{"path":"research/microtasks-in-broad-daylight.md","content":"# Microtasks in broad daylight: how well defined are the keeper, the splitter and the judges?\n\n**Date**: 2026-09-29 · **Type**: research and proposal · **Status**: the principle this doc supports was\nadmitted as a canon line the same evening (§ 5). § 4 item 3 was built on 2026-09-30 at his question,\nwith the coverage check rewritten and measured (§ 3e); the other § 4 proposals are unruled. The error\nreadings in § 3 are Claude's reading of the ten proposals stored that afternoon, not human labels.\n§ 3f (2026-10-03) reads the two open questions again, against the stored data and the live graph.\n\n**Occasion.** The founder, VERBATIM: *\"What does 'hard text tasks' mean exactly? Are they as\nwell-defined as what we're talking about with the keeper keys of sameness and with the splitting /\nthe separability/decomposition floor (what prompting/rules is needed for the splitting and how\nreliable/error-prone is it really? How hard can it be if you look at the concrete task in broad\ndaylight and see what variance there can be in interpreting the instructions?) and the relevance\njudge etc?\"*\n\n**Corrects** two framings in [proving-a-pass-reliable.md](proving-a-pass-reliable.md) (§§ 0, 5, 7,\n9f) and one in [split-review-and-the-founders-intentions.md](split-review-and-the-founders-intentions.md)\n(§§ 3, 5). See § 1 and § 3c.\n\n---\n\n## 0. Answer first\n\n1. **\"Hard text tasks\" were not our kind of task.** They were zero-shot labelling of contested\n   social-science categories with a one-line instruction, where the label itself is disputed. The\n   14–28% saving measured there is a poor prior for our passes, which carry written definitions,\n   worked tests and examples (§ 1).\n2. **Our tasks are better specified, and the variance that remains is mostly ours to remove.** It\n   sits in three places: one key meaning different things to the machine, to the human sitting and to\n   the routing code; two passes given conflicting instructions about added context; and no rule at\n   all for a clause the parent concedes or denies, which is where the splitter's harmful errors come\n   from (§§ 2–3).\n3. **The splitter's harmful error has one shape, and one question catches it.** Read part by part,\n   6 of the 37 written parts fail the test *if this part were false, would the parent be weaker?*\n   Two clearly (a clause the parent denies or only concedes, written as a flat assertion), four\n   more mildly (a background concession written as if the parent rested on it). Because a written\n   part caps its parent, refuting any of the six would pull the parent down, which is backwards.\n   The parts judge flagged all six; the coverage check flagged none (§ 3). Rewritten on 2026-09-30\n   to ask that question of every part, it catches all six (§ 3e).\n4. **The fixes are cheap and testable**: aligned definitions, two explicit tests in the splitter's\n   rules, one deterministic invention law, and two differently framed checks paired rather than one\n   trusted alone (§ 4). This is the founder's *\"designing of sturdy microtasks\"* made concrete.\n\n---\n\n## 1. What the \"hard text tasks\" actually were\n\nSource: Candès, Ilyas, Zrnic, *Probably Approximately Correct Labels* (ICML 2026,\narXiv:2506.10908), Table 1. Three computational-social-science tasks, labels from GPT-4o collected by\nGligorić et al. (NAACL 2025, arXiv:2408.15204), uncertainty from the model's own stated confidence\n(asked zero-shot, then asked how sure it was):\n\n| Task | Labels | Expert labels saved at a 5% error target | Error if the machine labels alone |\n|---|---|---|---|\n| Political bias of a news article | left / center / right | 13.8% | 37.7% |\n| Headline stance on global warming | agree / neutral / disagree | 28.1% | 24.8% |\n| Misinformation | misinformation / real | 18.1% | 18.6% |\n| *ImageNet, for contrast* | *1,000 object classes* | *59.6%* | *21.7%* |\n\nGligorić et al. report human–model agreement on the stance task at Cohen's κ = 0.57, *\"the highest in\nthis setting\"*. These labels are judgments where careful people differ: whether a headline \"agrees\nthat global warming is a serious concern\" depends on the reader.\n\n**Why that does not transfer.** The saving depends on how well the machine's confidence separates\nits right answers from its wrong ones. On a disputed label, confidence cannot do that well, because\nthe \"right\" answer is itself a matter of which annotator you ask. Our passes are a different kind of\ntask: each carries a written definition, a named test and worked examples (§ 2). The published figure\nis a floor for *contested* microtasks, and says little about *specified* ones.\n\n**The nearest published analogue to the keeper is natural language inference**, the task of saying\nwhether one sentence entails, contradicts or is neutral to another. In SNLI's validation, five people\nper pair: 58.3% of pairs unanimous, 98% with a three-of-five majority, Fleiss κ 0.70 overall and 0.60\nfor \"neutral\" (Bowman et al., EMNLP 2015). When a hundred people re-labelled the pairs that had only\nthree-of-five agreement, the majority label changed in roughly a fifth to a quarter of them, and\nmodels were near-perfect on high-agreement pairs and near chance on low-agreement ones (Nie, Zhou,\nBansal, *ChaosNLI*, EMNLP 2020). **So even a well-defined relation task has a contested tail, and the\nmachine's errors live there.** That is the division of labour abstention buys: settle the clear pairs,\nsend the tail to people.\n\n---\n\n## 2. Each microtask, read against its own instructions\n\n**The keeper of the seven keys** (`deliberus/keeper.py`). Well specified: each key has a one-line\ndefinition, sameness is a consent test, *opposes* covers contradiction and contrariety, *overlaps* is\n\"judged by EVIDENCE, not by topic\", *contains* is asymmetric, and the reason is written before the\nkey. **The variance is a definition split across three places:**\n\n- In the machine's prompt, *overlaps* is evidential (establishing one would partly advance the other)\n  and *unrelated* means no evidential bearing.\n- In the human sitting (`deliberus/sameness.py` KEYS), *overlaps* reads \"they share part of a claim\"\n  and *unrelated* reads \"different claims\", which is semantic.\n- In the routing code, *overlaps* and *contains* go to separability proposals, which treats them as\n  bundling. But *contains* in the prompt means entailment (\"a hedged claim is contained by its\n  definitive version\"), and a hedged claim is not a bundle of the definitive one.\n\nA person scoring the keeper in the sitting is therefore grading a different question from the one\nthe machine answered. **The measured agreement would be partly definitional disagreement, not error.**\n\n**The splitter** (`deliberus/extraction/decompose_claim.py`). Two named tests, both with worked\nexamples: Russell's (could a reader affirm one part while denying another?) and the neo-Davidsonian\n(strip a modifier; if the rest survives, the modifier was its own conjunct). Rules against splitting\nrhetorical parallelism, against invention, and for `already_atomic`. **Two gaps:**\n\n- The stranger rewrite is told *\"adding context here is correct\"*, while the coverage check counts\n  added material as invention. The two instructions conflict, and the conflict produces both false\n  holds and missed glosses.\n- Nothing addresses a clause that is not asserted at all: *rather than X*, *instead of X*, *even if\n  X*, *not X*, or a presupposition. Content from such a clause can be written out as a standalone\n  claim the parent never made.\n\n**The relevance judge** (`deliberus/relevance_judge.py`). A gate (does it bear at all), a direction\n(raises or lowers) and a five-step magnitude, each with its reason first. Gate and direction are\ncrisp. Magnitude is a graded judgment, and the one place where two careful readers would differ by a\nstep. The first trial, with a bare 0–1 number, answered 0.5 on 8 of 12; the second, with the named\nsteps, answered all 12 confidently with no disputed filing. The proposed refinement is to ask for the\ntwo conditional probabilities instead of a magnitude and compute the difference\n([jev-and-the-judgement-layer.md](jev-and-the-judgement-layer.md) § 6).\n\n---\n\n## 3. The splitter's real output, read item by item\n\nThe ten proposals stored that afternoon, six written and four held for invention (an eleventh,\nerased and then restored, is § 3d). Each was labelled `gemini-3-flash-preview`, but the Claude\nrunner's own log shows Claude Sonnet 5 answered every split call of 2026-09-21: the label recorded\nthe model asked for, not the one that answered. New proposals record the one that answered, and the\nten were relabelled on his word the same evening (§ 3d). All 37 written parts were read against\ntheir parents on 2026-09-29, first in bulk and then again part by part when the founder asked,\nVERBATIM: *\"What are those 2-4 errors, exactly?\"*; the second reading is the\none below, and it revised the first (§ 3c).\n\n### 3a. Error classes, with instances\n\n| Class | What happens | Instances among the 37 written parts |\n|---|---|---|\n| **Polarity flip** | a clause the parent denies or only concedes is written as a flat assertion | `1a5d3f9d5d53` part 5, *\"the underlying records … are destroyed rather than retained\"*, cut from a claim that records are reliable *rather than* obscured by destruction · `61e048b6abce` part 4, *\"most low-wage workers who remain employed … end up earning higher wages\"*, cut from a claim that only concedes this (*\"even when\"*) |\n| **Conceded background written as a part** | the same shape, milder: the parent concedes it but does not rest on it | `1a5d3f9d5d53` part 3 (a legal rule on what death certificates list) · `0c27fdec0781` parts 5–7 (documented errors, biases and inconsistencies in executions), which also add examples the parent never gave (*\"race, class, or geography\"*) |\n| **Added contrast or commentary** | a part says slightly more than the parent | `6ad101f38b48` parts 3 and 6 each add a *\"rather than …\"* the parent never stated · `42f5365032c7` part 6 adds that the newer method is *\"implicitly more advanced or trustworthy\"* |\n| **Dropped detail** | a part loses something the parent said | `6ad101f38b48`: the study's institution (*\"University of Washington\"*); the authors are kept, so minor, and the coverage check flagged it |\n| **Near-duplicate** | one part restates another, or the whole | `0c27fdec0781` part 1 is nearly the whole claim and contains part 2 · `61e048b6abce` part 2 sharpens part 1 |\n\nThe held proposals show the same shapes before any writing: glosses added in the stranger rewrite\n(`eb27c825a45e`, `4d6779b25e02`) and unasserted clauses (`c444e59749ce` part 4; `14241d73113b`\npart 1, borderline).\n\n**One question separates the harmful rows from the rest: *if this part were false, would the parent\nbe weaker?*** A real part passes, since the parent falls with it. The six parts in the first two\nrows fail: if records are kept, the claim that records are reliable is helped, not hurt. That\nmatters because a written part caps its parent (a whole cannot read stronger than its weakest part,\nruled 2026-09-21), so refuting a flipped part pulls the parent down, and the strength computed for\nthat claim runs backwards. Both clear flips came back coverage-clean, so both cap their parents\ntoday. These are Claude's readings, not human labels, and not a proof in the sense of\n[proving-a-pass-reliable.md](proving-a-pass-reliable.md) § 2.\n\n**This is the answer to the question the founder asked when he ruled that written parts cap**:\n*\"We should expect good splits, and I have some trouble imagining wrong splits, what would that\neven look like?\"* (2026-09-21, commit `6ddc980`). The shapes recorded that day were loss, invention\nand drift into commentary. The flip is a fourth, and the coverage check as then written could not\nsee it: nothing was dropped and nothing was added, since every word came from the sentence, with its\npolarity changed. Since 2026-09-30 the check asks of every part whether the claim would be weaker\nwere it false, and it catches the flip (§ 3e).\n\n### 3b. The two checkers err differently\n\n- **The coverage check** (a separate call that reconstructs the parent from the parts, then\n  compares) caught the dropped institution and held the glossed proposals, and **flagged none of\n  the six harmful parts**, as written until 2026-09-30 (§ 3e).\n- **The parts judge** dissented on 14 of the 37 (7 `does_not_fit`, 7 `miscast_as_evidence`), and\n  **all six harmful parts are among them**. Of the other eight, five are the mild problems in the\n  table (added contrast, commentary, near-duplicates) and two or three are faithful parts it read as\n  evidence for a sibling rather than as parts (`6ad101f38b48` part 1, `42f5365032c7` parts 5 and 7).\n\n**Two differently framed checks are complementary**, and on this failure the parts judge is the one\nthat works: it asks, in effect, the question above.\n\n### 3c. Correction: what \"62–68% genuine parts\" measured\n\nThe first reading of 2026-09-29 said this figure was *\"never a splitting accuracy\"* because the\nparts judge holds a different idea of a part. **The part-by-part reading narrows that sharply.** Most\nof the judge's dissents are real problems, so its 62% is roughly the share of parts with no problem\nat all. What it does not say is how many parts do harm: that is 6 of 37 (16%), of which 2 (5%) are\nclear. Quote those two numbers for harm, and the judge's figure only as a problem-free rate.\n\n### 3d. The seventh written split was erased by a re-store, and nothing said so\n\nOn 2026-09-21 the writer reported eleven stored proposals, seven written and four held, and on\n2026-09-22 the census counted 44 links written by the splitter. Today the graph holds ten proposals,\nsix written, and 37 links. The missing one is `claim_aig_goal`, the goal claim of the hand-made\nAI-governance map (found in the 2026-09-21 session's own output, which lists it among the seven\nwrites). When that map was re-stored on 2026-09-22 with 24 of its 116 claims corrected,\n`cleanup_source_graph` deleted and recreated every machine-kind claim of the source, as it is built\nto do. The goal claim came back without its proposal, and its seven parts, which the splitter writes\nas `extracted` and links to the parent's source, went with the cleanup. The corpus went on saying\nseven proposals and 44 parts for a week.\n\nThe loss is cheap to undo (one or two model calls re-propose the split). What it shows is dearer: a\nre-store erases machine work done on a claim after it was first stored, and the work is not re-asked,\nwhich is the defect [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md)\nnames (a pass fires once, at one door, and nothing asks again). Filed in `TODO.md`.\n\n**Restored 2026-09-29 on the founder's word** (*\"restore the erased split\"*). The original parts\ncould not be recovered (the session that made them printed only their first hundred characters), so\nthe split was proposed again, and it took three attempts, each reported here because the selection\nmatters. The first came back with eight parts and one invented clause (*\"what counts as\n'sufficient' is itself a further judgment\"*), so the writer held it, as its rule requires. The\nsecond came back degenerate: every field set to the word \"test\", stored as a verdict that the claim\nwas one thing. The third was coverage-clean, every part passes the question in § 3a, and it is\nwritten: eight parts, 45 split links over seven claims, live on the claim page.\n\nThe three attempts are also a small, unplanned test of re-asking. The two usable answers read the\nsame phrase differently: the first took *\"broad human control\"* as control wide in scope, the third\nas control spread across many people. Both are defensible, which is the point: that is variance\nfrom an unsettled instruction (how to cut through a word with two senses), the kind the canon line\nof the same evening assigns to the prompting rules rather than to the model.\n\n**Three guards came out of it, each tested against the failure it answers:**\n\n- **A re-store keeps the split of a claim whose words are unchanged**, its review marks included,\n  and drops a split made of words the map no longer says (`scripts/store_hand_extraction.py`,\n  tested on a real embedded graph). Other machine work on a re-stored map (links to other sources,\n  question scaffolding) is still erased and stays open in `TODO.md`.\n- **A proposal records the model that answered**, not the one asked for, and so does its coverage\n  check (`deliberus/proposed_split.py`). The ten older proposals were relabelled the same evening\n  on his word (§ 5).\n- **A degenerate answer raises instead of being stored.** The splitter refuses a part that shares no\n  word with its claim, and the Claude backend refuses any answer whose every text field is one\n  short token and retries it, which covers every pass on that backend, not only the splitter.\n\n### 3e. 2026-09-30: the coverage check fixed, on his question\n\nOffered a resident warning line (that the coverage check cannot see a flip), the founder asked\ninstead, VERBATIM: *\"Should we rather not just fix the \"coverage check\"? What exactly does it ask,\nagain, refresh my memory?\"* The warning line is withdrawn; the check is fixed.\n\n**What it asked.** Blind, rebuild the claim from the parts alone *(⚠ 2026-10-10: blind in the order of\nits answers only. It is one call, and its prompt holds the claim a few lines below the instruction to\nrebuild: [quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md) § 1.3)*. Then, shown the claim: do the parts\ncover everything, what did they drop, and what did they invent, defined as *\"anything the parts\nassert that the original did not\"*. That last question is the right one. It was answered as\n*anything the parts mention*, and every word of a flipped part came from the sentence.\n\n**What changed** (pass version 2; `deliberus/extraction/decompose_claim.py`,\n`deliberus/proposed_split.py`, commit `85b65ee`):\n\n- The parts are numbered, and each gets the part test: *if this part were false, would the claim be\n  weaker?* (`if_false`), plus a label for how the claim holds what the part says (`held_as`:\n  asserts, denies, concedes, supposes, absent). **Only the part test gates**; the label is kept for\n  counting the error classes.\n- `covers` is answered last, after the lists it sums, since field order is generation order.\n- A part the claim does not assert holds the proposal the way invention does (`unasserted`), and\n  presumes nothing in the strength layer (`support_semantics.UNFAITHFUL_COVERAGE`).\n- The splitter keeps each clause's polarity (a denial, grant or condition stays in the claim's own\n  form, and background taken for granted inside such a clause is not a part), and **adds nothing\n  the claim does not contain**, because a written part is attributed to the claim's author.\n\n**Measured** with Claude Sonnet 5 as splitter and checker, read-only; data in\n[`microtasks-in-broad-daylight/`](microtasks-in-broad-daylight/), rerunnable with\n`scripts/measure_splitter.py`.\n\n1. **The check over the 45 written parts, twice.** The part test flagged the three clear flips on\n   both runs, plus one part carrying added commentary; 45 of 45 parts were judged the same way twice.\n   The three conceded-background parts of the death-penalty claim passed the part test (read as a\n   presupposition the claim commits to) but the invented list caught their added examples on both\n   runs. All six harmful parts would have been held.\n2. **The first gate was wrong.** It also required the label to agree. Re-split under the new rules,\n   the minimum-wage claim came back with its denial and its concession in the claim's own form, and\n   the check answered *weaker, denies* and *weaker, concedes*: the label described the content, the\n   part test described the part, and a faithful split was held. The gate now reads the part test\n   alone, and a test pins the real case.\n3. **The shipped prompts over all eleven stored proposals** hold all eleven, the seven written ones\n   included. On a reading the added material is real: first names the claim lacks, *rather than*\n   contrasts and examples it never gave, commentary, an *even if* it never conceded, each written\n   into a part attributed to the claim's author. Two of those holds (the first names, the *even\n   if*) came on two runs of three, so the invented list is less steady than the part test.\n4. **A fresh split of all eleven claims with the shipped prompts**: all eleven writable, nine clean\n   and two with something left out. One left out a denial the claim asserts, correctly flagged. The\n   other left out the death-penalty claim's *\"documented real-world errors, biases, and\n   inconsistencies\"*, which the check reads as asserted and the splitter as background. That is a\n   contested reading (a presupposition as a conjunct, Russell's view, or as background the claim does\n   not rest on), it only adds a note, and all six asks read it the same way, five of them under a\n   trial wording that told the check not to (dropped again, since it changed nothing). Read part by part (by Claude), the\n   52 fresh parts hold no flip and no conceded clause stated as fact; one mild added contrast got past\n   the check (`claim_c444e59749ce` part 3, *\"rather than only occasionally or unpredictably\"*), and\n   the goal claim's parts say *\"control over AI\"* throughout, a reading of *\"broad human control\"*.\n5. **Re-asks.** One faithful part was flagged once in six asks, a false hold, which errs safe.\n   Twice in 49 check calls a quoted phrase came back partly translated (*\"con sus …\"*, *\"con dess\n   dokumenterade …\"*): a defect in the quote, not the verdict, and open.\n\n**The caveat that matters**: splitter, checker and reader are the same model family, the same-hand\nbias of the residual taxonomy, and production splits on Gemini's free tier, where none of this was\nmeasured.\n\n**What the live site then did (read 2026-10-05).** Two splits were written there, on 3 and 4\nOctober, both cut by `gemini-3.5-flash`, the model the free tier falls back to when the first has no\nquota left. One reads as clean. The other cut *\"a positive or necessary outcome\"* into two parts\nstated as fact. Put to five models with these instructions, the same sentence was cut that way by\nthree of them, Claude Sonnet included: the instructions have no rule for an *or*, and the check's\nlabels have no place for an alternative\n([pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) §§ 4b–4c, ledger\nY5 and Y7).\n\n**The consequence.** Version 1 wrote seven of these eleven claims, and six of its 45 written parts cap\ntheir claims backwards. Version 2 would have written all eleven, from different cuts, with none. The\nseven version-1 splits were still in the graph; replacing them (re-split with version 2, the old parts\nsuperseded and kept visible, the shape of the 2026-09-13 housekeeping ruling) was a production write\nand waited for his word. He gave it on 2026-10-06 (*\"Fresh cuts blessed\"*) and it was done that\nevening: § 3g.\n\n### 3f. 2026-10-03: the two open questions, read again for a picture\n\nAsked to draw the two questions he is still weighing, the founder, VERBATIM: *\"Try to visualize the\ncentral crucial conundrums we stand before here please. Both of the issues I referred to before,\nprimarily. Make sure you ground extremely deeply, and think deeply about HOW to visualize this\nbest.\"* Reading the stored data and the live graph for the drawing changed how both questions\nshould be stated. Nothing here is ruled.\n\n**The background question (ledger T6) was two questions.** The 2026-09-30 rule treated three kinds\nof clause content alike: what a claim denies, what it grants, and what it says in passing. For the\nfirst two, both of the questions below answer *no*. For the third they come apart:\n\n1. *Does the claim say it?* The death-penalty claim does say that practice has *\"documented\n   real-world errors, biases, and inconsistencies\"*. It says so in passing, which is the *at-issue\n   versus backgrounded* row of [decomposition-axes.md](decomposition-axes.md) (Potts), and the\n   splitter's own axis list names `presupposition` as *\"content the claim commits to even under\n   denial\"*.\n2. *Does the claim lean on it?* Here the two passes disagree, and each is consistent. The splitter\n   leaves the phrase out. The check, asked the part test about the three background parts of the old\n   split, answered *weaker, asserts* on both runs, giving as its reason that the errors are *\"part of\n   the reason\"* actual practice is set aside (`2026-09-30-check-over-stored-parts-twice.json`). So\n   the ledger's sentence *\"the part test sides with Strawson here\"* was Claude's reading of the test,\n   and the check's reading of the same test was the opposite. Corrected on the ledger row and in\n   `TODO.md`.\n\n**A written part does three things at once, and leaving a phrase out does none of them.** It becomes\na claim of its own that a reader can dispute; it is recorded as the claim's (`attributed_to` is\ncopied from the claim); and it caps the claim at the presumed half weight (computed with\n`necessary_ceiling`: an unexamined part holds its claim at 0.75 at most, a part refuted to 0.1 at\n0.55). The phrase needs the first two and is disputed on the third, and neither choice on offer\ngives that. The claim page adds a pull one way: for a phrase the check lists as left out it offers\n*\"Add it as a part\"*, which is all three.\n\n**The names were loose.** Strawson's presupposition is a precondition without which a claim is\nneither true nor false (*that a tax was raised*, for a claim about what raising it did), so a claim\ndoes stand on it. The death-penalty phrase is a remark in passing, which the claim would survive.\nThe splitter's rule covers both kinds when they sit inside a denied, granted or supposed clause.\n\n**A third arrangement (Claude's proposal, unruled; designed on paper, not built).** Write what a\nclaim says in passing as its own part, in the claim's own words and marked as said in passing,\npresumed to require nothing of the claim. Whether the claim leans on it is then argued in the graph,\nthrough the requirement claim that already exists. The pieces are in place: the splitter labels each\npart's `axis`, `presumed_requirement_weight` already returns zero for some parts, and\n`POST /claims/{id}/parts/{part}/requirement` raises a cap by argument. This replaces *leave it out,\nthen reword the check* as Claude's recommendation, for three reasons: it meets the separability\nfloor (a reader can accept the claim and dispute the remark), it drops nothing the check would\nreport, and it puts the disputed question where two readers can hold it. One thing is unchecked:\nwhether a part of zero requirement weight still pulls its claim through the weakest-part energy\nterm (`qbaf.subclaim_energy`).\n\n**The sameness checker's question (ledger T5) turns on one word.** The blessed table of 2026-08-31\ndefines *overlaps* as *\"establishing one would partly advance the other\"*, gives its ground as\n*\"compatible, shared content\"* and its consumer as the novelty dial\n([soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md),\naddendum of Aug 31). *Partly* reads two ways:\n\n- **A part of it**: the evidence settles one piece of the other claim, which therefore has pieces.\n  The founder's ruling of 2026-09-13 (*\"Wouldn't 'one contains the other (either way)' or 'they\n  overlap' … mean that the separability floor hasn't been reached?\"*) fits this reading, and so\n  does the sitting's wording, *\"they share part of a claim\"*.\n- **To some degree**: one claim is a reason for the other, which may have no pieces at all. This is\n  what a support link says, and it is how the keeper's instructions are read (*\"judged by EVIDENCE,\n  not by topic\"*).\n\n**Measured on the live graph, 2026-10-03.** The keeper has recorded 100 split proposals on\nresemblance links: 89 from *overlaps*, 7 and 4 from the two *contains* answers. Nothing reads them:\n`separability_proposal` has a writer and no reader. Of the 89 *overlaps* pairs, 42 also carry a\nsupport link, 23 a parts link, 6 an attack, 4 a qualifying link and 15 nothing. Claude's reading of\nthe keeper's own stored reasons for all 100 follows; it is one model family reading itself, from the\nclosing lines of each reason, and no human label:\n\n| The keeper's answer, and what its reason describes | Count | Is there something to split? |\n|---|---|---|\n| *overlaps*: one claim is a reason for the other as a whole | about 64 | no |\n| *overlaps*: the evidence lands on one piece of a compound claim | about 15 | yes |\n| *overlaps*: the two claims assert a common piece | about 5 | yes |\n| *overlaps*: no real bearing, in the keeper's own words | about 5 | no; by its instructions the answer was *unrelated* |\n| *contains*: a general claim and a specific case of it | 5 | no |\n| *contains*: a stronger and a weaker version | 5 | no |\n| *contains*: a precondition of the other | 1 | no |\n\nSo about four in five of the recorded split proposals point at a claim with nothing to split, and\nnone of the eleven *contains* answers is a claim holding the other as a separable piece. *Contains*\nis used three ways: a piece (the ruling's sense), a weaker version (the instructions' *\"cures\"* and\n*\"may help\"*), and a specific case of a general claim, which the instructions do not count as\ncontainment at all.\n\n**What this adds to § 4 items 1 and 2.** One definition per answer means choosing between the two\nreadings of *partly*, and the reading not chosen still needs a home. For *to some degree* the home\nexists: the relationship step writes the support link. Under that choice *overlaps* would be kept\nfor a shared or singled-out piece, with the piece written out so the answer can be checked, and the\nruling of 2026-09-13 would fire only there. The two pairs drawn for him: `claim_aig_slow_case` with\n`claim_4d4af2f8946b` (the evidence lands on the first half of a two-part claim) and\n`claim_aig_take_seriously` with `claim_doom_i2` (one claim is a reason for the other).\n\n**Measured across models, 2026-10-04**\n([whose-tokens-pay-for-mapping.md](whose-tokens-pay-for-mapping.md) § 3e). *Overlaps* is the answer\nthat changes with the reader. On twenty pairs where the keeper had recorded it, Claude Sonnet 5\nasked again agrees 16 times and two Gemini models agree 6 and 7 times. On twenty pairs that the\nrelationship classifier, on the same Claude model, had called unrelated, Claude as keeper answers\n*overlaps* 13 times and the Gemini models answer *unrelated* 15 and 17 times. On twenty random\npairs every model answers *unrelated*. So the keeper's two firm answers travel between models and\nthe loose one does not, which is what one definition per key (§ 4, item 1) would be expected to\nrepair, and it gives that proposal a test: after the rewrite, the models should agree.\n\n### 3g. 2026-10-06: the claims cut again, on his word\n\nAsked what *\"the 7 splits that could be replaced by fresh ones\"* were, and given the recap of § 3a\nand § 3e, the founder, VERBATIM: *\"Fresh cuts blessed\"*. A few minutes later, on the suggestion that\nthe decision wait for his review of the parts: *\"We can always review the cuts later after the\nreviews that's fine\"*. Ledger T7 is ruled, and the cuts were made the same evening.\n\n**What ran.** `scripts/recut_splits.py`, from the keeper's machine, on `claude-sonnet-5`. That is\nthe exact model the measurement of 30 September ran on: the Claude runner keeps its own log of which\nmodel answered each call, and every call of 21, 29 and 30 September in it was answered by\n`claude-sonnet-5`. The engine's default until this evening was the alias `sonnet`, which the runner\nresolves to whatever it currently ships. It still reached the same model that day (tested), and the\ndefault now names the model itself, so that a later change in what the alias means cannot alter the\nmodel under a measured pass while every label goes on saying `sonnet`.\n\n**It cut eleven claims, not seven.** The seven that version 1 wrote, and the four whose version-1\nproposals had been held for invention and never written (§ 3). The four had nothing to replace.\nTheir stored proposals asked a person to rule on parts the current version would not have made, and\nthe ledger row he ruled on says the fresh cuts settle them as well.\n\n**The rule for choosing among cuts, fixed before the run**, because choosing is a selection: up to\nthree attempts per claim; the first cut the check lets through is written; a cut the check holds is\nkept in the record and never written; if none passes, the earlier split stays as it was. A cut\nanswered by another model family than the one asked for is not written at all.\n\n**What happened.** Every attempt, held or written, is in\n[`2026-10-06-fresh-cuts-of-the-eleven.json`](microtasks-in-broad-daylight/2026-10-06-fresh-cuts-of-the-eleven.json),\nwith what each claim's split looked like before.\n\n| Claim | Earlier parts | Written on attempt | Fresh parts | The check's verdict |\n|---|---|---|---|---|\n| `claim_1a5d3f9d5d53` official outcome data is trustworthy | 6 | 1 | 4 | clean |\n| `claim_61e048b6abce` any job loss against net benefit | 4 | 1 | 4 | clean |\n| `claim_0c27fdec0781` evaluate the ideal, not the actual system | 7 | 2 | 4 | clean |\n| `claim_36633a07f210` executions' claimed effects lack evidence | 7 | 1 | 8 | clean |\n| `claim_42f5365032c7` 138 wage rises, jobs unchanged | 7 | 1 | 9 | clean |\n| `claim_6ad101f38b48` survey and Seattle study show harm | 6 | 1 | 8 | clean |\n| `claim_aig_goal` preserve broad human control | 8 | 1 | 8 | clean |\n| `claim_14241d73113b` Palestinian territory as another entity's | none (held) | 1 | 5 | clean |\n| `claim_c444e59749ce` written safeguards work in practice | none (held) | 1 | 3 | clean |\n| `claim_eb27c825a45e` competitive market against institutions | none (held) | 1 | 5 | something left out |\n| `claim_4d6779b25e02` self-defence law against occupation law | none (held) | 3 | 2 | clean |\n\nAll 45 version-1 parts are retired and 60 fresh parts are written. **The check held no\ncut**, so the rule above never had to choose between two cuts of one claim. The repeated attempts\nwere all after an answer that was no cut at all (below); the self-defence claim's third also sat\nthrough one call that ran out its five minutes and was asked again inside the engine.\n\n**What \"replaced\" means in the graph.** Nothing was removed.\n\n- The link from a claim to a replaced part carries `superseded_by` = `recut:decompose_claim@2:`\n  and the moment the fresh parts were written, which is also the moment on the fresh links, so\n  what replaced a part is one query away. Everything that reads strength, the map, the horizon or\n  the tail already leaves such a link out.\n- A replaced part's own claim moves to the lifecycle stage `superseded`, so the feed and retrieval\n  hold it back. Its page still opens, says it was replaced and links to the claim it was cut from.\n  A replaced part that other claims are linked to would keep its stage, since it is still in use;\n  none of the 45 had any link but the one to its claim.\n- The proposal a fresh cut replaced is kept on the claim (`split_proposal_before`), and the model\n  that made the earlier cut is now named on its links, which was until then recorded only in that\n  proposal.\n- `proposed_split.undo_recut` puts a claim's earlier split back, and `proposed_split.out_of_step`\n  names any claim whose live machine-cut parts are not the ones its stored proposal lists, which is\n  what a run that stopped between writing and retiring would leave. It named none.\n\nA claim where a person kept or reworded a part is not cut again by this operation: which of two\nrenderings stands is then a person's call. No part had been touched by anyone.\n\n**Read part by part, by Claude Opus 5.5 in the session. A model's reading, of the same family as\nthe splitter and its check, and not a person's.**\n\n- **No fresh part states as fact what its claim denies or only grants**, which is the fault the\n  cut was for. The six harmful parts of § 3a have these in their place: for the records claim, *\"is\n  not … systematically obscured by destruction of the underlying records\"*; for the minimum-wage\n  claim, the concession kept inside its part (*\"Even granting that the great majority … end up\n  earning more …\"*); for the death-penalty claim, the documented errors, biases and\n  inconsistencies kept inside the denial, with none of the examples version 1 added.\n- **One *or* was cut into two required parts.** The safeguards claim says the written rules\n  *\"reliably translate into actual prevention or detection\"*; parts 1 and 2 state prevention and\n  detection each as fact. That is the gap already on file (§ 3e's last paragraph, ledger Y5 and\n  Y7): the instructions have no rule for an *or* and the check has no label for an alternative. It\n  caps too strictly, never backwards: refuting one alternative would pull the claim down though the\n  other still carries it. Where an *or* sat inside a denial (the self-defence claim's *\"displaced\n  by, or confined to\"*), the cut made two denials of it, which is what the sentence says.\n- **Nine parts add a small gloss or contrast the claim did not state**, and the check passed\n  each: *\"under international law\"* and *\"legally\"* (the Palestinian-territory claim), *\"not\n  only large-scale or significant job losses\"* (the minimum-wage claim), *\"not workers in general\"*\n  (the survey claim), *\"not narrowly held\"* as a reading of *broad* (the goal claim), *\"with nothing\n  left out\"* (the records claim), and on the safeguards claim *\"not merely existing as provisions on\n  paper\"* twice and *\"rather than only occasionally, unpredictably, or in a small minority of\n  cases\"*, the same addition the reading of 30 September found on this claim.\n- **Some parts overlap.** The wage-study claim came back as nine parts, two pairs of which say\n  nearly the same thing, and one part of the survey claim (*some* studies find negative effects) is\n  contained in the next (*most* do).\n- **One presupposition was written as its own part**: *\"Seattle's minimum wage reached $13 per\n  hour.\"* On 30 September the splitter left presupposed background out and the check asked for it\n  (§ 3e item 4, ledger T6). Here it wrote one. The instructions are read both ways.\n\nSo the fault that made a cap run backwards is gone from these claims, and the smaller ones of § 3a\nremain at roughly the rate the September reading found. They are his to label in the sitting.\n\n**Three of the fourteen cut requests came back as no cut at all, and one of them said why.** The\nanswer was the word *test* in every field (twice, each beside one word naming an axis) and then\n*\"test\"*, *\"test atomic\"*, *\"test decontextualized\"* under the reasoning *\"Test minimal case to\nisolate schema issue before resubmitting full answer.\"* Each took two to three minutes. Read with\nthat sentence, the placeholder of 29 September (§ 3d, cause then unknown) is the model's own probe:\nits full answer was refused by the runner's structured-output step, it sent a minimal one to learn\nwhat the step accepts, and the runner returned the probe as the answer. The engine's guard missed\nall three, because it asked for every text field to be alike, and the splitter's own check caught\nall three (a part must share a word with its claim), which is what turned each into a repeat and\nnot into a stored verdict. The guard now covers the three shapes, with the three real answers as\nits tests; run over the 27 real answers in the record of machine judgments it fires on those three\nand no other. **What is still unknown is which part of a full answer the runner refuses.** The\nguard logs the model's closing text when it fires and the runner's log now keeps the number of\nturns, so the next one arrives with evidence.\n\n**What a reader sees.** A claim's map shows the fresh parts only. Its split panel lists the parts\nthat count and folds the replaced ones under them. A source's page for machines lists replaced\nparts last, under their own heading. Both needed a change, since each had listed every machine-cut\npart of a claim alike (`get_split_proposal`, `_claims_for_source`).\n\n**And one thing the first deploy still got wrong, found by looking at the live page.** Under the\nparts, the claim page offered *\"Still bundled deeper down\"* with a replaced part among its three\nlinks. Eight readers counted every part-of link, replaced or not: the count of a claim's parts in\nfive places, the walk that finds what to deepen next, the premises of the discursive-dilemma check,\nand the feed of claims that need help. They had been wrong since the first part-of links were\nretired on 27 September, and nothing showed it until forty-five more were. All eight now leave\nreplaced links out, each tested on a real graph engine with the count before as its control, and\n`tests/test_part_of_readers.py` reads the code for any query that matches a part-of link and never\nmentions supersession: such a reader is on a short list with its reason, or the test fails. Two\nare on that list as known gaps (the list of claims a horizon reading may be made for, and the walk\nbehind the question box, which follows superseded links of every kind); the replaced parts' own\nclaims are held back from retrieval by their stage, which covers these cases, and both are in\n`TODO.md`.\n\n**The fresh parts have headlines, written by hand**: 59 of them, one part being short enough to\nbe its own title (`scripts/titles/2026-10-06-fresh-cuts.json`, applied with\n`scripts/apply_titles.py`, which can also put the old titles back). The splitter gives a part its opening words as a title,\nparts of one claim share their opening, and so they read alike on a map: the defect he called *\"not\nacceptable\"* on 5 October, arriving again with every cut. Asking the splitter for a headline with\neach part, as extraction does, would end it; that changes the pass's reply form and so is a new\nversion to measure (`TODO.md`).\n\n**What it does to his sitting.** The sitting for machine-cut parts is built on each request\n([pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) § 6b), so it\nholds 111 parts now: the 45 replaced ones, which stay as the cases with known faults\nthat any checker must catch; the 6 cut on the live site by a stand-in model; and the 60\nfresh ones. Version 2 on the model it was measured on had no parts in the sitting and now has\n60, about the sixty that § 6b says a proof of it needs. The page does not tell a reviewer\nwhich parts were replaced, so his label on one is not led by that.\n\n**Every call of the run is in the record of machine judgments**\n([storing-the-record-of-machine-judgments.md](storing-the-record-of-machine-judgments.md) § 7a),\nrows 5 to 29, each asked of and answered by `claude:claude-sonnet-5`. It was the first batch\nof real work the record held, and the diagnosis above was read out of it: the placeholder answers\nwere found by listing its rows, and the sentence that explains them was sitting in a reply.\n\n---\n\n## 4. Proposals: sturdy microtasks (unruled)\n\n1. **One definition per key, in all three places.** Rewrite the sitting's key texts from the\n   keeper's prompt, and route by what each key means: *contains* as entailment, not as bundling.\n2. **Derive the keeper's key from four yes/no questions** instead of asking for it directly: does A\n   entail B? does B entail A? can both be true? do they share content? Each is closer to a checkable\n   question than the seven-way choice.\n   *(2026-10-06: a fixed table from the four answers to a key, and two changes it needs, in\n   [what-a-record-linkage-project-teaches-deliberus.md](what-a-record-linkage-project-teaches-deliberus.md) § 4c. The fourth question is asked as *is there a statement both assert*, and\n   relatedness is asked once, by the relation classifier. Ledger Z4.)*\n3. **Two rules and one test for the splitter**: rewrite for a stranger *in the parent's own words\n   only*, with no example, definition or name the parent lacks; content from a *rather than*,\n   *instead of*, *even if*, *even when* or *not* clause keeps its polarity, entering a part as\n   denied, conceded or conditional but never as a flat assertion (`61e048b6abce` part 3 does this\n   right, part 4 wrong); and before any part is written, ask of it *if this were false, would the\n   parent be weaker?*, holding the part when the answer is no. **Built 2026-09-30** (pass version 2,\n   § 3e): both rules in the splitter's instructions, and the question asked of every part by the\n   coverage check, which holds the proposal when the answer is no.\n4. **A deterministic invention law** beside the model check: a part may not contain a number, a date\n   or a proper name absent from its parent.\n5. **Pair the coverage check with the parts judge**: hold a written part when the judge dissents and\n   the part fails the question in item 3. The judge caught all six harmful parts the coverage check\n   passed. Give it a definition for split conjuncts as well, for the two or three faithful parts it\n   misread as evidence. *Since 2026-09-30 the question now runs inside the coverage check itself;\n   pairing with the parts judge as a second framing is still unbuilt.*\n6. **Measure self-consistency** before any human sitting: run the same inputs twice through the keeper\n   and the splitter. A task a model answers differently on re-asking cannot be proven reliable on any\n   sample, and the test costs minutes. *Measured 2026-09-30 for the coverage check: the part test\n   answered 45 of 45 parts the same way twice, the invented list less steadily (§ 3e). The keeper is\n   not yet measured.*\n\n**How this composes with asking in several framings.** The working rules this project's sessions\nrun under say something that sounds opposite: non-determinism is an ally, so ask in several framings\nand merge, because different framings catch different aspects. Both hold, at different levels. The\ncanon line of § 5 asks each question to give a stable answer, which is a property of its\ninstructions; item 5 is itself two framings, the coverage check and the parts judge, and it is the\npairing that caught the flips. Stability within each question, diversity across questions.\n\nEach of these lowers disagreement that comes from the instructions rather than from the claim, which\nis the variance a human sitting should never be spent on.\n\n---\n\n## 5. State\n\n- **Documented** here, with TODO items for § 4 and § 3d under the splitter entries in `TODO.md`,\n  and told as a thirteen-minute walkthrough\n  ([2026-09-29-microtasks-in-broad-daylight.md](../listening/2026-09-29-microtasks-in-broad-daylight.md)).\n- **Done on his word the same evening**: the erased split restored (§ 3d), and the ten older\n  proposals relabelled as answered by Claude Sonnet 5 (*\"Yes, relabel them\"*).\n- **Built 2026-09-30 on his question** (*\"Should we rather not just fix the \"coverage check\"?\"*):\n  the check asks the part test of every part, the splitter keeps polarity and adds nothing the claim\n  lacks (§ 3e, commit `85b65ee`). Live by 2026-10-03 (this line said *not deployed* until\n  2026-10-05); production splits on Gemini's free-tier chain, where it is unmeasured, and it has\n  written two splits there (§ 3e, the note of 2026-10-05).\n- **Ruled and done 2026-10-06** (*\"Fresh cuts blessed\"*): the seven version-1 splits are replaced\n  by fresh version-2 splits, the old parts superseded and kept visible, and the claims whose\n  proposals had been held were cut with them (§ 3g). That settles the retraction of the six harmful\n  parts. His review of the fresh parts comes later, by his word (*\"We can always review the cuts\n  later after the reviews that's fine\"*).\n- **Open, his to rule** (each also in `TODO.md`): the depth rule's two remaining clauses (ledger\n  T1); whether to build the keeper's § 4 fixes before anyone scores it; and whether presupposed\n  background counts as a part (§ 3e item 4), which decides whether it caps its claim. On the last two, 2026-09-30, VERBATIM: *\"I'm still absorbing\n  everything on both issues, don't feel ready for a decision yet.\"* Both are in the ruling ledger\n  (T5, T6). Both were read again on 2026-10-03 for a\n  drawing he asked for (§ 3f): the background question is two questions, and the keeper's turns on\n  the word *partly*.\n- **Admitted as a canon line, 2026-09-29**: *An error rate belongs to the task as written before it\n  belongs to the model.* Founder: *\"Add it as written (except remove the instructions examples\n  parenthesis, instead just let it say generally that it's about the prompting rules for the machine\n  passes since we want as deterministic/idempotent responses as possible)\"*. Recorded with his\n  extension in [proving-a-pass-reliable.md](proving-a-pass-reliable.md) § 9.\n\n---\n\n## Cross-references\n\n[proving-a-pass-reliable.md](proving-a-pass-reliable.md) (the correction note at its top, § 4's\neighth way a proof misleads, § 9's ruling) · [earning-the-right-to-write.md](earning-the-right-to-write.md) (found words can\nstill be wrong) · [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) (the\nre-store as another instance of work never asked again) ·\n[the-jhu-decomposition-line.md](the-jhu-decomposition-line.md) (the failure its three dimensions do\nnot name) · [claude-code-as-extraction-engine.md](claude-code-as-extraction-engine.md) § 7.7 (the\nplaceholder answer) · [split-review-and-the-founders-intentions.md](split-review-and-the-founders-intentions.md)\n§ 5 · the walkthrough: [2026-09-29-microtasks-in-broad-daylight.md](../listening/2026-09-29-microtasks-in-broad-daylight.md) ·\n[listening-for-the-calculus-and-sturdy-microtasks.md](listening-for-the-calculus-and-sturdy-microtasks.md)\n§ 3 (what others say on making such tasks sturdy, found 2026-10-05; its § 6 reads two of them against § 4 here).\n· [the-evals-loop-and-our-machine-passes.md](the-evals-loop-and-our-machine-passes.md) (2026-10-05: outside support for § 4, item 2. The practitioners' rule for a machine\ncheck is one kind of fault, answered yes or no, and a question of degree replaced by several\nyes-or-no questions)\n"}