{"path":"research/a-tree-per-claim.md","content":"# A tree per claim: microtasks composed around the mind of Deliberus\n\n**Date**: 2026-10-28 · **Type**: his direction, worked through against the standing-work design and the rulings it\ncomposes · **Status**: Claude's reading and proposals; nothing here is ruled, and nothing is built. It waits behind\nthe 2 November reading of the keeper's held-out pairs and the hold on microtasks.\n\n## His words\n\nOn the design for the keeper ([overlaps-and-analogy.md](overlaps-and-analogy.md) §§ 12–14): *\"It should. It should\nbasically \"think like the mind of Deliberus\" within the microtask. As far as we can take it.\"* Then: *\"I want to go\ntowards being as stringent as possible in this so without making double model calls or any such copouts, how could we\nensure the LLM does what it should each time?\"* And on reading the answer: *\"I love the process we're making here.\nDocument all these ideas carefully. This kind of looping in on itself through a fractal graph-building might be\nnecessary for other tasks to do their work optimally also. Not sure. Perhaps... if the microtask confidently detects a\nlikely analogy/metaphor, it is handed off to be dealt with by the extraction microtask or the analogy daemon/microtask.\nJust to have the microtasks be DRY and refactor them a bit, maybe compose them, use them for different complementary\npurposes. But I might be getting ahead of myself.\"*\n\n## 1. The idea in one paragraph\n\nA microtask judging two claims does best when it reasons as the full graph would: each claim with its parts cut to the\nfloor, the generalization it states or is a case of, its reference class, the kinds of evidence that would settle it,\nand the premises in dispute. That structure belongs to the claim, not to any pair, so it is drafted once per claim\n(as against once in every pair the claim is in, about a dozen today) and read by every microtask that needs it. Relations\nbetween two claims are then read off their two trees, by code wherever the trees settle it. Each node of a tree is a\nclaim with a tree of its own, so the process repeats at every level, which is the self-similar decomposition principle\napplied to the machine's own work.\n\n## 2. Pair 1, the case that led here\n\nA, from the Anthropic text: *information becomes consciously accessible when it gains entry to a small shared channel,\nthe workspace*. B, Anil Seth's: *global workspace theory proposes that conscious experiences in humans occur when\ninformation is made widely available to other parts of the brain*. Both render one generalization, G: *in any system,\ninformation that enters the system's global workspace (a stage of limited capacity whose contents are broadcast to the\nsystem's specialised processes) is consciously accessible in that system.* A states G almost whole; B restricts it to\nhumans and says *experienced* for *accessible*. So A is wider in scope, B stronger in outcome, and neither entails the\nother (clingo: three premises from A to B, five from B to A). The analogy he saw runs from the human evidence under B to\nthe language model the Anthropic text is about, through G, and rests on the claim that the model belongs to G's\nreference class.\n\n## 2b. The same idea, earlier, in his own words\n\nThese are his requests for recursive decomposition **of the graph itself**, which he calls the base level (2026-10-28). His direction of that week is a different level, inside one microtask: the microtask pictures the full graph to reach its one verdict (§ 2c).\n\n- 2026-03-29: *\"I want people and AI agents surfing around in Deliberus to see exactly what the best minds on our planet\n  think in detail and what the best evidence shows - evidence that also is decomposed into its constituent\n  parts/arguments.\"* ([scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md))\n- 2026-08-22: *\"Create a bunch of synthetic examples and do recursive decomposition on their subgraphs/trees of\n  claims/evidence on all of them, feed it all into the live argument graph and reflect on the ontology afterwards.\"*\n  ([synthetic-stress-suite-and-ontology-reflection.md](synthetic-stress-suite-and-ontology-reflection.md))\n- 2026-08-25: *\"I want Deliberus to automatically recursively decompose all claims in each mapping-break. It should\n  find as much research of high quality as possible and feed everything into the graph.\"*\n  ([kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md))\n- 2026-10-10: *\"We must not make too-large leaps/steps, that would be irresponsible for any scientist. We must examine\n  and make it possible to review each step of the reasoning tree/chain/graph.\"* And: *\"We also must decide WHEN the\n  machine writes its guesses and when it widens and narrows and based on what (the subgraph? which subgraph exactly,\n  connected to which exact claims/links?)\"* ([session49-what-goes-under-a-link.md](session49-what-goes-under-a-link.md) § 1)\n- Ruled the same day (AG6): *\"It may and should write deeper steps. But the microtasks in charge of doing so need to\n  be quality-controlled, just like the other microtasks that are in play in Deliberus.\"*\n\nThe question of 10 October was asked of the steps a machine writes under a link, and was answered then by two named\nreading sets, the link's neighbourhood and the rule's record ([the-chain-inside-a-link.md](the-chain-inside-a-link.md) § 5.2).\nThe tree adds a third: each claim's own, drafted once and read by every microtask that relates claims.\n\n## 2c. Two levels, which the first version of this note ran together (2026-10-28)\n\nHis correction: *\"Well yes but that was for the \"base level\" recursive decomposition (what could we nickname this?\nMaybe acid something? Enzyme? Ideas?) not this meta / within a microtask level. Right\"*.\n\n- **The base level** writes the graph: claims cut to the floor, evidence decomposed, every level kept, read by\n  everyone, standing work filling it over months. His requests of March, August and October above are this level.\n- **The inner level** is reasoning inside one microtask: to reach its one verdict on a pair, the model pictures what the\n  full graph would hold around the two claims, as deep as one call allows (his *hundred years* graph), and reads the\n  relation off that picture. It is a way of thinking, like a written chain of reasoning, and nothing in it needs to be\n  stored.\n\n**What this note proposed is a bridge between the two, and the bridge is a separate decision.** Keeping what a\nmicrotask pictured (the tree per claim) turns inner-level work into base-level drafts. For: a claim's picture is\ndrawn once and read by every pair and every microtask, and the graph fills. Against: every kept piece is a draft\nnobody asserts (AG12) under T1's cap, a picture can go far deeper than anything worth storing, and a picture drawn to\nanswer one pair is shaped by that pair. The inner level stands without the bridge: a microtask can picture the graph\nand keep only its verdict and the few nodes the verdict rests on (AM1). Whether to keep more is his call (`TODO.md`,\nitem 6).\n\n**Names, blessed 2026-10-28** (*\"Blessing the names!\"*). For the base level, **the acid**, from the corpus's own *Reason as Universal Acid* (Dennett's\nterm, [depth.md](../depth.md)): reason applied to every claim until it reaches the floor. *Enzyme* is taken: the\ndaemon design's *enzyme constraint* says a background agent may change the rate of a change people would make, never\nits outcome ([graph-daemons-design-space.md](graph-daemons-design-space.md) § 2); a sweep that applies the acid obeys\nthat constraint, so the two words would sit in one sentence with different jobs. For the inner level, **the imagined\ngraph**, which says what it is. Both are names of our own artefacts, exempt from the rule that a concept takes its\ndiscipline's name (BD11); the nearest discipline names for the inner level are step-back and decomposed prompting.\n\n## 2d. How the imagined graph is made: reuse and creation, never picturing (2026-10-28)\n\nHis words: *\"The imagined graph could be simulated by creation/reuse, any other way muddles it, maybe? Would prompting\nand within-call harness tool use with clingo etc suffice? Read up on tool use in model calls for the microtasks,\nmentioned this in the past\"*, and: *\"simulating by actually doing it or like deferring to it would probably be the most\nreliable method since it would anyway in the case of the analogy have to maybe run another or even several of the\nmicrotasks/sweeps that eventually will make up part of foundation of the mind of Deliberus ... but then that couldn't\nbe run from inside the microtask and the model just picturing the graph feels risky so idk. But microtasks could be\ncomposable etc yeah. And a sweep could entail complex interplays between well evaluated microtasks I guess. Checked by\nlogic or other code as possible/necessary.\"*\n\nWhat the record already holds on tools inside a call: BD23, ruled (*\"Ruled, blessed, hallelujah\"*), names harnesses\n*\"woven into the microtasks and sweeps, grounding a pass dynamically\"*; AM16 (open) has code compile filled slots into\nclingo or ProbLog; AM17 (open) sends a refused inference back with the failing slot named, on Ishay and Lee's result\nthat solver feedback raised a model's accuracy from 54.6% to 78.0% at about two revisions, while self-checking\nwithout an outside signal does not help (Huang et al. 2024;\n[tools-for-checking-inferences/report-harness-design.md](tools-for-checking-inferences/report-harness-design.md));\nand his word on AM14's one round back, *\"The whole thing about making repeated model calls feels like a copout to me\"*.\n\n**Proposed, unruled: three roles, kept apart.**\n\n1. **A microtask is one judgment by one model.** Inside its call it may use tools that are code, never another\n   model: a read of the graph around its claims (reuse: what is already written), the reference-class counts, and\n   clingo or ProbLog on what it has filled. A tool answer is information the model lacked, which is not asking the\n   same question twice; a model inside a model is, and it would also escape the record and the per-field measurement.\n2. **The imagined graph is made by reuse and by creation, never by picturing.** What exists is read. What is missing\n   is created by the microtask that owns it, run earlier in the sweep. Where something is still missing, the microtask\n   writes *owed* (the condition another microtask's sweep reads) and gives no verdict that rests on it, which is the\n   confession channel. Picturing stays allowed only as the choice of what to look up.\n3. **A sweep composes microtasks.** It runs them in data order (standing work, principle 5): a claim's acid first,\n   then the pair. Code checks between them (the form checks, the engines). The interplay of well-evaluated microtasks,\n   checked by code, is the part of the mind of Deliberus that machines run.\n\n**An analogy, so built**: the pair's microtask reads both claims' acid from the graph; the generalization they share,\nthe cases under it and the membership claim are created by the microtasks that own them, earlier in the sweep; the\npair microtask then names the relation, and code checks it; the analogy daemon's second step owes the mechanism test.\n\n**His search for a brain part**: the corpus already has a better picture, Hofstadter's Copycat, which the analogy\ndaemon was found to re-describe ([perceiving-possibility-and-the-analogy-organ.md](perceiving-possibility-and-the-analogy-organ.md)):\nmany small specialised *codelets*, each doing one thing, scheduled from a shared *coderack*, with testers checking\nwhat builders propose. Microtasks are codelets; a sweep is the coderack; code is the tester.\n\n**A strange loop?** His words, 2026-10-28: *\"So the analogy daemon is a very Hofstadter like idea then. A strange loop in\nthe graph.\"* Hofstadter's idea, yes; a strange loop, not by itself. A strange loop is a system whose levels cross\nback on themselves, so that what a higher level says acts on the level it is about (Gödel's sentence about its own\nprovability; the self, in *I Am a Strange Loop*). The daemon finds likenesses between claims and writes them as claims:\none level up, not yet back down. The loop closes in two places the corpus already names, both held: Deliberus's own\nrules argued as claims and those claims then governing the passes ([claims-and-code-one-loop.md](claims-and-code-one-loop.md),\nheld 2026-08-30), and a fractal prior about Deliberus stored in Deliberus (same hold). A small one is live today: the\nkeeper's prompt is rendered from ledger rows, so a ruling written in our docs changes how the pass reads the graph\n(ledger AM8). Unruled.\n\n**Built 2026-11-06, under his word of that night** (*\"Build all these out to the best of your ability. You can test and hone using credits. I trust your judgment.\"*): `deliberus/workup.py` drafts a claim's workup from what earlier passes wrote (the scope reading, the hedge, a conditional's half or a minted report, the live parts of a cut, the reasons taken off) and the keeper receives both claims' workups as named inputs of its call (`keeper.build_prompt(a, b, about)`, the version one fingerprint across pairs); what no pass has written yet is carried as `owed` on the pair's reading, the confession channel of the second role above. The sweep order that fills the workups before the pair is read is the standing-work order (scope first, then the cut, then pairs), run source by source that night. Not built: a workup read by any pass but the keeper; the handoff of a likely analogy to the daemon (the keeper records it on the resemblance link instead, `scripts/analogy_candidates.py`).\n\n## 3. What the corpus already holds that this composes\n\n| Already held | Where | What it gives the design |\n|---|---|---|\n| A microtask writes the piece of the subgraph its verdict rests on | ledger AM1; [three-minds-and-the-leap.md](three-minds-and-the-leap.md) § 4 | the tree is that piece, made whole |\n| Taught the method, never the graph as a lesson | AM12 | the tree is drafted by method; written structure is read by code, never shown as an example (AX2) |\n| Cut first, sameness among the pieces, links last | the floor-first order ([one-challenge-one-home.md](one-challenge-one-home.md)) | the order inside a claim's workup |\n| Found before made | AG2; match before mint, AL16 | every drafted node is matched against the graph before it is written |\n| A machine's sentence is a claim from the moment it is written, asserted by nobody, moving no strength | AG12 | the tree costs nothing in strength until it is examined |\n| The condition owns the work; running an item twice writes nothing twice; order follows the data | standing work, principles 1, 4, 5 ([machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) § 7) | handoffs are conditions on the graph, not calls between microtasks |\n| Merge microtasks, calibrate per field, version per prompt | [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) § 2c | how microtasks may be composed without losing measurement |\n| Machine work built on machine work is capped by a measured error rate | T1 | how deep the recursion may run |\n| A microtask writes nothing twice | AM21 | one tree per claim per version |\n| One definition per key, identical in the instructions, the review page and the code | T5, ruled ([microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md)) | the relations read off trees use the same definitions |\n| Derive the key from yes/no questions: does A entail B, does B entail A, can both be true, is there a statement both assert | [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) § 4 item 2; ledger Z4 | with trees, code answers all four |\n| The keeper's *overlaps* was mostly one claim being a reason for the other (about 64 of 89) | same note, § 3f | on trees, a reason is a link and an overlap a shared node; they stop sharing a word |\n| The splitter's harmful fault: a clause the claim denies or concedes written as fact; the part test *if this part were false, would the claim be weaker?* | same note, § 3a, § 3e | stage 1 of a workup keeps each clause's polarity and asks the part test |\n| Stability within each question, diversity across questions | same note, § 4 | two differently framed checks are two questions, not asking one twice (AD14, which he never blessed, stays open) |\n| Filled slots compile to the engines; a small trusted checker | AM16; [claims-and-code-one-loop.md](claims-and-code-one-loop.md) § 4 | the code that reads trees is the small trusted part |\n| Deterministic code as a microtask | his direction of 2026-10-24 (same note, § 7) | relations read off trees by code |\n\n## 4. The workup of one claim, as composed microtasks\n\nProposed. Each stage is a microtask that exists or is designed; the composition is the new part.\n\n| Stage | What it writes | Existing or designed microtask |\n|---|---|---|\n| 1. Cut to the floor | the claim's parts, each standing alone | `split.propose`, `coverage.check` |\n| 2. Reasons and conditionals off | inline reasons as support links; an *if* as a link | `reasons_off.read`, `reasons_off.check_conditional` |\n| 3. Generalization and reference class | the generalization the claim states or is a case of, its kind (strict, normic, statistical), and for a case the membership claim | the reference-class membership microtask (ledger BC3, BD2), designed |\n| 4. What would settle it | the kinds of evidence, and the premises in dispute, each a draft claim | the seeding microtask (BC6), designed |\n| 5. Reported speech | whose claim it is and what is reported | `reported_speech.split`, `attribution` |\n\nEach stage matches before it mints, so what exists is found. Its output is the claim's tree, kept in the graph, read by\nevery later microtask, and redrafted only when the claim's text or the stage's version changes.\n\n## 5. Microtasks that would read the trees\n\n| Microtask | What the trees let code read | Model still needed for |\n|---|---|---|\n| The keeper (`keeper.pair`) | one node (same); a part inside the other's tree (entails); contrary nodes (opposes); a shared premise both assert (overlaps); two cases under one generalization (analogy) | relations the trees do not settle |\n| The independence reading (`independence.pair`, ledger Q40) | a shared rival explanation is a shared node | rivals not yet written |\n| The relevance judge (`relevance_judge`) | for a case read through a generalization, the generalization's strength times the membership's (⚠ 2026-11-07: a product of two numbers is not Q50's relevance, the difference the reason makes; a conjunction, a rate and confidence in a generalization are different quantities, and the product needs a stated model before code may compute it: Codex's review, `.private/collab/codex-analogy-generalization-review-2026-11-07.md` § 6) | links whose generalization is not written |\n| Prior seeding (BC1, BC3) | the reference class is on the tree | membership where it is disputed |\n| Folds and routes on the map (AX4) | the route between two claims runs through tree nodes | nothing |\n\nPredicted, not measured: that each microtask gets better and cheaper reading trees. The test is one microtask at a time, on\ncases with known answers read with him (the credit's evals, `TODO.md`).\n\n## 6. Handoffs, as conditions\n\nHis proposal that a microtask hand a likely analogy or metaphor on fits the standing-work principles exactly, with one\nchange: a microtask never calls another. It writes a state, and the other microtask's condition owes it.\n\n| A microtask finds | It writes | Which microtask then owes it |\n|---|---|---|\n| a likeness with a shared mechanism | a draft analogy: the generalization, the two cases, the membership claim, the disanalogy as an attack | the analogy daemon's second microtask (S6's test) |\n| a likeness with no mechanism | a metaphor, kept as a seed | the daemon's human-seeded channel |\n| a part still bundled | a separability proposal | `split.propose` (already so) |\n| a case with no reference class | a case missing its class | the membership microtask |\n| a claim with no tree | nothing; the claim owes a workup | the workup's first stage |\n\nTermination is given by the principles already ruled: each item is owed once per version, and a stop answer counts as\nread, so no loop runs twice on one item.\n\n## 7. The fractal loop, and where it stops\n\nEach node of a tree is a claim, so it owes a workup of its own; each relation found between trees can write a new\ngeneralization, which owes one too. The recursion stops at three bounds the corpus already has: the separability floor\n(B2, AG14), the measured error rate that caps how deep machine work may build on machine work (T1), and the brake on\nsweeps (R8: a few calls per microtask until a microtask is proven). As microtasks are proven the bound moves down the tree; that is\nthe graph filling, and it is what makes later microtasks cheaper, since what a tree holds is read by code.\n\n## 8. DRY and composition\n\nThe same one-item functions serve several purposes: the splitter's parts are the keeper's evidence of containment, the\nmembership microtask's reference class is the prior's and the relevance judge's input, the seeding microtask's contested premises\nare the independence reading's candidate shared rivals. Composing microtasks into one call follows § 2c's rule: calibrate\neach output field on its own, version the whole prompt, and accept that any prompt change re-reads every field.\n\n## 9. What could go wrong\n\n- **Compounded machine error.** A tree drafted by a model and read by another model is machine work on machine work;\n  T1's cap binds, and the trees move no strength until examined (AG12).\n- **One frame everywhere.** A tree is the model's frame; trees drafted by one model family share its omissions. What\n  no source holds (coherent absence) and the frame itself stay with people, by the canon.\n- **Coarser re-sweeps.** A composed microtask re-reads every field on any change (§ 2c's priced cost).\n- **Cost up front.** A workup is more output than a pair's key; it pays back only if trees are reused across pairs and\n  microtasks, which is the measurement to make first.\n\n## 10. Order of work, proposed\n\nAfter the 2 November rulings: the keeper's likeness drafted as tree structure (overlaps-and-analogy.md § 10, § 12),\nthen one claim's workup on pair 1's two claims by hand, then the workup as composed microtasks, one microtask at a time under\nR8's brake, each measured on cases read with him.\n\n## How this was done\n\nRead whole for this note: [overlaps-and-analogy.md](overlaps-and-analogy.md) (its author's), [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) § 2c and § 7,\n[microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md), [claims-and-code-one-loop.md](claims-and-code-one-loop.md),\n[three-minds-and-the-leap.md](three-minds-and-the-leap.md) § 3–4, `deliberus/keeper.py` (answer form, `resolve`, routing),\n`standing_work.MICROTASKS`. Read in part: [session49-what-goes-under-a-link.md](session49-what-goes-under-a-link.md) (§ 1, § 8),\n[what-counts-as-evidence/corpus-map-founder-voice.md](what-counts-as-evidence/corpus-map-founder-voice.md) (§ 0, § 1h,\n§ 1cc to 1ee). Headings only: [what-deliberus-is-for.md](what-deliberus-is-for.md),\n[shared-sensemaking-what-deliberus-holds-and-owes.md](shared-sensemaking-what-deliberus-holds-and-owes.md). The hook\n`research-read-first-nudge` named the last six when the note was first written. No helpers, no outside check, no tests\nrun; the one case is pair 1.\n\n## Cross-references\n\n[overlaps-and-analogy.md](overlaps-and-analogy.md) · [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md)\n· [three-minds-and-the-leap.md](three-minds-and-the-leap.md) · [one-inference-per-link.md](one-inference-per-link.md)\n· [the-analogy-daemon.md](the-analogy-daemon.md) · [priors-as-standing-work.md](priors-as-standing-work.md)\n· [self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md)\n"}