{"path":"research/earning-the-right-to-write.md","content":"# Earning the right to write: a trust ladder for machine passes\n\n**Date**: 2026-09-21 · **Status**: **proposal, unruled**, except where a founder ruling is quoted ·\n**Sequel to**: [the-tidying-line.md](the-tidying-line.md) (which found the tidying already ruled and\nthe *showing* unruled) · **Homes of what it builds on**:\n[graph-daemons-design-space.md](graph-daemons-design-space.md) (tiers, enzyme constraint) ·\n[cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md) (the three warnings) ·\n[the-ratification-record-and-the-triage-feed.md](the-ratification-record-and-the-triage-feed.md)\n\n**Origin.** The founder, on being shown that three warnings stand between eager machine work and\nshowing it: *\"I think that there should be an exception here. Or rather, a discernment is necessary.\nWe need to consider that the warnings are actually most relevant to cases where an AI/machine\npre-digests/summarizes/makes a judgment call that goes deeper/has more inferential/analytical\ndepth/scope than what we can be confident results in an 'objectively' good, unbiased job (as would\nhave been judged by universal(ish) intersubjective human judgment, human logical intuition, etc) and\nI very much think accuracy in this way can be increased enough to be relied upon if we make sure\nAI/machine use is targeted to scoped/narrow enough passes/microtasks that we can develop trust in\nover time. I want to start by testing the assumption that such a microtask of splitting claims here\ncan be done accurately and in a risk-minimizing way… We can always mark machine-split claims in some\nsubtle way in order to know their provenance in order to review the accuracy continually/in the near\nfuture after we've let the system do its thing for a little while. The cost of something like this\nwould be rather low and the benefit would be high in my mind, IFF it works reliably, which I suspect\nit will, otherwise we will reconsider. This general approach I also want applied to other\nmachine-vs-human tasks and ratification-dependent processes/workflows etc.\"*\n\n---\n\n## 1. The discernment he is reaching for is sharper than narrow-versus-deep\n\nHis axis is scope: a narrow microtask can be trusted, a deep judgment cannot. That is roughly right\nand it does not survive contact with cases. *Classify this residue* is narrow and must stay\npropose-only; *find every premise this source leaves implicit* is wide and creates nothing a reader\nmust accept. Scope is a proxy for the thing that matters.\n\n**The thing that matters is the project's own core value, and it makes the discrimination\nmechanical: does the machine's output give a reader MORE places to disagree, or fewer?**\n\n- A summary takes twenty addressable claims and returns one paragraph. **Fewer.**\n- A terminus classification takes an open question and returns a settled kind. **Fewer.**\n- A restatement of someone's proposal makes it the anchor everyone argues about. **Fewer** — and\n  this is precisely the DeepMind arm that steered.\n- A **split** takes one sentence nobody can attack in part and returns six parts each of which can\n  be attacked alone. **More.**\n- An auto-connect, an implicit premise surfaced, a critical question opened, a staleness flag: all\n  **more**.\n\n**The warnings all fire on the first group and none of them fires on the second**, and this explains\nwhy rather than asserting it. Interpassivity is the machine doing the scrutiny so the human need\nnot; an output that *creates* objection sites has not removed any scrutiny, it has made scrutiny\npossible where it was not. The measured finding that question-framed assistance is developmental\nwhile finished output attenuates is the same fact stated from the psychology side: a question is\npure addressability.\n\n**And the counter-case is the strongest part of the argument.** An unsplit bundle is not a preserved\nopportunity for scrutiny. Nobody can attack the third conjunct of *\"AI governance should preserve\nbroad human control through bounded machine authority, distributed human capability, and plural\ninstitutions\"* without first doing the splitting themselves, and almost nobody will. **The bundle is\nthe barrier; the split is the invitation.** Leaving it unsplit to protect the reader's engagement\nprotects an engagement that does not happen.\n\nSo the corpus's own conviction supplies the test, and the enzyme constraint agrees with it from the\nother side: a split changes the **rate** at which a move a reader would have made arrives, and\nchanges nothing about what the graph converges to, because the parts were already in the sentence.\n\n### 1a. The test is a gate, never an objective — his correction, and it is not a nitpick\n\nFounder, on reading § 1: *\"Agreed on everything here, except ofc you need to consider the floor of\nseparability/atomicity/decomposability and not just blindly maximize for 'more places to disagree'…\nImportant context should at least be preserved in each splitting.\"*\n\nRight, and the distinction is worth stating sharply because the two are easy to run together.\n**The addressability test decides WHICH passes may enter the ladder. It never decides HOW FAR one\nof them goes.** Maximised as an objective it would license shredding a claim into fragments, each\ntechnically addressable and none of them a claim anybody could hold.\n\nThree bounds already exist and each answers a different half of his worry:\n\n- **The separability floor** (his own criterion, ruled): decompose until nothing separable is still\n  bundled, **and no further**. A part that cannot be affirmed while its sibling is denied was never\n  a separate claim.\n- **Each part stands alone** — the decontextualisation requirement every other pass in this pipeline\n  enforces. This is the context-preservation half: a part that only makes sense while you are still\n  reading the parent is a fragment, and a fragment in the graph is worse than no claim, because it\n  looks like a claim and cannot be argued with.\n- **Coverage**, checked by a separate call because a model must not grade its own omissions. It is\n  what catches the context that went missing rather than the context that reads oddly, and it has\n  already earned its place once: it found a dropped University-of-Washington attribution on the\n  first real run.\n\nSo the full statement is two-sided: **a pass earns entry by raising addressability, and then cuts\nonly to the separability floor with every part standing on its own.** Neither half is optional, and\nthe second is the one an eager machine would drop first.\n\n**Unruled, and the one thing in this document most worth his eye.** It is proposed as the\ndiscriminator for the committable tier, replacing *nothing that mints a claim* with *nothing that\nlowers addressability*.\n\n---\n\n## 2. Three things are already true in the code, and they retire the old objections\n\nThe propose-only tier's stated reasons for keeping decomposition out were that it *\"either authors\ncontent nobody asserted or moves strengths.\"* Both were true when written. Measured 2026-09-21:\n\n| Objection | State |\n|---|---|\n| **It moves strengths** | ⚠ **Corrected 2026-09-21 — it does, and saying otherwise overstated the protection.** What `confirmed` controls is only the **cap**: an unconfirmed machine link returns requirement weight **0.0**, so a weak part cannot drag its whole down to its own level. The part's own strength still reaches the parent, because the decomposition gather does not filter on `confirmed` at all. Measured on three children at 0.30 / 0.80 / 0.80: unconfirmed **0.498**, a person's parts **0.498**, argued-necessary **0.300**. See § 2a. |\n| **Splitting pays, so it is gameable** | It does not. The padding ruling (2026-08-31) made untagged decomposition children weakest-part-governed, so splitting **cannot raise** a parent's strength and the CORE-style padding attack earns zero by default. |\n| **A daemon that writes is an unattributed author** | This one stands, and it is exactly what his marking proposal answers: an attributed machine author is not an unattributed one. **Unbuilt**: a claim carries `claim_kind`, and nothing says *which pass* wrote it at *which version*. |\n\n**One thing was fixed on reading this, 2026-09-21.** The first objection held **by accident**. The\nlist of machine provenances had exactly one entry, `auto_connect`, so the next pass that learned to\nwrite would have been presumed constitutive on its very first row and would have moved strengths\nsilently from the start. It is named positively now — *extraction* and *human* are presumed, and\n**any other named provenance is a proposal until somebody confirms it**. That is the encode-the-axis\nrule rather than a list of the values seen so far, and it is also the *guard that holds only by\naccident* rule: the behaviour was right and rested on a property nobody had chosen.\n\n**One sub-case was nearly got wrong in the fixing**, and an existing test caught it. The first\nversion also dropped a **missing** provenance to zero, which reads as the safe default and is the\nopposite: the cap runs in the conservative direction, so presuming a part constitutive makes its\nwhole look *weaker*, and failing to presume makes it look *stronger* — the flattering direction.\nA missing provenance therefore keeps the presumption, and the residual hazard is named rather than\nhidden: a pass that writes without setting one is presumed on material it has not earned. Measured\nthe same day: **zero** live edges of any argument type carry a null provenance, so this records\nintent rather than behaviour.\n\nSo the remaining objection is the one he already solved in the asking. **The gap between \"designed\"\nand \"allowed\" here is a provenance mark and a review path, not a principle.**\n\n### 2a. What a machine-written part actually does to its parent, measured\n\nFounder, on being told a machine-written part *\"caps nothing\"*: *\"If the newly-split-out part is\nrightfully split-out, then if it's weak, it's weak, if it's strong, it's strong… It should simply be\njudged on its own merits/evidence/subgraph, right? Even when it's machine-split/minted I'm open to\nletting energy flow through it. Is that blocked, as the system currently stands?\"*\n\n**It is not blocked, and my earlier sentence implied it was.** Two quantities ride on a\ndecomposition edge and I ran them together:\n\n| | What it is | Does confirmation gate it? |\n|---|---|---|\n| **The part's own energy** | its strength, computed from its own supports, attacks and votes, reaching the parent | **No.** `gather_badge_inputs` selects every non-superseded `DECOMPOSES_INTO` child and never reads `confirmed` for this. |\n| **The cap** | how hard a *weak* part drags the whole down toward its own level | **Yes.** `edge_requirement_weight` → 0.0 for an unconfirmed machine link. |\n\nMeasured on one parent with children at 0.30 / 0.80 / 0.80:\n\n| Requirement weight | Parent |\n|---|---|\n| 0.0 — machine-written, unconfirmed | **0.498** |\n| 0.5 — a person kept the parts | **0.498** |\n| 1.0 — somebody argued the whole requires them | **0.300** |\n\nSo his instinct is already the shipped behaviour: a machine-split part is judged on its own merits\nand what it earns reaches its parent. And the presumed level barely differs from zero here — the cap\nonly bites where it would bind, which is why the argued-requirement row is the one that moves.\n\n### 2b. RULED 2026-09-21: a split part caps like a person's, and what a wrong split looks like\n\nFounder: *\"Yes it should. We should expect good splits, and I have some trouble imagining wrong\nsplits, what would that even look like? I guess in some cases they could lose some words/context\nthat is important to the claim being specific enough? What else could go wrong?\"*\n\n**Shipped**: `split` joins the presumed provenances, so a part cut by the splitter caps its whole at\nthe same modest weight a person's part does. Two things still presume nothing — a pass nobody has\nblessed, and `auto_connect`, which *guesses* that one author's claim is part of another's rather\nthan cutting one sentence into its parts. Different act, different risk.\n\n**And the answer to his question comes from the real output rather than from imagination.** Reading\nthe first four stored proposals, three failure shapes are visible, and only one of them is the one\nhe guessed:\n\n1. **Loss, which he named.** The Neumark–Wascher proposal came back coverage `False`, and the\n   dropped item was *\"the University of Washington study of Seattle\"* — the parent identified\n   Jardim et al. as the University of Washington study and no part carried that attribution. Every\n   part true, the provenance of a study gone. **Already caught**, by the separate coverage call.\n2. **Invention, which he did not name and which is the dangerous one.** The occupation-law proposal\n   came back coverage-clean and still reported two inventions: the parts distinguish self-defence\n   law *\"being capable of applying at all\"* from being *\"triggered only by attacks attributable to\n   another state\"*, and add *\"customary-law necessity and proportionality requirements\"* — doctrinal\n   specificity the parent never carried. **A wrong split that reads as a better one**, because each\n   part is sharper and more expert-sounding than the sentence it came from, and nobody said it.\n3. **Drift into commentary**, watched rather than asserted: two parts of the death-penalty proposal\n   read as the author's argument *about* the claim (*\"lacks evidentiary support\"*, *\"it is plausible\n   that the public's horror arose first\"*) rather than as parts of it. Coverage called it clean, and\n   whether the parent carries that material was not checked here.\n4. **An alternative written as a required part** (added 2026-10-05, from a split on the live graph;\n   narrowed the same day: the strength code can hold alternatives as parts of which the strongest\n   decides, and the splitter tags none that way, so each counts as required). A claim\n   that says *\"a positive or necessary outcome\"* asserts neither alternative alone, and its split\n   holds one part for each, stated as fact. Every word is the claim's own, nothing is lost and\n   nothing invented, so neither coverage nor the part test objects: losing an alternative does\n   leave the claim weaker. Three of five models made this cut, and the splitter's instructions have\n   no rule for it ([pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md)\n   § 4b, ledger Y7).\n\n**So the cap follows the coverage verdict, and that shipped with the ruling.** A part from a\nproposal whose coverage reported invention presumes nothing, whoever wrote it. Loss is visible to a\nreader and already instrumented; invention is the shape that makes a split look better than its\nparent, and an invented part capping a real claim is the one way a bad cut does damage rather than\nmerely adding a claim nobody wanted.\n\nThe pass's own flag for a fourth shape is working: the minimum-wage proposal marked\n*\"represents a modern, causally rigorous research design\"* as a fused value judgement, which is a\nverdict wearing a part's clothes.\n\n---\n\n## 3. The warning that genuinely survives, and why \"review it later\" is not enough\n\n**The ratification tether** bites: every propose-only instrument is safe *because* a ratifier can\ntell a good proposal from a plausible-wrong one, and the failure shape is that volume metrics stay\nhealthy while validation quietly goes surface. Auto-committing more machine structure grows the pile\nnobody has read.\n\nHis answer — *\"review the accuracy continually/in the near future after we've let the system do its\nthing for a little while\"* — is the right shape and the wrong mechanism, by this project's own\nstanding rule: **an observe-then-decide plan has a review date, and the date does not live in a\nsession report.** A diary entry for *check whether the splitter is still good* is the exact artifact\nthat came due at two weeks and sat unread for eighteen.\n\n**Said plainly, because the name is doing no work.** *Propose, then ratify* is only safe while the\nperson doing the ratifying can still tell a good proposal from a plausible-but-wrong one. The worry\nis that this ability erodes — through AI use generally, not through anything here — and then\nratification quietly turns into rubber-stamping. **The failure looks exactly like success**: the\ncounts stay healthy, everything is marked confirmed, and the graph fills with structure nobody\nactually checked. A machine that writes more makes that pile grow faster.\n\nFounder, reading the above and arriving at the answer before finishing the sentence: *\"yes I'm\npretty sure I agree that 'review it later' is worse than simply weaving in the reviewing naturally\nin the system for users to ratify — is this what you're trying to say?\"* Yes, and with one honest\nqualification below.\n\n**The mechanism that needs no diary: every human touch of a machine-written unit is already a\nlabelled datapoint.** The split page ships keep / reword / drop today. Keep is agreement. Reword is\npartially wrong. Drop is wrong. Splitting it differently is a wrong cut. Nobody is asked to review\nanything; the review is what using the thing produces.\n\nThat is a **correction record**: one row per machine-written unit, which pass wrote it at which\nversion, and what the next person did with it. It measures the human-plus-machine system rather than\nthe model alone, which is the quantity that will actually run — the same argument that made the\nreview flow the instrument in the sameness sitting.\n\n**The qualification, and it matters.** Weaving review into use *measures* the tether; it does not\ndissolve it. If people rubber-stamp, the correction rate looks superb and the pass gets promoted on\nfalse evidence — the same reason the sameness sitting mixes in seeded wrongs. The general version\ncannot use that trick, because planting fakes in live structure is deception by design, and the\ncorpus already ruled it out in favour of measuring the machine's **natural** mistakes by **time to\nfirst objection**. A pass whose real errors sit unchallenged for weeks is not a pass with a good\ncorrection rate; it is a surface nobody is reading, and the two are distinguishable only by that\nclock.\n\n---\n\n### 3a. Who the reviewers actually are, and what that changes\n\nFounder, 2026-09-21: *\"In practice, it will most likely be me and other semi-deeply-invested\nnerds/experts/users I recruit for this purpose that would go through the claims graph DB broadly and\nsearch for unsplit/decomposable claims and try to systematically ensure that the system works\nsmoothly without constant human ratification/revision/fixing.\"*\n\n**That population largely retires the rubber-stamping worry — by selection rather than by\nmechanism.** The tether's failure mode needs diffuse, low-investment ratifiers clicking through; a\nsmall recruited group doing a deliberate sweep is the opposite, since they are looking *for*\nmistakes, which is what the seeded-wrongs trick exists to simulate. So the measurement stays,\ncheap and quiet, and is simply not expected to fire. It matters that it stays anyway: the retirement\nis a property of who shows up, and it expires the moment the population widens.\n\n**A different risk takes its place, and it is the one his sentence implies.** A handful of recruits\nwho share his intuitions produce high agreement that measures *their agreement with each other*\nrather than accuracy — same-hand bias with people instead of models, where two instances of one\nframe are one detector sampled twice. The friends layer of the sameness sitting exists to measure\nexactly that divergence, and a like-minded panel drives it toward zero while looking like\nvalidation.\n\n**Consequence for recruiting — sharpened 2026-09-21 by his challenge**, *\"I think logic would reign\nhere. I don't expect different cuts to be a matter of taste, this is probably more a science than an\nart?\"* Mostly right, and the split is worth making precisely:\n\n- **Whether two things are separable is logic**, not taste: can one be affirmed while the other is\n  denied? Two careful people should agree, and where they do not, one of them is wrong.\n- **How far down to go is neither logic nor taste — it is a function of purpose**, and that is where\n  the measured record says agreement is not free. The closest measurement the corpus holds is sense\n  granularity: human agreement **56% fine-grained against 91% coarse**, with no reliable middle for\n  96 of 191 words — which is what set WordNet's and OntoNotes's granularity in the first place. And\n  the decomposition literature says the same from the other side: **every verifier has its own\n  atomicity optimum**, so there is no single right grain independent of what the split is for.\n\nSo divergence between two careful reviewers is **not a difference of taste and not noise — it is\na signal that the purpose was unstated**, and it is cheap to measure. That is a better reason to\nrecruit for it than the one written first: not *find people with different intuitions*, but *use\ndisagreement to find the places where nobody has said what the split is for.*\n\n**And the surface he describes does not exist.** *Go through the graph broadly and search for\nunsplit claims* is a corpus-wide sweep; today the split page works one claim at a time and the\nsitting works from a pre-built list. The blocker is the **ranking**: measured 2026-09-21,\n**1,530 of 1,824** unsplit claims trip the looks-bundled heuristic. A reviewer handed 1,530 rows has\nbeen handed the corpus back.\n\n⚠ **I first wrote that the 84% meant the heuristic was too loose, and the founder asked the right\nquestion: \"Do you have reason to suspect the looks-bundled flagging is wrong?\" No, and the\ninference was mine rather than a finding.** Measured: of the **11** claims the reader has actually\nread, **0** came back *already atomic* — no false positive has been observed at all. And 84% is\nplausible on its face, because claims extracted from prose usually are bundled, which is what this\nproject believes anyway: nothing is permanently atomic, so a flag firing on most claims may be\ndescribing the corpus rather than misfiring.\n\n**The correct statement is narrower and survives: a flag is not a ranking.** Even a perfectly\nprecise flag that fires on 84% leaves a reviewer no order to work in, and ordering is what a sweep\nneeds. What is genuinely unmeasured is the flag's precision on the *thin* end — **167 claims are\nflagged on a single conjunction and nothing else**, and 246 on length alone — where a false positive\nwould live if one does. The 11 read so far were taken longest-first, which is the easiest slice.\nCheap next check: read a dozen single-reason claims and count the atomic verdicts.\n\nThe natural ranker is the proposer's own output, which already carries what a sweep needs: how many\nparts it found, whether any is a fused value judgement, and whether the separate coverage call\nflagged a drop. A claim whose proposal is six parts with two flagged is a better use of a reviewer's\nminute than one with two clean parts, and neither is visible from the heuristic alone.\n\n## 4. The ladder\n\nThree states per pass, each with a countable entry condition, none of them a judgement call:\n\n| State | What the pass may do | How it gets there |\n|---|---|---|\n| **proposing** | Nothing enters the graph until a person acts on it. Where every pass starts. | default |\n| **writing-marked** | It writes, marked with its own name and version, and every surface shows the mark beside a one-click correction. | a measured agreement rate over a stated minimum number of human touches — **or a founder ruling, which is how the splitter got there** |\n| **trusted** | Same writes; the review becomes sampled rather than exhaustive, and the triage feed stops routing attention there. | a much longer record at that rate |\n\n**Demotion is automatic and needs no meeting**: a rising correction rate drops a pass a rung, which\nis the rejection-rate pause the daemon design already specifies, applied to a pass instead of a\ndaemon.\n\n**The numbers are not mine to set and should not be guessed.** They come out of the first review\nsitting, the same way the sameness bars do. What is proposed here is the shape.\n\n**The arithmetic those numbers come out of** (2026-09-29): about 60 blind, randomly drawn items with\nnone wrong bound a pass's error under 5% at 95% confidence, about 150 under 2%, and a sequential\ntest keeps the bound true with a trickle; review never has to keep pace with volume, because what\ncompounds is depth ([proving-a-pass-reliable.md](proving-a-pass-reliable.md) §§ 1–2).\n\n**Two invariants the ladder never relaxes, whatever a pass's record:**\n\n1. **Nothing a person wrote or confirmed is touched.** Untouched machine structure is the only thing\n   in scope, which is the 2026-09-13 tier boundary exactly as ruled.\n2. **A pass that lowers addressability never enters the ladder at all** — no accuracy record earns a\n   summary the right to write. That is § 1's test doing its work.\n\n### 4a. The splitter went to rung two on 2026-09-21, and the mark is not what this document said\n\nFounder, reading the claim that the machine may propose and may not write: *\"I thought we (in last\nsession) ended up in a place where the discernment here was different? That maybe it makes sense to\ntrust the splitting and we can as you mention always mark them subtly to be able to reverse if need\nbe, although I suspect we ultimately can trust the splitting-machinery?\"* — and then, asked whether\nto wire the mark before or after a review sitting: ***\"Wire and switch first, sitting later.\"***\n\nHe was right about where the last session landed. Propose-only was where the **code** sat; § 2 of\nthis document had already concluded that the gap between designed and allowed was a provenance mark\nand a review path rather than a principle, and the sentence that said otherwise was mine.\n\n**Three things the build found that the design above got wrong or left out.**\n\n1. **`confirmed=false` is the wrong mark, and the ladder said to use it.** That field already means\n   *a person looked and refused* — `presumed_requirement_weight` reads `False` as exactly that, with\n   a test named `test_an_explicit_refusal_outranks_any_provenance`. An untouched machine edge must\n   not borrow a refusal it never received, so a written part leaves `confirmed` **unset**, which\n   both keeps the cap his 2026-09-21 ruling gave a split part and leaves the refusal flag free for\n   the *drop* action to spend. Nothing needed reordering; the two meanings simply must not share a\n   field. Measured before assuming: **zero** live edges carry a presumed provenance with\n   `confirmed=false`, so the distinction governs only what the splitter writes from here.\n2. **The splitter may write what it FOUND and not what it ADDED.** A split earns the right to be\n   written because its parts were already in the sentence — that *is* the addressability test's\n   justification — so a proposal whose separate coverage call reports **invention** is authorship\n   rather than separation, and holds for a person. Loss does not hold: a part that drops something\n   is still a true separation of what it carried, and the dropped item is now shown with a button\n   that hands it to the write-your-own field, which is *no gap displayed without a move offered*\n   applied to a standing violation on this very panel.\n3. **The invention rate is far higher than the first four proposals suggested.** Run over all\n   eleven stored proposals, **four hold for invention and seven write** — 36%, against the 25% the\n   first sample showed. So the refusal in (2) is load-bearing rather than theoretical, and the one\n   number worth watching as the sitting approaches is whether it stays near a third.\n\n**What a written part carries**: `provenance='split'` and the coverage verdict on the edge (which\nis what stops `support_semantics`'s invention rule being a guard that never fires — measured that\nmorning at **0 of 150** decomposition edges carrying any verdict), the pass name and version on\nboth edge and child, and attribution to **whoever the parent was attributed to**, because the\ncontent was theirs and only the cutting is ours.\n\n**And the correction record is the edge itself.** `human_action` ∈ kept · reworded · dropped, with\nits timestamp and who did it, so the correction rate and *time to first objection* are one query\naway. Leaving a part alone writes nothing, because silence is not agreement. ⚠ **Measured 2026-09-28: the page never\nshowed these buttons.** It kept a stored split only if its status was `proposed` or it could still be asked for, so\nevery `written` split was discarded on load; from the build to the fix, *time to first objection* could not start.\nFixed in `web/src/lib/split.ts`, with a test. Dropping supersedes\nrather than deletes, per the 2026-09-13 ruling that the old link stays visible.\n\n**⚠ Found words can still be wrong (2026-09-29).** Writing what it FOUND and holding what it ADDED\ndoes not catch the error that matters most: a part built from the sentence's own words that states\nas fact a clause the claim denies or concedes. Six of 37 written parts do this, two of them clearly,\nand all six came from coverage-clean proposals, so all six cap their claims, backwards. The parts\njudge caught them and the coverage check, as then written, could not\n([microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) § 3). **Fixed 2026-09-30**: the\ncheck now asks of every part whether the claim would be weaker were it false, and the splitter keeps\neach clause's polarity and adds nothing the claim lacks (§ 3e there). The seven splits written before\nthe fix are still in the graph.\n\nOne thing deliberately **not** built: the expensive enrichment (scheme detection, critical\nquestions) does not run on a written part, because it costs a model call per edge and a part with\nno critical questions still reaches its parent through the subclaim walk. `scripts/write_splits.py`\nis the operator sweep.\n\n---\n\n## 5. Why this generalises, which is what he actually asked for\n\n*\"This general approach I also want applied to other machine-vs-human tasks and\nratification-dependent processes/workflows etc.\"*\n\nIt generalises because the review instrument already exists in general form. **A sitting** —\nmachine proposals, a keyboard flow, seeded wrongs so agreement can be told apart from\nrubber-stamping, a blind subset so anchoring is measured, and a score that sets no bar — was built\n2026-09-21 for claim sameness and is not specific to it. Point it at a different pass's proposals and\nit measures that pass.\n\nSo one instrument, one record, one ladder, and every machine pass in the system takes the same road:\n\n| Pass | Addressability | Today | Next |\n|---|---|---|---|\n| **split proposer** | raises | **writing-marked since 2026-09-21** (below) | its own sitting, then the record it is now producing |\n| **auto-connect** | raises | writes already, `confirmed=false` | the entailment re-check; a sitting of its own |\n| **cross-source premises** | raises | propose-only, operator-gated | a sitting |\n| **staleness daemon** | raises (a flag is a question) | flag-tier | already effectively rung two |\n| **claim merge** | **lowers** — two objects become one | designed, gated | the sameness sitting, and it stays gated on a person's word per merge until the record is long |\n| **terminus / horizon reading** | neither: it is a reading, not a write | computed, dated, conservative | unchanged |\n| **synthesis** | **lowers** | inspectable, ledgered | never enters the ladder |\n\n---\n\n## 6. The first test, which is runnable now\n\nHe asked to start by testing whether the split microtask can be done accurately. The honest state:\n**11 claims, 64 parts, 9 flagged as fused value judgements, and one coverage catch** — a dropped\nUniversity-of-Washington attribution the separate coverage call found. That is the instrument\nworking and it is **n = 11**, which licenses nothing.\n\nThe run that would license something is small: propose over ~60 more claims (two calls each, minutes\non the operator backend), then review them in a keyboard flow with three questions per part — *is\nthis a real part of the parent · does it stand alone · did anything get dropped* — with the coverage\ncall's verdict shown beside. Same shape as the sameness sitting, most of the code shared.\n\n**One quality dimension deserves naming before anything is switched on.** Of decomposition's three —\ncoverage, coherence, atomicity — coherence and atomicity are checkable from the parent sentence, and\n**coverage fails silently**: nothing downstream can see what was never written down. It is the only\nplace a marked, reviewable split can go wrong invisibly, which is why the separate coverage call\n(a model must not grade its own omissions) is not optional on any auto-committed split.\n\n---\n\n## 7. What is separately broken, found while answering\n\nThe founder, looking at the first pair of the sameness sitting — two renderings of the same reported\ncriticism of basic income, both attributed to the same author, one *\"(representing a common worry)\"*\nand one *\"(attributing it to critics)\"* — asked whether reported speech is not caught automatically,\nand whether these predate the pass.\n\n**They predate it by four and a half months** (minted 2026-03-31; the pass shipped 2026-08-15/16).\nAnd the pass has barely run: **1 `REPORTS` edge exists in the whole graph**, and the stance field\ndesigned in the stance-guard work is on **0 claims**. So the always-mint machinery is built and\nswitched on for new extractions and has never been run over what was already stored.\n\n**Both ran 2026-09-21** (*\"Please do the reported-speech backfill now.\"*), and the numbers are a\nfirst base rate for a pass that will go on the ladder:\n\n| | |\n|---|---|\n| Substantive claims | 1,859 |\n| Flagged by the pre-filter, read | **153** |\n| Minted | **46** |\n| Refused | 107 |\n| Failed | 0 |\n| `REPORTS` edges | **1 → 53** |\n\n**A 30% mint rate is the pre-filter working as designed, not a poor one.** It errs toward reading:\na false flag costs one model call and is refused by `validate_split`, while a missed report is a\nposition nobody can attack. The refusals break down the same way in both runs — most are the model\njudging the sentence not reported speech at all, then sentences below the confidence bar, then\nfragments the validator caught. And the second run's rate was far lower than the first (2 of 39\nagainst 44 of 114) because candidates are read longest-first: a long sentence is likelier to carry\na report, so the tail is thin by construction.\n\n**The refusals are recorded now**, keyed on the claim's text, so a repeat run reads only what is\ngenuinely new and a reworded claim gets a fresh reading. Without that every later run would have\nspent 107 calls to reach verdicts already written down.\n\nThe pre-filter gained the signal his question exposed: **the attribution string often says outright\nthat the claim is a report**, and reading it costs nothing. It adds 30 claims the syntax signal\nnever saw — *\"ACLU (quoting Justice John Marshall Harlan)\"*, *\"Justices Stevens and Breyer (cited by\nACLU)\"*, *\"longtermism (reported by the author, who rejects it)\"*.\n\nOne item remains, and it is the root fix rather than a symptom watch:\n- **The stance field**, designed and unbuilt: one enum on the extracted claim, *asserts · rejects ·\n  reports*, which is the root fix rather than a symptom watch.\n\nBy § 1's test both raise addressability, so both belong on the ladder's road rather than in a queue\nof things a person must do by hand.\n\n---\n\n## Cross-references\n\n[the-tidying-line.md](the-tidying-line.md) (the sweep this answers) ·\n[graph-daemons-design-space.md](graph-daemons-design-space.md) (the tier table this proposes to\nre-cut, and the rejection-rate pause the ladder borrows) ·\n[cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md) (the three warnings, and\nthe ratification tether that survives) ·\n[the-jhu-decomposition-line.md](the-jhu-decomposition-line.md) (coverage, coherence, atomicity; the\nseparate coverage call) ·\n[soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md)\n(the sitting's protocol, which generalises) ·\n[the-remaining-case-for-the-stance-guard.md](the-remaining-case-for-the-stance-guard.md) (the stance\nfield, § 7) ·\n[the-evals-loop-and-our-machine-passes.md](the-evals-loop-and-our-machine-passes.md) (2026-10-05: what the ladder's *measured agreement* would be measured with. Two\nrates against a person's labels on a sealed share, and for automatic demotion, a live sample with\nits margin of error) ·\n[what-a-link-rests-on.md](what-a-link-rests-on.md) (2026-10-07: the discriminator applied to links\nbetween texts. A named further claim under a link gives a reader one more place to disagree, and a\nmachine's verdict on the link gives none; four ways of having a machine judge a link moved with the\nquestion, so the proposal there is a pass that names and never judges)\n"}