{"path":"research/signposts-for-the-whole-project.md","content":"# Signposts for the whole project: what Deliberus assumes, and the signs that would tell us an assumption is failing\n\n**Date:** 2026-10-22\n**Standing:** a proposal. None of the assumptions, signposts or triggers below is ruled. They are to be\ndeliberated one at a time, starting with the case § 4 picks and says why.\n\n**His words, verbatim:** *\"Another idea from listening: we should list all the \"signposts\" of the whole\nproject of Deliberus. But they deserve careful deliberation and operationalization.\"*\n\n**What he was listening to** (`docs/listening/2026-10-22-what-an-argument-is-made-of.md`, Myrddin): *\"A plan\nnames what it hopes. A good plan also names what it assumes, and watches for the first sign that an\nassumption is failing. The planners at RAND called those signposts. The humility is in the watching.\"*\n\n**What this builds on.** `assumption-ranking.md` (March 2026) ranked fifteen load-bearing assumptions by\nevidence and set none of them a sign to watch; several of its rows are overtaken (its weakest, two-axis\nvoting, was decided against: agreement is voted, quality is computed, there is no second vote). The\npurposes note (`what-deliberus-is-for.md`, unruled, ledger AH1 to AH3) says what the work is for, with a\n*today* column that is already half a set of signposts. The success prior (the always-loaded block, and\n`fractal-priors-for-success.md`) names the four conditions under which projects of this kind have won and the\nfour counter-lenses. The convergence wager has its own falsifier (`../convergence.md`, *What Would Disprove\nThis*). Ledger AT10 (*load-bearing and vulnerable*) already gives a consequence trial's assumptions\nsignposts. This note turns the same method on the project itself.\n\n---\n\n## 0. In short\n\n- A **signpost** is the information to track to know whether an assumption the plan rests on is still\n  holding; a **trigger** is the reading at which a planned action starts. Planning under deep uncertainty\n  has used the pair for thirty years (§ 1).\n- Here a signpost is operational only when it carries ten things, each taken from a rule the corpus\n  already holds (§ 2): among them a trigger in *both* directions, a review date with a home, one check\n  against ground truth before its readings are trusted, and the answer to *what would move this sign\n  without the assumption holding?*\n- The register lists twenty-five assumptions in seven groups, each with a draft signpost, the instrument\n  that would read it and its reading today where one exists (§ 3).\n- The readings that exist say what the purposes note said: what one person and the machine can do is far\n  along; what needs a second mind is near zero. And one adoption-side caution governs every trigger in the\n  people group: a flat phase is not evidence against the threshold (§ 2, the last guard).\n- **First case, chosen on purpose: the outreach gate** (§ 4), what waits before people are asked to try\n  the platform. As first drafted it was a count of collaborators' reports; he corrected it the same evening:\n  he already tells people widely, and what waits is readiness, which he named in eight parts.\n\n## 1. What a signpost is, in the planning literature\n\n**Assumption-based planning** (Dewar, Builder, Hix and Levin, RAND 1993; Dewar 2002) runs five steps: list\nthe assumptions the plan rests on; find those that are *load-bearing* (the plan fails if they do) and\n*vulnerable* (they could fail within the plan's lifetime); give each such assumption **signposts**, early\nsigns that it is failing; and plan **shaping actions** (make the assumption hold) and **hedging actions**\n(prepare for its failure).\n\n**Adaptive policymaking** (Walker, Rahman and Cave 2001) adds the **trigger**: a critical value of a\nsignpost beyond which a pre-specified action starts. It names four kinds of triggered action: *defensive*\n(keep the plan, protect its benefits), *corrective* (adjust the plan), *capitalizing* (seize an opportunity\nthe sign reveals), and *reassessment* (when the assumptions critical to the plan have clearly lost\nvalidity). **Dynamic adaptive policy pathways** (Haasnoot, Kwakkel, Walker and ter Maat 2013) joins this\nwith *adaptation tipping points*, where an action stops serving the plan's aims. Read through search\nsummaries and the papers' own abstracts and excerpts on 2026-10-22.\n\n**Two refinements from the monitoring literature.** Haasnoot, van 't Klooster and van Alphen (2018) separate\n**performance signposts** (is the plan still meeting its aims?) from **context signposts** (is the world\naround it changing?), and **weak signals** (early, uncertain: time to prepare) from **strong** ones (act).\nRaso, Kwakkel, Timmermans and Panthou (2019) give the criterion every signpost here is held to: detection\nmust not be slower than the change it detects, since a sign that arrives after the decision was due has\nwatched nothing.\n\nWhat carries over and what does not. The planners watch a world that does not answer back. Deliberus's\nassumptions are partly about people and partly about its own machine, so two hazards come in that their\nliterature treats lightly: a sign the project can move without the assumption holding (§ 2, field 10), and\na sign read too early in a slow phase (§ 2, the last guard).\n\n## 2. What a signpost must carry here\n\nEach field is taken from a rule the corpus already holds; the rule is named beside it.\n\n| # | Field | Why, and the rule it comes from |\n|---|---|---|\n| 1 | **The assumption**, one sentence, worded so it can be false | a wager that cannot lose is not a wager (`../convergence.md`, the accounting) |\n| 2 | **What fails if it fails**: which purpose or part of the plan | load-bearing, in the hinge's sense (ledger AT10) |\n| 3 | **Why it might fail within a year**, and what the corpus already holds against it | vulnerable (AT10); read the record first |\n| 4 | **The signpost**: the observable, defined so two people would count the same thing, and the instrument that reads it, with its state (built, designed, none) | say where a thing stands |\n| 5 | **The reading today**, measured and dated, or *no instrument* | every published count labels its provenance |\n| 6 | **The trigger in both directions**: the reading that means act, and the reading that would mean the assumption holds | counter-instruments must keep the power to falsify the wager they guard (`constellation-candidates.md`), and a sign that can only alarm is as blind as one that can only reassure |\n| 7 | **The action each trigger starts**, of the four kinds, and **who decides** | a sign informs and a person acts: monitoring metrics inform, never trigger on their own (Goodhart), and every machine judgment on contested ground is a proposal |\n| 8 | **A review date and a home**: a calendar entry or `TODO.md`, never a session note | an observe-then-decide plan has a review date, and the date does not live in a session report |\n| 9 | **One check against ground truth** before the readings are trusted | an observability instrument is validated once before its data is believed; a guard that never fires looks exactly like one that works |\n| 10 | **The proxy check**: what would move this sign without the assumption holding? | the winner of a proxy contest is where the broken assumption lives |\n\n**Three guards that bind the people group in particular.**\n\n- **Warm feedback is near-worthless as evidence about deliberative quality.** Participants preferred an LLM\n  facilitator that did not improve consensus and shifted allocations (the standing threat model, Google\n  DeepMind 2026). Any sign of the form *testers liked it* is stated as such before a round produces it.\n- **Momentary engagement is worthless; what counts is engagement that recurs unprompted.** His definition of\n  pointing, 2026-08-19 (`structure-versus-scale.md`): *\"If we tell them to interact with a claims-oriented\n  graph UI I would expect them to momentarily engage.\"* The real measures are whether engagement recurs\n  over sessions and the *decay curve* within one.\n- **An early null indicts the door, not the substrate, and a flat phase is not evidence against the\n  threshold.** His ruling the same day: a null from a session whose interface never offered the handle\n  redirects the modality and never puts the substrate on trial. And the success prior's fifth consequence:\n  absence of traction in the flat phase is not evidence against approaching it. So every adoption trigger\n  is conditional on the door being built and carries a horizon, never a single reading.\n\n## 3. The register\n\nTwenty-five assumptions. *Serves* names the purpose of `what-deliberus-is-for.md` § 8.6 it serves (unruled\nthere too). *Reading today* quotes a measured figure with its date and source, or says there is none.\nTriggers are left for deliberation, case by case, except in § 4.\n\n### 3a. The bet: convergence and legibility\n\n| Name | Assumption | Draft signpost | Instrument | Reading today |\n|---|---|---|---|---|\n| **The wager can be read** | Descents reach horizons the horizon reading can classify, so the map can report either answer | Classified horizons on the live graph, by domain, split into those counting against convergence (hinge clash, parity, permissivism) and those dissolving into shared or empirical ground; a shared bottom counts only where it rules out a position someone holds (ledger AJ5) | Horizon reading, built (`GET /claims/{id}/horizon`, 2026-09-16) | Two classified termini, split one and one (July 2026); none readable from the live graph (2026-09-16, purposes note) |\n| **The descent is traversable** | A reader can follow a contested claim down to where the parties part, in an hour of attention | Share of descents ending at a horizon or shared bottom against those stalling with no move offered; time per descent | None as a series; dogfood runs and the deep-descent experiment | Not counted as a series |\n| **Seeing another's view helps** | People who meet a disagreement on the map understand the other side better, without having to agree | Holder-accepted accuracy: each holder accepts the rendering of their own view, and the other judges it as theirs (`what-if-and-lenses.md` § 8, proposal 8; ledger AU8) | Designed, unbuilt | None |\n| **The machine does not steer** | Passes do not move disagreeing people toward a middle before the graph measures them | Ingestion balance per debate cluster; drift between a claim and its source span; disagreements kept against disagreements found | Disagreement-preservation instrument, built (its denominator is existing conflict links); ingestion balance proposed, unbuilt | Not measured |\n\n### 3b. People: contact, contribution, their own terms\n\n| Name | Assumption | Draft signpost | Instrument | Reading today |\n|---|---|---|---|---|\n| **The outreach gate** | The platform is ready to be tried: a newcomer gets through the landing page and the maps to something worth having (his eight parts, § 4) | § 4 | Partly (§ 4) | § 4 |\n| **People come back by themselves** | Engagement with claims recurs unprompted, across sessions | Claim interactions per returning person over time; the decay curve within a session | Server logs; the per-session trace is designed (`structure-versus-scale.md`), not built | Twenty people have ever signed in (2026-10-11) |\n| **People add without being asked** | Others contribute reasons, objections and links of their own | Contributions per month by anyone but him and the machine | Countable from the graph | Four claims (2026-10-03) |\n| **People work in their own terms** | Taking part needs no ontology vocabulary (condition (d) of the success prior; the inversion principle) | A stranger, after the front page, can say what Deliberus is *for*, not only what it does; share of contributions made without the project's words | The stranger test, by a friend, intermittently | One outside reading (2026-09-29): the mechanism understood, the purposes mostly not (`what-deliberus-is-for.md` § 8.8) |\n| **Readers open the other side** | Readers are curious enough to look where they disagree (Kahan's precondition, `../convergence.md`) | Views of opposing branches against own-side branches; take-up of the return edge once built | None | None |\n\n### 3c. The machine\n\n| Name | Assumption | Draft signpost | Instrument | Reading today |\n|---|---|---|---|---|\n| **Passes can be made reliable** | A pass's error rate is first the task as written, and falls as the task is made sturdier | Passes with a proven bound and scope; error rate per instruction version | The record of machine judgments (since 2026-10-06); the review pages | No person had read one (2026-10-11); not re-read since |\n| **Holders recognise their own claims** | Extraction is faithful enough that a source's author would accept the claims as theirs | Share of extracted claims a holder accepts; corrections per claim | None as a series | None |\n| **Independence and sameness can be read** | A pass can tell two reasons that count twice from two that count once | Agreement with labelled pairs; the independence wording test | The keeper's evaluation sitting; `retrofitting-the-reasons-channel/independence_wording_run.py` | On nine pairs written for the test: the current wording 5 of 9, the most-likely-rival wording 9 of 9, ruled the same day (2026-10-22, `retrofitting-the-reasons-channel.md` § 9k); none on real pairs |\n| **Mapping stays affordable** | The model calls a mapped source needs can be paid for at the scale the work reaches | Spend per mapped source and per month, by account; quota exhaustion events | The router backend's spending cap; the Claude backend's usage log | Not re-read; the Max plan's monthly API credit found 2026-10-22 (`whose-tokens-pay-for-mapping.md` § 3h) |\n\n### 3d. The number on a claim\n\n| Name | Assumption | Draft signpost | Instrument | Reading today |\n|---|---|---|---|---|\n| **The number tells a stranger something true** | A claim's number reads as how likely it is in the world (ledger Q41's target) | Calibration on claims whose question later resolves (testable bets, Arc AO); what a stranger says a number means | Testable bets proposed (AO3), unbuilt | None |\n| **The arithmetic does what the rulings say** | The shipped strength code matches the ruled calculus | Ruled lines marked *unbuilt*; scripts that pin his expected numbers | The canon lines carry *Unbuilt*; the gate script | Q38 and Q34 unbuilt (canon, 2026-10-22) |\n\n### 3e. Integrity: adversaries and absences\n\n| Name | Assumption | Draft signpost | Instrument | Reading today |\n|---|---|---|---|---|\n| **Gaming does not pay** | Padding, claim-farming and brigading gain nothing (payoff removal before detection) | Attempts found; share of a claim's strength from one actor | None for actors: the graph measures a claim's exposure, never an actor's | No instrument has met a motivated stakeholder (conviction) |\n| **Power does not capture it** | Legibility that wins is not turned into control (Porter's dark rider) | Requests to remove or promote; a written constitution | None | The constitution is unwritten (success prior) |\n| **Obfuscated arguments are caught** | Error spread thin over many steps is found | Stability under re-decomposition; rate of *undecided* verdicts | Managed defences, built | Not re-read |\n| **People reach what the machine cannot** | Coherent absence and frame lock-in are reached by people positioned differently | *Nobody here has said X* contributions; frame gaps voiced | The contribution path for a frame gap is a build item | Near zero (few contributions at all) |\n\n### 3f. Keeping it alive\n\n| Name | Assumption | Draft signpost | Instrument | Reading today |\n|---|---|---|---|---|\n| **Maintenance can be paid for** | Hosting, models and upkeep are covered before they lapse (the category's documented failure mode) | Months of operation covered; applications sent and answered; funders' cycles (context) | The funding pipeline (private) | Private |\n| **A newcomer can pick it up** | A person or agent without context can do useful work from the docs alone | Time to a first useful contribution by a newcomer; an outside reader's account of the purposes | Outside readers (`an-outside-reading-of-the-thesis.md`) | One reading: what it does, yes; what it is for, mostly no |\n| **Carriers exist and will adopt** | Institutions whose standing depends on legibility take it up (condition (c)) | Carriers in conversation; a pilot | `carriers.md` ranks the nearest doors | Not re-read |\n| **Illegibility keeps getting costlier** | The trust and verification crisis keeps raising the price of opaque reasoning (condition (a); context) | Public incidents of machine persuasion; funders' calls on epistemic security | None; a context watch | *Arriving loudly* (success prior) |\n\n### 3g. The work itself\n\n| Name | Assumption | Draft signpost | Instrument | Reading today |\n|---|---|---|---|---|\n| **Decisions close faster than they open** | The corpus becomes his: rulings keep pace with proposals (session purpose B) | Ledger rows ruled against rows added, per week; positions awaiting his word | `scripts/flywheel_checkin.py` (weekly, rows open, ruled and with a canon decision) | 530 rows, 397 open, 112 ruled (its first reading, 2026-10-22 19:29); 335 awaiting against 133 a month before (2026-10-11) |\n| **The notes stay squared** | Corrections reach every note that repeats the corrected line | Corpus-graph findings of *overtaken* or *later correction* opened against fixed, per week | The corpus graph's found log | Not counted |\n\n## 4. The first case, chosen on purpose: the outreach gate\n\n**⚠ His correction, the same evening: the gate is readiness, not a count of reports.** Told that the hold\nlasts *\"until two or three close collaborators have used it and said, unprompted, that it gave them an\ninsight\"*, he answered: *\"I've started telling ppl about it more widely now. For example, look at the recent\nconvos in PauseAI's group chats, which have many ppl in them. Also I've talked about it here at [home] and\nshown [a housemate] the recent AI risk maps and he studied it for a minute or two and said it was cool :) and I'm\ntelling ppl left and right here and there. But few ppl are making it through the disordered landing page and\nthe whole platform still isn't quite ready for primetime. I've made it clear I wanna settle the ontology and\nUX and feel more ready and proud of what Deliberus is and be confident in that a wow factor will be there\nwhen ppl go to the landing page and the map views and try it out for a spin. It should be self-explanatory,\nit should match my vision closely enough, and it should be interactive and with a low enough error frequency\nand I don't wanna feel that the graph or features or rules of Deliberus muddles things to the degree that I\nstill feel they do. We need to get all microtasks / standing work reliable enough and humming along daily, we\nneed to extract more sources and fill up and saturate the graph to show what this can be and how it can\nchange things around.\"*\n\nHis record bears him out: in the PauseAI Sweden group chat (75 members) he linked deliberus.com and two maps\non 4 October, showed two sketches of the wheel map view on the Sweden-and-frontier-AI questions on 20\nOctober (marked in the images as an illustrative mockup), said what he builds it for\non 21 and 22 October, and on 22 October promised the group an intuitive demo on the AI-risk theme soon\n(the Beeper index).\n\nSo talking about Deliberus is not held; asking people to try it is, and what it waits on is readiness. The\nalways-loaded rule was restated in these terms the same evening. His readiness has eight parts.\n\n**What kind of thing they are (his question the same evening: *\"Are these 'signposts' in the sense that I\ninitially meant? We SHOULD keep an eye on these regularly, but just curious.\"*).** Mostly not. A signpost, in\nthe planning sense of § 1, watches an *assumption* the plan rests on: an early sign that something believed\nabout the world is failing. Six of the eight parts are goals the project builds toward, and together they are\nthe *condition of a trigger* (§ 1: the reading at which a planned action starts; here, asking people to try\nthe platform). Watching them is progress tracking, which is worth doing regularly, as he says. Two of them\nare also signposts, since each tests an assumption the plan rests on: that a newcomer finds the way in\n(*few make it through the landing page*), and that the experience speaks for itself (*a wow factor*).\n\n| His words | Kind | What to watch | Instrument today |\n|---|---|---|---|\n| *settle the ontology and UX* | goal | Open ledger rows on the ontology and the interface, ruled against added | The ledger counts in the weekly flywheel check-in |\n| *self-explanatory*; *few ppl are making it through the disordered landing page* | goal and signpost (a newcomer finds the way in) | Share of first visits that reach a map and do something there | Server logs; nothing reads this yet |\n| *a wow factor … the landing page and the map views* | goal and signpost (the experience speaks for itself) | His own judgement, and a newcomer's first words after a minute | None |\n| *match my vision closely enough* | goal | His judgement, map by map | None |\n| *interactive and with a low enough error frequency* | goal | Errors a visitor meets per session | None counted |\n| *the graph or features or rules … muddles things* | goal | The muddles he still meets when reading a map | None counted |\n| *microtasks / standing work reliable enough and humming along daily* | goal | Machine passes standing, of those registered | The flywheel check-in: 1 of 24 standing (2026-10-22) |\n| *extract more sources and fill up and saturate the graph* | goal | Sources mapped, and claims per topic | Countable from the graph |\n\nWhat stays from the draft below: the definition of a close collaborator (question 1, answered) and the\ncounted report, which is now one possible sign of the wow factor rather than the gate's condition. Whether a\ncount of reports remains a condition at all is not asked.\n\n**As first drafted (superseded in its framing by the correction above).** **Why this one first.** It is the only signpost in the register that is already acting: wider outreach has\nbeen held since May until *two or three close collaborators report unprompted insight-generation* (the\ndogfooding rule in the always-loaded memory, written for May and June and still standing). It governs the\nmost consequential near-term choice (who is invited next, which feeds funding and carriers). And the sign it\nwaits for has no definition, so every week it holds or lifts on an impression. Two other candidates were\nweighed: *decisions close faster than they open* (already measured, cheap, but it governs the sessions and\nnot the project's contact with the world) and *the wager can be read* (the project's own falsifier, but its\ninstrument cannot read anything yet, so deliberating its trigger now sets a number nothing will meet for\nmonths).\n\n**The ten fields, drafted.**\n\n1. **Assumption.** Close collaborators, using Deliberus on a question of their own, gain an insight they\n   would not otherwise have had.\n2. **What fails if it fails.** Purpose 1 and purpose 6 at first contact; the testimony funding applications\n   lean on; and the outreach that follows would invite people to something that does not yet pay.\n3. **Why it might fail.** Twenty people have ever signed in; four claims came from anyone else\n   (2026-10-03); the gradient that shows a person's words thickening into structure is unbuilt, so a\n   session today tests the door more than the substrate.\n4. **Signpost (proposed definition).** A report counts when all four hold:\n   - **their own question**: the insight is about a question the person brought, not about Deliberus;\n   - **in their words**: they say what they now see, and it is written down verbatim;\n   - **unprompted**: volunteered, not an answer to *did you learn anything?*;\n   - **pointable**: they can point at the claim or link where it happened.\n\n   And separately, **recurrence**: the same person returns to that question or another within a set\n   number of days without being asked (his pointing definition).\n   Instrument: a one-line record per report (date, person, words, the claim pointed at, prompted or not),\n   kept in `.private/` because it names people, with only the count public; plus the server's record of\n   returning sessions.\n5. **Reading today.** Not counted. The reports, if any, are in his messages.\n6. **Trigger, both ways.** Open: a number of distinct people, each with at least one counted report and one\n   recurrence, within a window. Weakening: after the gradient is built and a set number of real sessions,\n   engagement spikes at instruction and decays (his falsifier arm, `structure-versus-scale.md`).\n7. **Actions.** Open: *capitalizing*, outreach to the next circle. Weakening before the gradient: *corrective*,\n   change the door. Weakening after it: *reassessment* of the first-contact design.\n8. **Review date and home.** A date he sets, in `TODO.md` and the calendar.\n9. **Check against ground truth.** The first counted report is read back to the person: *is this what\n   happened?*\n10. **Proxy check.** What moves the count without the assumption holding: friends being kind; his own\n    presence prompting the insight; insight about the tool rather than the question. The four conditions in\n    field 4 are written against those three.\n\n**The questions this case raises, for him.**\n\n- Who counts as a close collaborator for the gate? **Answered 2026-10-22:** *\"Anyone who understands and\n  helps or has consciously helped or prompted me more than once to think about the project in some helpful\n  and insightful way during any phase of its ideation or development feels like a close collaborator to\n  me.\"* He rejected a definition built only on contributions in the graph as too weak (*\"it needs to be\n  multi-faceted\"*). Five tests follow from his words: understands the project; helped or prompted his\n  thinking; consciously; more than once; helpfully, not just warmly. Any phase counts, so the pool\n  reaches back to the 2011 group. He named seven people; the register of who meets the tests, with dated\n  evidence, is kept privately because it names people. What it changes about the gate: the people whose\n  reports count include some who shaped the idea before the site existed, so whether each has used the\n  site is part of what the register records.\n- Does a question he hands someone count as their own, or only one they bring?\n- How many people, each with a counted report and a return, open the gate, and within what window?\n- Is the record kept privately, with only the count public?\n\n## 5. What is not settled\n\n- **Every assumption sentence is Claude's wording.** Several are built from purposes that are themselves\n  unruled (AH1).\n- **No trigger value is proposed except in § 4.** A number set before its case is deliberated is the\n  premature menu the decision rules forbid.\n- **Twenty-five is too many to watch.** The planning literature watches only what is both load-bearing and\n  vulnerable; ranking the register on those two is the step after the first case, and some rows will drop.\n- **Which signposts can be read by code.** A weekly readout exists since the same evening: the flywheel\n  check-in (`flywheels-in-building-deliberus.md` § 6) runs every Monday, its script reads the ledger,\n  glossary, connections and passes, and the session that judges the readings reviews this register (a\n  reading that crossed a trigger, a review date due, a row whose reading the counts now give). Rows whose\n  reading is a graph count (contributions by others, horizon readings) are not yet among its readings.\n- **In the ledger as Arc AV**: the ten fields (AV1), the outreach gate first (AV2, its four questions open)\n  and the ranking step (AV3); all unruled. The deliberation is in `TODO.md`.\n- **Where the register lives once ruled.** A short block of the always-loaded memory is one candidate, as the\n  purposes note proposes for its own list; a page of its own is another.\n\n## Sources and read depth\n\nDewar, Builder, Hix and Levin, *Assumption-Based Planning* (RAND MR-114, 1993) and Dewar (2002): known through\nledger AT10, the formal reading (`what-if-and-lenses/outside-reading-formal.md` § 6.4, abstract and chapter\ntitles) and the monitoring papers' summaries. Walker, Rahman and Cave (2001) and Haasnoot, Kwakkel, Walker and\nter Maat, *Dynamic adaptive policy pathways*, Global Environmental Change (2013): abstract and excerpts.\nHaasnoot, van 't Klooster and van Alphen, *Designing a monitoring system to detect signals to adapt to\nuncertain climate change*, Global Environmental Change (2018), and Raso, Kwakkel, Timmermans and Panthou,\n*How to evaluate a monitoring system for adaptive policies*, Climatic Change (2019): excerpts. The corpus\nnotes named in each row were read in the parts cited on 2026-10-22; the success prior and the convictions\nwere read whole.\n\n## Cross-references\n\n`assumption-ranking.md` · `what-deliberus-is-for.md` (§ 3, § 8.6, § 8.8) · `fractal-priors-for-success.md`\n(§ 16, § 17) · `../convergence.md` (*What Would Disprove This*) · `structure-versus-scale.md` (pointing) ·\n`normic-rules-normative-reasons-and-trying-rules-out.md` § 4c and ledger AT10 · `what-if-and-lenses.md` § 8 ·\n`world-completeness-and-its-relatives.md` · `retrofitting-the-reasons-channel.md` § 9k ·\n`whose-tokens-pay-for-mapping.md` · `carriers.md` · `flywheels-in-building-deliberus.md` § 6 (the weekly\ncheck-in that reviews this register).\n"}