{"path":"research/how-the-project-evolved-read-from-the-record.md","content":"# How the project evolved, read from the record\n\n**Date:** 2026-11-02, read from 31 October to 2 November in a cloud session, while the production host\nwas in the van and `deliberus.com` did not answer.\n**Standing:** a session's reading, made on his ask (*\"Reflect on the project's evolution. Ground extremely\ndeeply first. Ultrathink\"*) and on his two questions the next day (*\"What's your assessment of the readiness\ngate? And what was the fourth row again?\"*). Nothing here is ruled. Three proposals are ledger rows AV4 to\nAV6. Each reading is marked as a reading; his words are quoted as the record holds them. A sentence about\na cost or a state says whether it speaks of Deliberus today or as designed.\n\n**What was read for it.** `SESSION-HISTORY.md` whole; `../history.md`; `the-kind-of-mistake-we-make.md`\nwhole; `what-deliberus-is-for.md` whole; `signposts-for-the-whole-project.md` whole;\n`the-canon-distillation-pass.md` whole; `session27-the-canon-pass-and-the-corpus-blind-spot.md`;\n`../vision.md`; `../frontier.md`; `../open-problems.md`; `what-he-held-in-the-group-2011-2013.md`;\n`what-the-walkthroughs-and-the-reviews-are-for.md`; `the-disempowerment-objection.md`; the flywheel\ncheck-ins and readings; the walkthrough *Inching toward the vision* (2026-10-27); `a-tree-per-claim.md`\n§ 2c and § 2d; the top of `TODO.md`; `scripts/log_arc.py` over the 32 days to 31 October; the commit log,\nunshallowed. **Not read:** the other three hundred research notes, except where the corpus graph's `around`\npointed at one. **And one thing this session had wrong about its own reach**: it told him `.private/` was out\nof reach, with the private repository cloned beside the checkout as `../deliberus-private`, the gotcha the\nalways-loaded file gained on 31 October and this session's copy did not yet carry (§ 6).\n\n---\n\n## 0. In short\n\n- The design was his by 2011. What 2026 added is the parts (extraction by a language model dissolved the\n  typing friction that killed the prototype), the discipline (a position moves asserted, ruled, compressed,\n  resident, in his words; every model call is recorded; a muddle becomes a test the day it is found) and,\n  on this reading, the beginning of a rejoining of the fork he took in 2014, at each link and not yet\n  across the graph (§ 2c).\n- The binding constraint moved three times: typing the structure (2012 to 2014), the honesty of the\n  instruments (March to August 2026), his rulings (late August to October). The check-in of 26 October reads\n  a fourth, the always-loaded file itself. The next, on this reading, is contact (§ 3).\n- The project's thesis was proven on its builders before it was proven on its passes: the costliest\n  mistakes on the record were muddles that failed silently, and each stopped recurring the day the\n  distinction became an artifact in view (§ 3).\n- The readiness gate of 22 October is sound where it guards the corpus and soft where it waits on his\n  feeling. By his own signpost form it has no trigger in either direction, no review date and no check\n  against ground truth, and *done-ish*, which the vision page asks to be written down before the review,\n  never was (§ 5).\n- One correction made on the way: the daily digest does count stranger page views by path. What nothing\n  reads is a visitor's path from the landing to a map (§ 5f).\n\n## 1. The long arc, 2009 to 2014\n\nBy his own dating the first inklings came in 2009 (*\"2009 is more like it\"*), by March 2010 he was asking\nfriends how to design *\"ett så rationellt forum som möjligt\"*, and by April 2010 he was coding the Cerebrate\nin Rails (`../history.md` § Genesis). By 2011 and 2012 he had said in the group, in sentences, most of what\nthe ledger ruled this autumn (`what-he-held-in-the-group-2011-2013.md`): strength is judged on the arrow and\nnot on the claim; a claim that *\"claim X is relevant to Y\"* is auto-created for each link and argued on a\nmeta level, which is today's relevance claim on the link; one argument carries many wordings; a concept is\ndefined by the statements that use it; automation first and microtasks for what is left; one claim in the\ncentre with a few inferences around it; private spaces beside a public commons; science as a rough diamond\nto be sharpened by making every inference checkable. Three positions from then he has since reversed by\nruling, and two of them he had already left by 2013: that the inference needs no statement of its own,\nthat evidence stays outside the system as a reference, and the validator ladder.\n\nThe prototype died on the one obstacle nobody could remove then, that someone had to type the structure\n(`../vision.md` § Why Now). His own diagnosis in 2012 names the other half: *\"the data model has not\nfollowed my realizations for the last year, which is kinda part of why I stopped coding on it\"*, and\n*\"Armchair philosophy only goes so far.\"* In 2014 he parted from argument graphs altogether, when a friend\nasked whether he was still on them: *\"Nej, bayesianska grafer nu, mer empiriskt fokuserat, mindre\nanalytiskt. Kausal inferens. Judea Pearl\"* (`../history.md` § Dormancy).\n\n## 2. What 2026 added\n\nNone of the three is the design.\n\n### 2a. The parts\n\nExtraction by a language model dissolved the typing friction, which is the second of the success prior's\nfour conditions arriving from outside (`fractal-priors-for-success.md`). That is why five days in March\n2026 reached a live site where two years in 2012 and 2013 did not (`SESSION-HISTORY.md`, sessions 1 to 8).\n\n### 2b. The discipline\n\nA position now moves asserted, ruled, compressed, resident, in his words (`the-canon-distillation-pass.md`\n§ 2; the rule that a ruling is recorded in his words, 2026-08-29). Every model call is written down\n(`storing-the-record-of-machine-judgments.md`). A muddle becomes a test the day it is found: the\nreason-before-verdict test, the gate-before-field repair, the readers-of-part-of-links test, the property\nguards, the retired-field test (`the-kind-of-mistake-we-make.md` § 6). Research and code land the same day,\ndoc first, in four of five measured pairs (the commit-vocabulary ruling of 2026-08-27), which inverts his\n2012 complaint that the data model lagged his realizations.\n\n### 2c. The 2014 fork rejoined (a reading, unruled)\n\nOctober's rulings make the argument graph's numbers Bayesian: a claim's prior from what it says through\nits reference class (BC1, *the claim says what it says and before anything else it has a prior*); a reason's\npull measured from its own prior by Jeffrey's rule (Q50, what relevance means); independent arguments adding\nin log-odds (Q46); independence read by rival explanations (Q40); every case moving a statistical\ngeneralization (AY1); and an observed common effect, Pearl's collider, next in the review with Codex\n(`TODO.md`, the Q48 item). The two things he parted into in 2014, argument graphs and Bayesian graphs, are\none object now, and the record shows he was near it in 2013 too: the link with its own confidence was his\nposition by May 2012 (`what-he-held-in-the-group-2011-2013.md` § 1.7), and a Bayesian network for arguments\ncame from two friends before his own line of 2013 (`SESSION-HISTORY.md`, the recordings session of\n23 to 25 October). This reading is the session's. It changes no row.\n\n**Checked against the code on the Mac, 2026-11-02, and narrowed.** *One object* overstates. What the\nrulings made Bayesian is the arithmetic at each link: a prior, a pull measured by Jeffrey's rule, independent\nreasons adding in log-odds. What a Bayesian graph does across links the code does not do, in its own words:\n`deliberus/graph/accrual.py` says that news about a conclusion does not flow back to its reasons to explain\nthem away, and the common-effect case (Pearl's collider, two causes coupled once their effect is observed) is\na model in `tests/test_combination_models.py` that the code is marked as not yet passing (Q48). So the\nsentence is: the argument graph and the Bayesian graph share one calculus at each link, and the operations\nonly a Bayesian graph has are written as models the code must still meet. By the resident rule that two\nthings are called one only when every operation applies to both (BD13), the rejoining is begun, not done.\n\n## 3. The bottleneck moved three times\n\n| Period | What bound the work | What moved it |\n|---|---|---|\n| 2012 to 2014 | typing the structure, and a data model behind his realizations | nothing then; the models of 2023 to 2026 |\n| March to August 2026 | the honesty of the instruments: silent failures reported success | the April outage, the confession principle, the strength audit, the residual-error taxonomy |\n| late August to October 2026 | his rulings: sessions' assertions hardening unruled | the ledger (2026-08-27), the canon in his words, the corpus graph, tests as laws |\n| 26 October, by the check-in | the always-loaded file, grown while no line was retired | the review of resident lines against checks (`the-canon-distillation-pass.md` § 4l) |\n| next, on this reading | contact: readiness by his eight parts, then a second mind | his call (§ 5) |\n\n**The middle row is the one the record proves on its own authors.** Every longest-lived mistake was a muddle\nthat failed silently: attacks counted as supports (March to 17 August), part-of links stored backwards (April\nto 7 October), reasons filed as parts (April to 14 October), a field named `confirmed` that nobody confirmed\n(30 March to 19 October), one sentence of Claude's standing as his position (29 March to 7 October). His\nquestion of 14 October got a straight answer, and the finding underneath it is the project's thesis turned\ninward: *the variable is whether the distinction is an artifact in view or a recollection*\n(`the-kind-of-mistake-we-make.md` § 5). The harness conviction, that machine reasoning's quality is a\nfunction of harness and grounding, was demonstrated on the sessions before it was demonstrated on the\npasses (§ 4 of the same note). The sessions grew a harness in the same months the passes did: hooks, gates,\nthe outside check, the corpus graph, and a ruling that carries its canon decision in the reply that records\nit.\n\n**The relation between him and the sessions changed with it.** March said research surfaces and never\nprescribes, which was right about authority and became the relativist shrug; 15 October blessed the\nanalytic register with his rider, *\"I'm a metamodernist at heart I guess\"*. Rulings went from paraphrase\n(six of thirty-one in his words on 2026-08-29) to quotation. The always-loaded file went from a format that\ngrew by itself (*the format was the instruction*, `session27-the-canon-pass-and-the-corpus-blind-spot.md`\n§ 7) to a fixed budget with a filter and tests. And the check-in of 26 October says the constraint moved\nagain: his ruling nearly kept pace with proposals in the measured days (98 ruled against 106 added, 22 to\n26 October), while the always-loaded file grew 11 % with no line retired\n(`flywheels-in-building-deliberus/checkins.md`).\n\n**The rhythm, by the log.** A build burst in March and April, a quiet spring, a thinking wave from July\nthat grew more disciplined each month. Over the 32 days to 31 October `scripts/log_arc.py` classifies the\ncommits as research nearly three to one against code, with more ruling markers than any month before and a\nrising correction share; the mistake record reads that rise as finding, not making, since October is when\nreaders went through the notes. Cite the command rather than these figures, which are dated as written.\n\n## 4. Where it stood, 31 October to 2 November\n\n### As Deliberus runs today\n\n- **Live before the van**: the wheel view, the institute map in both languages, the ledger as a wheel, the\n  record of machine judgments, the retired `confirmed` field, and the strength code for most of October's\n  rulings (the refuted-reason build went live on 27 October).\n- **Off or empty**: every machine pass off in production on his word (*\"Microtasks should still be turned\n  off in production.\"*, 2026-10-24); no live claim has a prior yet; the horizon reading can read nothing;\n  the keeper owes its pairs; cut-before-link is off.\n- **The word *together***: twenty people have ever signed in (2026-10-11) and four claims came from anyone\n  but him and the machine (2026-10-03). That line has not moved since March, and it is the one the root\n  purpose rests on (`what-deliberus-is-for.md` § 3, § 5).\n- **The reviews**: most ledger rows open; the constellation at fourteen ruled sentences; the canon's cluster\n  review at clusters one and two of twelve, with the lines in his own words begun on 25 October; the\n  verbosity audit unstarted; twenty walkthrough decisions open (the counts decay; the READ FIRST block and\n  the decisions register carry the current ones).\n- **The meta-layer gate, by its own criterion**: *\"The gate is until reading stops being what improves\n  it.\"* (`../vision.md` § Why Now). October's readers raised the correction share, which measures finding\n  rather than making, and by the page's own test reading is still what improves it. So the gate is not\n  discharged, and the instrument that would show it discharging exists.\n- **What the success prior names as the killer, maintenance**: one person's plan pays for everything, the\n  monthly credit expires on 6 November, the Middle Powers application is the one sent, Foresight closed on\n  1 November unsent (his word, *\"Nej, hann inte skicka Foresight\"*), and the host is on a mobile connection\n  behind which nothing outside can reach it until the relay exists (`TODO.md`, the order of work;\n  `CLAUDE.md` § Where this session is running).\n\n### As designed, and what the record says is now possible\n\n- **A number with a sentence behind it.** In March every claim sat at the middle, with every attack in the\n  graph counted as a support. Now a card's fill is a posterior, from a prior the claim earns by what it\n  says, moved by reasons by the difference each makes, with parts carrying nothing. His words: *\"The claim\n  says what it says and before ANYTHING else it has a prior, for anyone reading it wherever it appears.\"*\n- **Minds that take small steps.** One stated rule applied once, checked by code (AG14, AM14); a microtask\n  writes nothing twice (AM21); machine work as standing work (R8); and the principle that review need not\n  keep pace with a proven pass, since depth is what compounds (T1). The third mind he named on 18 October,\n  *\"more honest and autistic than most of us\"*, has a floor to stand on now (`three-minds-and-the-leap.md`).\n- **The canon in his own words**, line by line since 25 October, with his own verdict on the week: *\"I'm\n  proud of that progress and of all the new insights that have locked in place recently and the new canon\n  distillations that are ongoing, and deeply hope it'll result in snowball/flywheel effects in the coming\n  weeks and months.\"* (`the-canon-distillation-pass.md` § 4j).\n- **A strange loop, small and live.** The keeper's prompt is rendered from ledger rows (AM8), so a ruling\n  written in the docs changes how a pass reads the graph. The two larger closures, the project's rules\n  argued as claims and a prior about Deliberus stored in Deliberus, stay held on his word until the reviews\n  are nearer done (`a-tree-per-claim.md` § 2d; `claims-and-code-one-loop.md`).\n- **The demo on 21 November** is the first show-don't-tell on the record, and his own requirement for it\n  names the fold, since *\"most people don't think in individual inference steps\"* (`TODO.md`, item 2).\n\n## 5. The readiness gate, assessed\n\n### 5a. Two gates stacked, and the fourth row\n\nThe inner gate is the meta-layer gate his sentence set: *\"I NEED to get all the philosophy, epistemology,\nontology, process, method, evidence, logic, psychology, all the meta stuff right before we launch this, or\nit'll fall like a house of cards in a puff of wind.\"* Its exit criterion is reading no longer improving\nthings, and what lifts it is a table of four questions with four answerers (`../vision.md` § Why Now; the\ndivision Claude's, approved 2026-08-28 after his correction that deep readers can audit the frame, the\ncriterion itself not a ruled position):\n\n| Question | Who answers it |\n|---|---|\n| Is reading still overturning things? | the discovery rate, corrections per period from `scripts/log_arc.py` |\n| What does all this assume? | anyone reading from outside the frame |\n| Do the arguments actually hold? | a deep reader with patience for the corpus |\n| Is it enough? | the founder, and nobody else |\n\nThe outer gate is his readiness of 22 October, in eight parts, which governs how we work, namely whom we ask\nto try the platform (`signposts-for-the-whole-project.md` § 4). His words: *\"It should be self-explanatory,\nit should match my vision closely enough, and it should be interactive and with a low enough error\nfrequency and I don't wanna feel that the graph or features or rules of Deliberus muddles things to the\ndegree that I still feel they do. We need to get all microtasks / standing work reliable enough and humming\nalong daily, we need to extract more sources and fill up and saturate the graph to show what this can be and\nhow it can change things around.\"* The signposts note sorted the eight: six are goals, the condition of a\ntrigger, and two are signposts in the planning sense, that a newcomer finds the way in and that the\nexperience speaks for itself.\n\n### 5b. The eight parts, as Deliberus runs today (read 2026-11-02)\n\n| His part | Today |\n|---|---|\n| settle the ontology and UX | the ontology core settled in October, with the theory revisit of the refuted-reason build due after the move; UX partly, the wheel live and folds at part one |\n| self-explanatory; few get through the landing | not met: the landing opens on 45 stars, 31 of them drafts, and the overhaul is last in the order of work |\n| a wow factor on landing and maps | partly, his judgement: the wheel, and the demo prep week booked |\n| match my vision closely enough | his alone; *\"inching closer\"* on 27 October |\n| low enough error frequency | unmeasured: October's faults were found by sweeps, and nothing counts what a visitor meets |\n| nothing muddling | the words made exact on 25 October; the interface-words audit waits on him |\n| standing work humming daily | furthest off: two of twenty-five registered passes standing (2026-10-26), every pass off in production on his word |\n| fill and saturate the graph | not met: extraction paused on the site, the two large maps made by hand |\n\n### 5c. Three findings\n\n1. **The gate is right in its core, for the project's own reasons.** A corpus cannot be un-thought, frame\n   lock-in makes re-framing dearer with every extraction (`the-residual-error-taxonomy.md` class 7), and\n   warm feedback is near-worthless as evidence, so testers liking it would lift nothing anyway (the\n   DeepMind facilitation study, the standing threat model). A null from a door that offers no handle would\n   indict the door and not the substrate (`structure-versus-scale.md`; the signposts note's third guard).\n   None of that is perfectionism in disguise, and the discovery rate has not fallen.\n2. **By his own signpost form it is not operational.** Of the ten fields a signpost must carry (AV1), it\n   lacks the three that bite: a trigger in either direction (field 6), a review date with a home (field 8)\n   and one check against ground truth (field 9). Its two true signposts have no instrument that answers\n   their question (§ 5f), and *wow* has no reading but his. So the gate holds or lifts on an impression,\n   which the note's own first draft warned of. The vision page names the remedy: *\"done-ish gets written\n   down before the review, while it can still be a surprise, rather than fitted afterwards to wherever the\n   energy ran out.\"* Searched on 2026-11-02, *done-ish* is written nowhere but on that page and in two\n   readers' reports, one of which proposed on 2026-10-11 that the discovery-rate criterion be defined\n   before it is read (`corpus-graph/readers/2026-10-11-pages-open-question-three-vision-pages-as-a-reader-returned-it.md`).\n3. **The binding part is standing work, and the ruled order already puts it first.** Six of the eight\n   parts depend on it: the graph fills only if passes run, errors fall only if passes are proven, and\n   muddle leaves only if the keeper and the splitter write what the rulings say. Its dependency is the one\n   he set: *\"Microtasks should still be turned off in production … after theory is more set in stone\n   grounding the prompting for the microtasks.\"* That is items 5 and 6 of the order of work, not a new\n   item. The fill part also waits on money for calls, which he named as today's brake\n   (`whose-tokens-pay-for-mapping.md`).\n\n### 5d. The demo as a trigger outside the gate\n\nThe demo on 21 November is a dated trigger that does not wait on the gate, and the meta-layer gate's own\ntext wants exactly that: *\"the smallest contact that answers what only contact answers\"*, then wider. In\npractice the gate already runs per circle: a housemate read the AI-risk maps, a friend read his own map and\ncorrected its typing, members of a community chat saw two maps and one asked for shapes before detail\n(`signposts-for-the-whole-project.md` § 4; `how-a-map-is-cut-and-what-it-shows.md`). So *asking people to\ntry* has rungs, a friend side by side, a room, the public, and the first rung needs a door that offers the\nhandle (P20, P21 in `../ux-principles.md`) rather than all eight parts. The demo's question is substantive,\nwhich the matching rule wants; its room is the insider register, which that rule puts second\n(`CLAUDE.md`, *Match test audiences like test material*), so it tests the wow factor and not whether people\npoint at claims.\n\n### 5e. Proposals, his to rule (ledger AV4 to AV6)\n\n1. **The gate is per circle** (AV4): friend, room, public; the eight parts gate the public rung, and the\n   friend rung lifts on the door alone. This is his own correction of 22 October read one step further:\n   talking is not held, only asking people to try, and asking has rungs.\n2. **Done-ish written before the room** (AV5): before 21 November he writes which of the eight parts must be\n   met that day and which may stay open, so *enough* is on paper before the energy decides it. The gate's\n   first written trigger, and the vision page's own device. **Narrowed on the Mac, 2026-11-02:** a bar for\n   the room already stands in his own words of 2026-10-27 (`TODO.md`, the demo item): a fresh example graph,\n   folds and routes, two maps that show their connection at depth, both saturated, and the sweeps running as\n   standing work; three of the five are unbuilt today. So what AV5 still asks is the split, which of the\n   five must hold on the day and which may wait, and whether the sweeps may run from the Mac rather than on\n   the site, which is behind the relay.\n3. **A visitor's path, counted and unnamed** (AV6): one instrument for the two signposts, a per-visit\n   sequence from landing to map to a first act, aggregated and stored as a series, within the read log's\n   privacy rules. **Sequencing, a Mac session's reading 2026-11-02:** the instrument reads the door, and the\n   door is the landing that item 9 of the order of work replaces; built first, the series would record the\n   old door. So it is built with the landing overhaul and reads the new one from its first day.\n\n**Checked against the private repository afterwards (2026-11-02), on his question.** Nothing above rests on\nit, and nothing in it overturns the three claims: *done-ish* is written nowhere there either, and a reader\nof 11 October had already looked for a written threshold in the public files and not found it\n(`.private/reader-reports/2026-10-11-pages-open-question-three-vision-pages-word-for-word.md`); the\noperations note describes no visitor analytics beyond the digest; and the readiness note written for the\nMiddle Powers application (`.private/funding/pipeline-2026/middle-powers-work-and-readiness-2026-10-24.md`)\nsays what § 5a says, that readiness governs invitations and that its candidate observations are *\"candidates\nfor deliberation, not operational signposts\"*. That note adds two things worth carrying: readiness, signposts\nand consequences are three distinct questions (have we built the experience he requires; will the work fare\nas assumed; what happened in one use), and a bounded first use should distinguish *\"reading the map, checking\none's representation and live collaborative inquiry\"*, which is a cut by kind of use that AV4's cut by circle\nof people should sit beside, not replace.\n\n### 5f. What the digest counts, and the correction\n\nThis session first told him that nothing counts a stranger's visits. He answered: *\"But I get ntfy pushes\nevery day summarizing visits\"*. Read in `deliberus/notifications.py` and the tracking middleware in\n`deliberus/api.py`: every production GET on a page route by a non-owner, non-bot visitor is tallied by path\n(`record_page_view`), and the daily push lists the total and the six busiest paths, with doc reads, queries,\nextractions, contributions and new users. What it cannot do: tell visitors apart (no visitor id, no\nsession, no sequence, so five loads of the landing by one person are five views), or keep a series (the\ntally lives in process memory and resets after each push, so a deploy drops the day). The read log\n(`deliberus/reading.py`) names signed-in readers only and drops anonymous ones by design. So the corrected\nsentence is: *the digest counts stranger views by path; nothing follows a visitor from landing to map*. The\nsignposts note's instrument cell, which said *nothing reads this yet*, is corrected in place, and AV6 asks\nfor the two steps the tally is short of.\n\n## 6. What this session got wrong\n\n- **Nothing counts stranger visits**: read off `reading.py` with `notifications.py` unopened; the shape is\n  the mistake record's third row, memory for the record (`the-kind-of-mistake-we-make.md` § 13).\n- **`.private/` out of reach**: the two-path check at the session's start missed `../deliberus-private`,\n  the mistake the always-loaded file names from 31 October; this session's copy of that file was the\n  28 October version. Nothing in this note needed the private repository, and nothing from it is here.\n- **The fourth row, named by position**: a reply referred to *the fourth row of the gate's table* with no\n  self-decoding label, and he had to ask; the row is *Is it enough? the founder, and nobody else*.\n\n*Added 2026-11-05: the constraint read as moving to contact has Krauss's sentence for it, from his 2026 paper on 761 discoveries: \"A new tool not only needs to be developed but also applied; innovation comes from application, not just invention.\" His clock runs from a tool's first application to a new problem, which here is the gate on asking people to try the platform, ruled per circle on 2 November (AV4) ([tools-and-analogies-as-engines-of-discovery.md](tools-and-analogies-as-engines-of-discovery.md) § 4g).*\n\n## Cross-references\n\n`../vision.md` § Why the meta layer is the gate, § Why Now (the two gates and the four rows) ·\n`signposts-for-the-whole-project.md` § 4 (the eight parts, corrected 2026-11-02) ·\n`what-deliberus-is-for.md` (the purposes and the *today* column) ·\n`the-kind-of-mistake-we-make.md` (the muddles, and § 13) ·\n`what-he-held-in-the-group-2011-2013.md` (his 2011 positions) ·\n`../history.md` (the timeline, with a pointer here under 2014) ·\n`the-canon-distillation-pass.md` § 4i to § 4l (the canon in his words; tests before resident lines) ·\n`flywheels-in-building-deliberus.md` § 6 (the check-in that read the constraint moving) ·\n`session26-open-rulings.md` Arc AV (AV1 to AV6) ·\n`the-disempowerment-objection.md` (his position on control, in his words).\n\n## From the Topic Index\n\n*The `CLAUDE.md` Topic Index entry that pointed here, word for word as it stood on 2026-11-15, before the index was trimmed to the warnings only it carries. The index keeps a short entry; the full text lives here.*\n\n- **\"How did the project evolve, and what did 2026 add?\" / the long arc from 2009 / where the binding constraint moved / the readiness gate assessed** → `docs/research/how-the-project-evolved-read-from-the-record.md` (2026-11-02; a session's reading, unruled). The design was his by 2011 (`what-he-held-in-the-group-2011-2013.md`); 2026 added the parts, the discipline and, on that reading, the rejoining of his 2014 turn to Bayesian graphs; the binding constraint moved from typing the structure to the instruments' honesty to his rulings, and is read as moving to contact; the costliest mistakes were muddles that failed silently, the project's thesis proven on its builders. The readiness gate of 22 October, assessed against its own ten fields: no trigger either way, no review date, no check against ground truth, *done-ish* never written though `docs/vision.md` § Why Now asks for it; three proposals, ledger AV4 to AV6. **Do not say nothing counts stranger visits**: the daily digest tallies every non-owner, non-bot page view by path (`deliberus/notifications.py`) and follows no visitor from landing to map.\n"}