{"path":"research/three-minds-and-the-leap.md","content":"# Three minds, and the leap a microtask makes: humans, models, and the mind of Deliberus\n\n**Date**: 2026-10-18 · **Type**: the founder's reflection, recorded whole, and a session's reading of it ·\n**State**: his reflection is his; the generalisation in § 4 (amended on his push the same night, § 4a) and the\ncast's purposes in § 5 are proposals (ledger Arc AM, rows AM1 to AM7), unruled; the keeper half of AM1 is built.\n\n## 0. Plain language first\n\nA person's mind, a language model's mind, and the mind that Deliberus would be are three different\nthings. Each person reasons from a worldview assembled by a life: what they have rehearsed, what they\nwere rewarded for believing, who they belong to. A model reasons from the average of a written record,\nfast, fluent, and prone to the failures the research names (hallucination, wiggling, sycophancy, the\nbiases it inherited). The mind Deliberus is meant to become is neither: it would hold every claim and\nevery inference written out, and it would be more careful and more honest than either, because that is\nwhat its code does. The founder saw this juxtaposition in one chapter of a walkthrough, where a machine\npass had called two claims overlapping because they shared a topic, though they rested on different\nevidence; the full mind of Deliberus, with each claim split to its floor and every premise written out,\nwould have seen at once that they were not the same. § 4 asks what follows for that pass and for every\nsmall machine task like it, and proposes one rule: a machine's verdict is not accepted without the\nobject the verdict depends on.\n\n## 1. His reflection, whole (2026-10-18)\n\nThe sentences are his, in the order he wrote them, cut only where they addressed the session's\nhousekeeping. He had been reading a summary of the research on model failures (hallucination as fluent\nunsupported text; \"wiggling\" as drift that compounds once a wrong token is in the context; a table\nmapping model flaws to human biases: confabulation, automation bias, social desirability) when he wrote\nit.\n\n> ...for example, the fact that the [semantic-meaning space and worldview and inner world of\n> well-traversed and rehearsed ideas and reasoning and liked/preferred conclusions due to for example\n> varying life experiences/paths and differing/opposing ideologies and religious and spiritual\n> identifications, identity in general psychologically and motivated reasoning and groupthink and so\n> on vs more unfamiliar or out-there ideas or thinking judged as irrational or dangerous or confused\n> or as being due to aforementioned motivated/life-path-dependent reasoning etc] differs quite a lot\n> between humans, more than one might imagine, certainly more than LLMs imagine,\n>\n> and it certainly differs between human and LLM \"minds\" and semantic/dimensional-embedding spaces,\n> which are now arising and suffusing through our society and lives as a new species living right\n> next to us in our daily lives and shaping our thinking and lives, while being susceptible to\n> hallucinations and wiggling and sycophancy and biases of different kinds, similarly to human minds\n> but also having distinct quirks...\n>\n> Whatever \"worldview\" ends up being expressed/argued for through use of Deliberus and through\n> betterment of its ontology and processes and through its many different optimized collaborative\n> epistemic virtues and features...\n>\n> That resulting \"mind\" as a whole will be quite unlike each human or LLM mind, certainly in relation\n> to all the aforementioned aspects. It will most certainly be more thought-through, if we end up\n> getting enough people and agents to use it fruitfully and to scale it up to be able to weave our\n> minds together and become successfully interwoven in our seeing of the world through this\n> structured intersubjective scrutiny-machinery of which we'll all be a crucial part, both human and\n> LLM and code. It will be an interweaving of the code of life and the code of computers, and will\n> hopefully be very synergistic.\n>\n> But it will be unlike the worldviews and minds of humans and LLMs with their ingrained\n> emotional/reward-habits and well-trodden intellectual habits and inference structures and familiar\n> neural pathways. It will be a new type of alien mind, even more alien than the AIs we've built thus\n> far, even more crystallized-like in its intelligence, and more honest and autistic than most of us.\n> It will know all the inferences and all the claims in the world intimately, because that's part of\n> its code.\n>\n> Now, attunement and the worldview-lenses is an important part of Deliberus, and should fractally be\n> applied and reflected on at each scale of its creation and structure, but for now, let's imagine\n> that more as a capability of the mind of Deliberus to perform a Vulcan mind-meld between two other\n> minds, bridging their perceptions and beliefs about the world, showing what a worldview looks like\n> from the inside, to hopefully increase mutual understanding and curiosity, which can fuel a mutual\n> discovery of increasingly shared ground epistemically through engagement with Deliberus, etc.\n>\n> Anyway I was reminded of this juxtaposition between the minds of humans, LLMs and \"the mind of\n> Deliberus\" recently.... when reflecting on the chapter in the recent walkthrough called \"Sentient\n> silicon and a civilisation on machines: overlaps by topic, not evidence.\"\n>\n> That chapter compared a pair of claims that were clumped together by the sameness keeper because\n> they were vaguely similar by topic, but NOT by evidence, as pointed out in the walkthrough.\n>\n> You could imagine the LLM employed as the sameness-keeper thought about the two claims and saw\n> enough \"sameness\"/overlap there. But if it could for a moment possess the final form of the mind of\n> Deliberus and split and decompose each claim down to the separability floor and research every\n> premise and assumption and grounds for each claim, if it were to go through every explicated\n> inferential step (read up on the vertical chain-in-the-link and all that) of all the\n> evidence-as-subgraphs (read up on that too) of each claim, then it would... obviously see\n> excruciatingly clearly that these are not the same claim.\n>\n> And I wonder where to go from this insight. Will the keeper need to be given an awareness of the\n> complexity of all this, in its prompting? Or would it suffice to be quite pragmatic and tighten its\n> instructions to not clump together claims that perhaps are \"roughly the same kinda topic\" but depend\n> on completely different evidence and argumentation?\n>\n> How does this generalize to other microtasks used in Deliberus now and in the future? How can we\n> avoid them making leaps that are irresponsibly-large and so on, how can we avoid the muddling they\n> tend towards, due to their nature as aggressively-evolved-and-trained-to-be-useful,\n> fast-and-furious, not-so-careful, latent-space, motivated-reasoning thinkers? Same issue goes for\n> humans I guess.\n>\n> But both humans and LLMs can produce good intellectual work. I mean, me and you are making progress,\n> slowly... :)\n>\n> Realizing now this long rant should probably be folded into one of the higher level docs in\n> Deliberus, perhaps partly put directly in CLAUDE.md, and then merely reminded-of in the\n> instructions for the registers in the audio walkthroughs, lest these insights become siloed for\n> only my daily personal-Deliberus-podcast walks, haha\n\n## 2. Where this already stands in the corpus, and what is new\n\nThree of its parts were in the record before this reflection, under other names; one is new.\n\n- **That human minds differ more than models imagine, and that a model's space is the mean of them.**\n  [why-representations-converge-and-whether-minds-do.md](why-representations-converge-and-whether-minds-do.md)\n  measured it: human minds converge within a community of transmission and diverge between them, while\n  a model averages across them, so its vector for a contested word is *nobody's meaning, and the same\n  mean in every model*; and [core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md)\n  put the divergence where he puts it, in the periphery assembled by a life rather than in the shared\n  core. His sentence adds the cause in human terms (life path, identity, reward habit, groupthink) and\n  the warning that the model does not see how wide the spread is.\n- **That the model's failures are slop, and the remedy is the harness.**\n  [the-harness-conviction.md](the-harness-conviction.md) names the same failures (sycophancy, wiggling,\n  hallucination, forgotten premises) as what selection produces without a harness, and its law is\n  *remove the constraint, don't repair the reasoner*. His phrase for the reasoner, *fast-and-furious,\n  not-so-careful, latent-space, motivated-reasoning thinkers*, is that conviction's description of the\n  unharnessed model, with the honest rider that humans are the same kind of thinker.\n- **That the interwoven mind is the destination.** The 2011 formulation *argument map + human society =\n  hybrid AGI* and the 2012 line that finer discourse granularity *could be likened to the advent of\n  language itself* ([civilizational-vision.md](../civilizational-vision.md) § The Next Evolutionary\n  Transition); Wright's *brain of brains of brains* with the synapse layer missing. His sentence names\n  the connective tissue's product: a mind made of *the code of life and the code of computers*.\n- **New: the character of that mind.** *More honest and autistic than most of us*, *more\n  crystallized-like*, *even more alien than the AIs we've built thus far*, knowing *all the inferences\n  and all the claims in the world intimately, because that's part of its code*. Nothing in the corpus\n  had described the whole from the inside. It is the right description: the thing's intelligence is\n  crystallised in the sense that every step is written down and stays written, which is exactly what\n  neither a person nor a model does; and *autistic* in his sense (the sense he gives his own cognition\n  elsewhere) is a virtue here, a mind that does not round a distinction off to be agreeable.\n- **New, as a framing: attunement as the mind-meld.** The worldview lens\n  ([worldview-lenses.md](../worldview-lenses.md)) was built as something a person puts on. His frame\n  makes it something the mind of Deliberus *does* between two other minds: shows each what the world\n  looks like from inside the other, so that curiosity and understanding grow and shared ground is\n  found through the engagement. The capability is the same; the subject changes from the reader to the\n  system, and that is the frame under which the attunement pole scales (the dialectic applied\n  *fractally at each scale*, his words).\n\n## 3. The keeper, and what the mind of Deliberus would have seen\n\nThe case is the twenty-first listening piece's part two: two claims from the machine-consciousness\ndebate, one about sentient silicon and one about a civilisation running on machines, that the keeper\nof the seven keys judged *overlaps*. The keeper's own instructions say what overlaps means and how it\nis judged:\n\n> overlaps — establishing one would partly advance the other, but neither contains the other\n> …\n> OVERLAPS is judged by EVIDENCE, not by topic. Two claims about the same subject with no evidential\n> bearing on each other are unrelated. The question is whether establishing one would partly establish\n> the other.\n\nSo the instruction he proposes tightening is already there, in capitals. The keeper read it and still\nanswered by topic. That is the finding of\n[pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) in one more instance\n(seventeen of nineteen measured failures trace to how the question was posed or how the answer was\nreceived, not to the model's tier), and it is also the limit of that finding: an instruction can be\nclear and still be answered badly when the answer needs an object the reader does not have.\n\nWhat the keeper lacks is what he names: the parts of each claim down to the separability floor, the\npremises and grounds under each, the steps written out under every link\n([the-chain-inside-a-link.md](the-chain-inside-a-link.md), [writing-out-the-unwritten.md](writing-out-the-unwritten.md)),\nthe evidence as a subgraph rather than a word\n([scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md)).\nGiven those, *does establishing one partly establish the other* is a question about whether two\nsubgraphs share a node, which a second pass can check against the graph. Without them it is a question\nthe keeper answers from the two sentences and its own sense of the topic, which is to say from the\nlatent space he describes.\n\n## 4. The generalisation, proposed: no verdict without its object\n\nThe answer to his either-or (awareness of the complexity in the prompt, or a pragmatic tightening) is\nneither, and the corpus already has the shape, in three places that had not been joined:\n\n- a machine's sentence under a link is a claim the moment it is written (ledger AG12, ruled);\n- a link is written out in full rather than judged ([writing-out-the-unwritten.md](writing-out-the-unwritten.md), his *\"weak sauce\"* on naming alone);\n- a part stands alone or it is not a part (the separability floor, ledger B2).\n\nJoined, they say: **a machine pass answers by producing the object its answer depends on, and a\nverdict that comes with no object is refused.** For the keeper, *overlaps* is accepted only with the\nshared statement written out, a sentence that establishing either claim would help establish, which is\nthen itself a claim and can be attacked; *same* only with the consent-test restatement; *contains* only\nwith the part named. For a link, the step. For a split, the parts. For a scheme, the critical question\nthat the scheme would ask of this inference. The object is what the full mind would have had in front of\nit; asking the pass to write it is asking it to do, on one pair, the one slice of the descent its verdict\nrests on. A pass that cannot write the object has told you its verdict was a leap, which is the\nconfession channel this class of error lacked: the error rate of a leap is invisible (every answer is\nplausible), the error rate of a written object is countable by code and by a second machine reading (a shared statement either sits in both subgraphs or does not), with no person going through pairs.\n\nWhat it costs: one more field per pass, read before the verdict (the field order rule, reasoning before\nthe verdict it explains), and more tokens, since the object is longer than a key. What it does not do:\nit does not make the pass see the whole subgraph. It makes the pass write the piece of it the verdict\nneeds, which is the piece the mind of Deliberus would then keep. The written object is stored as a\nmachine's sentence, a draft claim nobody asserts, under the link or the proposal it belongs to; the\ncode for that exists (`deliberus/machine_sentences.py`).\n\nAwareness of the complexity in the prompt is the wrong tool for the reason the harness conviction\ngives: telling the reasoner to be careful repairs the reasoner. Pragmatic tightening is the wrong tool\nfor the reason § 3 shows: the tightening was already there. Requiring the object removes the\nconstraint, since the pass can no longer answer from topic, and it leaves the verdict where a person\ncan see what it rests on.\n\nHis other question, how to keep every microtask from irresponsibly large leaps, has the same answer\none level up: a leap is a verdict whose object was never written, so the standing test for a new pass\nis *what object does this verdict rest on, and does the pass write it?* The smallest-step rule for\nwritten steps (AG14, ruled) is this test applied to the chain; this note applies it to the keeper and\nproposes it for every pass.\n\n### 4a. His objection the same evening, and what it changes\n\nTwo sentences in § 4 did not hold, and he said so (2026-10-18): *\"This doesn't hold up. Poke holes in\nthis reasoning for me.\"* The holes, found on his push:\n\n- **The object is also written from latent space.** Asked for a statement that establishing either\n  claim would help establish, a model produces one for apples and zebras (*both are living things*).\n  Writing the object does not remove the sloppy answer; it relocates the slop into the object, where\n  field order only makes the key fit it. His own formulation of the risk: *\"If we ask 'how are apples\n  and zebras the same/overlapping, what's your take?' it'll always come up with some weird slop\n  justifying it.\"*\n- **\"Both of his options repair the reasoner\" was a false dichotomy.** The harness conviction's own\n  content is harness *and grounding*. A prompt that carries the rulings bearing on sameness (the\n  separability floor, the seven keys' definitions, evidence not topic, the one-step invariant, the\n  stranger test) and tells the pass how far apart two claims can be while sharing a topic is\n  grounding: it hands the pass the slice of the corpus's crystallised mind that bears on its\n  judgment. That is the harness, not a repair of the model. His words: *\"perhaps we can prompt it\n  skillfully enough to glean enough awareness of what it doesn't know and how careful it must be\n  with inferential steps and with judging sameness in a sloppy manner. Have it realize that there IS\n  vast inferential distance between these claims. Jog its imagination by auto-including and referring\n  to and reminding of rulings that are relevant… I think this is our only chance.\"*\n- **The right object is the chain, not a shared statement.** `overlaps` means establishing A partly\n  establishes B, which is a claim about a step from A to B; what shows distance is the number of\n  steps and what each rests on (AG14's smallest step, applied to sameness). Four steps is a far link,\n  not an overlap. So *write the object* survives in this form: write the steps between the two claims,\n  and read the count.\n- **The keeper's prompt has `unrelated` and `cant_tell` on the menu beside the answers**\n  (`deliberus/keeper.py`), which the ruling of 2026-09-22 forbids for every pass: a no-answer belongs\n  in a gate asked first. Read 2026-10-18; a build item.\n- **\"A person's reading catches the rest\" was wrong, and he has ruled against it more than once.**\n  Microtasks do their work without people; a person is not the graph's janitor (2026-09-13: *\"spare\n  human users from having to later clean up and organize the graph manually etc, to instead free up\n  their mental cycles for the uniquely human-needing tasks\"*; 2026-09-22: *\"so humans ideally don't\n  have to do reviews/housekeeping/microtasks that AI could be tasked with confidently\"*; T1, that\n  review need not keep pace with a proven pass). What a person does is calibrate a pass once (T8),\n  contribute what argues for itself, and the two reserved classes. The twenty-first listening piece\n  said *goes to a person first* in two parts and predicted he would want a sample read by a person;\n  that was the session's stance, not his, and the piece's header now says so.\n\n**Built the same night** (his word: *\"Implement the pragmatic inkling as you propose now. Then we can always revise and improve on it.\"*): `deliberus/keeper.py`, header and `resolve`; the pair run through the pipeline's own rules by hand, with the count: `the-chain-inside-a-link/sentient-silicon-pair-through-the-pipeline.md`.\n\n**His next move the same night, after the virtual run** (ledger AM5): the conditional in claim A is itself a link that the splitter should cut out first (consequent as a conclusion asserted by nobody, antecedent as an unasserted premise, the conditional as the author's relevance claim under the link, the hedge on the link), so that the keeper compares the cut conclusion with B, and what is left between them is three sense questions for the concept layer. The virtual run's three learnings: `the-chain-inside-a-link/sentient-silicon-pair-through-the-pipeline.md` § 7.\n\n**And two more of his the same night** (ledger AM6, AM7, his words there): a hedge is the author's stance and goes to telemetry, never into strength or into a minted conclusion's wording; and a dispute about what a concept means has one entry point, the claim, since every use is a vote on the meaning and the concept layer is read over uses as a horizon kind is read over arrangements. For the keeper that means a stop on a sense-bound pair names the word pairs it hinges on.\n\nWhat AM1 says after this: a pass is given the record that bears on its question (the relevant\nrulings, the parts and reasons already under each claim) and writes the chain its verdict rests on;\nthe chain is checked by code and by a second reading, never by a person per item.\n\n## 5. What this says to the walkthroughs, and the cast's purposes\n\nHe asked that the reflection be reminded of in the register instructions rather than left in them. The\nreminder is this: a walkthrough's cast is a small version of the interweaving. Each voice stands for one\nof the virtues the whole mind is meant to compose, and a piece that uses only the spine has left the\nothers out. The purposes, one per register, proposed for his word:\n\n| Voice | Register | What it is for, the value to hold in mind |\n|---|---|---|\n| Lucy | the spine | **Fidelity to the record.** Says what is so, in the plainest words, with the date and the source, and extrapolates nothing. She is the ledger of the piece: what the harness conviction calls *steel* when it is in a voice. |\n| His clone | his own words | **Sole authorship.** His position is rendered only in his words; a prediction of him is a prediction, built from his record and the project's aims, and is heard as him so that he can hear whether it is. |\n| Elliot | philosophy | **Distinctions.** Takes a question past its textbook statement to the tradition that treats it, and names the distinction that dissolves a confusion. His own method ruling: a term is defined before it is used, and a more-correct answer is derived from the purpose and stated beside what it beats. Every longest-lived mistake in the record was a distinction collapsed. |\n| Scotty | the irreverent | **The adversary inside the house.** Argues against the piece's own drift, says what a motivated critic would say, borrows a thinker's voice when that earns it (marked as borrowed, rule 28). He is the attempt to stand outside the frame the piece is written inside, which no mind does for itself. |\n| Anna | the vision, the interface, the studies | **Hope examined with stringency.** What this would open, in the civilizational frame, and what the literature actually says; the metamodern pairing he named, the vision held at full size and the evidence beside it. |\n| Myrddin | the elder | **Proportion.** The long view: what matters across years, what to let go of, said inside what can be stood behind (rule 29). The attunement pole's own voice in the piece, and so the one that holds his hand rather than his attention. |\n| Hope | the stranger | **Legibility to someone with none of the context.** Hears one of our sentences for the first time and says what she took it to mean. She is the stranger test in person, and condition (d) of the success prior at the doorstep: every place she must *learn* a word is a place the constructed code leaked. |\n\nSet against § 1: Lucy and his clone are the record and the person; Elliot and Scotty are the two\nmovements of scrutiny, the distinction and the challenge; Anna and Myrddin are the two reaches of the\nattunement pole, forward into what could be and outward into what matters; Hope is the outside. That\nis the mind-meld in miniature, which is why a long piece with one voice has lost more than variety.\n\n## 5a. The morning after (2026-10-19): one case at a time, and a pair read in its game\n\n**How a person reads a pass.** Asked whether yesterday's nine-pair sample of the keeper counted as a\nperson's reading, he said the canon line (*a person reads what a machine pass produced before anyone\nproves how often it is wrong*) does not carry his gist, and gave it:\n\n> was not saying a human needs to go through more than 1-2 samples/cases before you can identify\n> useful regularities and insights, and ofc ideally subsequently more samples than that, but just like\n> when listening to walkthroughs it helps if I can deep-dive one case at a time (I think that's the\n> ideal for me, I can discern very early what's going on and I think many (philosophically trained)\n> humans are like this) and let it soak in my brain and think about why I have the\n> intuition/reflections/opinion/take I have for that case, BEFORE I move on to think deeply about the\n> next. This is the way I feel like it is in the vast majority of situations for me. So I think I wanna\n> do it this way instead of using the sitting/review-pages, however amazing it was that they were\n> created for me. One thing we should pay attention to is how we select which cases I take a look at\n> first, because the whole process is obviously going to be quite sensitive to any bias in that\n> selection, but I'm sure you can pre-filter for any seemingly weird outliers and either serve them up\n> to me or serve up seemingly-obviously-easy-to-judge cases/samples, I guess it depends, we'll see.\n\nEarlier, on 8 October, the same position from the other side: *\"Agreeing on rules comes late, yes. By\nreviewing cases you reflect on them and see, distill, regularities, learnings.\"* A re-cut of the canon\nline was put to him the same morning and blessed (*\"Blessed\"*; ledger Y9, where his sentence stands).\nWhat it changes: a review is a conversation on one case, chosen on purpose, whose take and reason are\nwritten down before the next; the sitting pages stay as instruments, not as the way he reads. What the\nsample file of 18 October lacks is any field for his take: his reflection of that night (§ 1) is the\nonly person's reading on record, and it read one pair (§ 3).\n\n**Three reflections the same day, when the rule was made global** (`~/.claude/CLAUDE.md` § *Case review starts one case at a time and ends in a rule he can bless*). A case may carry more than one question: *\"There may be more than one question/angle to ponder on the case.\"* A take that answers more than was asked shows the question was too narrow (on a narration-judge sheet he labelled two cases by where he was, and that was the right answer). And the reading has an end beyond itself:\n\n> The ideal is also to elucidate and clarify regularities that can help sift through and winnow out what remains to be ruled about the other cases, so that I can merely bless the application of some kind(s) of automated judgment of (ideally) all cases in the future, whether by machine microtasks or by automated deterministic code tools added to a codebase or to a harness or whatever is applicable to the situation at hand.\n\nSo each reason is turned into a candidate rule and tried on the cases not yet read, and he is shown only what no rule decides yet. The canon line carries all three since that day.\n\n**Four fresh keeper answers, read the same morning.** Drawn from the record of model calls, the four\nthe keeper gave last under its new gate-first prompt (no cosine draw, no new call):\n\n| Pair | Keeper |\n|---|---|\n| *The normative priority of distributed human authority can be defeated under certain conditions* · *Serious loss of human control over AI systems is a coherent possibility* | no bearing, confident |\n| Anthropic found Claude's J-space more powerful in some respects than the human global workspace · that workspace showed traces of step-by-step reasoning | no bearing, confident |\n| the J-space indicates such a workspace is not merely a feature of human wiring · Claude's features resemble global workspace theory | overlaps, confident, two-step chain, both steps asserted |\n| *Human brains and bodies will soon be superseded by machines* (the consequent cut from a conditional) · *Future human civilization will … run on … AI* | overlaps, unconfident, two-step chain nobody asserts, refused by the gate |\n\n**His reading of the first** (the first case read one at a time, by the method above):\n\n> What does \"no bearing\" mean exactly? A is normative, and \"distributed human authority\" needs badly\n> to be defined, I mean, in relation to what? If you see these claims juxtaposed you might understand\n> the context of that it's talking about general authority/control over infrastructure/society and in\n> particular relating to AI. This is honestly a case in which the (annoyingly fluid) context for the\n> concept-sense-use matters (read up on this), and saying \"distributed human authority\" is sort of a\n> type of move in a language-game, no?\n\n*No bearing* means the keeper wrote no chain under which establishing one would move the other. Read\nin their sources (a structured position on AI governance, the structured doom argument), B is the risk\nthat the priority in A exists to guard against, and A says the priority can be overridden: B supports\nthe priority and A qualifies it, two moves in one game. The stranger rewrite strips the game that fixes\nwhat *distributed human authority* means, and the keeper sees only the rewritten sentences. So its\nanswer is true of the two sentences and false of the two moves. The corpus already holds the pieces:\nsameness is use-relative (`claim-sameness-philosophical-readings.md`, the late Wittgenstein line), a\nuse is a vote on the meaning (AM7), and a sense is somebody's difference from a shown reference\n(`core-periphery-and-the-concept-layer.md` § 5b). What is new is that the keeper's *bearing* question\nis a use-relative question too. Proposed as AM9: hand the keeper each claim's quoted passage and the\nsenses in play for a shared term.\n\nThe other three pairs were not yet read by him. **The third, read by the session on his request**\n(*\"Well, chain of two steps? Try to paint me a picture of the inferential chain/distance yourself\"*):\nthe keeper answered *overlaps*, confident, on a two-step chain with both steps asserted. Read\nwhole, its first step restates claim A (Claude's features resemble the global workspace) and its\nsecond restates claim B (the J-space shows such a workspace is not merely human wiring), so the\nchain holds nothing between them. The step it leaves unwritten is the leap: from *resembling a\ntheory's posit* to *possessing the structure the theory is about*, carried by the word\n*workspace*. So the chain was produced in form and not in substance, which AM1's gate cannot\ncatch by counting steps: a chain whose steps only restate its two ends passes. **Proposed check,\nunruled**: a chain step that is the same claim as either end (by the keeper's own *same*) does not\ncount as a step. His position on the same pair is filed as AM10 (a theory is defined like a concept,\nby the claims that use its name and survive), in his words on the ledger row.\n\n**AM3 and the word *autistic*, as the morning walkthrough put it to him (unruled).** The word does\ntwo jobs in his own record: on 2026-10-18 a virtue (*\"more honest and autistic than most of us\"*),\non 2026-08-30 a vice (daemons to reason about the graph *\"in a way that isn't autistic in some\nsense, too detail-focused\"*). One word, two senses, both his: the collision test fires. *Crystallized*\ncollides too, with Cattell's crystallized intelligence (stored knowledge), where his sense is a\nlattice whose every step stays in place. Claude's advisory stranger read: a diagnosis used as a\ncompliment. Candidates that describe what the mind does instead of labelling it: *a mind that keeps\nevery distinction* · *a mind that rounds nothing off to be agreeable* · *a mind that takes each\nsentence as written and remembers it*. The two senses map onto the two poles: the virtue is the\nanalytical pole, the vice is what the mind-meld (attunement) is in his sentence to balance.\nPredicted reception: he keeps the word in his own sentence and lets the resident line describe.\nWhat the research says about both senses (2026-10-20): the vice as a deficit does not hold (global processing is slower, not missing), and the virtue's best-supported form is a smaller framing effect; [careful-and-broad-minds.md](careful-and-broad-minds.md).\n\n**R12 meets AM8: a proof of the keeper lasts until the next change to its rendered block.** A proof\nis per pass version, the keeper's version is its instructions' fingerprint, and since AM8 those\ninstructions carry a block rendered from five listed ledger rows and three canon lines, so a ruling\nor a reworded headline on a listed row voids the proof, and a ruling on an unlisted row does not.\n⚠ Corrected the same morning: first written as *every ruling voids the proof*, which the code does\nnot do (`proving-a-pass-reliable.md` § 11 has the reading and his question on keeping an accurate\nhistory). Two ways out, neither ruled: the version excludes the grounding block (rulings treated as data, not instructions), or\nproofs are short-lived by design and the sweeps re-sample after each ruling. Also from the same\nreading: links were retired as *unrelated* on the keeper's own confident flag, which R12 calls a\nself-report, so the house already acted on the number the open row distrusts.\n\n## 5b. His proposal of 19 October: take stance hedges out of every claim (RULED 2026-10-20, built)\n\n**Ruled 2026-10-20**, founder: *\"Proceed. We can always refine this later. Still in a dogfooding and\ntesting and refinement phase after all.\"* Built the same night in pass 2b (`extraction/focused.py`,\n`AtomicClaim.hedge`, `hedge_reading`); tried on nine real passages, `a-sentence-as-a-pointer-into-a-landscape.md`\n§ 6e. With it he named the aim: *\"I want claims as close to testable as possible, testable bets that can be\nchallenged and tried whenever possible/applicable. Kinda similar to prediction markets like Polymarket I\nguess.\"* (research in `docs/research/testable-bets/`).\n\nHis words, after a rewrite turned an author's *a second issue* into *a significant technical hurdle*:\n*\"Couldn't we simply rule to strip out certainty estimates/hedges from original phrasings, since the\ncertainty for a claim in Deliberus will simply be shown over time. If a claim ends up overwhelmingly\nstrong after a month of collective deliberation and research, \"may\" or \"probably\" or \"could\" just looks\nlike a humblebrag in retrospect kinda, and if it drops dead after intense debate then it would never have\nheld up anyway. And if it changes a month later when new data comes along, then so be it. These\nprobability terms play a role in the language game that is better played by the\nprobability-estimation-machinery of Deliberus itself, is my intuition, rather than belonging inside the\nclaim itself, does that make sense? Pros/cons?\"*\n\nIt is AM6 widened from minted conclusions to every claim, and it is also his open revisit of *may*.\n\n**For it.**\n1. *One proposition, many stances.* \"X may cause Y\", \"X probably causes Y\" and \"X causes Y\" become one\n   core with three uses (AM7: a use is a vote on the meaning; core-periphery § 5). Sameness gets easier.\n2. *A hedged claim can hardly be wrong.* \"X may cause Y\" survives almost any evidence, so the graph's\n   weighing has nothing to bite on. The definite form is the one evidence can move.\n3. *No double count.* The author's hedge and the graph's strength would both speak to certainty, and a\n   reader anchors on the first.\n4. *A hedge is often a move, not an estimate*: politeness, a shield, legal caution. That is the function he\n   names, played better by the graph's own estimate, as the language-game reading of AM9 says.\n5. *One rule instead of two*, since AM6 already does this for minted conclusions.\n\n**Against it, and what each needs.**\n1. *Strengthening is a straw man.* The author said *may*; the graph shows *will*; an attack that defeats\n   *will* looks like a defeat of the author, who never said it. This reverses the hand-mapping discipline\n   *only ever lower the asserted certainty, never raise it*, and the principle of charity. Needs: the hedge\n   kept on the author's use and shown wherever the claim is attributed to them, and a defeat of the\n   definite form never shown as a defeat of the hedged use.\n2. *Some modals are content.* \"This policy could cut emissions\" is often a feasibility claim, and\n   feasibility is one of the epistemic kinds the extractor already names. AM6's two-substitution test\n   (*I think* against *it is possible that*) separates them, and its first live run read a *may* as\n   *ambiguous*. Needs: strip only what reads as stance; keep what reads as content.\n3. *Scope words are not hedges.* \"Agents tend to praise their work\" is not \"agents praise their work\";\n   *tend to, often, most, some* fix what the claim covers. Needs: never strip scope.\n4. *A stated probability is the claim.* \"A 70% chance of X\" is a forecast; the number is its content.\n5. *The hedge is information about the author*: calibration, how a source speaks. Kept as stance, it is\n   not lost; this is a con only if it were deleted.\n6. *A machine does the sorting* and will err. The author's own sentence is kept on every new claim\n   (`author_wording`), so a wrong strip can be seen and undone.\n\n**Reading.** Adopt it for stance only, as AM6 for every claim, with four guards: scope words, possibility\nas content and stated probabilities stay; the hedge is stored on the author's use and shown with every\nattribution; a defeat of the definite form is shown against the proposition, never as a defeat of the\nauthor's hedged use; the original sentence stays beside the claim. One interaction to know: the rewrite\ninstruction ruled the same evening says *do not add certainty the author did not state*. As built (2026-10-20), that line stands and the hedge instruction follows it in the same list: a stance or ambiguous hedge leaves the wording for the `hedge` field, so the definite form is not counted as certainty added (`extraction/focused.py`, the decontextualisation prompt, item 3).\n\n## 6. What is not settled\n\n- Whether the keeper is given the object rule first, or every pass at once (ledger AM1, AM2).\n- Whether *the mind of Deliberus* becomes a resident line; a candidate is written in the ledger (AM3)\n  and nothing enters always-loaded memory without his word.\n- Whether the cast's purposes stand as written (AM4).\n- His three proposals of the same night were ruled 2026-10-19 (*\"Build AM5+6.\"*, AM7 *\"seems reasonable\"*): a\n  conditional is a link cut out before the splitter (AM5) and a hedge is stance and not strength (AM6) are\n  built in `deliberus/reasons_off.py` and ran once on this note's own pair\n  (`the-chain-inside-a-link/sentient-silicon-pair-through-the-pipeline.md` § 9); AM6 holds *for now*, his revisit\n  of *may* open. That revisit is now ledger AO1: keep a hedge that reads both ways (*this policy could cut\n  emissions*) in the claim's wording, flagged, since stripping it stores *will* for an author who may have meant\n  *can*. Both research reports recommend it (`testable-bets/README.md`); he wants more perspectives first\n  (2026-10-20: *\"Also feels connected to \\\"may\\\" from before. I guess I need some more perspectives on this.\"*). A use is a vote on the meaning (AM7) is ruled and unbuilt.\n- The critical questions of a scheme as distance tests for the keeper (in AM1's amendment; he has not absorbed it yet).\n- What the three minds' shared failure, reasoning by motive, should become in the graph (his question of 2026-10-19, ledger Arc AN): causes of belief enter in three ways, each with one rule, and the mind of Deliberus is honest in the measurable sense that its critical questions do not know which side a claim serves. [causes-of-belief-and-the-three-minds.md](causes-of-belief-and-the-three-minds.md).\n- Grounding every pass in the rulings that bear on its question (AM8, his words: *\"I think this is our only chance\"*). RULED 2026-10-19, *\"Rendered! But will that replace the whole prompt for the pass?\"*: the block is rendered from the ledger and the canon, the task stays hand-written. Built that night (`deliberus/grounding.py`, `docs/research/grounding/keeper.pair.md`); the lists for the other passes are open.\n\n## Cross-references\n\n- [the-harness-conviction.md](the-harness-conviction.md) (slop and the harness; remove the constraint)\n- [why-representations-converge-and-whether-minds-do.md](why-representations-converge-and-whether-minds-do.md), [core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md) (how far minds differ, and where)\n- [auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c (the keeper), [the-chain-inside-a-link.md](the-chain-inside-a-link.md), [writing-out-the-unwritten.md](writing-out-the-unwritten.md), [scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md) (what the full mind would have in front of it)\n- [pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md), [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) (an error rate belongs to the task as written)\n- [../civilizational-vision.md](../civilizational-vision.md) § The Next Evolutionary Transition, [../worldview-lenses.md](../worldview-lenses.md), [../analysis-and-attunement.md](../analysis-and-attunement.md)\n- [../listening/how-to-write-for-listening.md](../listening/how-to-write-for-listening.md) rule 57 (the cast)\n- Ledger: [session26-open-rulings.md](session26-open-rulings.md) Arc AM\n"}