{"path":"research/flywheels-in-building-deliberus.md","content":"# Flywheels in building Deliberus\n\n**Date**: 2026-10-22 · **Type**: definition first, inventory after · **Status**: § 1 proposed by\nClaude, unruled; § 2 not written yet, waiting on his reading of § 1.\n\nHis idea, 2026-10-22:\n\n> *\"Extractions and learnings from inspecting them and their maps and playing around with them in the\n> UI and thinking about better UX and then extracting better is one sort of flywheel in Deliberus. All\n> standing work and its maturation seem to me to be similar flywheels that interact synergistically.\n> We should name all the flywheels that can be identified in the work of building Deliberus.\"*\n\nAnd before any naming: *\"Think deeply about the definition of a flywheel before we start.\"*\n\nThe corpus already names one: the reuse flywheel (each mapped claim makes the next source cheaper to\nmap), with its condition (sameness must be cheap to decide), its measured shape (paired, not\ncumulative) and its shadow (frame lock-in): [the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md),\n[lean-deliberus-analogies.md § 2](lean-deliberus-analogies.md#2-the-simp-flywheel-accumulated-knowledge-that-automates-reasoning),\n[lowering-the-cost.md](lowering-the-cost.md). That flywheel lives in the product. His question is\nabout the work of building it.\n\n---\n\n## 1. What a flywheel is (proposed)\n\n**Two meanings travel under the word, and both matter here.**\n\n- **The machine part** stores energy in a heavy wheel. It does not amplify anything. It is hard to\n  start, hard to stop, and it keeps a machine turning smoothly between irregular pushes.\n- **The business metaphor** (Jim Collins, *Good to Great*, 2001; Amazon's \"virtuous cycle\" drawn in\n  2001) is a reinforcing loop: many pushes in one direction add up, each turn makes the next one\n  easier, and no single push is the breakthrough. In systems dynamics this is a reinforcing feedback\n  loop (Senge's *virtuous cycle*), and its failure twin is the vicious cycle.\n\nHis two examples fall one on each side. The loop from extraction to inspection to better extraction\nis the reinforcing kind. Standing work that keeps running while he is away is the storing kind: the\nwork goes on between his pushes.\n\n**Proposed test. A loop in the work counts as a flywheel when all five hold:**\n\n1. **It closes.** What the last step produces is what the first step needs. A pipeline that runs once\n   is not one.\n2. **Something is left behind each turn, outside anyone's head.** A store that carries over: claims in\n   the graph, a labelled case, a ruling in the ledger, a pass's instructions, a doc. A lesson that\n   lives only in a chat or in his memory leaks away; that leak is the loop's friction.\n3. **The next turn costs less or yields more because of it.** The gain has to be named and, sooner or\n   later, measured. A loop whose turns cost the same each time is a routine, not a flywheel.\n4. **It keeps some motion without a push.** Part of the store does work on its own, a standing pass\n   or a rule applied by code, so progress continues while he is elsewhere.\n5. **What it saves is named.** In this project the scarce input is his attention, so the natural\n   measure is how much of his reading and ruling a unit of progress costs, and whether that falls.\n\n**What every flywheel also has, and should be named with it:**\n\n- **A condition**: what must hold for the gain to appear (for reuse: cheap sameness).\n- **Friction**: what bleeds momentum (docs that go stale, an answer key voided when a model is\n  replaced, findings left in chat).\n- **A limit**: where more turns stop paying (saturation; the head of a topic is narrow and heavy).\n- **A shadow**: momentum resists a change of direction (frame lock-in is the named case).\n- **Propagation, the shadow's twin**: a flywheel compounds whatever enters it, errors included.\n  Sameness is transitive, so reuse spreads a wrong merge as efficiently as a right one: on the\n  Semantic Web `owl:sameAs` ran 3–20 % wrong links, and one closure unified 177,000 names (CLAUDE.md,\n  Academic Foundations). The canon pass holds it as waiting candidate 5, *\"the flywheel and the error\n  corridor are one mechanism\"* ([the-canon-distillation-pass.md](the-canon-distillation-pass.md) § 6).\n  The building work has its own case: eight confident steps, each built on the last\n  ([the-kind-of-mistake-we-make.md](the-kind-of-mistake-we-make.md) § 9a).\n- **A governor**: what stops an error riding the wheel. For any wheel whose store is machine output\n  it is already ruled, ledger T1: *\"A machine pass may build on machine output only as deep as its\n  measured error rate allows; with no measured rate, a person engages first.\"* Criterion 4, motion\n  without a push, is exactly where T1 applies.\n- **Gearing, and its sign**: which other flywheels its output drives. Every wheel here draws on one\n  scarce input, his reading. Two wheels that both draw on it compete, unless one's output lowers the\n  other's draw on him; that is why the four reviews were put in order (CLAUDE.md, READ FIRST). His\n  *\"interact synergistically\"* holds where one wheel's store is another's input or lightens its\n  demand on him, and fails for wheels that merely run side by side on his attention.\n\n**The failure of criteria 1 and 2 already has a name**: the uncollected idea, correct and filed and\nchanging nothing, because nothing depends on it\n([the-kind-of-mistake-we-make.md](the-kind-of-mistake-we-make.md) § 6). The same section names a\nbuilding-work wheel with ten instances, a mistake turned into a check (*\"Each of those is a muddle\nthat stopped recurring the day it became a check\"*), and another in the ruling ledger (*\"the\nhammering made durable: each row ruled is one fewer thing to restate\"*).\n\n**Two wheels already carry the measure criterion 3 asks for.** The corpus graph: ask it before the\nreaders go, then count how many of the notes they found it already held; the first count was 33 of\n77, and *\"If that number does not rise as connections are kept, the graph is a record and no help\"*\n([corpus-graph/README.md](corpus-graph/README.md) § 8). The canon wheel, built 2026-10-22 (commits\n`83ab9dc`, `b32c63d`, `038cf98`): every ruling carries its canon decision, the resident lines are\nmeant to cut his corrections of things the corpus already held, and the measure proposed beside\nAK3 and AK4 in TODO.md is the share of his messages that make such corrections (19 % of his\nDeliberus messages 10–19 October is the ceiling). *Points from this paragraph and the four bullets\nabove came from the session \"Claude app self-update\" the same day, checked against the repo here.*\n\n**Why the test is strict.** Calling every repeated activity a flywheel would make the word mean\n\"work\". The measured shape of the reuse flywheel (paired, not cumulative: the yield rose only where an\nopposing source was present) is the warning: a loop can look like a flywheel and have a gain of about\none. Naming a loop is a hypothesis about its gain, to be checked like any other.\n\n---\n\n**His ruling on § 1, 2026-10-22:** *\"Definition is excellent and thought-through now. Now name the\nflywheels.\"*\n\n**A word kept apart.** In this project *the wheel* already names a map view (the wheel view,\n`docs/sketches.md`). So this note always says *flywheel* in full, never *wheel*.\n\n## 2. The flywheels (Claude's naming, unruled)\n\nTwo kinds, by what they turn. **First-order** flywheels do the work: they map sources, prove passes,\nfill the graph. **Second-order** flywheels improve the sessions that run all the others, which is\nEngelbart's \"improving the improvement\" (§ 3). His question, *\"which ones lower the other demands?\"*,\nis answered by the second kind.\n\n| # | Flywheel | The store it leaves | What the next turn gains | Turns without him? | Order |\n|---|---|---|---|---|---|\n| 1 | Ruling-to-canon | ledger rows, their Canon field, resident lines | sessions start already knowing his positions; fewer of his turns restate | partly: lines load by themselves | second |\n| 2 | Mistake-to-check | tests, hooks, checks | a muddle cannot recur | yes: a check runs every time | second |\n| 3 | Connection memory (the corpus graph) | kept connections, the connections register | a session is handed what bears on a note before writing | yes: the graph hands lines out at the note | second |\n| 4 | Shared vocabulary (the glossary) | blessed terms | shorter, plainer chat; fewer glosses to read | yes, once blessed | second |\n| 5 | Answer key | his labelled cases, the rules they became | the next case is decided by the rule; only the residue reaches him; a new model is admitted in minutes | yes, after the rule | first, with a second-order effect |\n| 6 | Standing work | the record of machine judgments, coverage per pass | the graph fills while he is away | yes, within T1 and the sweep brake | first |\n| 7 | Mapping (his example) | pass instructions, the UI, friction logs | the next source maps with fewer faults and reads faster | no: it needs his eyes on maps | first |\n| 8 | Reuse (in the product) | claims in the graph | the next source is cheaper to map | yes | first, already named |\n| 9 | Listening | his voiced reactions and rulings | his attention is spent away from the screen | no, but it adds to his attention rather than drawing on it | raises supply |\n| 10 | Outside reading | research notes read against Deliberus | later questions answered from the corpus, not read again | no | candidate: gain unmeasured |\n| 11 | Collaborators | other people's reactions and contributions | more minds than his | not yet | parked: the dogfooding hold |\n\n**1. Ruling-to-canon, the one he pointed at.** A position written in a note is ruled in the ledger,\ncompressed to one line, and made resident in the always-loaded file, so every later session reasons\nfrom it instead of from the template, and his turns go to new rulings instead of restating old ones.\n*\"The ledger is the hammering made durable: each row ruled is one fewer thing to restate\"*\n([the-kind-of-mistake-we-make.md](the-kind-of-mistake-we-make.md) § 6). Until 2026-10-22 it did not\nclose: rulings stopped at the ledger, and his question that day (*\"Idk why this doesn't end up in the\nalways-loaded file after I have ruled something?\"*) found the missing step; the Canon field on every\nrow now closes it ([the-canon-distillation-pass.md](the-canon-distillation-pass.md) § 4d).\nMeasure: the share of his messages that correct something the corpus already held (ceiling 19 %,\n10–19 October; TODO.md beside AK3 and AK4). Condition: a line is in view when it matters (AK3).\nLimit: the always-loaded budget. Shadow: a resident line hardens a position he holds fallibly.\nGovernors: the wrong-by-default filter, his blessing, the outside check. Gearing: positive with\nevery other flywheel, because every session that turns any of them reads the canon.\n\n**2. Mistake-to-check.** A muddle he catches becomes a test, a hook or a check, and stops recurring\n*\"the day it became a check\"*; ten instances listed in\n[the-kind-of-mistake-we-make.md](the-kind-of-mistake-we-make.md) § 6. The same flywheel runs\nfleet-wide in the global rules and hooks. Measure: recurrences of a named muddle after its check\n(zero for polarity and backwards links). Shadow: a check written to a wrong rule enforces the\nwrong rule. Gearing: positive with all.\n\n**3. Connection memory.** Sessions and readers find what bears on what; kept connections are handed\nto the next session at the note it opens. Measure: the share of a reader's findings the graph already\nheld (33 of 77 at first; *\"If that number does not rise as connections are kept, the graph is a\nrecord and no help\"*, [corpus-graph/README.md](corpus-graph/README.md) § 8). His reason for it:\n*\"attention relates everything to everything\"*. His standing wish, AK3: he is never the retrieval\nmechanism. Propagation: a wrong connection is handed out as readily as a right one.\n\n**4. Shared vocabulary.** A term he blesses is used bare; one he has not is glossed. Store: the\nglossary. Gain: chat he reads faster. Measure: the counter of unglossed terms per reply.\n\n**5. Answer key.** He judges cases one at a time, each verdict becomes a candidate rule, the rule is\ntried on the remaining cases, and only the residue comes back\n(global rule *Case review starts one case at a time*, generalised from a Deliberus ruling). A\nlabelled set is paid for once and admits or refuses every later model in minutes\n([pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) § 5e). Measure:\nthe share of cases a rule decides; the bound a pass earns\n([proving-a-pass-reliable.md](proving-a-pass-reliable.md)). Governor: T1. Friction: a model or\nprompt change resets a bound. The same flywheel turned this week outside Deliberus, on Ruby's\nnarration judge. Gearing: positive with 6 (a proven pass may run unattended).\n\n**6. Standing work.** Every pass registered with a coverage watermark; sweeps read what was never\nread, so the graph fills between his pushes (`deliberus/standing_work.py`,\n[machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md)). The machine-part\nmeaning of the word. Governor: T1 and the sweep brake (R8, three or four calls a pass until he\nraises it). Fuel: the Max plan's API credits from 2026-10-22\n([whose-tokens-pay-for-mapping.md](whose-tokens-pay-for-mapping.md) § 3h). **Gearing is negative\nwith 7 unless 5 has run first**: more machine output is more for him to read, until a proven pass\nlets him read a sample instead.\n\n**7. Mapping.** Extract a source, look at its map in the site, find the faults and the friction, fix\nthe instructions and the screens, extract again (his example; the friction log\n[dogfood-run-1-friction-log.md](dogfood-run-1-friction-log.md) is one turn of it). Friction: what he\nsees while playing stays in chat. Gain to measure: faults he finds per source, and minutes to read a\nmap.\n\n**8. Reuse.** Already named and measured: condition cheap sameness, shape paired, shadow frame\nlock-in, propagation of wrong merges ([the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md)).\nA product flywheel; the building work serves it through the sameness keeper.\n\n**9. Listening.** Walkthroughs he hears away from the screen, his voiced reactions, rulings from\nthem (`docs/listening/`). It does not lower the draw on his attention; it adds hours in which his\nattention can be spent. The player that pauses when he speaks (TODO.md, *Listening with feedback in\nthe moment*) would remove its main friction, reactions given without their place in the walkthrough.\n\n**10. Outside reading** and **11. Collaborators** are named so they are not forgotten. The first has\nno measured gain yet. The second is held on purpose until two or three close collaborators report\ninsight unprompted.\n\n**Not flywheels, and why**: the outside check, T1, the wrong-by-default filter and his blessing are\ngovernors (they keep errors from riding a flywheel). The cost of model calls (cheaper models,\ncaching, credits) is fuel for 5 and 6, not a loop of its own.\n\n## 3. Which flywheels lower the others' draw on him (his question)\n\nHis question, 2026-10-22: *\"Which ones lower the other demands? They sort of have a leverage\nmultiplier effect, don't they, if we could identify those?\"* Yes, and two bodies of theory say which,\nin agreement with each other ([literature report](flywheels-in-building-deliberus/literature-report-2026-10-22.md)\n§ 1.3, § 1.4).\n\n**Goldratt's theory of constraints.** A system goes as fast as its one constraint; improving anything\nelse is wasted, and can do harm by building a queue in front of the constraint. Here the constraint is\nhis attention. Goldratt's steps sort the flywheels:\n\n- **Exploit the constraint** (waste nothing it already does): ruling-to-canon, connection memory and\n  shared vocabulary. Each removes a kind of reading or saying he does twice: restating a position,\n  retrieving a note (AK3: *he is never the retrieval mechanism*), decoding a word.\n- **Subordinate everything to it**: standing work, which must run at the pace his review can absorb\n  and stop when its check fails (Toyota's *stop the line*). Run faster, it does not add capacity; it\n  adds a pile on his desk. This is why § 2 gives standing work a negative sign with mapping until the\n  answer key has proven the pass.\n- **Elevate it** (add capacity once the first two are done): the answer key, where a rule decides in\n  his place; listening, which adds hours in which his attention can be spent; collaborators, later,\n  which add minds.\n- **Watch for inertia**: the store of rules itself becomes the constraint when rules pile up faster\n  than they are retired. The always-loaded budget is where that would show first.\n\n**Engelbart's improvement ladder.** A is the work (mapping, standing work, reuse). B improves how A is\ndone (answer key, mistake-to-check, ruling-to-canon). C improves how B improves: checking whether each\nB flywheel is paying, and changing it where it is not. Engelbart's claim is that C pays most and is the\nlevel organisations skip. **Deliberus has no C flywheel yet**: nobody looks, on a cadence, at whether\nany flywheel's measure is moving. A B flywheel counts as a multiplier only if a C look finds that it\nlowered the A flywheels' draw on him.\n\n**So, ranked by leverage (Claude's reading):**\n\n1. **Ruling-to-canon together with mistake-to-check**, read as one ladder. The literature's first\n   practice: a blessed ruling goes into the cheapest durable form that acts without being read. That\n   means a check if code can enforce it, a resident line if only a session can apply it, and the\n   ledger row alone only if neither can. Every session in every flywheel benefits.\n2. **The answer key**, with its cases chosen by what they teach (where two passes or two models\n   disagree, or a pass is unsure), as active learning does: one of his verdicts then decides many cases.\n3. **Connection memory and shared vocabulary**: smaller gains each, but on every turn.\n4. **Listening**: it lowers no one's draw, yet it adds attention. Its value grows with everything above.\n\n## 4. What the literature changes in § 1 (proposed, his to rule)\n\n1. **Criterion 5 becomes the gate, and is measured once.** A loop counts only if the saving falls on\n   his attention, shown by a measure at least once. METR's 2025 trial is the warning: developers felt\n   about 20 % faster with AI tools and were measured about 19 % slower.\n2. **Criterion 3 says what makes a store pay.** Not its size: being refined each time it is used\n   (precedent, Levi), surprising the reader (Luhmann's slip box), predicting something not yet seen\n   (Lakatos: a progressive versus a degenerating programme), and being in view at the moment of action.\n3. **A stable direction, with a versioned criterion.** Collins' doom loop is a direction changed after\n   each disappointment; Shankar's *criteria drift* is labels whose meaning changes as people grade.\n   Both hold when the criterion is changed on purpose and each label records which version it used.\n4. **Limit and governor are separate.** A limit is a resource running out; a governor is a control\n   someone designed. The strongest governors are an outside check and a line that stops when the check\n   fails; machine self-critique is not an outside check (Huang et al.). The deepest review asks whether\n   the criterion itself is right (Argyris and Schön's double loop).\n5. **Every store has a carrying cost.** Mathlib keeps a backlog of debt; a rule loaded into every\n   session costs context and attention; pruning belongs inside the loop, not after it.\n\n## 5. Proposals (Claude's, unruled, in order)\n\n1. **Add a destination to every ruling**: *enforced by* a check, a resident line, or the row alone,\n   chosen by the ladder in § 3. The Canon field built 2026-10-22 is half of this.\n   **BLESSED 2026-10-22**, his words: *\"Bless the ladder.\"* It extends the Canon field (ledger\n   discipline 4, built the same day in another session): the field says where a ruling lives among the\n   things a session reads (resident, the core pack, with its note, declined); the ladder adds the rung\n   before those, a check in code when code can enforce the ruling. Wired the same evening by the canon\n   session (commit `77c1f33`): `check` is the first place, naming its test, hook or guard by path, and\n   `tests/test_ledger_canon.py` fails on a named path that does not exist.\n2. **One C look on a cadence**: each flywheel's measure this month against last (the correction\n   share, the corpus graph's held share, the share of cases a rule decides, recurrences after a\n   check), and which flywheel is now the constraint. It could ride on `/reflect`.\n3. **Choose his next cases by disagreement**: between passes, between models, or where a pass is\n   unsure. The Ruby judge review this week already worked this way, picking cases where its\n   variants disagreed.\n4. **Pace standing work to his review**, with a stop when its check fails; the sweep brake (R8) is\n   the start of this.\n5. **Retire canon lines that never fire**, once the always-loaded budget is felt.\n\nSources: the literature report beside this note, written by a research helper the same day and kept\nverbatim. Its § 6 (iv) lists what was read at a primary source and what was not; the papers by Huang,\nShankar, Sinha and Shumailov were read only at abstract level.\n\n\n## 6. The weekly flywheel check-in (built 2026-10-22)\n\nHis request: *\"come up with one powerful scheduled flywheel check-in. To start with. Then we'll do\nanother one and another one after that, later. The check-in should also review what's up with\nsignposts.\"* This is Engelbart's C level (§ 3) made routine, and it is a flywheel itself: its store is\nthe series of readings and check-ins, so each week compares against the last.\n\n**Two halves, split by what each is good at.**\n\n- **Code reads** (`scripts/flywheel_checkin.py`): every measure in § 2 that can be counted today, with\n  the change since the last reading, appended to\n  [`flywheels-in-building-deliberus/readings.jsonl`](flywheels-in-building-deliberus/readings.jsonl).\n  Ledger rows open, ruled and with a canon decision; glossary terms known and not yet; connections kept\n  by readers and by sessions; his own connections; walkthroughs; tests; ruling, canon and fix commits\n  in seven days; passes registered and standing, and the keeper's owed pairs. A reading that fails says\n  *no reading*, never zero.\n- **A session judges**, every Monday at 10:00 (a scheduled task in the Claude app, *Flywheel\n  check-in*, stored at `~/.claude/scheduled-tasks/flywheel-checkin/SKILL.md`): which measures moved, which\n  flywheel is now the constraint (Goldratt's fifth step), what the signposts register says (a reading that\n  crossed a trigger, a review date due, a row whose reading the counts now give), and **one decision for\n  him**, the one that blocks the others. It appends at most 150 words to\n  [`flywheels-in-building-deliberus/checkins.md`](flywheels-in-building-deliberus/checkins.md) and\n  commits only the two files of the check-in. It rules nothing and changes no code.\n\n**What it cannot read yet** (said so it is not mistaken for zero): the correction share (ruling-to-canon's\nreal measure, unoperationalised, TODO.md beside AK3 and AK4); the corpus graph's held share (read only\nwhen readers run); recurrences after a check; the share of cases a rule decides (the review tables);\nfaults per mapped source. Each becomes a reading when its instrument exists, and the check-in names the\none whose absence costs most.\n\n**First reading, 2026-10-22**: 530 ledger rows, 397 open, 112 ruled, 16 with a canon decision (96\nwaiting); 50 of 399 glossary terms known; 873 connections kept (52 in seven days); 22 walkthroughs;\n3,000 tests; 1 of 24 registered passes standing, its pairs 715 owed.\n\n**How to stop it**: disable or delete *Flywheel check-in* in the app's scheduled tasks. It runs while the\nClaude app is open; a run missed while the app was closed runs at the next launch.\n"}