{"path":"research/the-number-in-the-world-and-the-repertoire.md","content":"# The number in the world, and the repertoire: what the probability reading needs, who supplies each piece, and in what order interventions come\n\n*2026-10-15. The founder, on the number note's pick (examined-case confidence, best reason as the\ndefault, probability as the limit): \"I do want R1 I believe because I want the number/strength to\ncommunicate to a stranger (who hasn't taken in any of the surrounding/preceding/resulting\nargumentation EXCEPT for reading a singular claim and its strength/probability/truthiness) to be\nable to trust that that is the strength/probability/truthiness IN THE WORLD of the claim in\nquestion. Do you understand?\" And: \"What else does R1 require? … I have explicitly said I want\nsensible priors as part of Deliberus in the past, right? What did the earlier work on priors say\nabout this? Is there any principled reason why we couldn't have both of these requirements?\" And:\n\"'Removing the payoff for adding' is a mindless intervention and should be lower priority than\nsimply reflecting on what is going on and trying our best to get it all right … Let's come up with\na priority hierarchy of suitable intervention classes that are philosophy-first.\" And, mid-turn, on\nattention and the corpus: whether loading more of each note into the always-loaded file or the\nlarge context window would give the interconnection of everything he has been missing. The companion\nnote of the same day is `retrofitting-the-reasons-channel.md`. Nothing here is ruled.*\n\n## 0. The short version\n\n- **Yes, understood, and the honest form of it is two numbers, not one.** A number computed from\n  the record can promise a stranger two things: that it is *calibrated* (of claims shown 0.8 that\n  later resolve, about eight in ten come out true), and that it carries its *weight* (how much of\n  what would bear on the claim has been examined, and how evenly). Keynes showed in 1921 that\n  weight is a dimension apart from probability; one number cannot carry both. The badge's bands\n  already carry a crude weight. So \"the probability in the world\" is reachable as *the probability\n  given the record, shown with the record's weight, checked against resolutions* (§ 1, § 3).\n- **R1 requires seven things, and the corpus has designed six of them.** Rivals sharing one\n  partition, honest starts as reference-class claims, independence between reasons, a per-link\n  likelihood, a joint computation, calibration against resolutions: all designed, some built. The\n  seventh, that the record is a fair sample of the world, is the open one and no reading escapes it\n  (§ 1).\n- **Both requirements at once, yes.** A coherent assignment and non-indifference priors are one\n  object, a Bayesian network over the claim graph; the starting-values note's recommendation 1 and\n  recommendation 9 compose. The only principled obstacle, that the system would be asserting a\n  belief, is answered by *the prior is a claim* (§ 2). His words on this go back to 2011, 2012 and\n  2013 (§ 2).\n- **The pick stands as today's operating number and R1 is the target, reached claim by claim.** A\n  claim whose start, rivals and reasons' independence are written down reads as R1 today under\n  Q37's rules; \"limit\" in the number note meant that, and \"target\" says it better (§ 3).\n- **A philosophy-first hierarchy of intervention classes**, seven rungs with culture under them,\n  and one measured demonstration that payoff removal ahead of the distinction removes the wrong\n  payoff: the 24 reasons filed as parts sit under min today, so adding a reason to those claims\n  lowers them (§ 4).\n- **On attention: more in the window is the cheapest-looking lever and the weakest.** What makes\n  a session connect everything to everything is which whole notes are in view at the moment of\n  writing and what keys them, not how many one-liners sit in the standing file; the corpus graph is\n  the key, the window is the budget for whole notes, and the standing file's job is the few lines a\n  session would get wrong without (§ 5).\n\n## 1. What R1 requires, item by item, with the instrument that supplies it\n\nR1 as the number note defines it: *the probability the claim is true, given everything recorded*.\n\n| It requires | Why | Supplied by | State |\n|---|---|---|---|\n| 1. **Rivals share one partition** | three rival explanations cannot each be 0.5 | the keeper's `opposes` (built, routed to an attack candidate, not stored as a partition); the weighing questions' *option set* (built as a critical question); the horizon reading's *option sets cross-attacked* (built; reads 0 on the live graph); normalisation over marked rival sets (`starting-values-and-the-problem-of-the-prior.md` § 6.4, recommendation 9) | pieces built; the partition as an object designed |\n| 2. **A start that is not indifference** | 0.5 for all is incoherent as a prior over all claims (Q35) | the start as a reference-class claim, computed from the class's own subgraph, never the claim's; stored as rate plus strength; nested kinds pooled; dated and scored (same note, § 6.1 to 6.3, § 10.3, recommendations 1 to 7); who asserted a claim is its first evidence, not its start (§ 6.2) | designed; § 9 question 1 is the root and is unruled |\n| 3. **Independence between reasons** | a joint distribution, not two marginals | the keeper's `same` / `contains` / `overlaps` for content; source nodes and channels for origin (Q4); the lift as a claim whose critical question is independence (Q37); the companion note § 4 and § 5 | two thirds built, not run on sibling reasons; origin unbuilt |\n| 4. **A likelihood per link** | how much a reason moves its claim when it holds, and when it fails | today: relevance × (standing − 0.5), which is Jeffrey's rule with every start at 0.5 (`strength-calculus-literature/bayesian-and-informal-logic.md` § 2.1); proposed: two conditionals stored, the number computed (`relevance-and-the-two-factors.md` § 6); honest starts change one symbol, acceptability − start (§ 6.8) | one factor built; the pair designed |\n| 5. **A joint computation** | a probability over all claims is a fixed point, not one pass | Potyka's semantics is a fixed point and the code runs one pass; the live argument links hold no two-cycles (measured 2026-09-28), so a single pass over a directed acyclic graph is exact today | adequate until cycles appear |\n| 6. **Calibration** | a number nobody checks against the world is a report, not a probability | claims that resolve (replications, events, retractions) score the starts and the number, per kind and per proposer (starting-values recommendation 7; ledger P5's oracle bridge; `forecasting-room-at-the-top-and-deliberus.md`) | designed; *calibration starvation* named: most claims never resolve |\n| 7. **The record as a fair sample of the world** | P(claim given the record) is P(claim in the world) only if what was recorded was not chosen | nothing can test it; the search can be recorded (moves tried, standing questions, ingestion balance, the provenance-labelled count): `world-completeness-and-its-relatives.md` § 0, § 3 | open, by the one law: absence of a record is not a record of absence |\n\nRows 1 to 6 are owed work with a design. Row 7 is why even a finished R1 carries a second number.\n\n## 2. Priors: what the earlier work said, and whether both can hold\n\n**His record.** The starting-values note (2026-09-30 to 10-03) was written because he said it:\n*\"every claim in Deliberus will ultimately have a subgraph which is equivalent to the already-guess\nin the probability/Bayes perspective - no? This has always been my intuition for Deliberus\"*\n(ledger Q35); and earlier, ledger P2: *\"if there's inductive evidence for any claim, that's what\ndetermines the probability assessment\"*. His 2013 board says *\"Probability estimate for each\nstatement is calculated\"* (Proposal FB-01, January 2013; `docs/history.md`). In January 2012 Curtis\nwrote that probability estimation across all claims *\"can get hairy\"*, which is requirement 1 of\n§ 1 named early; in 2011 Ben Hyink raised *radically different priors*, which is the\nprior-dependent state below. So yes: sensible priors are his position, on the record since the\nproject's first years, and the note of 30 September is the corpus's answer to it.\n\n**What that note answered**, in one paragraph. There is no evidence-free prior; a start is always\nrelative to a reference class; every field that built evidence-to-number software converged on\none move, a visible, contestable starting value taken from a reference class and kept apart from\nthe evidence. Deliberus's native form: **the start is a claim too**, a reference-class claim (*hypotheses\ntested in this field turn out true one time in ten*) whose own strength is computed from its own\nsubgraph, joined to the claims it applies to by a membership link anyone can attack. It cannot\ndouble count (the class's evidence is about the class). Who asserted a claim is its first evidence,\nthrough a source node, not its start. Kinds pool rather than compete. A start carries a strength\n(how many cases the class rests on) so a thin class yields to the claim's own evidence. And\nworldviews differ largely in their starts, which gives two readable states beside *unexamined* and\n*unsettled*: **robust** (holds under every reasonable start) and **prior-dependent** (flips), the\nsecond being both a named home for the worldview lens and a measurement of the convergence wager.\n\n**Both at once.** Coherence across rivals (requirement 1) and honest starts (requirement 2) are not\ntwo requirements in tension; a Bayesian network over the graph has both by construction, and the\nnote's recommendation 9 (*normalise starts over rival sets marked by the keeper's opposes answer or\na stated set of alternatives*) is where they meet: within a partition the rivals' reference-class\nrates are renormalised to sum to at most one. The one principled objection, that a prior is a\nbelief and *Deliberus maps reasoning, it does not endorse conclusions*, is answered by the prior\nbeing a claim: the system asserts nothing, it computes from what was contributed, and the display\nrule (§ 6.7) shows a best guess only when its start is a named, attributed claim. What remains is\nnot principle but work and two risks the note names: gaming the reference class (the membership\nlink is attackable; pooling; the robust verdict needs every reasonable start) and anchoring (a\ndisplayed start pulls later contributions, § 3.8).\n\n**His sensing that rival detection is possible.** It is, and the pieces exist: the keeper's\n`opposes` key *covers both \"cannot both be true\" shapes*, a claim and its contradiction and two\nrivals that exclude each other without exhausting the options (its prompt says to look for this one\ndeliberately); the weighing pass asks for the *option set*; the horizon reading looks for option sets\ncross-attacked. What is missing is the partition as a stored object that the start computation\nreads. That is a microtask of the closed kind (*can these both be true?*), two short items, which\nthe pass-quality note found cheap models read alike.\n\n## 3. The stranger's number: what can be promised, and the pick restated\n\nHe wants a stranger who reads one claim and its number to trust the number *in the world*. Taken\nstringently, a number computed from a record can promise exactly two things, and the second is the\none usually left out:\n\n1. **Calibration.** Of claims shown at 0.8 that are later settled, about eight in ten come out true.\n   This is testable only on claims that resolve, and it is what forecasting means by a probability.\n   The starting-values note's recommendation 7 and the oracle bridge of ledger P5 are the design;\n   nothing scores anything today.\n2. **Weight.** How much of what would bear on the claim has been examined, and how evenly. Keynes\n   (1921) separated the weight of an argument from its probability: a coin judged fair from one toss\n   and from a thousand carry one probability and different weights. The badge's bands (*unexamined*,\n   *no data*, *unsettled*, *nothing stands (N checked)*) are a crude weight already; the derivable\n   requirement space (companion note § 2) gives a finer one, *fraction of what would bear on this\n   claim that has been answered*; ingestion balance gives the evenness. A stranger who sees 0.8 at\n   low weight and 0.8 at high weight should see two different things, and today sees one.\n\nSo \"the probability in the world\" is reachable in this form: **the probability given the record,\ncalibrated against resolutions, shown with the record's weight.** No reading escapes row 7 of § 1,\nand the second number is how the badge tells the stranger so without hedging the first.\n\n**The pick, restated.** The number note said *R2, with R4 as its default and R1 as its limit*. The\nword *limit* read as an asymptote the graph never reaches; the better word is **target**, because it\nis reached one claim at a time: a claim whose start is a reference-class claim, whose rivals are in\na stored partition, and whose reasons' independence the keeper has read, reads as R1 today under the\nrules Q37 gives (the lift as a claim; convergent reasons typed as adding, Q22). Where none of the\nthree is written, the same rules read as R4, the best reason. The operating number and the target\nare one calculus at two states of the record, not two readings. Corrected in the number note today.\n\n## 4. The repertoire: a philosophy-first hierarchy of intervention classes\n\nHe asks for a standing repertoire of reusable approaches (*\"what deterministic code can possibly\nhelp with, what LLMs can help with, how to maximize reliability … how humans can help\"*) and for\nthe classes ordered philosophy-first, with payoff removal demoted from first move to last. The\ncorpus holds every rung somewhere; what it lacked was the order. One table, then the demonstration\nthat makes the order a finding rather than a preference.\n\n| Rung | Class | The question it answers | The test that it was done | Where the corpus holds it |\n|---|---|---|---|---|\n| 1 | **Ontology**: the right objects, told apart | what things are there (part, reason, line, rule, sense, origin, kind, partition) | the sorting test: can two users disagree about it and the system keep working | `what-belongs-in-the-ontology.md`; `the-kind-of-mistake-we-make.md` § 3 (every longest-lived mistake was a distinction collapsed) |\n| 2 | **Process**: the right question, in order, with the right menu | what is asked first, what a menu may not contain | the downstream menu holds no answer already given upstream; the no-answer is a gate | the four menu gotchas in the always-loaded file; sameness before relation; reason before verdict |\n| 3 | **Representation**: the answer computable from the record | what must be stored for a stranger to re-derive the number | can a reader rebuild the number from stored objects alone | written-out steps (AG14); a machine's sentence as a claim (AG12); sense per use; source nodes (Q4); the lift as a claim (Q37); the start as a claim |\n| 4 | **Deterministic code** | what follows from the record by arithmetic or id comparison | a run with a positive control; a guard verified by firing | the strength arithmetic; coverage by id; pair counts; sense-id comparison once senses are stored |\n| 5 | **Machine microtasks** | what needs reading, one short item against one closed question | reason before verdict; asked a second way where it matters; proven on faults a person found first; logged | `microtasks-in-broad-daylight.md`; `proving-a-pass-reliable.md`; `the-evals-loop-and-our-machine-passes.md`; `pass-quality-and-the-choice-of-model.md` (closed questions read alike on cheap models; a degree reads alike on none) |\n| 6 | **People** | the two reserved classes (coherent absence, the frame) and consent over one's own position; corroboration by readers of other positions before a report is loud | a reasoned contribution is a claim, a bare judgment is telemetry; positional contributions are verified positionally | `what-human-judgment-is-for.md`; `the-ratification-record-and-the-triage-feed.md` § 3; `derivability-of-missing-considerations.md` § 9; `checks-on-machines-and-people.md` |\n| 7 | **Incentives**: payoff removal, then detection | what still pays after rungs 1 to 6 | does the move still pay; publish the removal, never the detector | `fractal-scales-and-temporal-frame.md` § three classes of adversary |\n| 0 | **Culture**, under all of it | what holds where no instrument reaches | his 2026-08-30 statement: *security, transparency, integrity and a strong culture, where our convictions and entire corpus are in the very air we breathe, grown side by side with the software platform* | the same note § 0a |\n\n**Why payoff removal is rung 7 and not rung 1, demonstrated.** The weakest-part rule of 2026-08-31\nwas a payoff removal: splitting into twelve parts stopped paying. It was made with Dr Smith's\ncredentials as the example of *jointly required* parts, and those were reasons. The rule was scoped\nto the parts channel, so the reasons channel escaped min; but the 24 reasons people filed as parts\n(companion note § 1) sit under min today, and for those claims adding a reason lowers the claim.\nThe payoff that was removed was the one a reason-giver had for adding a good reason. **A payoff\nremoval applied before the objects are told apart removes the wrong payoff**, and it does so\nsilently, which is the mark of every rung-1 mistake. He is right that it is mindless: that is its\nvirtue against an adversary who adapts (rung 7 is where the mindlessness is wanted, since a\ndetector an adversary can read is armed by publication) and its vice everywhere above.\n\n**What the corpus already said that this does not overturn.** *Prefer payoff removal to detection*\n(the three-classes note) orders rung 7 internally and stays. *Anti-gaming is the core work* (the\nsuccess prior) says rung 7 is not optional, and stays. What this table adds is that rung 7 is\nreached last, and that most payoffs are gone before it is reached: with lines told apart and max\nacross them, the reformulation attack has nothing to buy; with the part test run, the padding attack\nhas nothing to buy.\n\n**The queue for reframe reports**, his *\"perhaps first appearing in a queue of incoming such\nreports\"*. Under *contributions are added, never overwritten, nothing needs permission to exist*,\na report that a region is framed too narrowly exists the moment it is filed. What may wait is its\n*prominence*: shown loud after readers who do not share the reporter's position have said it bears,\nwhich is Community Notes' gate (cross-tribal agreement) applied to the one act the corpus calls the\nmost valuable and the hardest to check. Existence at once, loudness after corroboration, and the\ncorroborators' positions on record so that a report corroborated only by allies reads as such.\nFiled as a proposal (ledger AK2); he said it needs deeper thought, and it does: a reframe report\nfrom a lone position is exactly the case positional verification cannot reach.\n\n## 5. Attention, the corpus graph, and whether more in the window is the lever\n\nHe asked, mid-turn, whether loading more of each note (its contributions, counter-intuitive\ninsights, building blocks, analogies) into the always-loaded file or into the large context window\nwould move the sessions toward the interconnection of everything he has been missing, given that\nattention relates everything to everything. The Quanta piece he woke into (Wolchover, 2026-10-12,\nread as a summary) says what attention does: for each next word, weight every word in the window by\nlearned relevance and sum; the 2017 architecture kept only that; and its co-author now says it is\ninsufficient, since people converse without having read the whole internet. Thinking clearly about\nhis question means separating three levers that the phrase *more in context* runs together.\n\n1. **What is in the window.** Attention can only relate what is there. Two ways to put a thing\n   there: standing (the always-loaded file, paid by every session and every helper; measured at\n   about 323k tokens for a general-purpose helper's first request, against 38k for one that skips\n   it) and per task (whole notes opened for this work). A one-liner in the standing file is not the\n   note; it is a pointer the session decompresses from memory at the moment of use, and the\n   kind-of-mistake note's finding is that decompression from recollection regresses to the template,\n   which *is* the muddle. So cramming insights as one-liners trades the failure of not finding a note\n   for the failure of fusing it with its nearest neighbour. The ruled filter for the standing file,\n   *wrong by default*, is the principled version of this: a line stays only where a session lacking\n   it would assert the opposite; everything else is retrieved whole when it bears.\n2. **What keys it.** Attention is keyed by the sentence being written. A note in the window that the\n   current sentence does not key to is as absent as one on disk; the lost-in-the-middle measurements\n   the global file cites are this. The corpus graph is the keying mechanism built for this corpus: it\n   answers *what bears on this note or row* from what readers and sessions judged, and the hook hands\n   that out when a note is opened. The gap he feels is that the hook fires on opening a note and not\n   on a subject arising in the session's own writing. The mechanism that would close it is a check\n   at reply time: run the graph's `search` on the reply's own key claims and hand back the notes that\n   bear before the reply ships, the way the connection watch already runs on his messages by meaning.\n   That is a hook, cheap, and it fires on the session's sentence, which is where attention is keyed.\n3. **Whether the retrieved thing is held apart.** This is the analytic-discipline half. A whole note\n   with its definitions in view is an artifact; a line is a recollection. The hammering that worked\n   became a test (the kind-of-mistake note § 6); the interconnection that works will be the graph's\n   judgments in view, whole, at the moment of writing.\n\nSo: the large window changes the *budget* for lever 1 per task (more whole notes can be opened for\none piece of work) and changes nothing about lever 2, which is the one he is missing. The standing\nfile is the wrong place for interconnection, since it is paid everywhere and keyed by nothing; the\ncorpus graph plus whole-note reads plus a reply-time search is the right one. His own sentence about\nattention is the argument: attention relates everything *in the window* to everything in the window,\nkeyed by the present word, and a corpus is interconnected only when the right whole notes are in the\nwindow at the right word. The Quanta co-author's second thought is the same point from the other\nside: intelligence is not having read everything, it is bringing in what bears.\n\n### 5b. The same afternoon: his correction, and the two sessions composed\n\nHe read § 5 in another session and corrected it, in words the other session filed beside AK3\n(`TODO.md` § RELEVANCE): *\"if Claude makes the same mistakes every time I discuss something which\nwould benefit from the distilled learnings and ideas from that note (or from all of the (most\nimportant to what's left to spec/build) notes) then it's better to always include it rather than\nfor me to have to remind Claude about everything that could be relevant each time we stand before\na decision.\"* And, asked to think more deeply before the walkthrough: *\"Think more deeply about\nthis before making the walkthrough also.\"*\n\n**The correction holds, and it does not reverse § 5; it names the condition § 5 left out.** The\ncomparison in § 5 (a whole note pulled in when it bears is paid once; a standing line is paid by\nevery reply) assumes the pull happens without him. Today it mostly does not: the search of lever 2\nis unbuilt, and the hook that hands out what bears fires on a note being opened, not on a subject\narising. Until the pull is automatic, **he is the retrieval mechanism, and that is the one cost\nnothing here may charge him.** Measured from the connection hook's own record: in this project it\nhas judged about one message in six to be him pointing at something that exists or joining two\nthings (32 of 174 messages it read). That rate is his reminding, counted.\n\n**So three layers, not a choice**, which the other session put the same way:\n\n1. **Always loaded, as whole notes, kept out of the always-loaded file.** A pack of the core notes\n   for what is left to spec and build, loaded at session start. Whole notes, since a one-liner is a\n   recollection (§ 5, lever 3); out of the instruction file, so a helper does not carry it and an\n   edit does not copy it into every running session (`claude_plan_usage_and_what_spends_it.md`\n   § 5). His own stance on what is loaded, the same day: *\"I don't know how to judge what part of\n   the instruction files and the always loaded memory … is redundant. Doesn't feel like it's\n   redundant at all.\"* The wrong-by-default filter (ruled 2026-08-29) still governs the instruction\n   file's own lines; the pack is a second surface with a second test, *would a session stand before\n   a decision this note has already made*.\n2. **Pulled in automatically**: the reply-time search of the corpus graph on the session's own\n   sentences (AK3). Conditional-silent, no model call, so it is a nudge and not a standing line,\n   which is why the 2026-08-31 rejection of a per-turn hook line (J14, on its token cost) does not\n   reach it.\n3. **A check per mistake that keeps coming back**: the hammering that worked (the kind-of-mistake\n   note § 6), which fires on the word and needs no memory at all.\n\n**Why all three, said once more from the side of attention.** Attention retrieves by key: the\nsentence being written is the query, and what it finds is whatever in the window carries a matching\nkey. The pack supplies the window with whole notes, so the model's own keys can reach them, and it\nfails where keys do not match (lost in the middle). The search supplies keys the model does not\nhave: the connections readers and sessions found, which similarity alone misses (measured\n2026-10-10: no ranking computed from the files finds the unexpected connections). It fails where no\nreader connected the two. The check supplies neither notes nor keys; it fires on a word. Each fails\ndifferently, which is the reason to run all three rather than to pick.\n\n**The test that settles how much goes in the pack**, proposed by the other session and sharpened\nhere with the cases named: replay five past decisions that a session got wrong when the corpus had\nthe answer, under three setups (today's · the pack · the pack with the search), and count which\nreach the corpus's answer without him. The five, each with a recorded wrong answer and a recorded\nright one: Dr Smith's credentials filed as parts (April to 2026-10-14) · the accrual school read as\nthe sum (August to 2026-10-14) · *never generating premises, only finding real ones* read as his\nposition (March to 2026-10-07) · the keeper left unnamed when independence was called\n*contributed* (2026-10-14, corrected the next morning) · the filing of his *\"make humanity more hive\nmindey\"* as declining the hive mind. Cheap helpers can run it, once his hold on readers is lifted.\nFiled as AK4; the decision is his.\n\n## 6. Open after this (proposed, unruled)\n\n- The pack of whole core notes at session start, and the replay test that sizes it (AK4).\n\n- The target restated: R1 reached per claim as its start, partition and independence are written\n  down; the badge to show weight beside the number (Q41).\n- The philosophy-first hierarchy as a standing repertoire; whether it earns a resident line is his\n  (AK1).\n- Reframe reports: exist at once, loud after corroboration by readers of other positions (AK2).\n- A reply-time search of the corpus graph on the session's own sentences (AK3; a hook, built on the\n  keeper's machine).\n\n## Cross-references\n\n`what-the-number-on-a-claim-means.md` · `starting-values-and-the-problem-of-the-prior.md` ·\n`retrofitting-the-reasons-channel.md` · `world-completeness-and-its-relatives.md` ·\n`the-kind-of-mistake-we-make.md` · `fractal-scales-and-temporal-frame.md` ·\n`the-ratification-record-and-the-triage-feed.md` · `what-human-judgment-is-for.md` ·\n`derivability-of-missing-considerations.md` · `what-agent-memory-research-says-about-agent-surfaces.md` ·\n`docs/research/corpus-graph/README.md` · `docs/history.md` (the 2013 board).\n"}