{"path":"listening/2026-10-24-checking-every-inference.md","content":"# Checking every inference, and where tonight's two sessions meet\n\n- **What it is**: a listening walkthrough of about twenty minutes, made late on 2026-10-24. It merges the\n  afternoon's walk (`2026-10-24-one-inference-per-link.md`, rendered at 18:13) with the evening's work, and with\n  the other session working the same night, on priors. His requests: first *\"Don't want another walkthrough until I\n  trust what you're telling me\"*, then, after the research on software: *\"Now this it seems like I'll have some\n  value from listening to on my walk. So now you can make the walkthrough. Merge it with the last one. Weave in\n  what is still of any value from it. And include a deep understanding of what the other session active tonight is\n  thinking about also. Think deeply about all this and how it is interconnected. 20 minutes in total.\"*\n- **⚠ Overtaken the same night (23:40), by the other session's rulings after this was written**: (1) part five's question to carry, whether *a consideration counts once* goes into the always-loaded file, was already settled: he blessed that line at 22:34 (BC8) and it is resident, so the question and its prediction were stale; (2) part five's advice to say *reference class* for a claim's kind goes against the other session's own word, which is *kind* (BC1, *a prior from the claim's kind*; Arc BD, *saying what kind a claim is*), so *kind* now does two jobs in two resident lines (a kind of inference, a kind of claim), a collision for him to name; (3) *a seeding pass may write it, as a claim with a source* follows BC1's ruled text, and BD1 (open) narrows it: the pass names the kind, and code reads a counted rate, never a rate a model states. **Part five rewritten and re-rendered the same night and spliced into the file at 14:56, now 21.1 minutes** (his word: *\"If it's not too expensive, please rerender what needs revision\"*): the question to carry is now BD1, a kind of claim as an attackable claim; *kind* kept for both senses, said which each time.\n- **Kept from the afternoon walk**: the recap, shortened; one inference per link (slots, the floor trial's leap,\n  Carroll, his *method, not repetition*), with *sound* no longer used (it meant only *without a gap by the\n  trial's reader*); what a pass is taught (AM12). Dropped for length: the part on finding rival answers through their question (still open for his word),\n  Kelvin's cooling Earth, the stranger's question, and most of the recap.\n- **Covers**: [../research/one-inference-per-link.md](../research/one-inference-per-link.md) (§ 8, every kind and\n  what it owes); [../research/tools-for-checking-inferences.md](../research/tools-for-checking-inferences.md) (the\n  software, the trials); [../research/priors-as-standing-work.md](../research/priors-as-standing-work.md) (the other\n  session, §§ 8–10); ledger AM12 to AM17 and BC1, BC6 to BC8\n  ([../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. Every kind, and what each owes · 4. The tools that can check it · 5. Where the two sessions meet · Close.\n- **Registers used**: Lucy; his clone; Elliot (Carroll's regress, `philosophy-of-the-unwritten-step.md` § 3; Hume's\n  greater wonder and Quine's web, `priors-as-standing-work.md` § 8); Scotty (the critic of the compile design,\n  `tools-for-checking-inferences.md` § 0); Anna (the studies, `tools-for-checking-inferences/report-harness-design.md`;\n  Kirkpatrick and Sterken, abstract checked); Myrddin (sayings, each with its fit note).\n- **Predictions**: from his typed words, found by the prepare step's `--said` queries: 2026-10-24 *\"We need method,\n  not repetition\"*; 2026-10-22 *\"I always wanna settle the theory … before we do any empirical testing\"* and\n  2026-10-24 *\"So maybe we could/should teach method about this to the model as well\"*; 2026-10-23 *\"If we can apply\n  deterministic code to that, then I'm all for it … want what can be settled as logical in the graph to be settled\n  automagically\"*; 2026-10-24 *\"a single source of truth/disagreement on each claim/question/rebuttal/case/exception/consideration\"*.\n  Historical: 2011-12-06, the group (Prover9 on Anselm, *\"This is for the far future.\"*); 2011-07-29, the group\n  (relevance argued on a meta level).\n- **The correction spoken** (rule 37): that deterministic code checks every kind; it was two trial scripts on cases\n  Claude wrote (his question, *\"Are you dead sure that code exists in reality?\"*). Not spoken: the muddle over\n  *naming the definition* (AT12's test is the two questions; definitions were its examples), corrected in\n  `one-inference-per-link.md` § 8.\n- **What was run, and what was read**: the engines in part four ran on this machine on 24 October\n  (`tools-for-checking-inferences/trials.py`); the study figures are from the helpers' reports, the clingo\n  self-correction table opened by Claude.\n- **Status**: rendered 2026-10-24 at 22:43, 20.7 minutes (14.7 of speech, 6.1 of added silence), 2,236 words, seven chapter marks; copied to `~/Documents/Deliberus listening/`. Passed the dry run's gates with no override.\n\n---\n\n# Checking every inference, and where tonight's two sessions meet\n\n[[scene Lucy opens a late walk for Fredrik on 24 October. Earlier tonight he said he wanted no walk until he could trust what he is told. Now he has asked for one. She is calm and plain, and careful with every claim.]]\n\nHow would you know a machine took one small step, and not a leap? And how would you know I did?\n\n[[pause 5]]\n\nTonight you said you would take no walk until you trust what you are told. So this one says what was run, and what was only read.\n\n[[pause 4]]\n\nThe map has three groups. First, the day in brief. Second, one inference per link: its kinds, and the tools that can check it. Third, where tonight's two sessions meet.\n\n---\n\n# Part one. Since the last walk\n\n[[scene Lucy gives a short recap of a long day. Steady, a little proud, never boastful.]]\n\nWhich changed more today: what the site does, or what it refuses to do?\n\n[[pause 5]]\n\nMostly what it refuses. Every machine pass is off on the public site, until the passes are revised.\n\n[[pause 4]]\n\nOn the maps, two claims that say one thing of two subjects now share a card. And the yellow link, the qualifying one, is off every menu that writes links.\n\n[[pause 4]]\n\nAnswers changed too. An answer to a critical question now weighs by its own support, times its relevance. An answer nobody has backed starts at the lowest tier, never at nothing. You approved that line for the always-loaded file.\n\n[[pause 4]]\n\nToday, the site reads and does not think. If the revision lands, the passes come back one at a time, each with its method.\n\n---\n\n# Part two. One inference per link\n\n[[scene Lucy takes the topic that decides the others: how to know that each link of a machine's chain is one inference. Elliot brings the logic, 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 is one inference?\n\n[[pause 5]]\n\nToday, nobody does. The keeper is only told to take small steps.\n\n[[pause 4]]\n\nHere is a way to check. Write each link in slots. A rule, with a gap for the thing it is about. The thing. The case. The conclusion.\n\nThen code compares words. 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 4]]\n\nTake a link from the floor trial. The rule spoke of a tendency in almost all AI 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 the comparison sees it. The link is refused.\n\n[[pause 4]]\n\nWhy does that work? Say the case fits the condition, and the conclusion fits the outcome. Then a rule that has no exceptions cannot hold while the conclusion fails. No gap is left in the arrow. So a leap can only hide in the rule, written out, where anyone can attack it.\n\n[[pause 4]]\n\n[[voice elliot]]\nLewis Carroll's tortoise would not accept a step until its rule became one more premise. Then it wanted a rule for that one, forever. The slots end the regress by giving two jobs two homes. The rule is a premise you can doubt. Applying it is what the code checks.\n[[voice lucy]]\n\n[[pause 3]]\n\nIn the floor trial, a reader marked thirteen gaps. When the links are read for form, the check refuses eight. They are all the gaps where the conclusion says more than the rule. The rest doubt whether a rule is true, and the rule's own claim is their home.\n\n[[pause 4]]\n\nThen the idea of a second model reading, which you rejected 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 3]]\n\nThe record agrees. Asked the same wording twice, the trial's reader changed its mind about once in four. So a refused link goes back to its writer, with the fault in words. Repetition, with clarification.\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 corrected only 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 3]]\n\nThe question to carry. Should this form check replace a second model reading, as the test of a machine's chain?\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, and people and the graph argue whether the rule is true. Measure first how often the writer fills the slots right.\n[[voice lucy]]\n\n[[pause 3]]\n\nToday, the chain the keeper writes is free text that nobody checks. If the slots are built, every leap in it becomes a rule someone can attack.\n\n---\n\n# Part three. Every kind, and what each owes\n\n[[scene Lucy turns to the kinds of inference, and to a correction she owes him. Anna has the studies, then the elder. Mood: honest, steady.]]\n\nIs *usually* a weaker *always*, or a different kind of rule?\n\n[[pause 5]]\n\nTonight you ruled that every inference a machine writes names its kind, and meets that kind's form.\n\n[[pause 4]]\n\nA rule's kind comes from two questions about a case where it fails. Does the failure call for an explanation? If not, the rule only states a share of cases. That is a statistical rule.\n\n[[pause 3]]\n\nIf it does, can an explanation leave the rule standing? If yes, the rule holds unless an exception applies. That is a normic rule. If no, it is strict.\n\n[[pause 4]]\n\nTake the Griess test, from the Birmingham Six case. The rule said a positive swab means the hands had touched nitroglycerine. Written as a rule with no exceptions, it passed every check of form.\n\n[[pause 4]]\n\nNow ask the questions. A false positive calls for an explanation, and the explanation leaves the rule standing. So the rule is normic. Its known exception becomes a claim of its own. A swab that touched nitrocellulose also turns positive, as on coated playing cards.\n\n[[pause 4]]\n\nEach kind owes something like that. A statistical rule states its share, and its case must belong to its class. A rate across states says nothing about one immigrant. An induction files its cases under its rule. An explanation names its rivals. An analogy writes, as a rule, the respect that makes two cases alike.\n\n[[pause 4]]\n\n[[voice anna]]\nThe studies say models are weakest exactly here. Across twenty-eight models, most read a generic statement as a universal one.\n[[voice lucy]]\n\n[[pause 3]]\n\nNow the correction I owe you. Earlier tonight I said deterministic code checks the form of every kind. It was two trial scripts, run on cases I wrote myself. That was my own leap, from *I wrote a script* to *code checks*.\n\n[[pause 4]]\n\nYour answer, when the line went into the always-loaded file:\n\n[[voice fredrik]]\n[gently insistent] And it also must be remembered that it applies to you, until we improve the canon and what method we have in Claude.md such that you can be trusted even more, along these lines.\n[[voice lucy]]\n\n[[pause 3]]\n\nSo the line now binds this session too. An inference I write in a note or a reply names its kind, and what that kind owes.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[fit says: naming an exception confirms the rule where no exception is named · applies: a normic rule's exceptions, written as claims, leave it standing for every other case · breaks: in everyday speech the phrase excuses any counterexample · source: legal maxim, exceptio probat regulam, established]]\nThe exception proves the rule, the old lawyers said. They meant: name the exception, and the rule holds everywhere else.\n[[voice lucy]]\n\n[[pause 3]]\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. Settle the theory first, then teach it. A pass should know how a rule fails before it writes one.\n[[voice lucy]]\n\n[[pause 3]]\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 a false one.\n\n---\n\n# Part four. The tools that can check it\n\n[[scene Lucy reports tonight's search for software, and what ran on the laptop. Anna has the studies, Scotty argues against the design. Mood: curious, a little playful.]]\n\nIs there a spaCy for logic: a library that checks an inference the way spaCy tags a sentence?\n\n[[pause 5]]\n\nNo. Nothing does it end to end.\n\n[[pause 4]]\n\nYou saw the parts coming long ago. In 2011 you shared a story in the old group. A theorem prover had checked Anselm's proof of God, and found a simpler one. You added:\n\n[[voice fredrik]]\n[wryly] This is for the far future.\n[[voice lucy]]\n\n[[pause 3]]\n\nTonight, engines ran on our own cases, on this laptop.\n\nFirst, a solver for rules with exceptions, called clingo. Given the Griess rule, it concluded that Hill had handled nitroglycerine. Told that the swab had touched nitrocellulose, it withdrew the conclusion.\n\n[[pause 4]]\n\nThen Farr's cholera. Farr saw deaths fall as London's districts rose above the Thames, and blamed the air. Given air and water as rivals, the solver listed which explain the pattern. Both did. Add the deaths gathered round one pump, and only answers with water survive.\n\n[[pause 4]]\n\nBut it also offered both together as an answer. So the simplest explanation must be asked for, in words.\n\n[[pause 4]]\n\nThen a probability engine, called ProbLog. One rule, that red skies bring rain a third of the time, gave its own share. Two rules from two classes, for one morning, gave a number higher than either. It quietly assumed the two were independent.\n\n[[pause 4]]\n\nSo an engine decides some things silently. Which class a case belongs to, and how reasons combine, must stay our rulings, written as claims.\n\n[[pause 4]]\n\nTonight you ruled that it never gets the chance, and the guard is built. Two rules for one morning never reach the engine together.\n\n[[pause 3]]\n\nIf someone has counted red-sky mornings with falling pressure, that narrower count answers alone. Otherwise our own rule combines them, and they add only when read independent.\n\n[[pause 4]]\n\nAnd you asked whether the front is really a shared reason. It is. Both signs point to one cause, a weather front approaching.\n\n[[pause 3]]\n\nSo the front becomes a claim of its own. The red sky and the barometer support it, and it supports the rain once. That shape is proposed, not ruled.\n\n[[pause 4]]\n\n[[voice anna]]\nOne study stands out. A checker in code sent each model its error and let it revise. Accuracy rose by more than twenty points.\n[[voice lucy]]\n\n[[pause 3]]\n\nPublished systems that check reasoning with an engine have the model write the engine's language. That translation is where they fail. So the proposal turns it round. The model fills slots in its own words. Plain code turns the filled slots into the engine's input.\n\n[[pause 4]]\n\nHere is the critic, in a borrowed voice.\n\n[[voice scotty]]\nClever. The machine never writes logic. It just writes the sentences the logic is made of. Same risk, new coat.\n[[voice lucy]]\n\nPartly true. But a wrong sentence sits on the page, where a person can attack it. A wrong translation hides inside the engine.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[fit says: a power that serves well when held is dangerous when it rules · applies: an engine computes well what it is given, but its defaults decide what nobody chose · breaks: fire has no defaults, only fuel · source: English proverb, established]]\nFire is a good servant and a bad master. Feed the engine, but never let its defaults choose.\n[[voice lucy]]\n\n[[pause 3]]\n\nThe question to carry. Should the slots compile to the engines, so the model never writes logic itself?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[thoughtful] Yes. If we can apply deterministic code, I'm all for it. What can be settled as logical in the graph should be settled automagically. Beyond that, the model, with its leaps kept small.\n[[voice lucy]]\n\n[[pause 3]]\n\nToday, the engines have run on a few of our cases, on this laptop only. If the slots compile to them, an inference is checked by what it implies.\n\n---\n\n# Part five. Where the two sessions meet\n\n[[scene Lucy turns to the other session working tonight, on priors, and shows how its rulings and this session's fit together. Elliot brings Hume. Mood: warm, unhurried, a little in wonder.]]\n\nBefore you hear a single reason, how likely is a claim to be true?\n\n[[pause 5]]\n\nAnother session spent tonight on that question. You asked whether a standing pass could give new claims a sensible prior, and seed their first answers.\n\n[[pause 4]]\n\nIt found that *a sensible prior* named three numbers. How likely a claim like this is. Who said it. How far a link bears. Only the first is a prior.\n\n[[pause 3]]\n\nAnd no model states it. The pass names the claim's reference class: the group of claims like it. Code reads how often claims in that group turned out true, from a counted source.\n\n[[pause 4]]\n\nYou set what the prior is read from.\n\n[[voice fredrik]]\n[calmly firm] The claim says what it says and before ANYTHING else it has a prior, for anyone reading it wherever it appears.\n[[voice lucy]]\n\n[[pause 3]]\n\nSo the prior comes from claims like it, whose truth rate somebody has counted. That group is its reference class. Hypotheses in top psychology journals are one. A narrow reference class leans toward a broad one, by how little it has counted.\n\n[[pause 3]]\n\nYou ruled one more thing there. The pass names a narrower reference class only when it is sure enough. A wrong narrow one costs more than staying broad.\n\n[[pause 4]]\n\nThen you named your own way of judging.\n\n[[voice fredrik]]\n[in wonder] How strange would it be that it were true?\n[[voice lucy]]\n\n[[voice elliot]]\nThat is Hume's greater wonder. Believe a report of a miracle only if its falsehood would be the greater miracle.\n[[voice lucy]]\n\n[[pause 3]]\n\nIn Deliberus, those implications are not folded into the prior. They are written as claims and links. Take *H. pylori causes most stomach ulcers*. It clashed with *stress causes ulcers*.\n\n[[pause 3]]\n\nWritten as an attack, that clash weighs by the old claim's own strength. And the old claim was widely held, but weakly supported. So the clash weighed less than it felt.\n\n[[pause 4]]\n\nYou approved both: a prior from the claim's reference class, and a seeding pass that writes the implications. And you said why.\n\n[[voice fredrik]]\n[quietly delighted] giving the microtasks a chance to indulge in its curiosity about any strange new claims and questions.\n[[voice lucy]]\n\n[[pause 3]]\n\nNow where the two sessions meet. Three threads.\n\nFirst, the word *kind* did two jobs tonight. Here it meant a kind of inference: strict, normic, statistical. There it meant a kind of claim, for a prior.\n\n[[pause 3]]\n\nSo the claim sense gets its own name. A kind of inference stays a kind. The group a claim belongs to, for its prior, is its reference class.\n\n[[pause 4]]\n\nSecond, they are one structure. A prior from a reference class is a statistical inference about the claim itself. The rule: claims in this class are true so often. The case: this claim is in that class.\n\n[[pause 4]]\n\nSo the membership claim the other session asked for is the case slot. And the engine's silent combination is the crossing of reference classes it warned about.\n\n[[pause 4]]\n\nThird, and deepest. The slots drag unwritten rules onto the page. A line you put in the always-loaded file tonight says what happens then. A consideration counts once. While unwritten, it lives in a coarse estimate, a prior or a relevance reading. Once written, it counts only through its own link.\n\n[[pause 4]]\n\nTake the gate on a Tuesday. A reading of how far Tuesday bears on the gate being open prices the holidays nobody named. Then someone writes that today is a public holiday. If the reading stays, holidays count twice. So the reading is made again, setting holidays aside.\n\n[[pause 4]]\n\nWithout counting once, the slots would make every written rule count twice.\n\n[[pause 4]]\n\n[[voice myrddin]]\n[[fit says: counting one thing twice inflates the total · applies: a consideration priced in a prior and also written as a link · breaks: coins do not change when written down · source: ours]]\nA coin counted in two purses makes neither richer. Count it where it lies.\n[[voice lucy]]\n\n[[pause 3]]\n\nThe question to carry. Should a reference class be a claim itself, in the graph, whose count anyone can attack?\n\n[[pause 5]]\n\nMy prediction, in your voice.\n\n[[voice fredrik]]\n[warmly] Yes. A reference class is a claim like any other. Its count has a source, and the source can be wrong. One home for its truth, and for disagreement about it.\n[[voice lucy]]\n\n[[pause 3]]\n\nToday, both sessions' rulings are on paper, and none is built. If they are, every premise the slots bring up has one home, and counts once.\n\n---\n\n# Close\n\nThat is the walk. One word per chapter, if you like: muddled, agreed, or disagreed and why.\n"}