{"path":"listening/2026-10-24-one-inference-per-link.md","content":"# One inference per link, and what a pass is taught\n\n- **⚠ Part two says the small model rated the gaps above *the sound links* (noted 2026-10-24, on his question)**: *sound* there meant only the inferences in which the trial's reader found no gap, not sound in the logician's sense (the premises true as well). The spoken text is kept as it was rendered.\n- **What it is**: a listening walkthrough of about twenty minutes, made the afternoon of 2026-10-24. His\n  request: *\"Make a 20-min walkthrough on all this when you're done. Ultrathink\"*, and *\"Include in it a recap of\n  everything that happened since the last walkthrough, about 5 mins or so\"*. It opens on his report on the last\n  piece (*\"This is... muddled. Don't agree with the prediction.\"*, on part one of the walk of 23 October), recaps\n  what was committed since that walk, then takes three topics: making each link of the keeper's chain one\n  inference (the topic that decides the others), what a pass is taught, and how rival answers are found.\n- **Covers**: [../research/one-inference-per-link.md](../research/one-inference-per-link.md);\n  [../research/resonance-db-and-negation.md](../research/resonance-db-and-negation.md);\n  [../research/the-chain-inside-a-link.md](../research/the-chain-inside-a-link.md) § 10, § 11;\n  [../research/three-minds-and-the-leap.md](../research/three-minds-and-the-leap.md);\n  ledger AM11 to AM14, T5, Z6, AM2, AX2, Z3, AL2, AP1 ([../research/session26-open-rulings.md](../research/session26-open-rulings.md)).\n- **Chapter list for noting** (rule 33): 0. Opening · 1. Since the last walk · 2. One inference per link ·\n  3. What a pass is taught · 4. Finding rival answers · Close.\n- **Registers used**: Lucy; his clone; Elliot (Carroll's regress, from `philosophy-of-the-unwritten-step.md` § 3 and\n  `claim-parts-and-inferential-reasons-2026-10-14.md` § 10); Scotty (the critic's case, from\n  `one-inference-per-link.md` § 4); Anna (the studies, from `one-inference-per-link/report-methods-for-one-inference.md`);\n  Hope (the stranger, on the sentence as it stands in § 0 of the note); Myrddin (sayings, each with its fit note).\n- **Predictions**: each from his typed words found by the prepare step's `--said` queries (the queries are in the\n  session; the hits used): 2026-10-24 *\"We need method, not repetition\"*; 2026-09-02 *\"machines check\n  parts-against-sentence, humans remain the only check of sentence-against-world\"*; 2026-10-24 *\"So maybe we\n  could/should teach method about this to the model as well\"*; 2026-10-20 *\"It's own node. We have questions in the\n  data model for good reasons\"*; 2026-09-27 *\"Questions are a first hand entity in the ontology of Deliberus, no?\"*.\n- **Numbers kept**: the trial's eight refused gaps (the result the topic turns on); fully atomic chains under\n  prompting, a few in a hundred (why the slots will be refused often); the two similarity readings under the bar\n  (the recall gap itself).\n- **Status**: rendered 2026-10-24 after his top-up, 18.7 minutes (12.8 of speech, 5.9 of added silence), 5,116 credits, in 78 seconds with four pieces at a time (the first render with the concurrent renderer). An earlier attempt the same afternoon was refused for want of credits.\n\n---\n\n# One inference per link, and what a pass is taught\n\n[[scene Lucy opens an afternoon walk for Fredrik on 24 October. He asked for twenty minutes and a recap. Yesterday he called the last walk's first part muddled. She is calm and direct, and keeps the words plain.]]\n\nHow would you know that a machine took one small step, and not a leap?\n\n[[pause 5]]\n\nThat question runs through this walk.\n\nLast time you called the first part muddled. You disagreed with its prediction. So today the parts are shorter, and each prediction comes from your own typed words.\n\n[[pause 4]]\n\nThe map has three groups. First, what happened since the last walk. Second, making each link of the keeper's chain one inference. Third, what a pass is taught, and how rival answers are found.\n\nThe silences are deliberate. They are room to think.\n\n---\n\n# Part one. Since the last walk\n\n[[scene Lucy gives a brisk recap of one long night and day of work across several sessions. Mood: steady, a little proud, never boastful.]]\n\nWhat would you guess changed most since last night: the map, the passes, or the move?\n\n[[pause 5]]\n\nAll three moved. Start with the passes.\n\nEvery machine pass is now switched off on the public site. A call to a model refuses there. A reason someone adds is stored and marked as waiting for a reading. You asked for that, so that the passes can be revised before they run again.\n\n[[pause 4]]\n\nThen the move. Darwin leaves the farm's fibre at the end of October. So there is now a read-only mode. A copy of the site can serve reading from the relay, and refuse changes politely. The banner wording is the one you blessed.\n\n[[pause 4]]\n\nThen the money. The Middle Powers application went out early on Saturday morning. Its map of the Swedish safety institute is now online, filled by hand to the rulings. Each act on it is kept as a test case for its pass.\n\n[[pause 4]]\n\nThen the map itself. Two claims that say one thing of two subjects now share one card. A claim that only relays one reason folds into its line. Every link is drawn as a band, with faint chevrons for its direction. And the badges that only said nothing was weighed are hidden.\n\n[[pause 4]]\n\nThe map also got faster. A query now starts at the node it names. A claim's strength tags arrive in under a second, where they took nine.\n\n[[pause 4]]\n\nThe yellow link went too. A qualifying link is no longer on any menu that writes links. The old ones stay, for the keeper to read again.\n\n[[pause 4]]\n\nAnd this session went through a paper on wave-shaped retrieval. Its drop-in score turned out to be ordinary similarity under another name. It also tested denials. Every flat denial was found as a candidate. And the keeper on Gemini answered *opposes* to each.\n\n[[pause 4]]\n\n[[voice anna]]\nAnd one model that never left the laptop. A small inference model, trained on millions of sentence pairs, caught almost every opposition and invented none. So it can sort the keeper's queue. It cannot replace the keeper.\n[[voice lucy]]\n\n[[pause 4]]\n\nThen you ruled six proposals on the keeper for now. One definition per answer. Look-alikes in its test cases. A contradiction as its own outcome. And you set a hold. No keeper runs until the theory is settled.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[sound stove]]\n[[fit says: a man sharpens his axe before he fells the tree · applies: the passes are revised before they run again · breaks: an axe is sharpened once, the passes keep changing · source: proverb, often given to Lincoln, established]]\nSharpen the axe before you fell the tree. A sharp axe cuts once.\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the site reads but does not think. If the revision lands, the passes come back one at a time, each with its method written down.\n\n---\n\n# Part two. One inference per link\n\n[[scene Lucy takes the topic that decides the others: how to know each link of a machine's chain is one small inference. Elliot brings the logic, Scotty argues against the drift, Hope asks as a stranger. Mood: precise, curious.]]\n\nThe keeper writes a chain between two claims before it answers. How could anyone check that each link in it is one inference?\n\n[[pause 5]]\n\nToday nobody checks. The keeper is told to take one inference per link. Code only refuses one answer, *overlaps*, when the chain runs long.\n\n[[pause 4]]\n\nYou asked whether it is provable. Here is a way that is.\n\n[[pause 4]]\n\nWrite each link in slots. A rule with a gap for the thing it is about. The thing. The case. The conclusion. Then code checks two matches. The case must be the rule's condition, said of the thing. The conclusion must be the rule's outcome, said of the thing.\n\n[[pause 5]]\n\nTake a link from the floor trial of the thirteenth. The rule said a tendency of almost all AI systems holds for today's systems. The case said almost all AI systems will pursue self-improvement. The conclusion said today's systems are pursuing it.\n\n[[pause 4]]\n\n*Will pursue* became *are pursuing*. That is a leap, and code sees it. The conclusion is not the rule's outcome. So the link is refused.\n\n[[pause 4]]\n\nWhy does this work? Because of what the slots force. If the case matches the condition, and the conclusion matches the outcome, then a strict rule cannot be true with the conclusion false. No gap is left in the arrow.\n\n[[pause 4]]\n\nSo a leap has only one place to go. It goes into the rule, written out, where anyone can attack it.\n\n[[pause 5]]\n\nHere is that in history. In 1863 Kelvin dated the Earth by its cooling. His rule needed the Earth to cool by conduction alone. His case never said so. The check refuses that link until the sentence is written.\n\n[[pause 4]]\n\nAnd that sentence is exactly what John Perry attacked in 1895. A fluid interior carries heat faster, so the Earth is older.\n\n[[pause 4]]\n\nFarr's case went the other way. He saw cholera deaths fall as districts rose above the Thames. His rule said such a disease spreads through the air. That link passes the check. The leap is the rule itself, written in the open. Snow's water-company figures attack that rule.\n\n[[pause 5]]\n\nSo the check does not know which premise is false. It only refuses to let the premise hide.\n\n[[pause 4]]\n\n[[voice elliot]]\nThere is an old trap here. Lewis Carroll's tortoise would not accept a step until its rule was written as one more premise. Then it wanted a rule for that, and so on, forever.\n\n[[pause 4]]\n\nThe slots avoid the regress. The rule is a premise you can doubt. Applying it is what the code checks. Those are different jobs.\n[[voice lucy]]\n\n[[pause 5]]\n\nHow well does it fit what was recorded? In the floor trial, a reader marked thirteen gaps among twenty-five links. When the links are read for form, the check refuses eight of them. Those are all the gaps where the conclusion says more than the rule.\n\n[[pause 4]]\n\nThe other five are doubts about whether a written rule is true. The floor you ruled gives those a home already: the rule's own claim.\n\n[[pause 5]]\n\n[[voice anna]]\nA small inference model was tried on the same links. It rated the reader's gaps as following better than the sound links. The gaps are near-misses in wording, and such models read near-misses as matches.\n[[voice lucy]]\n\n[[pause 4]]\n\nThen your objection to a second reading. You wrote it today.\n\n[[voice fredrik]]\n[with quiet conviction] We need method, not repetition, without clarification, the answer is not liable to change much.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe record agrees. Asked the same wording twice, the trial's reader changed side about once in four. Voting helps a reader accept good steps, and barely helps it reject bad ones.\n\nSo a refused link goes back to its writer once, with the fault in words. That is repetition with clarification.\n\n[[pause 5]]\n\n[[voice scotty]]\nLet me argue against the house. You have not removed the leap. You moved it into the rule and called the problem solved.\n[[voice lucy]]\n\nThat is true, and it is the point. A leap inside an arrow cannot be attacked. A rule on the page can, and its cases count against it.\n\n[[pause 4]]\n\n[[voice anna]]\nOne warning from the studies. Asked to use one rule per step, models wrote fully atomic chains only a few times in a hundred. So the check will refuse a lot at first.\n[[voice lucy]]\n\n[[pause 4]]\n\n[[voice hope]]\nSo the computer checks that the sentences fit together, not that they are true?\n[[voice lucy]]\n\nYes. Exactly that.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[fit says: what is exposed to light can be cleaned · applies: a rule written in the open can be attacked and corrected · breaks: sunlight cleans by itself, a written rule is only corrected when someone attacks it · source: Louis Brandeis, 1913, established]]\nSunlight is said to be the best of disinfectants. The check cleans nothing. It only opens the shutters.\n[[voice lucy]]\n\n[[pause 4]]\n\nHere is the question to carry. Should this form check replace a second model reading, as the test of a machine's chain, with the refutation question kept for people?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[calmly firm] Yes. Method, not repetition. Code checks the shape of each inference, and people and the graph argue about whether the rule is true. But measure first how often the writer fills the slots right.\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the keeper's chain is free text that nobody checks. If the slots are built, every leap in a machine's chain becomes a rule someone can attack.\n\n---\n\n# Part three. What a pass is taught\n\n[[scene Lucy moves to what a pass should be given in its instructions. His own analogy carries the part. Mood: warm, thoughtful.]]\n\nIf the graph held every argument ever made, could a pass learn to reason just by reading it?\n\n[[pause 5]]\n\nYou answered that yourself today, with an analogy.\n\n[[voice fredrik]]\n[amused] It's like the difference between educating a person about spycraft vs showing them a spy thriller. They'll get some of the gist from the movie, but they won't understand everything from inside, from the first principles, formal logic, analytic rigor, all that.\n[[voice lucy]]\n\n[[pause 4]]\n\nYou ruled it, and it now stands in the project's always-loaded file. A pass is taught the method. The graph's contents are never its lesson.\n\n[[pause 4]]\n\nWhy it matters here: the slots move every leap into a rule. So a pass that writes rules must know how rules behave.\n\n[[pause 4]]\n\nThe studies say models are weakest exactly there. On rules with exceptions, the best scored about two in three.\n\n[[pause 4]]\n\nTake the Griess test from the Birmingham Six case. The rule said a positive swab means the hands touched nitroglycerine. Later, other substances were shown to turn the test positive too.\n\n[[pause 4]]\n\nIs that failure one that calls for an explanation? Yes. So by your rulings the rule is a usually-rule. It is written as the rule plus a named exception, and the exception is a claim of its own.\n\n[[pause 5]]\n\nThat is the proposal still open. Put those rulings on rules into the keeper's instructions, as its grounding.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[fit says: copying a master's finished work does not teach why each stroke was made · applies: a pass shown the graph's arguments copies their look without their reasons · breaks: an apprentice learns by copying too, in time; the point is what comes first · source: ours]]\nCopy the master's chairs, and you learn their shape. Learn why he cut each joint, and you can build a chair he never made.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Should the keeper's instructions carry the rulings on rules and their exceptions?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[earnest] Yes. It was my own idea today. Teach the method from our rulings, distilled, so the pass knows how careful to be.\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, the keeper is told what each answer means, and not how a rule fails. If the rulings go in, a pass can tell a rule with exceptions from one that is simply false.\n\n---\n\n# Part four. Finding rival answers\n\n[[scene Lucy takes the last topic: two oppositions the machine never sees, because they do not sound alike. Mood: curious, a little playful.]]\n\nTwo claims answer one question in opposite ways. Why might the machine never put them side by side?\n\n[[pause 5]]\n\nBecause the machine looks for claims that sound alike. It fetches a pair only above a similarity bar. Two real oppositions fell under it.\n\n[[pause 4]]\n\nOne is from the institute map. One answer says Sweden should rely on its existing agencies. The other says Sweden should build a new institute. The other pair is to raise the minimum wage, against freezing it.\n\nNeither pair shares a negation word. So their wording drifts apart.\n\n[[pause 5]]\n\nYour theory was distance. These are different takes, resting on partly different premises. So you suggested a lower bar, and counting the chain between them.\n\n[[pause 4]]\n\nHalf of that holds well. Their grounds do part, somewhere deep. But as answers they are one step apart. They exclude each other.\n\n[[pause 4]]\n\nSo a lower bar is the costly route. It fetches many pairs that share nothing. The cheaper route is the question they both answer.\n\n[[pause 4]]\n\nYou ruled that a question is its own node in the graph. Two answers to one question are rivals by definition. Find them there. Then write down what each answer rests on, and fold what they share.\n\n[[pause 5]]\n\n[[voice myrddin]]\n[[fit says: you recognise rivals by the prize they compete for, not by their looks · applies: two answers are rivals because they answer one question, however unlike they sound · breaks: rivals can also look alike, which similarity already finds · source: ours]]\nYou will not know rivals by their faces. Look at what they are both reaching for.\n[[voice lucy]]\n\n[[pause 4]]\n\nThe question to carry. Should rival answers be found through the question they answer, rather than by a lower similarity bar?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[curious] Through the question, yes. Questions are a first hand entity in the ontology. Then spell out each smallest step between the answers' grounds, and fold what they share.\n[[voice lucy]]\n\n[[pause 4]]\n\nToday, two real oppositions stay unread, because they sound unalike. If questions gather their answers, rivals meet wherever they are worded.\n\n---\n\n# Close\n\nThat is the walk. One word per chapter, if you like: muddled, agreed, or disagreed and why.\n"}