{"path":"research/retrofitting-the-reasons-channel.md","content":"# Retrofitting the reasons channel: what the parts rulings meant, read for reasons\n\n*2026-10-15. The founder, the morning after the Dr Smith redraw: \"reasons were mistaken for parts\nearly on, right? So any decisions that have been made for the 'parts' 'channel' that were\nsemantically confused should actually be retrofitted with all applicable features/rulings if that\nholds after some serious deep thought.\" Then four more questions in the same breath: whether the\ncompleteness check for parts fooled us about world completeness; whether concept tracking can\ncarry relevance and completeness; whether the sameness keeper is the independence detector Pollock\nsaid no reasoner has; and how the shared-origin check works. The corpus graph was asked first\n(`around` on the number, priors, evidence, keeper and concept notes; `search` on independence),\nand every passage cited below was opened. Nothing here is ruled.*\n\n## 0. The short version\n\n- **Four rulings on the parts channel were made with a reason as the example.** Each carries to\n  the reasons channel, but not as it stands: the weakest-part rule carries as *weakest premise\n  within one line of reasoning, best line across lines*, which is Pollock's arithmetic and his own\n  2013 sketch; the tag vocabulary (`necessary` / `corroborative` / `alternative`) was borrowed\n  from support semantics in the first place and wants its three words re-defined on reasons; the\n  requirement cap already crossed over as Q34's relevance; the fair-share discount does not cross\n  at all (§ 1).\n- **The \"needed\" sensing answered two closed questions and never the open one.** Whether a part\n  is needed is decidable from a well-defined claim; whether a premise is needed is decidable from a\n  written-out step; whether every consideration that exists is in the record is not decidable from\n  anything. What a well-defined mother claim buys is the *space* of what would bear on it, not\n  whether anything in that space does (§ 2).\n- **Concept tracking is a relevance input, the cheapest one, and an equivocation detector; it is\n  not a relevance judge.** His 2013 board already had it (*concepts are defined by all statements\n  using the concept*), the keeper's prompt already has the sense clause, and the one missing piece\n  is that which sense a claim uses is modelled and not stored (§ 3).\n- **Two thirds of the independence detector exist and run.** The keeper's `same` catches\n  reformulation and its `contains` / `overlaps` catch dependence in content; both are live in\n  auto-connect and neither is run between the sibling reasons of one claim. The third kind, shared\n  origin, is provenance, not sameness: ledger Q4, unbuilt. The fair-share discount is none of these\n  (§ 4).\n- **The lift's independence question is the keeper's microtask**, closed, two short items, swap\n  checked: *would establishing one of these, apart from what it does for the claim, advance the\n  other?* (§ 5).\n\n## 1. The parts rulings, each read for what its example was\n\nWords, as the number note defines them (`what-the-number-on-a-claim-means.md` § 1): a **part** says\nsome of what its claim says; a **reason** argues for a claim without being part of what it says;\na **channel** is one of the two sets of links a claim's number is computed from. One word this note\nneeds that the number note did not define: a **line** (his 2013 *argument bundle*,\n`docs/object-model.md`) is several reasons that together make *one* argument for a claim, so that\nremoving one breaks the line; Pollock's one argument with several premises, Prakken's accrual read\nas a linked argument. A claim can have several lines. Today the graph stores lines nowhere: every\nsupport link hangs off the claim on its own, which is the part of the 2013 sketch the map note\ncalls *not built* (`how-a-map-is-cut-and-what-it-shows.md` § on bundles).\n\n| Ruling, date | The example it was made on | Part or reason? | On the reasons channel it becomes |\n|---|---|---|---|\n| **B10, the weakest part governs** (2026-08-31) | two: padding (three parts against twelve), and Dr Smith's credentials as *jointly required* | padding: parts, genuine; Dr Smith: reasons, misfiled | **min within a line, max across lines.** Min across all of a claim's reasons would treat them as jointly required, so an added weak reason lowers the claim: the wrong payoff the 24 misfiled reasons sit under today. Within one line min is Pollock's weakest link, and product within a line would be the legal conjunction paradox, honest depth punished (`weakest-link-arithmetic-and-the-merge-hunch.md` § 1). Max across lines is Q37. The padding defence on reasons is the max, not the min: a reformulation added as a second line cannot beat the first by more than its own standing, and under max a weaker copy adds nothing |\n| **`support_semantics` tags** (2026-08-20): `necessary` (min), `corroborative` (sum), `alternative` (max) | Dr Smith (necessary) against three studies (corroborative) | both reasons | the vocabulary came from Nouioua and Risch's *necessary support*, a relation between arguments, and was applied to parts. On reasons the three words have exact jobs: `necessary` = premises of one line (min within it); `alternative` = separate lines (max across, Q37); `corroborative` = several reports of one event, which **saturate** (Schum; ledger Q22), never add. A reader found the collision on 2026-10-10: in code `corroborative` adds, in the evidence note it saturates. The retrofit settles it by giving the word the evidence note's meaning on reasons |\n| **O13, the argued-requirement cap** (2026-09-17) | a part caps its whole by relevance strength, 0.5 presumed on a confirmed part | parts | already carried: Q34 (2026-09-30) put relevance on the link with the same presumption and the same way to argue it. One asymmetry is right and stays: a part can cap its claim, a reason cannot (a weak reason drops out), except that a weak premise caps *its line* |\n| **B2, separability as the floor** (2026-08-31) | how far to split | parts | carried: AG14's third clause (2026-10-14), one stated rule, once, to one stated case; the same floor in the other direction (`the-chain-inside-a-link.md` § 10) |\n| **The rebuild check** (coverage, 2026-08-27; writing since 2026-09-21) | does a cut drop or invent | parts | no analogue is owed: the reasons-side check is the chain's no-gap and removal test (AG14). The pair is already parallel |\n| **The fair-share discount** (`evidence_routing.py`, 2026-08-14, never adopted) | one piece of evidence behind several parts, reaching the mother several times | parts | does not transfer. Two reasons with one origin are a different object, and the discount never saw them (`pollock-against-accrual-and-the-overlap-check.md` § 5). The instruments for that are § 4 |\n| **The part test** (`checks-on-machines-and-people.md` § 11; T9) | 28 of the 39 parts people attached would be withheld; 24 are what the claim rests on | reasons, filed as parts through a box that asks *What's underneath?* | the refile. Under today's sum it raises 24 claims on no new evidence (the inflation fear); under Q37's max it raises nothing. So the refile is done after Q37 or with its effect measured first, and the box's wording is a separate fix |\n\nSo the retrofit is not \"apply the parts rulings to reasons\". It is: **the line is the object the\nrulings were about**, and once lines exist the parts arithmetic (min inside) and Pollock's (max\nacross) are the same rule at two scales. Q37 said *best reason* because lines do not exist yet; with\nlines it reads *best line, each line as strong as its weakest premise*. The 2013 bundles, Pollock's\nweakest link and Q37 are one picture.\n\n**What the retrofit asks to build, in order.** A grouping of a claim's reasons into lines (the\nlinked/convergent distinction; Yu 2025 reads it off scheme structure, our proxy is same target,\nsame scheme; Prakken 2019's accrual sets, supplied not computed, so a person or a pass proposes the\ngrouping and it is attackable); the three tag words re-defined on support links; max across lines\n(Q37); the refile of the 24.\n\n**A simpler representation to compare (founder, 2026-10-17):** let a reason state its combined case\nin its claim text, with separately challengeable parts. Other reasons join the same shared case for\nthe conclusion. This may represent a line using an ordinary composite claim; compare it with premises\nunder a relevance link before adding a new object type. This compares representations of the existing\ncase/rule distinction and AG14's inference floor, not a newly discovered need for joint premises.\n[One challenge, one home](one-challenge-one-home.md) § 2a; [recovered drawing and legacy audit](legacy-object-model-harvest.md).\n\n## 2. The completeness check, and whether we fooled ourselves\n\nThree questions wear the one word *needed*:\n\n| \"Is it needed?\" asked of | Means | Closed or open | Instrument |\n|---|---|---|---|\n| a **part** | the claim is false without it | closed, given a well-defined claim | the part test; the rebuild check |\n| a **premise** of a line | the line fails without it | closed, given the step written out | AG14's removal test |\n| a **reason** across lines | nothing: a line is one of several | not a question; Q34 lets a failed line drop out | |\n| **the record** | every line that exists is in it | **open** | none; only the search is recordable (`world-completeness-and-its-relatives.md` § 0) |\n\nThe sensing of need answered the first two, and both are real: a part that fails the part test\npresumes no need (checks note § 9 step 1, read-and-show for now), a premise that fails the removal\ntest is idle (the ten links of 2026-10-13, idle in 8 of 29 steps). Neither says anything about the\nfourth row. **Where the muddle entered**: from April to October a reason was stored as a part, so\n*every part is needed* (true, and checkable) sat on things that were reasons, where *needed* is not\neven a question. From there it is one step to reading the completeness oracle as saying what it\ncannot: `fully_exposed` at depth zero flatters by construction (dogfood run 4), and the oracle's\nown name invites the reading. So: nobody wrote that sensing need yields world completeness, and the\nfiling made it look as if it did.\n\n**What a well-defined mother claim does buy**, and this is the half of his intuition that holds: it\nfixes the *space* of what would bear on it. The derivable requirement space (`derivability-of-\nmissing-considerations.md` § 4, rungs 0 to 4) is enumerable from the claim alone: each part's\nnegation, the scheme's critical questions, and (§ 3 below) one definitional question per contested\nterm it uses. That space can be pre-populated for every claim alike, which is why selective\nscrutiny stops paying. What the claim cannot buy is whether anything in the space is *true*, and the\ntail outside it, evidence from a position (rung 5), which is corpus-absent by definition. So world\ncompleteness splits the way the harness conviction's ceiling does: the questions are derivable, the\nanswers and the tail are not. \"More feasible given a well-defined mother claim\" is right about the\nquestions and wrong about the answers, and a badge that showed *fraction of the derivable space\nanswered* would be the honest proxy, not a completeness.\n\n## 3. Concept tracking as a relevance input and an equivocation detector\n\n**Sharpened by the founder, 2026-10-16:** examine the senses at both ends of support and attack.\nThe comfort example distinguishes a definition, a defeasible indication and a changed context; a shared\nsense is not itself evidence of relevance, and different senses can be connected through an explicit\nbridge. [One challenge, one home](one-challenge-one-home.md) § 2 gives the tests and Lins comparisons.\n\n**His 2013 board had it.** *\"Tracking of concepts … Concepts are defined by all statements using\nthe concept\"* (Proposal FB-01, January 2013; the public record of the board is `docs/history.md`).\nThat is the position the corpus reached again on 2026-10-06 as *store the uses, read the meaning*.\nHis intuition this morning (*\"by tracking which concepts are used, we can indirectly judge relevance\nand it influences our completeness-sensing\"*) is the same board read one step further.\n\n**What is built.** The keeper's prompt carries the sense clause: *the same words carrying a\nshifted sense of a contested term are not the same claim* (`deliberus/keeper.py`, the SAME rule).\nIt reads sense from the text of the pair, and it stores nothing about which term: a field naming\nthe contested term an `opposes` pair turns on was deliberately left out, because nothing would have\nread it (the file's own note). Concept tags exist on 465 of 1,741 substantive claims, from\nextraction time only; which sense a claim uses is modelled and not stored\n(`core-periphery-and-the-concept-layer.md`, measured 2026-09-06); a sense is to be recorded as\nsomebody's delta from a reference, not as a description (ledger N4).\n\n**What concept tracking can do for relevance, stated so it cannot be over-read.**\n\n1. *A necessary condition, computed by code.* A reason that shares a contested term with its claim,\n   in the same sense, passes the first gate for free; a reason that shares no term bears on the\n   claim only through a rule that bridges them, which is what `what-a-rule-rests-on.md` says a step\n   rests on. With senses stored per use, this is an id comparison: no model, no steadiness problem.\n2. *An equivocation detector.* His car case: the claim's author uses *car* for a vehicle that can\n   drive itself; the attacker uses the ordinary sense. The attack lands on a different proposition.\n   Today the keeper would read the pair as `overlaps` or `unrelated` from the text and record nothing\n   about why. With the term stored, *shared term, different senses* is a query, and the answer is\n   not \"irrelevant\" but \"relevant only under a sense the claim does not use\", which is a reason\n   anyone can read, unless a stated bridge crosses the two senses (an ordinary, challengeable claim, AM7). This is the one attack class no scheme's critical questions catch except\n   Walton's verbal classification.\n3. *A definitional slot in the derivable space.* Each contested term a claim uses adds one question\n   to what would bear on it: *is this the sense at issue?* That is where concept tracking feeds the\n   completeness of § 2, and it is enumerable.\n4. *Why a sense tag beats an inline definition*, which he asks: a definition typed into every claim\n   is repeated per claim and still has to be matched; a sense is a pointer to one claim on the word\n   (ledger AC: a definition of a wanted word is a claim on it), attacked once, reused everywhere, and\n   the delta is the small object where the definition is the large one. The exception is the\n   stipulated technical term, whose delta from lay use *is* the definition; it is stored once, as a\n   definitional claim, and pointed at.\n\n**What it cannot do.** No unnoticed equivocation is necessary and not sufficient (a difference in sense\nis allowed where a stated, challengeable bridge crosses it, as the sharpening at the top of this section\nsays and AM7 rules; item 1 covers only a reason that shares no term): the squirrel step shares every\nconcept with its claim and still needs its rule. And today's sense tags are machine-proposed, so a\nwrong tag manufactures false irrelevance; the reading is shown and argued, never a switch (the checks\nnote's § 11.6 rule, applied one layer over).\n\n## 4. Independence: what the keeper already is, and what shared origin needs\n\nPollock's ground for refusing the probability reading was that a reasoner cannot tell from its own\nrecord whether two arguments are one argument twice (`pollock-against-accrual-and-the-overlap-\ncheck.md` § 1b). Three different things hide in \"two arguments are one\":\n\n| Kind | Example | Instrument | State |\n|---|---|---|---|\n| **Reformulation**: one proposition in two wordings | a reason restated | the keeper's `same` (a consent test: would the author accept the other as a restatement) | built, live in auto-connect since 2026-09-22; `same` routes to a merge candidate |\n| **Dependence in content**: one reason entails or partly establishes the other | *the inspection passed* beside *the inspector signed the form* | the keeper's `contains` and `overlaps` (*establishing one would partly advance the other*, judged by evidence, not topic) | built, live; the verdict routes to a separability proposal and nothing the strength reads |\n| **Shared origin**: two different propositions with one cause | two reports of one study; Pollock's two jokesters who conferred | sources as nodes, inputs from one origin as one channel, `SHARES_ORIGIN_WITH` proposed and keyed on shared errors (`what-counts-as-evidence.md` § 8c, ledger Q4); his words of 2026-10-14, *\"Overlapping roots of evidence obviously shall not be added up\"* | designed, unbuilt |\n\nSo he is right about two of three: the keeper is the reformulation and dependence detector, and\nthe question it asks for `overlaps` is the independence question. What is missing is a scheduling\nchange and a reading: **the keeper runs over pairs auto-connect proposes within a source, not over\nthe sibling reasons of one claim**, and its verdict is filed as a separability proposal rather than\nread as an independence fact. The third kind cannot be read from text at all, since two true reports\nof one study share no error; that is why Q4 keys copy detection on shared *errors* and why it needs\nprovenance, not wording.\n\n**How the shared-origin check that was \"never turned on\" works, exactly.** It is the fair-share\ndiscount in `evidence_routing.py`: one piece of evidence routed to *k* parts contributes 1/*k* to\neach, so the mother collects it once. It changes nothing when nothing is shared, can only lower a\nmother's number, and was never wired into scoring. It covers one object only: one evidence claim\nbehind several parts. It never saw two reasons with one origin, and after the 2026-08-31 min on parts\nit is moot for its own case (`pollock-against-accrual-and-the-overlap-check.md` § 5). So \"the overlap\ncheck was turned off\" overstates: a discount for a different problem was built and not adopted. The\ncheck he means, for reasons, has never existed; its pieces are the keeper (built) and Q4 (not).\n\n**Pollock's point survives in one form.** No detector is complete; here it does not need to be,\nbecause independence is a contribution: the keeper proposes, a person files *these are the same\nargument twice* as an ordinary claim, a source node makes shared origin a query. That is what the\nnumber note's § 4 meant by *the independence can be contributed*, and it should have named these\nthree; corrected in that note today.\n\n## 5. The microtask\n\nFor a claim with two or more reasons, the question is closed and has two short items:\n\n> R1 and R2 are each given as reasons for C. Setting C aside: would establishing R1 make R2 more\n> likely to be true, or the reverse? Answer one of: one restates the other · one contains the other\n> · they partly establish each other · neither bears on the other · cannot tell.\n\nThat is the keeper's seven keys with `opposes` removed (two reasons for one claim do not oppose) and\nC held out, so the judge cannot answer from their common conclusion. It has the properties the\npass-quality note found cheap models hold (`pass-quality-and-the-choice-of-model.md`): a closed\nquestion on short texts read the same on every model tried; the swap check the keeper already\ncarries (`judge_pair(..., swap_check=True)`) is a real consistency test for the asymmetric keys; and\nreason-before-verdict is the keeper's field order. What it feeds: under Q37 the answer is the first\nanswer to the lift's own critical question, *are these independent?* An answer of *neither bears on\nthe other* lets the lift claim stand on content grounds; *contains* or *restates* withholds it;\norigin is Q4's question and is not asked here. Cost: pairs per claim, bounded the way linking is\nbounded (`--max-partners`); on the live graph the claims with two or more person-added reasons are\nfew (the checks note counted 32 reasons and objections in all).\n\nWhat no microtask does: tell whether two reports came from one study. That is a fact about the\nworld, entered as provenance by whoever ingests the source, and it is where the crowdsourcing he\nnames does the work the model cannot: *everyone using Deliberus can help speed it up* is right for\norigin, since a reader who knows the two reports cite one trial can say so in one link.\n\n## 6. What the corpus and his record already had\n\nListed because he asked that it be used, and because the number note of yesterday failed to name\ntwo of these:\n\n- **January 2013, Proposal FB-01**: *per-statement probability estimates*; *concepts are defined by\n  all statements using the concept*; argument bundles (the lines of § 1).\n- **January 2012, Curtis**: probability estimation across all claims *can get hairy*, which is the\n  coherence requirement of R1 named fourteen years early.\n- **2011, Ben Hyink**: the *radically different priors* challenge, which is the prior-dependent state\n  of `starting-values-and-the-problem-of-the-prior.md` § 6.6.\n- **The keeper** (2026-09-22): the sense clause and the overlap question (§ 3, § 4).\n- **Q4 and Q22** (2026-09-26): origin as structure; typed combinations.\n- **The starting-values note** (2026-09-30 to 10-03): twelve recommendations, the rival partition, the\n  prior as a claim; its § 9 question 1 is the root of the whole number question, and it is unruled.\n\n## 7. Open after this (proposed, unruled; rows Q38 to Q40)\n\n- Lines as an object on the reasons channel; min within, max across; the three tag words re-defined\n  on support links (Q38).\n- Sense per use stored, the keeper recording the term a pair turns on, sense alignment as a\n  code-computed relevance input and equivocation reading, shown and argued (Q39).\n- The keeper run over a claim's sibling reasons, its verdict read as the lift's first independence\n  answer; origin through Q4 (Q40).\n- The refile of the 24, after Q37 or with its effect measured.\n\n## Cross-references\n\n`what-the-number-on-a-claim-means.md` (the words; corrected today to name the instruments) ·\n`pollock-against-accrual-and-the-overlap-check.md` · `claim-parts-and-inferential-reasons-2026-10-14.md`\n· `a-part-and-a-reason.md` · `checks-on-machines-and-people.md` § 11 · `world-completeness-and-its-\nrelatives.md` · `derivability-of-missing-considerations.md` § 4 · `core-periphery-and-the-concept-\nlayer.md` § 5b · `store-the-uses-read-the-meaning.md` · `what-counts-as-evidence.md` § 8c ·\n`auto-connect-upgrade.md` § 2c · `how-a-map-is-cut-and-what-it-shows.md` (bundles) ·\n`the-number-in-the-world-and-the-repertoire.md` (the same day's companion).\n"}