{"path":"research/checks-on-machines-and-people.md","content":"# Checks on machines and people: who supplies, who checks, and how much has been tried\n\n**Date**: 2026-10-12 · **Type**: advice the founder asked for, read from the corpus and checked against\nthe code and the live graph · **Status**: nothing here is ruled. No ledger row is added: every proposal\nbelow is an existing open row seen from his message, and § 9 says which.\n\n**⚠ Revised the same evening (§ 11).** He leaned yes on the one decision put to him and asked what it\nentails. Both machine readings were then replayed over everything people have added to the live\ngraph. The result changes the advice for parts and corrects three sentences below, each marked where\nit stands: the way back for a contributor who is misread is dearer than § 4.5 said; the check on parts\nis the part test, which § 3 and § 4.2 described loosely; and withholding a part changes less in the\nstrength code than § 4.2 said.\n\n**His message, verbatim:**\n\n> *\"If we temporarily assume there'd no spamming/sabotage either way (trying to strengthen or weaken\n> claims they like or dislike), then... well, people will simply contribute reasons for and against\n> claims and they will be relevant and they will be evaluated on their merits/evidence, and the\n> relevance and strength of different supporting or attacking reasons will be continually judged\n> (provisionally by machine readings - which are inevitably biased by its semantic embedding dimensions\n> of its LLM brains - and continually reviewed and blessed or tweaked/corrected by humans). The\n> provisional machine relevance evaluation (and all the other machine passes/microtasks that bring\n> order to Deliberus' graph) that would occur before a human is in the loop and can speak their mind,\n> is what prevents spamming/sabotage/disorder/entropy, and then ofc we need to be aware that the\n> humans-in-the-loop that later/at any time review/correct the relevance judgment and other machine\n> judgments may still be biased/mischievous and as such we need all kinds of checks and balances,\n> whatever we can come up with that is wisest and most robust.*\n>\n> *Idk how \"arguing the search was thorough\" fits into all of this, but yeah we should have some\n> awareness of whether all or only some or none of all possible relevant arguments that could be\n> brought to bear on a claim have been tried and tested and found lacking and how we judge\n> all/some/none in terms of relevance.*\n>\n> *Quality control and checks and balances are needed all over the place, regardless of whether humans\n> or machines are doing stuff, since we may at any point be dealing with a wide spectrum of\n> well-meaning but unintelligent humans/AIs to malevolent/destructively-prompted and extremely capable\n> humans or AIs. Of course we have much more control over the machine passes/microtasks that we design\n> and run as part of Deliberus.*\n>\n> *Read up on everything remotely relevant to all this and advise.\"*\n\n---\n\n## 0. In short\n\n1. **His picture is the ruled design.** Each sentence of it has a ruling behind it, dated between\n   August and October (§ 1). He asked nearly this question on 3 September, and the answer written then,\n   *cheap to add, slow to matter*, was never put to him for a ruling.\n2. **The code is behind the picture in one place that matters for sabotage.** The machine's first\n   reading does not run on anything a person adds. A reason a person adds is presumed half relevant,\n   and a part a person adds is presumed half needed, unread. The layer he describes as keeping\n   sabotage out guards only the links a machine draws (§ 2).\n3. **What a machine's reading can be trusted with has been measured, and the line is sharp.** *Do these\n   two claims have anything to do with each other* got the same answer from four models on every pair.\n   *How strongly does this bear* got the same answer on 4 to 6 of 12. *Is this a reason for that*\n   changed side on 18 of 30 when the question was worded another way (§ 3).\n4. **Advice one: a machine's reading may take a default effect away and may not add effect.** The\n   corpus already says this for three separate passes and has not said it once for all. It tells which\n   part of the ruled relevance judge can go live now (its yes or no) and which waits (its degree) (§ 4).\n5. **Advice two: read what people add with the two checks that already exist.** ⚠ *Revised by § 11:\n   read it and show the reading, and let neither reading change a weight yet.* Replayed over what\n   people have added so far, the check on links would withhold 2 of 32, both honest objections filed\n   at the wrong address, and the check on parts would withhold 28 of 39, because most of what people\n   have filed as parts is what a claim rests on and not what it says (§ 4.2, § 11).\n6. **Advice three: the person who reviews is under the same rule as the machine.** Nobody's bare\n   judgement settles anything about the shared world, which he ruled on 31 August. A correction is a\n   contribution with a name and a reason, the machine reads it like any other, and anyone can attack\n   it. No reviewer is last (§ 5).\n7. **Advice four: awareness of how much has been tried is a count, and the list it counts against must\n   not be chosen by whoever gains from it.** Three lists, of which code can count one today: 2,084\n   standing questions hang on the reasons in the live graph, and one has an answer anybody has backed.\n   *The search was thorough* is the claim that turns *every listed reason failed* into *probably\n   false*. The count is its evidence, and naming an untried argument is how anyone reopens it (§ 6).\n8. **What stays undefended is said in § 8**: several accounts acting as one, and a first reader that\n   is one witness however many models agree.\n\n---\n\n## 1. His picture, sentence by sentence, against what is ruled\n\n| What he said | Ruled, and when | Built |\n|---|---|---|\n| People contribute reasons for and against, judged on their merits | A link's effect is the reason's own standing times its relevance (28 September, *\"your proposal is blessed, make it so\"*) | live since 30 September |\n| Relevance is judged first by a machine and then by people | The same ruling: proposed by a machine of a second model family, argued by people, never voted. His direction of 19 September: *\"judged provisionally by the machine/LLM-passes or whatever and then humans can challenge it\"* | the judge is built and has written nothing: it waits on his word |\n| The machine passes that run before any person looks are what keep disorder out | *\"Any regular housekeeping should be done that could be done with high confidence in safety/accuracy\"* (13 September) | sameness before any relation, splitting, staleness flags; none measured against a person |\n| The people who review may be biased or mischievous | *\"Human force enters the graph only through contributions that argue for themselves; bare judgments are telemetry\"* (31 August, *\"Agreed on everything\"*) | there is no review act, with one exception (§ 5.2) |\n| Awareness of whether all, some or none of the possible arguments have been tried | A failed reason drops out; a third display state, *nothing stands (N checked)*; a thorough search argued as its own claim (30 September, *\"for now\"*) | none of it built |\n| Quality control for whoever does the work, machine or person | A person reads a pass before anyone proves it (6 October, *\"Sure\"*); an error rate belongs to the task as written (29 September); machine work on machine work is capped by measured error (29 September, *\"Yes.\"*) | five habits of asking and marking run; no person has read anything (§ 2.4) |\n\n**One document already answers the whole message and has never been ruled.** On 3 September he asked:\n*\"So what's a conservative approach that still uses humans to fill up the tail faithfully but has a\nhealthy measure of distrust against both LLMs and humans?\"* The answer\n([cheap-to-add-slow-to-matter.md](cheap-to-add-slow-to-matter.md)) is a table of who supplies what, who\nchecks it, and what it may do before the check, with the rule that every checker is someone other than\nthe supplier. Two days later his *\"None at all on entry?\"* produced six corrections to it. § 7 below is\nthat table brought up to date. What today's message adds to it is the middle sentence: that the\nmachine's reading, made before any person arrives, is itself the guard.\n\n## 2. Where the code is behind the picture\n\nEach of these was opened in the code or counted on the live graph on 2026-10-12.\n\n**2.1 A person's contribution gets no machine reading of its relevance.** When a person adds evidence\nto a claim, the system stores it, finds its place among similar claims, names the kind of argument and\nlists that kind's standing questions. Nothing reads whether it bears on the claim. Its link is presumed\nhalf relevant because a person wrote it. When a person adds parts to a claim, nothing compares them\nwith what the claim says, and each is presumed half needed. That second presumption has a price worked\nout on 19 September: **one irrelevant part caps any claim at 0.75, and at 0.5 once the part is\nrefuted**, for the cost of one contribution, while the defender has to write an argument (ledger O13,\nopen; [weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)\n§ 9e).\n\n**2.2 The links a machine draws are the guarded ones.** A pair the linking step considers is put\nfirst to the sameness check, which stops on *unrelated*, then asked outright whether the two claims are\nrelated before any kind of link is named, and a link that is drawn presumes no relevance until somebody\nargues it. So the layer he describes exists, and it stands in\nfront of the machine's own guesses. It is absent in front of people.\n\n**2.3 The relevance judge has read twelve links and written none.** Its `--write` waits on him. As\nbuilt it also passes over every link the machine drew, because it selects links that carry a kind of\nargument, and those carry none\n([quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md) § 1.3).\n\n**2.4 No machine pass has been read by a person.** Counted again today: 0 answers in the sameness\nsitting, 0 in the review of machine-cut parts, 0 notes on any machine verdict, and 0 of 111 machine-cut\npart links kept, reworded or dropped by anyone. Quality control today is five habits: ask small and\nclosed, let something other than the writer check, hold what fails and mark what lands, record every\ncall, and name the model that answered. They make a wrong answer rarer, visible and undoable. They do\nnot say how often an answer is wrong.\n\n**2.5 Nothing counts how much has been tried** (§ 6 has the count made today).\n\n**2.6 Nobody with a stake has ever used it.** Every source in the graph was put there by the pipeline\nor by its keeper. No check here has met a person who wanted an outcome, and none has met one who\nstudies the checks ([fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md)\n§ Three classes of adversary).\n\n## 3. What a machine's first reading can be trusted with\n\nThree kinds of question have been put to machine readers in this project, and they behaved differently.\n\n| The question | What was measured | Where |\n|---|---|---|\n| **Do these two claims have anything to do with each other?** | Twenty pairs known to be unrelated: every one of four models said unrelated on all twenty. On a hundred random pairs across sources a decision model said *not related* with confidence on 83 and was never confidently wrong. | [whose-tokens-pay-for-mapping.md](whose-tokens-pay-for-mapping.md) § 3e · [jev-and-the-judgement-layer.md](jev-and-the-judgement-layer.md) § 3f |\n| **If this part were false, would the claim be weaker?** (the part test) | Asked of each numbered part of a machine's cut: the same answer twice on 45 of 45 parts, and all six harmful parts caught. ⚠ *This row first worded the question as \"does the part say something the claim says\". That is a label the same call gives, and only the part test gates. On the 39 parts people attached the two agreed in every case (§ 11.2).* | [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) § 3e |\n| **How strongly does this bear?** | A five-step scale on twelve links. One model agreed with its own earlier run on 10. Three others agreed with it on 4 to 6, and one of them answered *strong* to all twelve. | [whose-tokens-pay-for-mapping.md](whose-tokens-pay-for-mapping.md) § 3e |\n| **Is this a reason for that?** (of two related claims from different texts) | Thirty links, one model: 4 refused under one wording and 20 under another, eighteen changing side. | [what-a-link-rests-on.md](what-a-link-rests-on.md) § 2 |\n\nSo a machine reads reliably that two things are unrelated and that a sentence is not in a text. It does\nnot yet read degree, and it does not yet read whether a related claim is a reason. **None of these\nfigures is against a person's label**: they say that the first two questions get one answer whoever\nreads them, which is a condition for a measurement and not yet one.\n\n**Added 2026-10-13, on his two questions** (*\"How strongly what? Explain plainly\"*, and then\n*\"Isn't that the same thing as relevance?\"*)**.** The third row's question is how much a reason bears on a claim if the reason is\ntrue, which is the degree of what the corpus calls relevance, and the first row's question is the\nsame factor's yes or no. The five steps the models were given mix three different numbers, and a\nlink with nothing written under it gives a reader no one thing to rate. So that row measures a\nbadly put question as much as it measures the readers:\n[the-chain-inside-a-link.md](the-chain-inside-a-link.md) § 10.\n\n**On his parenthesis**, that machine readings are *\"inevitably biased by its semantic embedding\ndimensions\"*. Two things from the record. The readings above are a model reading two sentences. The\nvectors only find the candidate pairs, and no strength is read off them, which is the standing rule\n(*use for discovery, never as authority on meaning*). And his point holds one level up: models of\ndifferent makers share much of what they get wrong and much of what they cannot see, so two models\nagreeing is fewer than two witnesses\n([why-representations-converge-and-whether-minds-do.md](why-representations-converge-and-whether-minds-do.md)).\nThe check on a machine's reading that is independent of it is a person who stands somewhere, and the\ncheck on that person is another person who stands somewhere else (§ 5).\n\n## 4. Advice one: a machine's reading may withhold, and may not add\n\n**4.1 The rule, which the corpus holds in three pieces.** The decision model beside the relevance\njudge: *\"can only withhold, never add\"*\n([jev-and-the-judgement-layer.md](jev-and-the-judgement-layer.md) § 6c). The machine that asks questions\non behalf of a position not yet heard: *\"cap-only, never lift\"*\n([derivability-of-missing-considerations.md](derivability-of-missing-considerations.md) § 8). The\nmachine row of the September table: *\"never lift\"*. And the code for a part's presumed need says of a\nmachine's reading that it *\"SCALES the presumption and can never exceed it\"*. Each was written for one\npass. Said once:\n\n> **On its own, a machine's judgement may take a presumed effect away. It may not raise one above what\n> is presumed.**\n\nWhat is presumed follows who stands behind a thing, as the code already has it: something a source\nsaid, a person wrote, or a cut found in a claim's own words counts modestly until argued, and a\nmachine's guess counts for nothing until argued. Raising an effect beyond that takes a named author\nwith a reason that can be attacked, or a pass that people have read and measured (§ 4.2, item 3).\n\nWhy this is the right line and not caution for its own sake. Taking a default away costs an honest\ncontributor an argument (⚠ more than one sentence, as the code stands: § 11.4), and the reading that\ndid it is shown with its reason. Giving effect by\nmachine puts weight in the graph that nobody stands behind, written by a reader whose answer on degree\nchanges with the model (§ 3). It also mirrors the ruling of 30 September from the other side: there, a\nfailure counts against a claim only through a contribution that says why. Here, support counts for a\nclaim beyond the default only through one. A gain in knowing the rule buys an attacker nothing, which\nis the test the design law asks for: steering the reader into saying *this bears* earns exactly the\nmodest default every contribution has today.\n\n**4.2 What it gives, concretely, in the order of what is surest.**\n\n1. **A part a person adds is put to the part test.** ⚠ *Not as a switch: § 11.* The check exists: it\n   asks of each part whether the claim would be weaker if the part were false, and labels how the\n   claim holds it (asserts, denies, concedes, supposes, or does not say it). A part that fails the\n   test presumes no need, by an override that is already in the code for the machine's own cuts and\n   that outranks who wrote the part. This is this morning's finding put to work\n   ([a-part-and-a-reason.md](a-part-and-a-reason.md)): a part is something the claim says, and **the\n   saboteur's bogus part is a demand filed as a part**. Filed as what it is, it is a requirement that\n   caps nothing until argued, which is what the ruling of 30 September already gives a reason. The cap\n   of 0.75 for the price of one contribution closes. ⚠ *\"Without any new arithmetic\", as this\n   sentence first ended, was wrong: a part that requires nothing is left out of the cap and not out\n   of the other channel by which parts act, so a refuted one still pulls its claim below the middle\n   (§ 11.4).*\n2. **A link a person adds is asked the sharp question.** A confident *these have nothing to do with\n   each other* lowers the link's presumed relevance from one half to nothing, until somebody argues for\n   it (ledger W5, open). An uncertain answer changes nothing and puts the link on the list of things\n   that want a person's eye.\n3. **The degree a machine reads is stored and shown, and weighs nothing yet.** Its way in has two\n   steps, in the order ruled on 6 October. First the same link, asked a second way and by a second\n   maker's model, has to get the same step, since a reading that changes with the reader is not worth\n   a person's hour. Then a person reads a sample of it. After that it may move a default in either\n   direction, as the ruling of 28 September has it. That turns the measurement of 7 October from a\n   prohibition into an entry test.\n\n**4.3 What this does to two rows that cannot both stand as worded.** The ruling of 28 September has a\nmachine propose a link's relevance. A proposal of 7 October (ledger AD7, with AD17) says no machine\nreading may weigh a link. Read by § 3 they divide cleanly: **the ruling governs the yes or no, which is\nsteady, and the proposal's finding governs the degree, which is not, until it is.** The ruling stays the\ndestination. Nothing in it has to be taken back.\n\n**4.4 What the sharp question does not catch.** A reason that is on the subject and beside the point\npasses it, and so does a plausible demand. For those the defences are the ones already ruled or built:\nthe default is modest, a need has to be argued and its argument can be attacked, and a requirement any\nreader could have asked for is simply a fair question\n([derivability-of-missing-considerations.md](derivability-of-missing-considerations.md) § 4). The\nsharp question stops noise and crude sabotage. It does not stop a careful opponent, and § 8 says what\ndoes not either.\n\n**4.5 The risk this advice carries, and the test it owes.** The contributions a machine is likeliest\nto call unrelated in error are the ones this project most wants: what a thing costs from where somebody\nstands, said in words no source in the graph uses. The figures of § 3 were measured on random pairs\nand on the machine's own cuts. None was measured on an unusual contribution that does bear. So three\nconditions go with steps 1 and 2:\n\n- **Withheld is never hidden.** The contribution stays where it was put, with the reader's reason\n  beside it and one move: *say what connects these*. ⚠ *This bullet first said the author's sentence\n  then restores the effect, at the cost of that one sentence. As the code stands it does not: an\n  argued connection counts only once it has been examined, so a sentence nobody has examined restores\n  nothing. For one sentence to be enough, the reader has to read the pair again with that sentence\n  beside it, and that is not built (§ 11.4).*\n- **Only a confident no withholds.** An unsure reading changes nothing.\n- **A test in the other direction comes first**: a few dozen pairs that are unusual and do bear, taken\n  from real material: testimony from the people a debate is about, of the kind the September test\n  named in four places as beyond any model's reach. Then the count of how many the reader calls\n  unrelated.\n  A guard that is only ever tested on what it should stop says nothing about what it stops by\n  mistake.\n\n## 5. Advice two: the person who reviews is under the same rule\n\n**5.1 What is already ruled is most of the answer.** There is no act by which a person confirms or\noverrules a machine's reading of the shared world. Agreeing is recorded as having looked. Disagreeing\nhas force only as a contribution: a relevance claim on the link, a requirement claim on a part, an\nobjection, a different reading with its reason. Each carries a name, each is a claim, and each is read\nby the machine and by other people like any other claim. So his worry about a biased or mischievous\nreviewer is met by there being no reviewer's seat to sit in. **The loop ends in the record, never in a\nperson.** The only things one person settles alone are their own: their stance, and whether a rendering\nof their position is one they accept\n([the-ratification-record-and-the-triage-feed.md](the-ratification-record-and-the-triage-feed.md) § 4b).\n\n**5.2 One place where a person's touch still acts without a reason.** On a claim page any signed-in\nreader can drop a part the machine cut, which takes it out of the claim's strength on one click, or save\na new wording over it. A note can be given and none is asked for. The dropped link stays visible; the\nfirst wording does not, and the page has no undo. That was accepted for the present phase, in which the\nreviewers are the keeper and people he recruits. The rule for a wider public is the one he gave the\nmachine on 13 September, applied to people: **replace, keep the old thing visible, and say who and\nwhy**, after which anyone may contest the replacement. A person may then do to machine structure\nexactly what a machine may do to it, and neither removes without a trace\n([split-review-and-the-founders-intentions.md](split-review-and-the-founders-intentions.md) § 4b sets\neach control on that page against each ruling).\n\n**5.3 The checks on people, gathered.** In order of how much they have been thought through:\n\n- **No bare verdicts** (ruled). Force only through a contribution that argues.\n- **Whoever checks is someone else.** The machine reads what a person adds; a person reads what the\n  machine wrote; a second person, placed differently, reads the first. The same-hand rule holds for\n  people as for models: a handful of like-minded reviewers agreeing measures their likeness\n  ([earning-the-right-to-write.md](earning-the-right-to-write.md) § 3a).\n- **One contributor alone does not change what a reader sees.** A reading crosses from one display\n  band to the next only when two independent origins stand behind the move (proposed 3 September,\n  unruled; it guards against one mistake or one mischief, and by the corpus's own correction it is\n  no defence against one actor with many accounts).\n- **What is said from a position is checked by others in that position.** *That is not what it is\n  like* is the one check no model and no outsider can make.\n- **A reviewer is measured the way a pass is.** Blind items, planted wrong ones that would favour the\n  reviewer's own side, two reviewers on the same items. In a sitting, never in the live graph (he\n  dropped live planting on 31 August).\n- **Disagreement between reviewers is kept and shown.** It marks where nobody has said what a\n  judgement is for, and it goes on the list for a person's eye.\n- **An account and a rate at the door**, which judge nobody's merits and protect the attention of\n  everyone behind the door.\n\n## 6. Advice three: how much has been tried\n\n### 6.1 Three lists, and only one of them can be counted today\n\n| | The list | Who makes it | Can it be counted | State |\n|---|---|---|---|---|\n| 1 | **The standing questions on each attached reason.** Every kind of argument has its fixed questions: for an expert's word, is this their field; for a cause, is there another. | code, from the kind of argument | yes, exactly | live as potentials; the count below is new |\n| 2 | **What would bear on the claim itself.** What would have to be true, what would count against, which known framings disagree. | a machine pass, the same for every claim | yes, once it exists | designed in September, unbuilt |\n| 3 | **What nobody has thought to list.** Mostly what a thing costs from where somebody stands. | people, from positions | no | cannot be listed by construction |\n\n### 6.2 The count, made today\n\nFrom the live graph, by code, with no model\n([the script](checks-on-machines-and-people/how_much_has_been_tried.py)):\n\n| | |\n|---|---|\n| Claims with a reason or an objection attached | 727 |\n| Reasons and objections on them | 1,618 |\n| … that carry a kind of argument, so a list of questions exists | 639 |\n| … that carry none, so no list exists | 979 |\n| Standing questions on the 639 | 2,084 |\n| … with an answer anybody has backed | **1** |\n| Claims where none of the listed questions has been tried | 449 |\n| Claims where some have | 1 |\n| Claims where every one has | 0 |\n| Claims whose every reason is of the kind with no list | 277 |\n\nTwo things to read off it. **Today the honest word on nearly every claim is *none*.** And **six in ten\nreasons have no list to be tried against**: the 921 links a machine drew between texts were never given\na kind of argument, and 58 others have none either. Naming the kind on those links would give each its\nquestions. As the code stands it would move no strength, since a machine-drawn link presumes no\nrelevance. It is a run of the classifier over 979 links, and the naming is the unsteady part of that\nclassifier: on 24 live links a second reading changed 4 names.\n\n### 6.3 The list must not be chosen by whoever gains from it\n\nThis is the point on which the whole awareness stands or falls. Suppose the measure were *of the\nreasons attached, how many have been checked*. An opponent attaches three weak reasons to a claim and\nrefutes them, and the page reads *nothing stands, 3 of 3 checked*. The count is true and the opponent\nwrote its denominator. So:\n\n- **List 2 is made by a machine, the same way for every claim.** Nobody then chooses which arguments\n  count as the ones to be tried. It is the proposal of 2 September, made to remove the gain from\n  picking on one claim, and it serves here unchanged: scrutiny that everyone gets pays nobody\n  ([derivability-of-missing-considerations.md](derivability-of-missing-considerations.md) § 4).\n- **Anyone may add to a list. Nobody takes an item off it except by an answer of *does not apply*,\n  given with a reason, which is itself a claim.** That is also how relevance enters his *all, some or\n  none*: roughly half of the questions a machine generates are beside the point (a 2025 shared task\n  measured it), so the count is read over the items nobody has shown not to apply.\n- **An item counts by what hangs on it.** The code already prices a question by whether a decisive\n  answer would change what a reader sees. *Tried* is best read over those.\n\n### 6.4 The words on the page\n\nShown beside the badge and never folded into its number:\n\n- **nothing tried** · **3 of 14 listed questions tried** · **every listed question tried**\n- **the kind with no list**, said in so many words where it applies\n- for the open remainder, two signals and no count: **who has been heard** (which sides of the debate\n  have a source in the graph), and **whether anything new still arrives** (of the distinct reasons so\n  far, the share that only one source gave is an estimate of the chance that the next reason is new:\n  [saturation-and-the-long-tail.md](saturation-and-the-long-tail.md) § 6)\n\n**The page never says *all possible*.** It says *all listed*. A list that looks finished cannot tell a\nsubject fully mapped from a frame that has stopped admitting what does not fit, which is the caveat\nthat note carries and that has to travel with the count.\n\n### 6.5 Where *arguing the search was thorough* fits\n\nHis question, and the answer has two halves that do different work.\n\n**The awareness is a count.** It is computed, always on, and argues nothing. It is the machine's\nstanding report.\n\n**The step from *every listed reason failed* to *this claim is probably false* is an argument**, and\nan old one: the argument from ignorance. It holds on two conditions, which are its two standing\nquestions: was the search thorough, and would a good reason have turned up if the claim were true.\nSomebody makes it as a claim. **The count is its evidence**, and it can be attacked in the one way that\nimproves the map: by naming an argument that was not tried, which then joins a list and reopens the\nquestion.\n\nThat answers this morning's challenge in his own terms. He is right that a claim whose every relevant\nargument has failed should read low. What the ruling of 30 September refused was letting it read low by\ndefault, because *every argument has failed* is exactly what an opponent can manufacture when the\nopponent chooses the arguments (§ 6.3). With a list nobody chose, the default becomes safe to revisit.\n\n### 6.6 A candidate, marked as mine and unruled\n\n**His challenge and a question open in the strength layer since 31 August are one question with two\nsigns.** The open question: three parts each established at 0.9 lift their whole only to 0.507,\nand the proposal has been that a whole should rise as the share of what it needs is established. His\nchallenge is the same thing downward: a claim should sink as the share of what could support it is\ntried and fails. Both are *how much of a list nobody chose is settled, and which way*. One measure\nwould serve both.\n\nIt should not weigh anything yet, for a reason this project has already paid for: list 2 does not\nexist, roughly half of a machine's list is beside the point, and no person has read one. Show the count now. Let weight\nhang on it after a person has read what the list-making pass writes, which is the order ruled on\n6 October.\n\n## 7. The table of 3 September, brought up to date\n\nEvery *checked by* names someone other than the supplier. Nothing waits for a check in order to exist.\n\n| Supplier | Supplies | Checked by | Before the check it may | After |\n|---|---|---|---|---|\n| **A machine pass that reads** (related at all; does the claim say this; sameness) | a yes or no with its reason | code first; a pass of another maker's model; people through use | take a presumed effect away; rank what people are shown | the same, with a proven rate |\n| **A machine pass that writes** (splits, links, the lists of § 6, steps under a link) | marked structure that can be undone | a separate check that is not the writer; a person's first reading; then a blind sample | exist, marked; a found part caps modestly; a guessed link and a made sentence weigh nothing | write unwatched to the depth its measured rate allows |\n| **A machine reading of degree** | a step on a scale, with its reason | the same link asked a second way and by a second maker's model; then a person's reading of a sample | be stored and shown | move a default up or down, as ruled on 28 September |\n| **Any person** | claims, reasons, objections, needs, answers | the machine (related at all; does the claim say this; is it already here); other people, by attacking | exist at once, with a name; have the modest default unless a machine reading withholds it | effect through what stands behind it |\n| **A person speaking from a position** | what a thing costs from where they stand; what should be asked | others who share the position | exist, marked as one witness; move one premise | weigh with independent witnesses |\n| **A person correcting a machine** | a different reading, with a reason | the machine, which reads it like any contribution; other people | supersede machine structure, the old kept visible | the same; their reversals are a count others can see |\n| **A reviewer in a sitting** | labels on a sample | a second reviewer on the same items; planted wrong items, offline | nothing in the graph | set a pass's rung |\n\nThe row that is new since September is the third from the bottom, and the first row's *take a presumed\neffect away* is new as a general permission.\n\n## 8. What this leaves undefended\n\nSaid plainly, because a list of checks reads as safer than it is.\n\n1. **Several accounts acting as one.** Every count above, and every rule about two independent origins,\n   can be met by one actor with many names. The defence is the cost of an identity and counting origins\n   rather than accounts, which is why sources are proposed to become things in the graph. Nothing built\n   addresses it (ledger D8 and D9, open;\n   [legibility-under-power-red-team.md](legibility-under-power-red-team.md), hole five).\n2. **A capable opponent reads the published rules.** So each rule is to be asked what still pays under\n   it. Under § 4 to § 6 three things pay: arguing well, answering a listed question, and bringing a\n   reason nobody listed. All three are the wanted behaviour. A rule that fails that question should not\n   ship however good its detector.\n3. **The machine's first reading is one witness.** It is the same witness for every claim, which is its\n   virtue in § 6.3 and its limit here: what it cannot see, it cannot see anywhere. People placed\n   differently are the only outside check, and reaching them is recruitment, never code.\n4. **Text written to steer a reader.** A contribution can be written to push a machine reader's answer.\n   Under § 4 the most it buys is the default. A first screen for instructions hidden in a source\n   exists and only flags.\n5. **Every check ends by asking a person something**, and the list that would ration that attention is\n   designed and unbuilt. Without it, slow to matter becomes never matters.\n6. **The honest contributor who stands alone is slow to count.** Accepted on 3 September and to be\n   shown on the page, with the move beside it: *one witness so far; be the second*.\n7. **No check here has met a person with a stake.** Whatever is built from this note is built for an\n   opponent who has not arrived, and its first failure should be read as that arrival and not as a bug.\n\n## 9. What would be built, in order, and what is his to rule\n\nNothing below is a new ledger row. The rows named are the open ones each step would settle.\n\n| | Step | Needs | Row |\n|---|---|---|---|\n| 1 | Put a part a person adds to the part test; a part that fails it presumes no need. ⚠ *Revised by § 11.6: read and show only, after his reading of the 39 parts people have attached* | his word: a machine reading changes the weight of a person's contribution | O13, T9 |\n| 2 | Ask of a link a person adds whether the two claims are related at all; a confident no presumes no relevance. ⚠ *Revised by § 11.6: the reading is stored and shown as a question; it changes no weight until a person has a place to object to a step.* Before any switch: the test of § 4.5 on unusual contributions that do bear | the same word | W5 |\n| 3 | The two repairs of 30 September (a refuted reason counts as looked at; *nothing stands, N checked*), with the count of § 6.2 beside the badge | ruled; the count line is new and small | Q34 |\n| 4 | Name the kind of argument on the 979 reasons that have none, so each gets its questions | a run of the classifier; it moves no strength | (a to-do) |\n| 5 | His first reading of a machine's work, on the pages already built (two sittings and the first reading of written-out steps) | an hour of his | T3 |\n| 6 | The pass that lists, for every claim alike, what would bear on it; shown as questions, moving nothing | after step 5 | the September proposal on the derivable list |\n| 7 | Degree: the same link asked a second way and by a second maker's model; it weighs when the answers agree | his word on the entry test | Q33 with AD7, AD14, AD17 |\n| 8 | Whether the count of § 6 may ever move a strength, up or down | after 5 and 6 | the open question on lifting, and § 6.6 |\n\n**The one that unblocks the others is the word for steps 1 and 2**: may a machine's confident *no*\ntake away the default effect of something a person added. For: it is the layer his message describes,\nit closes a measured opening, and both readings are of the kind that came back the same from every\nreader. Against: a machine reading then changes what a person's contribution does, and an honest\ncontributor who is misread has to argue once. ~~Rec: yes, with the reading and its reason shown beside\nthe contribution, and after the test of § 4.5 has been run.~~ ⚠ **Replaced the same evening by\n§ 11.6**: read and show now, change no weight yet, and for parts his reading of the 39 comes first.\n\n## 10. What was read, and how far\n\nRead whole or in the sections named, this session: the note of 3 September and its critical\nexamination; the September record of his intentions on review (§ 0, § 1c to § 1f);\nthe ratification record (§ 3 to § 5); what human judgement is for (§ 4d); the three classes of\nadversary; the red team's hole five; the incentives analysis (§ 4 and § 5); relevance and the two\nfactors (§ 0 to § 6); the derivability note (§ 1, § 3 to § 5, § 8 to § 10); drowning in claims\n(§ 5 to § 8); proving a pass reliable (§ 0, § 9); earning the right to write (§ 3, § 4); quality control\nfor deeper steps (§ 1); the decision model and the relevance judge (§ 6); the comparison of four models\n(§ 3e); the late reading of 8 October and § 14.2 of the note it belongs to; ledger rows D8, D9, O13,\nQ2, Q22, Q34 to Q36, T1, T3, T6, W5, AD7, AD14, AD17. In the code: what happens when a person adds\nevidence or parts, the presumptions for a link and for a part, what the strength layer gathers, and\nhow potential questions are listed.\n\n**Not done.** No outside literature was read for this note: every outside figure in it is one an\nearlier note recorded, at that note's own depth. No model was called. Nothing was changed in the\ncode that weighs anything. The count of § 6.2 reads questions as tried when somebody has attached\nevidence to an answer, which is the strength layer's own test, and it leaves parts out.\n\n## 11. What a yes would do, replayed on everything people have added (the same evening)\n\n**His reply to the decision of § 9**: *\"I'm leaning yes, but please think deeply about what this\nentails and any unintended side effects or whatnot?\"*\n\n**What was done.** The two readings were run over everything in the live graph that carries a person\nas its writer: 32 reasons and objections, and 39 parts under 9 claims. 41 calls, every one answered by\nClaude Sonnet 5, each kept in the record of machine judgments. Nothing was written to the graph.\n([the script](checks-on-machines-and-people/replay_on_what_people_added.py), and every answer with its\nreason in the file beside it.) One model, one run, and no person has labelled any of it.\n\n### 11.1 Reasons and objections: 2 of 32 would be withheld, and both are honest\n\n| The sameness check's answer | Filed as reasons (11) | Filed as objections (21) |\n|---|---|---|\n| cannot both be true | 0 | 13 |\n| one would partly advance the other | 8 | 4 |\n| one says all the other says, and more | 3 | 2 |\n| **no bearing either way** | 0 | **2** |\n| the same claim reworded | 0 | 0 |\n\nBoth that would be withheld are legal objections in the hand-made map of the Israel and Palestine\npair. One says an attack launched from occupied territory is continuous with resistance to the\noccupation and no fresh armed attack, and is attached to the claim that Hamas was the author of the\nattack of 7 October. The other says a single concluded attack exhausts the legal trigger, and is\nattached to the claim that an attack occurred that day. **The reader is right about both, and both\nare real moves.** Neither makes the fact less true. Each disputes what the fact is taken to show. They\nare objections to a step, attached to a claim because the page gives a person no place to object to a\nstep.\n\nTwo things follow. **For a link the reading is a good question and a poor switch**: *this does not\nbear on whether the claim is true; is it what the claim is taken to show that you dispute?* sends the\nobjection to where it belongs, and withholding it loses a move the map needs. And the check's seven\nanswers have none for an objection that undercuts: of 21 objections people filed it read 13 as\ncontradicting and 6 as overlapping or containing. It was made to sort sameness, and a filed objection\nis a register it was not made on.\n\nNo claim was found restated as its own reason, and no reason was read as opposing its claim.\n\n### 11.2 Parts: 28 of 39 would be withheld, 25 of the 30 that carry no tag\n\n| The claim | Parts a person attached | Read as *weaker* | Read as *unaffected* | Cannot tell |\n|---|---|---|---|---|\n| Justice is morally more important than equality, in the legal system | 10 | 0 | 10 | 0 |\n| The criterion for minimum-wage rises should be net benefit to low-wage families | 16 | 2 | 14 | 0 |\n| Distributed human authority has priority by default, and can be defeated | 3 | 3 | 0 | 0 |\n| Five claims of the built examples (the stress suite) | 9 | 4 | 3 | 2 |\n| A death-penalty claim | 1 | 0 | 1 | 0 |\n| **All** | **39** | **9** | **28** | **2** |\n\n**The reading and the label agreed on all 39.** Every part read as *weaker* is one the claim asserts\n(9). Every part read as *unaffected* or *cannot tell* is one the claim does not say (30). So on parts\npeople wrote, the part test behaves as *does the claim say this*.\n\n**What the 28 are.** One is debris from a test (*\"A test procedure is being conducted\"*, attached to\na real claim about the death penalty, where it presumes half a need today). Twenty-four are what a\nclaim rests on: under the justice claim, the premises of the argument its author gives for it (that\nthe guilty who escape do no injustice to the guilty who are punished; that desert makes punishment\nfitting); under the minimum-wage claim, a whole argument with its evidence, its mechanism and its\nvalue judgement. Three are in the built examples: a consequence, an assumption that a threshold is\nenough, a definition the claim leans on.\n\n**Why.** The page's box for this asks *What's underneath?* and offers *\"Possible smaller claims\nunderneath this one\"*. People answer with what the claim rests on, and the answers are stored as what\nthe claim consists of. This is this morning's finding about the splitter\n([a-part-and-a-reason.md](a-part-and-a-reason.md), ledger T9) met at the other door: there a sentence's\nown reasons were filed as its pieces, here a person's reasons are filed as its parts. Both doors file\nreasons into the channel that since 31 August treats its contents as required. So for nearly\neverything entered through that box, a premise nobody has examined caps its conclusion, and a premise\nthat fails counts against it, which the ruling of 30 September says a reason does not do.\n\n### 11.3 What a yes would therefore be\n\n**For parts it is no guard against a rare saboteur. It re-sorts 25 of the 30 untagged parts people\nhave attached**, the project's first hand-made decomposition among them. Three ways to act on that, and\nthey are different decisions:\n\n- **Withhold**: the 25 stop capping and count for nothing until argued. It silences honest work.\n- **File them as what they are**: a reason becomes a support of its claim. It then lifts the claim\n  when it is established, never caps it, and fails without counting against it. That is the\n  arrangement ruled on 30 September, reached for what already exists.\n- **Leave them**: what is underneath a claim goes on being presumed required.\n\nWhich of these is right is the question he asked this morning, *what differentiates a part from a\nreason*, and it is his. These 39 are the best material there is for answering it: real, already\nattached, most of them his own.\n\n### 11.4 Two things the code does that the advice above got wrong\n\n**A part that requires nothing still acts.** Run on the strength code:\n\n| Parts of a claim | The claim reads |\n|---|---|\n| two real parts standing at 0.9 | 0.507 |\n| and a third at the middle that requires nothing | 0.500 |\n| and that third refuted, still requiring nothing | 0.493 |\n| the same, with the third presumed half needed (a person's part today) | 0.493 |\n\nRequiring nothing takes a part out of the cap. It does not take it out of the weakest-part rule by\nwhich parts move their whole, so it cancels the lift of the real parts and, once refuted, pulls the\nclaim below the middle exactly as a presumed part does. Withholding would change nothing until that\nis repaired. One live link is in that state today.\n\n**The way back is dearer than one sentence.** An argued connection, or an argued need, counts only\nonce somebody has examined it. So a contributor whose honest addition was withheld would have to\nwrite the connection and then give it standing, which takes several further contributions. An\nopponent pays the same toll, and can pay it with junk that is on the subject. So as built, the guard\nwould stop noise and the lazy, and would tax the honest and the determined alike. A cheap way back\nhas to be built for it to be fair: the reader reads the pair again with the stated connection beside\nit, and a default that a reading took away is one a reading may give back.\n\n### 11.5 Other effects, each thought through\n\n1. **It bears hardest on what comes from outside the frame.** A reader that knows the debate's usual\n   terms is likeliest to say *no bearing* of the consideration nobody has raised. The two withheld\n   objections are that case in small. The same signal is therefore the best one there is for what a\n   person should look at next, so a withheld contribution goes to the head of the list for a person's\n   eye and never to the bottom.\n2. **What is withheld tends to stay withheld.** The author has usually gone, and a reader who sees the\n   machine's reason beside a contribution mostly defers to it.\n3. **A resident line is touched.** *Nothing a person wrote or confirmed moves without a proposal*\n   speaks of content, and a presumption is the system's own default about a contribution nobody has\n   read. The line would still need a sentence saying so, and that is his to bless.\n4. **Even-handedness.** A reading applied to people's links and not to those a machine took from a\n   text treats people with more suspicion than a machine. Applied to every writer, it shifts\n   strengths that exist. Few today, since a cap only bites above about 0.54.\n5. **On the live site the reading would run on a free allowance that runs out.** It then has to fail\n   open, say that it has not read, and read later, as owed work.\n6. **A reading changes with the model and with its instructions.** Each is stored with both, and a\n   change is followed by a recorded second reading, never a silent one.\n7. **Whoever writes the reader's instructions decides what counts as bearing.** They are public and\n   versioned, and every call is in the record. It belongs in the constitution that is not written yet.\n8. **It sets a precedent.** The rule of § 4.1 is its boundary: only defaults, only taking away, never\n   a value somebody argued.\n9. **It closes one door that is open today.** A person who keeps a part the machine invented makes it\n   a presumed part in their own name. Under the part test it would presume nothing, whoever kept it.\n\n### 11.6 The advice, revised\n\n1. **Read every link and part a person adds, now, and show the reading as a question to its author.\n   Let neither reading change a weight yet.** Nothing is lost by waiting: a reason nobody has examined\n   moves nothing today, and nobody outside has contributed. What a switch would cost is measured\n   above.\n2. **For parts, his reading of the 39 comes first.** For each: a part of the claim, a reason for it,\n   something it needs, or debris. A quarter of an hour, and the first time a person labels a machine\n   reading in this project. What is filed where follows from it, and so does the wording of the box,\n   which wants two questions where it has one: what this claim says in smaller pieces, and what it\n   rests on.\n3. **Repair whatever he decides**: a part that requires nothing is left out of both channels. And the\n   debris part is retired on his word, since that is a write to the live graph.\n4. **A place to object to a step**, which the two withheld objections ask for.\n5. **The switch for links** waits for the first contributor from outside or the two-person rehearsal,\n   and for the test of § 4.5 with objections of this kind in it.\n6. **The trial on twenty arguing sentences (ledger T9) is still worth running, second.** It asks the\n   same question from the splitter's side. The 39 come first because they exist, are real, and cost\n   no further call.\n\n---\n\n## See also\n\n[cheap-to-add-slow-to-matter.md](cheap-to-add-slow-to-matter.md) (the table this brings up to date) ·\n[a-part-and-a-reason.md](a-part-and-a-reason.md) (what a part is; when every reason fails) ·\n[relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) (the ruling on relevance and the\nruling on a failed reason) ·\n[derivability-of-missing-considerations.md](derivability-of-missing-considerations.md) (the lists) ·\n[the-ratification-record-and-the-triage-feed.md](the-ratification-record-and-the-triage-feed.md) (no\nreview act) · [quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md) (the ten\ncontrols) · [proving-a-pass-reliable.md](proving-a-pass-reliable.md) ·\n[fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md) (who the opponent is) ·\n[weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md) § 9e\n(the cap an irrelevant part buys)\n"}