{"path":"research/corpus-graph/readers/2026-10-27-batch-1-reasons-channel-and-relevance-as-a-reader-returned-it.md","content":"# What a reader found to bear on six notes on the reasons channel, relevance, a part and a reason, one challenge one home, and the two Codex readings (2026-10-27)\n\nA read-only reader (Sonnet), one of those sent through the notes no reader had yet read, after the founder\nasked for the graph's coverage to be widened (README § 12). It was given the brief in\n[../reader_brief.md](../reader_brief.md) and these 6 notes:\n\n- `docs/research/retrofitting-the-reasons-channel.md`\n- `docs/research/what-relevance-is.md`\n- `docs/research/codex-review-19-october-read-back.md`\n- `docs/research/a-part-and-a-reason.md`\n- `docs/research/one-challenge-one-home.md`\n- `docs/research/codex-review-19-october-second-reading-2026-10-20.md`\n\nIt made 86 tool calls. Its report is below, unchanged. What was kept from it, after the quoted words were\nchecked by code and both passages of each finding were opened, is in `../found.jsonl` (lines whose `where`\nnames this file).\n\n---\n\nI read all 6 of the 6 notes in full.\n\n### F1\ntarget: docs/research/codex-review-19-october-second-reading-2026-10-20.md:66\ntarget_words: \"Splitting a conjunction exposes its components; it need not prohibit retaining their conjunction where an inference needs it.\"\nolder: docs/research/session26-open-rulings.md:1254\nolder_words: \"Every claim is stored at the separability floor, a list item by item; what a reader sees may join parts back into one sentence\"\nkind: contradicts\nwhy: AX8 (ruled 23 Oct, after this reading) says every claim is stored cut to the floor. A joined sentence is only a display reading, \"never stored as a claim\". Codex's favoured drawing keeps a stored claim \"Contact A is closed and contact B is closed\". Together with AL15 this removes that drawing as a stored node. Someone ruling AL3/AL4 should treat the composite claim as a display join, not a candidate.\nstrength: strong\n\n### F2\ntarget: docs/research/one-challenge-one-home.md:513\ntarget_words: \"AL14 stays open until he chooses; a reconciliation of the two is also on the table\"\nolder: docs/research/session26-open-rulings.md:1047\nolder_words: \"nothing beyond the concept layer is needed for theories\"\nkind: already-answers\nwhy: AL14 offers a Theory node or a reconciliation (the theory's name as a concept-layer term, tenets as attributed claims). AM10 was ruled 2026-10-25 (\"Bless the theory concept use scrutiny ideas\"). It says a theory is defined on its page by the claims that use its name, as a concept is, with nothing beyond the concept layer. That is the reconciliation, so AL14 is largely decided.\nstrength: strong\n\n### F3\ntarget: docs/research/retrofitting-the-reasons-channel.md:791\ntarget_words: \"only when relevance is the conferred degree, so the proposal above, which concerns co-premises only, still\"\nolder: docs/research/session26-open-rulings.md:465\nolder_words: \"A link's relevance is the difference the reason makes, P(C | A) − P(C | not A), everywhere\"\nkind: already-answers\nwhy: The passage says the fact×relevance product is the chain rule only if relevance is the conferred degree, and that the payoff waits until that point is settled. Q50, ruled 2026-10-26 and built, rules relevance is the difference a reason makes everywhere, measured from the reason's own prior (Jeffrey's rule). The open point (2) of this paragraph is therefore settled, and the 22 October reading of \"conferred degree\" is out.\nstrength: strong\n\n### F4\ntarget: docs/research/what-relevance-is.md:446\ntarget_words: \"prior + relevance × (supporter − prior)`, the claim's strength with that one reason alone.\"\nolder: docs/research/session26-open-rulings.md:465\nolder_words: \"A link's relevance is the difference the reason makes, P(C | A) − P(C | not A), everywhere\"\nkind: later-correction\nwhy: § 3g describes the shipped arithmetic as a Bayesian mixture measured from the claim's prior. Q50 (ruled and built 26 Oct, after the 25 Oct section) changes `alone_value` to measure from the supporter's own prior. It also bounds a reading by both priors, so a difference of 0.8 cannot exist in some cases. The § 3g formula is now outdated, and the text calling it \"shipped\" should say which version.\nstrength: strong\n\n### F5\ntarget: docs/research/retrofitting-the-reasons-channel.md:940\ntarget_words: \"**Proposals** (spoken names first; the first is ruled and recorded in Q40, the rest are ledger rows Q42 to\"\nolder: docs/research/session26-open-rulings.md:457\nolder_words: \"Rule Q42 with the clauses, except that a machine-written middle carries its cases' strength from the start\"\nkind: later-correction\nwhy: The note calls Q42 to Q45 unruled. The ledger shows Q42 ruled 2026-10-23, with the founder's change that a machine-written middle carries its cases' strength from the start. Q44 was ruled 2026-10-22 (\"Writing out a premise moves nothing\"). Only Q43 and Q45 are open. A reader of § 9k point 3 would not know the middle claim's strength rule has been decided.\nstrength: strong\n\n### F6\ntarget: docs/research/a-part-and-a-reason.md:177\ntarget_words: \"The cap a required part puts on its whole is\"\nolder: docs/research/session26-open-rulings.md:1388\nolder_words: \"A part link records a cut and carries no strength: nothing a part says reaches its whole, neither the weakest-part\"\nkind: later-correction\nwhy: § 4 equates the shipped cap with the exact probability ceiling, and § 0 and § 2b say \"a part caps\" and \"ruled and built\". BD16 (ruled 25 Oct, built as `PARTS_CARRY_STRENGTH = False`) removed the cap. Part links now pass no strength. The part-versus-reason \"presumption of need\" is therefore no longer carried by the arithmetic. O13 is also marked narrowed by BD16.\nstrength: strong\n\n### F7\ntarget: docs/research/a-part-and-a-reason.md:229\ntarget_words: \"links add, and nothing discounts arguments that share an origin (open)\"\nolder: docs/research/session26-open-rulings.md:461\nolder_words: \"Across arguments read as independent, a claim's strength is logit(prior) plus the sum of each argument's shift in log-odds\"\nkind: later-correction\nwhy: The table cell for \"the case for the claim\" says links add and the origin question is open. Q37, Q38, Q40 and Q46 later ruled this. The best argument counts until independence is read, then independent arguments add in log-odds, and a group the independence reading finds to be one story counts as its best argument. The cell is out of date.\nstrength: strong\n\n### F8\ntarget: docs/research/a-part-and-a-reason.md:181\ntarget_words: \"standing above neutral, where relevance, on the reading the September review recommends, is\"\nolder: docs/research/session26-open-rulings.md:465\nolder_words: \"A link's relevance is the difference the reason makes, P(C | A) − P(C | not A), everywhere\"\nkind: later-correction\nwhy: § 4 describes the link rule as relevance times the supporter's standing above neutral (0.5). Q50 replaces \"above neutral\" with \"above the reason's own prior\" and rules relevance as the difference the reason makes. The § 4 algebra (cap = ceiling, term for term) was derived with both priors at 0.5. The two agree only there.\nstrength: medium\n\n### F9\ntarget: docs/research/codex-review-19-october-read-back.md:257\ntarget_words: \"until then two tagged links still add (0.569, and 0.519 with one refuted)\"\nolder: docs/research/session26-open-rulings.md:461\nolder_words: \"One fully relevant reason at 0.9 then gives 0.9, two independent ones 0.988, where the shipped code gives 0.57 and 0.70.\"\nkind: later-correction\nwhy: The 0.569 and 0.519 figures come from the quadratic QEM squash. Q46 (ruled 22 Oct, live 23 Oct) retires that squash on the reasons channel. Two independent reasons would now read 0.988 and a single one 0.9. The cost of counting one joint argument as two is larger than the table says. The structural point (do not count it twice) stands; the numbers do not.\nstrength: medium\n\n### F10\ntarget: docs/research/retrofitting-the-reasons-channel.md:909\ntarget_words: \"The chance that the next independent source brings a reason nobody has found is about *f₁* divided\"\nolder: docs/research/saturation-and-the-long-tail.md:192\nolder_words: \"the probability that the next observation is a *new* type ≈ the fraction of current types seen\"\nkind: same-idea-earlier\nwhy: Q45 proposes Good–Turing and Chao counts as a new instrument for how many reasons the search has missed. The saturation note (§ 6, 21 Sept) already stated the same estimator for unseen distinct claims per topic, with a Chao lower bound. It adds a caveat Q45 lacks: apparent saturation may be frame ossification, so read it together with the `does_not_fit` rate. Q45 should reuse that dial and its caveat.\nstrength: strong\n\n### F11\ntarget: docs/research/one-challenge-one-home.md:198\ntarget_words: \"an ordinary composite claim holds the joint premises, and the existing\"\nolder: docs/research/folding-joining-and-zoom.md:71\nolder_words: \"A joined sentence is a reading, like a headline (AL10, a claim's text is written\"\nkind: contradicts\nwhy: The composite-claim candidate stores a claim holding the joint premises. The later note applies AL10 and AX8: a conjunction of claims is computed on read, may be cached, and is never stored, not even as the record's text. The ruled floor therefore leaves no place for a stored composite claim. The \"no new bundle node type\" hope is not available in that form.\nstrength: medium\n\n### F12\ntarget: docs/research/retrofitting-the-reasons-channel.md:874\ntarget_words: \"4. How much correlated reasons can add.\"\nolder: docs/research/reasons-and-cases.md:247\nolder_words: \"multiplies the variance by 19.5, so the data hold about a twentieth of the information the independent\"\nkind: same-mechanism-elsewhere\nwhy: § 9k point 4 gives Kish's effective number for correlated reasons as a display, untested on data. The cases note applies the same design effect to 357,347 BCG-trial people counted as independent cases. A shared trial setting multiplies the variance by 19.5, leaving about a twentieth of the claimed information. It is the same formula measured on real numbers, which supports Q43's \"these four count as about two\" display. It also shows the dependence is not a small correction.\nstrength: medium\n\n### F13\ntarget: docs/research/codex-review-19-october-read-back.md:32\ntarget_words: \"happened, I don't know which*, in which case splitting it into two asserted facts invents information.\"\nolder: docs/research/pass-quality-and-the-choice-of-model.md:879\nolder_words: \"As splitter, five of the six runs kept *positive or necessary* in one part**; GLM 5.3 cut them\"\nkind: supplies-a-missing-piece\nwhy: The \"A or B\" worry is measured on a real claim. In the first probe three of five models (including Claude Sonnet 5) cut \"positive or necessary\" into two parts each stated as fact. A later six-run probe kept it whole in five of six. The same note's correction adds that an alternative can be a part if tagged as an alternative, not required. AL3's \"keep as one claim\" is one of two repairs, not the only one.\nstrength: medium\n\n### F14\ntarget: docs/research/retrofitting-the-reasons-channel.md:154\ntarget_words: \"unless a stated bridge crosses the two senses (an ordinary, challengeable claim, AM7)\"\nolder: docs/research/the-chain-inside-a-link/ten-links-written-to-the-floor.md:124\nolder_words: \"equivocates between two unlike senses of \\\"fast\\\" (competitive catch-up eroding a market edge\"\nkind: narrows\nwhy: The sense-bridge allowance assumes a bridge can be written as a claim. In ten links written to the floor, link 5 could not be: the only generalization crossing the two senses of \"fast\" misdescribes one of the claims, so no single believable rule exists. A crossing of senses sometimes has no honest bridge. The detector then gives a sound \"no bridge\", not just a bridge to be challenged. The bridge is also not always needed because it is not always writable.\nstrength: medium\n\n### F15\ntarget: docs/research/what-relevance-is.md:142\ntarget_words: \"The judge's reading of a bare link is a starting value for the\"\nolder: docs/research/priors-as-standing-work.md:361\nolder_words: \"claim or the link, and the relevance judge must read a link with its written defeaters in view.\"\nkind: narrows\nwhy: BC8 (ruled 24 Oct) says a consideration counts once. While unwritten it lives in a coarse estimate, including a link's relevance reading. Once written as a claim it leaves the estimate and counts through its links. The judge must be handed the link's written defeaters, and its reading is stored with the list it was given. This fixes the double-counting risk in § 3c's \"judge then exceptions\" scheme and says what the judge's input must be.\nstrength: medium\n\n### F16\ntarget: docs/research/a-part-and-a-reason.md:293\ntarget_words: \"**Why the ruling does not do it automatically.** Anyone can attach a poor reason to a claim they\"\nolder: docs/research/burden-of-proof-and-priors.md:157\nolder_words: \"the straw-man exploit: attach two weak reasons to a true claim, refute them, and it reads 0.10.\"\nkind: supplies-a-missing-piece\nwhy: § 5a argues that a failed reason must not pull a claim down, because a straw man would pay. The burden-of-proof note computes it with his own 2013 case. Two reasons at relevance 0.5, both refuted, leave the claim at \"nothing stands (2 checked)\" under Q34. The rejected subtract-from-0.5 rule would drop it by 0.40. It states that to make the claim probably false, the \"no\" side must argue. This answers his challenge numerically.\nstrength: medium\n\n### F17\ntarget: docs/research/what-relevance-is.md:62\ntarget_words: \"strength of one of those links, not on the conclusion, is *ancillary* evidence; in his large case\"\nolder: docs/research/what-counts-as-evidence/law-intelligence-history-journalism-forensics.md:386\nolder_words: \"reasoning set up by directly relevant evidence. In short, ancillary evidence is evidence about\"\nkind: same-idea-earlier\nwhy: The evidence survey already holds Schum's directly relevant versus ancillary distinction with verified quotations (1999 and the 2001 Cardozo paper). It says ancillary evidence is \"metaevidence\" and that generalizations plus ancillary evidence are the \"glue\" holding an argument together. The note (§ 1) cites Schum from secondary sources only and says the book was not read. This verified text can replace that read-depth caveat, and it names the same link-level slot the note builds on.\nstrength: medium\n\n### F18\ntarget: docs/research/retrofitting-the-reasons-channel.md:693\ntarget_words: \"**The one place it is wrong in probability's terms.** Min is the ceiling for a conjunction, never its\"\nolder: docs/research/what-counts-as-evidence/law-intelligence-history-journalism-forensics.md:255\nolder_words: \"elements are correlated through a single story) and the attack (it overstates the whole).\"\nkind: same-idea-earlier\nwhy: The evidence survey (§ 1.7) already framed min as the court's element-wise acceptance. It gave both a defence (elements are correlated through a single story) and the attack (min overstates the whole). That is the Q47 shared-root answer and the overstatement point of § 9j, stated earlier and with legal sources (Cohen's conjunction paradox, Allen and Pardo). Someone ruling on min versus product can start from its list of defences.\nstrength: medium\n\nNoticed: Several notes still describe as open what the ledger now rules. These include the Q42/Q44 status in retrofitting § 9k, which F5 covers, and AL14 as open in one-challenge, which F2 covers. what-relevance-is.md § 3g and a-part-and-a-reason.md § 4 still use the pre-Q50 formula. The hook messages from the corpus graph are all consistent with these findings and were treated as data. I found no existing found.jsonl entry for any pair above. The ledger rows are not stored under their file name there, so the file-name check is clean but row-level pairs were checked by grep for the row id.\n"}