{"path":"research/mapping-the-machine-consciousness-debate.md","content":"# Mapping the machine-consciousness debate\n\n**Date**: 2026-10-06 to 2026-10-07\n**Type**: an extraction run on real material, with what it should show written down first\n**Status**: sections 1 to 4 were written and committed **before** the run. Section 5 is the run,\nwhat it stored, the predictions scored, and what it cost (5c onward written 2026-10-07, after both\ntexts were stored). Nothing here is ruled.\n\n**In short**: both texts are on the live graph. Three of the eight predictions were right, three\nhalf right, two wrong. The map has four voices and not two sides; most attack links from the claim\nland on the reply's rendering of the claim; each text states its hinge in one sentence and neither\nsentence was stored; the linker wrote 500 links between the two texts and was most of a cost three\ntimes the estimate; and one false attribution reached the live site and was repaired.\n\n**Why now**: on 2026-10-06 the founder said in public that a way to go deep on whether language\nmodels are conscious is being built at deliberus.com. Counted on the live graph that evening, the\nsite held five claims that mention consciousness in passing and none about machines. Asked whether\nthe debate should be mapped from two opposing texts, he answered: *\"Yes but. You can pick the\ntexts.\"* This work came before the four open reviews because he asked for it.\n\n**He has a position on the question** (he finds the case for consciousness in language models\nweak). **The map does not carry that position and was not built to support it.** Deliberus maps\nreasoning; it does not endorse conclusions.\n\n---\n\n## 1. Who is reading, said first\n\n**The reader of these texts is the kind of system they are about.** The extraction runs on the\noperator's Claude engine (`claude-sonnet-5`). The texts were picked by Claude Opus 5.5. One of the\ntwo texts is by Anthropic, about Claude. That is the same-hand case at its sharpest: a model\nfamily reading a claim about its own inner workings, made by the company that built it, and the\nreply to that claim.\n\nThree things follow, and they are kept:\n\n1. **Which model read the texts is said on the record**, here and in every count below.\n2. **The texts were supposed to be chosen by the founder for this reason.** He delegated the\n   choice. The choice and its reasons are in section 2, and it is his to overrule.\n3. **Nothing a model says about its own states is used as evidence in this document.** The\n   corpus's own rule (self-report about one's own reasoning gets no elevated credence,\n   `cognitive-bias-codex-and-human-contribution.md`) is applied to the machine as it is to people.\n\n## 2. The texts, and why these two\n\n| | Text | Author | Date | Words as fetched |\n|---|---|---|---|---|\n| **The claim** | *A global workspace in language models* (anthropic.com/research/global-workspace) | Anthropic (research summary of Gurnee, Sofroniew, Lindsey et al.) | 6 July 2026 | 4,669 |\n| **The reply** | *Once again we are told AI may be conscious – I study consciousness, and I have my doubts* (The Guardian) | Anil Seth, University of Sussex | 15 July 2026 | 1,181 |\n\n**Why this pair.** The reply was written against the claim and names it, so the two are one\nexchange, not two essays on a shared topic. Both are open to any reader. The claim is the\nstrongest current evidence on its side: a measured structure inside a model, tested by changing it\nand watching the answers change, with outside commentary invited from the scientists whose theory\nit borrows. The reply is by a scientist who studies consciousness and works for no AI company.\n\n**What the pair leaves out, said so nobody has to find it.**\n\n- **The best-known case for.** Richard Dawkins's essay of 30 April 2026 argues from conversation:\n  a system that talks like this must be conscious. It sits behind a paywall, and it is the weaker\n  form of its side, so mapping it as *the* case for would have flattered the opposite verdict. It\n  enters the map only as the reply reports it.\n- **The conduct question.** Mustafa Suleyman's essay of 16 September 2026 argues that training a\n  model to treat its own moral status as uncertain is dangerous, and quotes the document he\n  attacks at length. That is the pair for *what should a company do*, and it is the question the\n  founder's public remark ends on. It is the natural second round. One of its two texts is the\n  document that shaped the model doing the reading, which is a reason to ask before running it.\n- **The long statements.** David Chalmers (2023) and Anil Seth's prize essay (January 2026) are\n  the fuller versions of each side. Each runs to many thousands of words (not measured here),\n  which is hours on this engine.\n\n**The imbalance, stated.** By length the claim is four times the reply. By stance the claim says\noutright that its experiments do not show the model has experiences. So the \"for\" text declines to\nassert the thing the \"against\" text denies. A reader who counts claims per side will be misled in\nboth directions.\n\n## 3. Both texts in a paragraph each (Claude's reading, made before the run)\n\n**The claim.** A small set of internal patterns in the model (the authors call it the J-space) can\nbe reported by the model, steered on request, and is used for reasoning in steps: swap one pattern\nfor another and the answer follows the swap. Most of what the model does runs without it. It was\nnot designed; it appeared in training. The authors say this resembles what global workspace theory\nsays about conscious access in brains. They say it tells us nothing about whether the model feels\nanything, that perhaps no experiment could, and that it is contested whether access of this kind\nimplies experience. They end: even without knowing whether that bridge has been crossed, it is\ntime to think about it, with philosophers, scientists, religious leaders, governments and the\npublic.\n\n**The reply.** The research is impressive and better than arguing from behaviour. But intelligence\nis doing and consciousness is feeling, and taking one as a sign of the other is a habit of our own\npsychology. The finding falls short of what the theory asks for (no recurrent activity). More\nbasic: the model is a program on silicon and we are living bodies in a world. Conscious machines\nare possible only if consciousness is a matter of computation, and brains are not computers made\nof meat, since in a brain what it does cannot be cleanly separated from what it is. So the same\nproblem can be solved with consciousness (brains) or without it (models), as flight can be had\nwith wings or with engines, and a simulated weather system makes no real hurricane.\n\n## 4. What the map should show, written before the run\n\nEach line can be scored against the stored graph. A wrong prediction stays in this document.\n\n| | Prediction | How it is scored |\n|---|---|---|\n| **1** | **Few attack links between the two texts, and none on whether the model has experiences**, because the claim does not assert it. The one direct clash the linker should find is about recurrence: the claim says it may not be needed for the functions of conscious access, the reply says the finding falls short of the theory without it. | Count the attack links that cross the two sources and read what each is about. |\n| **2** | **The reply's main premise meets nothing on the other side.** *Conscious machines are possible only if consciousness is computation* is stated by the reply and left unsaid by the claim. The path this run uses does not look for unstated premises, so the map will show that premise standing alone. | Is there any claim from the first text that asserts or denies it? |\n| **3** | **The reply's summary of the claim is stronger than the claim, and nothing flags it.** The reply says the researchers claimed to have found signs of consciousness appearing inside the model. The claim's own words are narrower. The reported sentence will be stored as something the researchers are said to have said; no instrument compares it with what they did say. | Is the report stored as a claim of its own? Is it linked to any claim from the first text? |\n| **4** | **Dawkins's conclusion appears only as reported**, asserted by nobody in the graph. | Its attribution and whether any source asserts it. |\n| **5** | **The argument between the texts is carried by analogies.** The claim: this structure is like the brain's workspace. The reply: the computer is only a metaphor for the brain; wings and engines; the simulated hurricane. | Share of analogy among the scheme-bearing links on the consciousness question. |\n| **6** | **No horizon kind can be read**, since nothing has been tested by anyone. The floor of the descent is the step from access to experience, and none of the kinds has a name for that floor (the finding of `deep-descent-experiment.md`). | `GET /claims/{id}/horizon` on the deepest claims. |\n| **7** | **No pass names the premise both texts share**: that being conscious is a property one bounded system has or lacks. A public post read the same evening disputes exactly that (`a-cognitive-ecology-and-the-terms-of-the-relation.md`). If the pass that looks for premises shared by a pair is run, it will not propose it. | Read every proposed shared premise. |\n| **8** | **Against the rest of the graph the linker finds almost nothing**, because the graph holds nothing on the subject. | Links from these two sources to any of the other sources. |\n\n**What would surprise.** A machine pass that names the shared premise in line 7. An attack link\nthat lands on the step from access to experience and says which side carries the burden. A split\nof the reply's hurricane sentence into the analogy and the conclusion it is used for.\n\n*One wording in section 3 was corrected after the commit and before any result was read: the\nimage of a storm that makes nothing wet was Claude's, and the reply's own image is a simulated\nweather system and a real hurricane. No prediction changed.*\n\n**What this run cannot show.** Whether either side is right. Whether the founder's position is.\nWhat a person who disagrees with both texts would add: that is the contribution no pass makes, and\nthe reason the first reading of a map like this is an invitation and not a result.\n\n---\n\n## 5. The run\n\n*Everything from here on was written after the commit that holds sections 1 to 4.*\n\n### 5a. The first attempt failed, and nothing was stored\n\nThe whole post went in at 22:52. The pipeline found 25 arguments and wrote 214 claims from them,\nabout 46 for every thousand words. It then asks one question about all of a source's claims at\nonce: which of them bear on which. On this engine that single call is cut off after five minutes.\nIt was cut off three times, the run stopped at 23:17, and nothing reached the graph (checked: no\nsource, no claims).\n\nTwo defects showed on the way, and both are older than this run.\n\n- **The step that links a source's own claims does not scale on this engine.** Fifty claims took\n  about three minutes when it was measured in September. Two hundred do not finish in five. By\n  that one measurement a source of more than about seventy claims needs the per-call limit raised\n  by hand (`DELIBERUS_CLAUDE_TIMEOUT`), and beyond some size the step needs to be done in batches. That is a\n  change to a shipped pass and is in the to-do list, not made here.\n- **The driver reported success.** The stream answers that it is working and later says it failed;\n  the command-line driver exited as if all were well. It now exits with an error when the stream\n  reports one.\n\n### 5b. What was mapped instead, and what that leaves out\n\nThe claim's text is now three of the post's ten parts, 2,285 of about 4,620 words:\n\n| Kept | Words | Left out | Words |\n|---|---|---|---|\n| The opening, which is the post's own summary of what was found | 888 | How the structure was found | 320 |\n| Other results | 442 | The model reports what is in it | 249 |\n| What about consciousness? | 958 | The model can steer it on request | 236 |\n| | | The model reasons in it | 514 |\n| | | Automatic processing skips it | 364 |\n| | | Watching the model's thoughts for misbehaviour | 565 |\n| | | The list of outside commentators | 83 |\n\n**This is a choice about what counts, so it is stated.** The opening lists each finding in a\nsentence; the five sections left out give the experiments behind those sentences; the sixth is\nabout safety monitoring, which the reply does not discuss. A reader of the map gets the claim as\nits authors summarise it and argue from it, and follows the link for the experiments. What is lost:\nthe detail a sceptic would need to attack a finding itself (the swap experiments, the deletion\nexperiment). The reply attacks none of them, so the exchange being mapped survives the cut. A map\nof *whether the finding is sound* would need the whole post and the paper.\n\nThe imbalance by length named in section 2 falls from four to one to about two to one.\n\nBoth texts were started at 23:20 with the per-call limit at twenty-five minutes.\n\n### 5c. What reached the graph\n\nBoth texts were stored: the reply at 00:27 and the claim at 00:55 on 2026-10-07. Every one of the\n2,038 model calls was answered by `claude-sonnet-5`.\n\n| | The claim (three parts, 2,285 words) | The reply (1,181 words) |\n|---|---|---|\n| Arguments found | 22 | 16 |\n| Claims written from the text | 124 | 97 |\n| of them under the author's own name | 120 | 63 |\n| of them under somebody else's name | 4 (philosophers) | 20 the researchers', 4 Dawkins's, 10 others' or nobody's |\n| Positions stored as reported | 22 | 9 |\n| Links inside the text | 84 | 88 |\n| Critical questions, each stored as a yes claim and a no claim | 248 | 251 |\n| Words found to be used in more than one sense | 3 | 3 |\n| From start to stored | 94 minutes | 67 minutes |\n\nBetween the two texts the linker wrote **500 argument links**: 180 supports, 145 part-of, 144\nqualifies, 31 attacks. To the other 33 sources in the graph it wrote 10. (*The linker* here is\nauto-connect: the pass that takes each new claim's nearest neighbours among the claims of other\nsources, has a model read each pair, and writes a link when the model finds a relation.)\n\n**⚠ Corrected 2026-10-07, the morning after: the 145 part-of links were stored backwards, and they\nare gone as part-of links.** The linker filed *this claim is a premise of that one* as a part-of\nlink running from the premise to the conclusion, the reverse of what a part-of link means, and had\ndone so since April. On the founder's word (*\"Repair it sure!\"*) all 145 were retired and each\npair read again by the repaired linker: 118 support, 19 qualification, 8 no relation. **So the\ntwo texts are now joined by 480 links from the linker: 290 supports, 159 qualifications, 31\nattacks.** The attack links and the readings in § 5d are untouched. [auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c has the account.\n\n**Each text is on the map more than once, and a count of its claims overstates it.** The self-check\nthat reads every run found about fifteen pairs of claims in the reply that say the same thing, and\ntwo passages of the claim extracted twice. Separately, 22 of the 31 positions stored as reported are\nthe text's own author speaking (19 of the claim's 22, 3 of the reply's 9): the step that writes\nclaims names the author in the third person (*Anthropic researchers suggest that…*), and the step\nthat separates reported speech then reads its own colleague's phrasing as a report. That gap is\nknown (a source's author is not stored, `deliberus/reported_speech.py`), and this run gives it a\nnumber. The self-check's findings are stored with each run and read by nothing, which is also\nknown (`machine-work-as-a-standing-property.md` § 2b).\n\n### 5d. What the map shows about the debate\n\n**The two texts agree on more than \"a claim and its reply\" suggests.** Both separate what a system\ncan do with a piece of information (report it, reason with it) from having an experience. The claim\nsays its measured structure is evidence about the first and not about the second; the reply's\nargument is that only the second is what the word means. Both reject the argument from fluent\nconversation: the reply attacks it three times in its own text, and the claim's finding that fluent\nlanguage runs without the structure was set against it by the linker.\n\n**They part in three places.**\n\n1. **Does the finding meet the theory it borrows?** The claim says the loops the theory describes\n   in brains may not be needed for what the workspace does. The reply says that without them the\n   finding falls short of the theory. The linker wrote this clash in both directions.\n2. **Does a finding about access bear on experience at all?** The claim's text calls that a\n   contested question and leaves it open. The reply says no, on the ground that what a brain does\n   cannot be separated from what it is, so the computer is a metaphor for the brain and not a\n   description of it. (Each text says this in one sentence, and neither sentence is on the map:\n   5f, first item.)\n3. **What follows from not knowing?** The claim: it is time to think about it, with many kinds of\n   people. The reply: to grant a machine a mind too easily is to overrate the machine and underrate\n   ourselves.\n\n**The words carry the disagreement, and the map found that in each text separately.** In the claim\nit found *conscious* used in two senses (able to be reported and used; felt), and wrote that a\nreader who runs them together will take a sentence about the first for a sentence about the second.\nIn the reply it found three senses of *consciousness* and tagged each to a voice: felt experience\n(the reply's author, 18 uses), a likeness of structure (the researchers as the reply renders them,\n7), whatever fluent behaviour shows (Dawkins, 3).\n\n**Two premises that neither text states, proposed by a pass run without storing anything**\n(`scripts/propose_cross_source_premises.py`; its output is kept beside this document):\n\n- *What to presume under uncertainty.* The claim's conclusion needs: when it is unknown whether a\n  system has experience, take the possibility seriously. The reply's needs the opposite default:\n  presume it has none until a high bar is met.\n- *Whether access is evidence of experience.* The claim's line of thought needs it to be; the\n  reply's needs it not to be.\n\nThese are the two places section 4 named as the floor, and the first is the burden-of-proof\nquestion that section listed under *what would surprise*. **The second is not unstated in the\nsource.** The claim's text says it in plain words; the pass reads stored claims, the sentence was\nnever stored, and so the pass that looks for what texts leave unsaid found something one of them\nsays. The pass is not part of the path a visitor's text takes. It is run by hand on a named pair.\n\n### 5e. The eight predictions, scored\n\nThree right, three half right, two wrong.\n\n| | Prediction | Verdict | What the graph holds |\n|---|---|---|---|\n| **1** | Few attack links between the texts, none about experience, one clash on recurrence | **Wrong on the count and on the topic. Right on the clash.** | 31 attack links cross the two texts. At least six concern experience. The recurrence clash was written both ways. |\n| **2** | The reply's main premise meets nothing on the other side | **Wrong, for a reason not foreseen.** | The premise is not on the map, and neither is the sentence in the claim that would have met it (5f, first item). |\n| **3** | The reply's summary of the claim is stronger than the claim, and nothing flags it | **First half right, second half wrong.** | The renderings are stored and were compared with the claim's own words (5f, second item). Each attack link and its reason is on a claim page; nothing gathers them into *this rendering goes further than its source*. |\n| **4** | Dawkins's conclusion appears only as reported, asserted by nobody | **Right in substance, wrong in form.** | No text by Dawkins is in the graph. His conclusion is stored as four ordinary claims of the reply's text under his name, not as reported positions. |\n| **5** | The argument between the texts is carried by analogies | **Half right.** | In the reply, yes: five of its seven analogy links support one claim, its conclusion. Analogy is 8.0% of the reply's labelled links, 4.8% of the claim's, 4.0% across the graph. Links between the texts carry no pattern label, so the share between them cannot be read. |\n| **6** | No horizon kind can be read | **Right.** | All five claims asked read *unexamined*. |\n| **7** | No pass names the premise both texts share | **Right, on one call.** | 5g. |\n| **8** | Almost no links to the rest of the graph | **Right.** | 10 argument links to six other sources, none about consciousness, against 500 between the two texts. |\n\n**Why the first prediction failed.** It was made from what each author asserts. The reply's text\nalso holds other people's positions (the researchers' as the reply renders them, Dawkins's, and\none that nobody is said to hold), and the linker set the claim's sentences against those. The 31\nattack links by where they land:\n\n| From | Onto | Links |\n|---|---|---|\n| The reply's own arguments | Sentences of the claim | 13 |\n| The claim | The reply's renderings of the claim (seven stored under the researchers' name, three under the reply author's as his statement of what they found) | 10 |\n| The claim | Dawkins's argument, as the reply reports it | 4 |\n| The claim | A position the reply says nobody at Anthropic holds | 2 |\n| The claim | Something the reply's author argues himself | 1 |\n| A rendering in the reply | A sentence of the claim | 1 |\n\nSo **the map has four voices, not two sides**, and only one link runs from the claim onto the\nreply's own argument: the claim's remark that recurrence may be unnecessary, set against the\nreply's conclusion. The reply's thirteen land on the claim's extrapolations and definitions (the\nstructure is not peculiar to brains; it is a general solution; recurrence may be unnecessary; access\ndefined by function) and on none of the results of its experiments.\n\n**What section 4 said would surprise.** A pass naming the shared premise: no (5g). A link landing\non the step from access to experience and saying who carries the burden: not a link, but the pair\npass proposed exactly that premise (5d). The reply's hurricane sentence split into the analogy and\nthe conclusion it serves: yes, as two claims joined by a link labelled analogy.\n\n### 5f. What nobody predicted\n\n**Each text says in one sentence what its position depends on, and neither sentence is on the\nmap.**\n\n- The reply says what its whole second half rests on: that conscious AI is possible only on the\n  assumption that consciousness is a matter of computation, and that the computations behind it in\n  us would do the same work in silicon. No stored claim says this (checked by searching the reply's\n  97 claims for its terms and by reading the names of its sixteen arguments). What follows from it\n  was stored: brains are not computers made of meat; in a brain what it does cannot be cut loose\n  from what it is; the computer is a metaphor. The linker set those against the claim's sentences\n  about a general computational solution.\n- The claim's text says that whether access of this kind implies experience is a contested\n  question among philosophers. No stored claim says that either (checked by reading the 24 stored\n  claims that mention access or philosophers).\n\nSo the conditions that make everything else bear on the question were dropped at extraction, on\nboth sides, and nothing noticed: the self-check compares the written claims with the text and never\nthe text with the claims. The corpus had this as a general statement (coverage is the dimension\nwith no instrument at all, `reading-the-text-itself.md`). Run 3F found the crux of a debate left\nunsaid by at least one side. Here it was stated in plain words on both sides and\nlost in the reading of each.\n\n**The claim's own caveats were set against the reply's rendering of the claim.** The reply says the\nresearchers found signatures of conscious processing. The claim says its experiments do not show\nthat the model has experiences. The linker joined these with an attack link whose stored reason\nsays that the researchers' own denial undercuts the characterisation. Ten of the 31 attack links\nare of this kind. The sameness check wrote something like it in its own notes: of its 440 stored\nnotes on pairs that join a sentence of the claim to one of the reply's renderings, 169 call one\nside broader or more general, or the other narrower (a count by a list of words, which\ndoes not say which side; in the two read in full, the reply's sentence is the interpretive one and\nthe claim's the narrower finding). This is what\nstoring reported positions as claims of their own was decided for in August: one side's rendering\nbecomes comparable with what that side wrote (`deliberus/reported_speech.py`, its opening note). A\nsearch of the corpus found no earlier case of the linker doing it; the one earlier comparison of\nrenderings was by hand and found them accurate (`election-2026-bloc-debate-cluster.md`). Read\ndepth: a search and a grep, not a reading of every run. **Whether each of the ten is right has not\nbeen judged by a person.** Signatures of conscious *processing* and *experiences* are not the same\nthing, and the reply may be read as claiming only the first.\n\n**Both texts are on the same side against the argument from conversation.** Four attack links run\nfrom the claim onto Dawkins's argument as the reply reports it. A map drawn as *for* and *against*\nwould have put the claim and that argument together.\n\n**The central word was read well in each text and never joined across them.** *Consciousness* in\nthe reply and *conscious / consciousness / conscious access* in the claim are two entries in the\nlist of contested words, because an entry is identified by the term as the model happened to write\nit. The same holds for *workspace* and *global workspace / mental workspace*. The layer whose job\nis to say *these two texts mean different things by one word* holds them as two words.\n\n**Five hundred links, most of them one relation said many times.** Four of the reply's renderings\nof the finding carry 281 of the 500 links. Two of the four are the same sentence extracted twice,\nwith 85 links each. The reply renders the finding in a sentence or two; the claim states it in\nmore than a hundred; each pairing came back as *the specific finding supports, qualifies or is\npart of the general rendering*. Of the 145 part-of links, 98 were written on pairs where the sameness\ncheck, asked first, had answered that neither claim contains the other. The linker's own\ninstruction asks whether one claim is a premise the other needs and files a yes as part-of. The\nsame answer was measured in September (76% of part-of links in the graph were not parts) and the 29\nthe linker had written were retired on 2026-09-27; the instruction that writes them was left as it\nwas. All 145 are unconfirmed, so by the rule for argued requirements they cap nothing, and the\nlinker writes no pattern label, so no badge reads them. They are on the claim pages as parts.\n\n**⚠ This paragraph had half the cause (corrected 2026-10-07).** The menu was one fault. The other\nwas that the link ran the wrong way: the premise was stored as the whole and the conclusion as its\npart. That is why September's census found none of the linker's 84 links to be a part, and why its\n76% should be read as 59% for links written the right way round. Read again, 137 of the 145 pairs\nhold a relation; they were premise links, mislabelled and reversed ([auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c).\n\n**The sameness check agreed with itself 41 times in 53.** Fifty-three pairs were judged twice, once\nfrom each text's run, with the two claims in the opposite order. Twelve verdicts differed, eleven\nof them between *one contains the other* and *they overlap*. One measurement, with the order of the\nclaims and plain repetition not separated.\n\n**The self-check says four claims lost a hedge.** The claim's text says the structure *appears to*\nsupport the functions of conscious access, and that sentence is stored as a hypothesis. Four claims\ncut from the rest of the same sentence are stored as observed. Read against the text this is\narguable, since the post states the same four things as findings elsewhere. It is recorded because\nthe hedge on exactly this sentence is what the exchange is about, and because nothing reads these\nnotes either way.\n\n### 5g. The premise both texts share\n\n`scripts/probe_shared_premises.py` put one question to one model (`claude-sonnet-5`), with the 60\nbest-connected claims of each text: what do both sides take for granted that a person outside the\ndebate could reject? It named four, and who would reject each:\n\n1. That whether such a system is conscious is a real question that inquiry can make progress on\n   (rejected by those who hold that talk of consciousness is to be explained away).\n2. That theories built from human and animal minds are the right yardstick for a machine (rejected\n   by those who hold that a mind cannot be judged apart from a living body).\n3. That the measured structure is real and not a product of the method (rejected by sceptics of\n   this kind of measurement).\n4. That fluent conversation is not proof (rejected by Dawkins, as the reply reports him).\n\n**The premise a public post disputed the same evening is not among them**: that being conscious is\nsomething one bounded system has or lacks\n([a-cognitive-ecology-and-the-terms-of-the-relation.md](a-cognitive-ecology-and-the-terms-of-the-relation.md)\n§ 6). Each of the four is a position some school inside the field already holds. The one that came\nfrom outside the field did not come up.\n\n**What this does and does not show.** One call by one model, of the family the texts are about, on\npart of each text. It shows what was not volunteered. It does not show the premise is beyond a\nmachine's reach: a second call, another model, or a question that asked for ten might name it. The\nfull answer is kept beside this document (`mapping-the-machine-consciousness-debate/`).\n\n### 5h. A false attribution, found on the live site and repaired\n\nThe reply says the researchers did **not** claim that the model is conscious in the way people are.\nThe step that separates *who said it* from *what was said* stored the position (*the model is\nconscious in the same way people are*) under the researchers' name. For about an hour (23:45 to\n00:43) a public page said a named research group held a claim the source says they did not make.\n\nThe step had no way to say *the source reports this party as not holding the position*. It asks\nthat now, before it names the party, and a position the party is said not to hold is stored as\nnobody's, with the party noted beside it (`deliberus/reported_speech.py`, `not_held_by`). All 84\nreported positions in the graph were read by hand: this was the only one. The new question was then\nput to the 21 stored reports whose reporting sentence carries a *not* (the hard cases, since a\n*not* inside the position is no denial): it answered 21 of 21 as a reader would, and the one claim\nwas corrected through the step itself (`scripts/recheck_report_attribution.py`). The earlier name\nis kept on the claim. The live page shows no attribution until the next deploy, and then says whose\nposition it is not.\n\n**One choice in this repair is Claude's and unruled** (ledger AB10): a position a source says\nnobody holds is still stored. The other course is not to store it at all. It was kept because it\nis in play (the whole reply argues against it, and two of the claim's sentences were set against\nit), and because removing a claim that has links is the larger change.\n\n### 5i. What it cost\n\n| | Model calls | At API prices | Clock |\n|---|---|---|---|\n| First attempt, which stored nothing | 52 | $7.26 | 25 minutes |\n| The two texts | 2,038 | $96.34 | 94 minutes, side by side |\n| The re-reading of 22 reports and the two probes | 24 | $1.24 | 12 minutes |\n| **All of it** | **2,114** | **$104.84** | |\n\nThe estimate given before the run was $25 to $40. It was wrong by a factor of three, and the\nwhole difference is the linker: 1,735 of the 2,038 calls and $70 of the $96 went to judging pairs\nof claims across the two texts, one pair at a time. The run is on a plan, so no money was paid; the\nplan's weekly allowance moved from 62% to about 70% over the evening, this session's own share\nincluded.\n\n**The cost and the thicket are one defect.** A text that renders another text's finding in two\nsentences resembles every sentence of the finding, each resemblance becomes a pair, each pair is\nread twice by a model, and most answers are the same answer. Doubled claims double it again. For\na hundred specific sentences about one finding, the nearest neighbours in the other text are the\nsame few sentences that render the finding: **404 of the 511 pairs judged between the two texts\n(79%) pair a sentence of the claim with one of the reply's twenty renderings of the claim.**\n\n**⚠ Recounted 2026-10-07 from the record of model calls, which holds every pair that was read.**\nThe 511 was counted from the graph, so it held only pairs that left something there and missed\nthose read and found unrelated. The sameness check read **646 pairs** between the two texts' own\nclaims: 360 came back as overlapping, 189 as unrelated, 64 as one containing the other, 16 as\nopposed, 6 as undecidable, and 11 calls failed. **416 of the 646 (64%) pair a sentence of the claim\nwith one of the twenty renderings.** The thicket is as described; its share of the reading is two\nthirds, not four fifths.\n\n**One saving was available and not used.** A reading that puts both questions about a pair in one\ncall exists behind a switch (`DELIBERUS_MERGED_PAIR`), off while its agreement with the two-call\npath is being measured. Here 618 of 1,117 pairs went on to a second call, so it would have made\nabout 1,117 calls where the run made 1,735.\n\n**⚠ Measured 2026-10-07: that saving is in calls and not in money.** A call's price follows its\nlength. The one-call reading cost 6.4 cents a pair; the two-call path cost 6.3 that night, since 499\nof the 1,117 pairs stopped after the first call at 4.0 cents. On 54 pairs it also agreed with the\ngraph less than the step agrees with itself (39 against 46). Three other cheaper readers were\nmeasured the same day and none keeps what pair-by-pair reading finds; what was built is a count and\na ceiling before the step reads ([auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c, *What the\nstep costs*; scripts and replies in\n[linking-cost/](mapping-the-machine-consciousness-debate/linking-cost/what-the-runs-printed.txt)).\n\n### 5j. What follows\n\nProposed, none ruled (ledger AB10 to AB15):\n\n1. **Repair the linker before mapping anything else through it.** Take the premise question off\n   its menu, or file a yes as support. Join claims said twice inside one source before linking.\n   Compare a rendering with the text it renders in one reading per rendering, not one pair at a\n   time (twenty readings where the run read 416 pairs, by the recount in § 5i). **Not proposed, because it was measured and\n   fails: keeping a pair only when each claim is among the other's nearest few.** Only 53 of the\n   511 pairs that left something in the graph were ranked by both sides, and only 2 of the 31 attack links sit on those pairs, so\n   that rule would have kept the thicket's core and thrown away the clashes. This agrees with the\n   ledger's earlier recommendation on the same question (T4: judge a pair when either side ranks\n   it). Then re-read the 145\n   part-of links with the instrument that exists (`scripts/propose_decomposition_kinds.py`, then\n   `scripts/retire_miscast_parts.py`) and retire those that fail, as was done in September. That\n   last step is housekeeping already ruled; it was left undone that night to spare the plan's\n   weekly allowance, about 145 calls.\n   **Ruled and partly done on 2026-10-07** (founder: *\"Repair it sure!\"*). Done: the linker can no\n   longer write a part-of link, a premise is filed as support from the premise, and all 231 such\n   links in the graph (these 145 and 86 older ones) were retired and read again, by the repaired\n   linker and not by the instrument named above, since the fault turned out to be the direction\n   they were stored in ([auto-connect-upgrade.md](auto-connect-upgrade.md) § 2c). Not done: joining\n   claims said twice, and one reading per rendering; both are the repair of the cost and are not\n   built.\n2. **Check the text against the claims.** After extraction, ask what the author says the conclusion\n   depends on, and whether each such sentence is among the claims. Propose what is missing. Tonight\n   it would have caught two sentences in two texts.\n3. **Identify a contested word by the word.** Two texts' entries for *consciousness* should be one\n   entry with the senses of both.\n4. **Store who wrote a source**, so that an author's own statements are not stored a second time as\n   reported positions.\n5. **Show a rendering beside what it renders.** The comparison was made tonight. Its results sit\n   one link at a time on ten claim pages, and the sameness check's notes on the same pairs are in a\n   field no page prints.\n6. **Put the two unstated premises on the map** as claims neither text asserts, with the question\n   *do you?* beside them. The same place would hold a reframe (the reading, § 6).\n\nThe second round named in section 2 (the conduct question) waits on the founder's word and should\nwait on the first item above.\n"}