arithmetic, pain); the per-scheme \"evidence tracking\" sections; auto-connect over auto-generate (Claude's inversion in that document, never his position: see the correction under 1l). Correction UX ships (`0ae4eb6`): **Add evidence** creates a new claim + SUPPORTS/ATTACKS edge with human provenance; **Answer CQ** creates an evidence claim linked to a CQ polarity claim. Option B: quality from CQ evidence, not votes.\n\n**Apr 1 (Session 9).** ClaimBase unity; mother claim pattern; `EvidenceTypeLit = citation | quantitative | anecdotal | none` lives on the claim.\n\n**Apr 5 (mobile) → Apr 8 merge.** Truth-graph read layer (`6dc27e8`): `methodology_tier`, `effect_size`, `sample_size`, `study_count`, `heterogeneity`, CI bounds on ClaimBase; `METHODOLOGY_BASE_WEIGHT` (meta-analysis 0.70 … expert opinion 0.30); four evidence kinds (testimonial, statistical, logical, observational — dropped in the Apr 8 rewrite, recovered 2026-09-26); proposed sample-size CQ weighting `log(N)/log(N_ref)`; `/query` synthesis. His P2/P3 positions and the audit directive — none ruled, headings mislabelled \"Decisions\" until 2026-09-26.\n\n**July 2026.** Terminus types induced from two hand descents (2026-07-10), including `empirical` (\"dissolves with evidence\"); weighing scheme; dogfood runs. `empirical` becomes the leg of the oracle gradient.\n\n**Aug 13–18.** The August honesty turn applied to evidence:\n- Aug 13: methodology-per-source defect (a review video's ten lessons all tiered alike); heterogeneity/study_count inert.\n- Aug 14: evidence routing (`evidence_routing.py`) — link never duplicate, abstain (`undecided` ≠ `irrelevant`), fair-share discount; built, **not adopted**. Scope, not similarity (47 papers bear on \"publication record\", not on \"labor economics is the relevant domain\").\n- Aug 15: his 47-papers correction — evidence must itself decompose; always-mint reported speech (reported content is a claim; a non-propagating REPORTS link).\n- Aug 16: \"evidence\" names three objects (properties / linked claim / document subgraph); first two are the same object at different grain; **there is no atomic evidence** (Claude's answer to 1h's question); heterogeneity is a missing node, not a missing formula.\n- Aug 17: badge fix (`5641547`): incoming only, signed by polarity, QEM energy; 1,567 attacks had counted as supports; badges 3,828 → 427. Keynes split (`ff42b21`, `evidential_weight.py`). \"Adding evidence cannot move a badge\" — an unexamined edge sits at 0.5 and contributes zero; attached ≠ scored.\n- Aug 18: fact-checking correction (1y); dispute structure around factual claims is non-commodity; fact layer first; `overall_quality_score` removed rather than calibrated.\n\n**Aug 21–22.** Evidence division (1bb): confirmation holism (Duhem–Quine); sum vs product vs min; `necessary` as the conjunction bound; correlation-exposure confession figure proposed; sheaf-gluing failures as findings; falsifier = strength distributions dominated by reasonable routing variation. Stress suite: minting a property buys contestability, not weight. Staleness daemon + `invalid_at` with grounds (the doctor's note ages); his challenge that the detector's inputs are judgment-laden.\n\n**Aug 25–31.** Fallible self-knowledge; no stop signs; kartpaus research-feeding (machine finds research, feeds draft tier). Residual-error taxonomy: *asymmetric ingestion becomes asymmetric strength* (an absent premise accumulates no attacks). Padding ruling (`necessary` default, option C); the backfill's three-way output adds `miscast-as-decomposition` — telling **parts from evidence items**. Weakest-link research: variety-of-evidence failures (a known shared source becomes a node, never a blind discount). No ratify act.\n\n**Sep 2–5.** Derivability ladder: the machine derived every consideration commenters raised on 13 arguments; what it could not supply was **evidence from a position** (what placement feels like to a placed child). Positional evidence; impact testimony via witness/position-to-know schemes; single-source ceiling and positional tiers proposed; sole-witness label.\n\n**Sep 13–22.** Impact testimony named; sameness-before-relation built; the keeper judges *overlaps* by evidence; decomposition-edge census: 76% of 197 not part-of relations (⚠ 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), **70 evidential relations filed as parts**. Terminus → horizon kind, read never set (empirical = \"subtree bottoms out in evidence edges\"). The argued-requirement cap (relevance strength, never existence). Daniel Lindqvist's epistemic ladder flattened by the map; no status for *open*.\n\n**Sep 26.** Arc P: April 5 recovered; methodology tiers declared unratified and inert (no badge reads a tier); `/query` crash fixed (claims now grouped borne-out / cut-against-or-split / nobody-has-weighed), pushed on a branch, not merged.\n\nArc of the concept in one sentence: **a crowd-voted citation link (2013) → a black-box attachment he refused (Mar 2026) → numeric properties on a claim (Apr) → claims all the way down, with direction and weight separated and every evidence property a node-to-be (Aug) → a web managed by linking rather than divided (Aug 21) → with the scarcest evidence being positional, testimony checked only by those who share the position (Sep).**\n\n---\n\n## 4. Open loops, unruled decisions and TODO items touching evidence\n\n### Ruling ledger (`docs/research/session26-open-rulings.md`)\n- **P2** same evidence → same probability (his position, OPEN). **P3** fuzzy-subjectivity fear, not yet a rule (OPEN). **P4** methodology-tier baseline: keep / mint as claims / drop (OPEN — decision; unratified, inert; stated direction is to mint methodology as a claim). **P6** user estimate as decomposition invitation (\"What do you know that the graph doesn't?\") (OPEN, Claude's, unbuilt). **P7** does Deliberus ever output a probability — (a) never, (b) only via the oracle bridge, (c) user estimates as failure detection (OPEN — decision). **P9** no copout axioms = property of the graph, usability of the response (OPEN). **P10** GRADE-like certainty from existing fields (OPEN; overstated). **P11** Rorty/Peirce; the *solidarity* reading of what a badge measures never developed (OPEN). **P12** concept-sense auto-merge (OPEN). **P13** confirm or mark superseded. **P14** when the graph has nothing: say so, vs background knowledge with a caveat — reversal never marked (OPEN — decision). **P15** each query as a convergence test (OPEN). **P16** weight CQ energy by sample size (OPEN; replication half superseded).\n- **B16** low implicit-premise rate not by itself evidence of under-extraction (OPEN). **C1–C3, C5–C7, C9, C10** (human-judgment analysis; C3 four acts incl. impact testimony OPEN). **C8** rename `confirmed` → `ratified` flagged superseded, wants re-ruling. **D9** weight defenses untested against a motivated stakeholder (OPEN). **E6/E7** no expert gate; no contributor-rationality scoring (OPEN). **H6** descent moderates in proportion to closeness to a checkable fact (OPEN); **H7** flat result = evidence against the enum's cut; **H14** position slider (OPEN — decision). **I10** impact testimony as the concession channel (OPEN). **O4/O7/O8/O9/O11/O12/O13** horizon readings, bucketing by who could settle (the world = empirical), dated readings, lead with the record, relevance from a machine proposal (OPEN).\n\n### `TODO.md`\n- § April 5 recovered: merge/deploy branch `claude/truth-graph-evidence-system-akm9a`; **a tier on an edge reads as examined when nothing was examined** (precondition on P4 \"keep\"); **quantitative fields arrive only by accident** (1 of 57); first-take screen drops `unexamined` claims; **the oracle bridge is specced nowhere** (P5, P7); thesis vs wager in headline copy.\n- § Long-form sources: **move methodology from source to claim** (open); **is flat `meta_analysis: 0.70` too generous** (57% dependent effect sizes, 53% analysed as independent); **mechanistic evidence has no tier**; **should an aggregate source have an independent tier at all**; **publication bias, replication status, funding conflicts have no representation** (each naturally a CQ).\n- **Mint the evidence properties as claims** (2026-08-16, open — cheapest first half of evidence decomposition). **Decompose the evidence, not only the claims** (the missing evidence-decomposition pass).\n- **Adopt fair-share discounting BEFORE evidence routing ships** (ordering is the whole decision; founder decision, changes every strength).\n- **⚠ Evidence-alignment question before building more decomposition** (decompose-then-verify degrades without per-sub-claim evidence). **Recomposition check** (coverage).\n- **Rule on re-filing the 70 `miscast_as_evidence` edges as support links** (a ruling, moves strength).\n- **Methodology damping** (a tier reaches a claim damped — decide if right). **Adding evidence cannot move a badge** — P4 copy vs carrying the user to the second step (UX, open). **Bridging feed does not read the badge** (founder call, still open).\n- **Provenance-split strength readout** (machine-derived vs human-engaged); wire `count_safe_summary` into `/graph/stats`; provenance for non-human authors.\n- Agents: **fact-layer-first, dispute-shaped MCP tool set**; group-attributed value positions as the intermediate rung; **what does an AI reader owe the graph** (founder decision); substrate-run pre-registration; agent shipping gated on the strength layer.\n- Machine-only path to \"the whole is established\": fill derivable requirements **only from graph evidence via auto-connect/evidence routing, never from model opinion**; explicit badge label for **ABSENT POSITIONAL EVIDENCE**.\n- **Single-source ceiling** and **positional tiers for testimony** (awaiting ruling; amended 2026-09-05); **position grouping + sole-witness label**; **two-sided derivability test** with people from the affected positions.\n- **Simulated-lens pass** (simulate to ask, never to testify; `simulated` provenance; cap-only) — unruled. **Backfill lived experience from online testimony** (his idea, parked with three catches: online testimony is the head not the tail; no consent channel for private people; yields evidence with one channel missing, never attunement).\n- **Who records a support interpretation** — \"guessing retroactively would be inventing evidence, so no backfill\" (founder decision).\n- Kartpaus enrichment: designed, unbuilt; normal-use enrichment his lean.\n- Reported-speech attribution/content split (founder ontology decision); always-mint's two hard requirements (non-propagating link — done for REPORTS 2026-09-17; `count_safe_summary` still uncalled).\n- Epistemic status gap: none of the six can say *this is open* (founder's call, not taken).\n\n### Other homes\n- `truth-graph-evidence-system.md` § Open Questions **4: should methodology tier attach to evidence claims, edges, or both?**\n- `evidence-division-and-the-foundation.md` § 4–5: measure alignment on the live corpus; correlation-exposure figure beside the badge; generalize the gluing check; the routing-variation falsifier (unmeasured).\n- `open-problems.md` § 3 (can evidence be divided honestly — \"progress looks like the shared-evidence note beside every score\") and § 5 (**is step-by-step examination blind to testimony, long practice, lived experience?** — the midwife case; \"either an honest account … or a second style of examination\").\n- `weakest-link…` § 9e: attack-proofness of the requirement cap open; the up-transmission whisper (coverage-gated lift) open.\n- `supersession-and-bitemporal-lifecycles.md`: content time (`valid_at`) populated by nothing; LLM tier gated; terminus expiry open.\n- `drowning-in-claims.md` / `saturation-and-the-long-tail.md`: head/tail citation balance; per-cluster ingestion-balance count (asymmetric ingestion → asymmetric strength).\n- `the-ratification-record…`: endorsement and identity claims as ordinary factual claims — **held for later by his direction** (no verification ticks; *account-is-person* is attackable content).\n\n---\n\n## 5. What he has said he wants evidence handling to achieve\n\n1. **Full visibility of the best evidence, decomposed**: \"people and AI agents surfing around in Deliberus [should] see exactly what the best minds on our planet think in detail and what the best evidence shows - evidence that also is decomposed into its constituent parts/arguments\" (1h).\n2. **Meta-analyses and other aggregates decomposed \"as far as possible — maximum granularity that is helpful to the rest of humanity\"** (1i); long-form reviews analysable \"both as a whole … and as its constituent parts\" (1w).\n3. **Credibility and expertise decomposed** into checkable sub-questions (1j) — expertise is never a black box.\n4. **Evidence per domain for generalizing claims**, with mother-claim strength flowing up \"transitively/programmatically\" (1n); and **honest, well-calibrated division and aggregation that honors fuzzy domain boundaries** (1bb).\n5. **No subjectivity moving assessments without decomposition, analysis and cross-checking** (1r); ideally the **same evidence → the same assessment for everyone**, while admitting it may be irreducibly subjective (1q).\n6. **Evidence itself decomposes** — the 47 papers are not one unit; everything relevant in the graph bears on them (1v).\n7. **Agentic fact-checking as infrastructure, alongside values-checking, and sooner** (1y); agents check facts, conclusions and reasoning via an MCP/service (1y context, 1m).\n8. **Machine research feeds the graph**: in live tests \"find as much research of high quality as possible and feed everything into the graph\", unstated premises scaffolded, weighing questions tentatively answered \"as faithfully as possible according to our reading of the source\" — and, capped, in normal use (1ee).\n9. **Offload the work of epistemic virtue to the substrate**, with humans invited, never conscripted, to fix anything (1aa); a harness producing \"always-grounded\" reasoners from \"the most accurate priors, evidence, inferences and analogies that humanity can muster\" (1nn).\n10. **Human attention spent on what only humans can do** — a triage feed for machine-uncertain discernments and human-reserved error classes, not re-checking what is already confident (1mm, 1yy).\n11. **No motivated voting on anything but stance** (1mm); human force only through contributions that argue for themselves.\n12. **Lived experience and impact brought in** — \"Very important\" (1rr); impact testimony named and valued (1xx); and a conservative design that \"uses humans to fill up the tail faithfully but has a healthy measure of distrust against both LLMs and humans\" (1vv).\n13. **Humans are fallible about causes of their experience**, so testimony about causes is ordinary, challengeable evidence (1ff).\n14. **The empirical case stays open**: evidence is a subgraph dealt with in depth, never a terminus (1aaa); readings of settledness are provisional and not over-emphasised (1ccc, 1ddd).\n15. **Get evidence right before launch** — named in the meta-layer gate (1jj).\n\n---\n\n## Caveats on this report\n\n- Dates for Session-7 quotes are the doc's date (2026-03-29) for a Mar 29–31 session; exact turn dates are not recorded.\n- 1b's \"icky issues\", 1v's 47-papers correction, the Keynes \"stringent fix\" request, the staleness challenge and the Option B decision are **paraphrase — his words not recorded** in the repo.\n- 1d is Tom Boniecki's, not the founder's (checked).\n- Commits read in full: `3c7f82d`, `0ae4eb6`, `6dc27e8`, `a9d2c8c`, `970346b`, `b8fb65c`, `5641547`, `ff42b21`, `0c2b121`, `730c0c3`, `792794e`, `37f251c`; `6470761`, `833a1e0`, `f090d86` (bodies empty or brief, TODO-only).\n"}