{"path":"research/quality-control-for-deeper-steps/corpus-bearings-people-and-what-checks-cannot-see-as-a-reader-returned-it.md","content":"# People, showing, and what no check can see (2026-10-10)\n\nOne of three read-only readers (Sonnet 5.5) sent on 2026-10-10, the hour the founder ruled on how deep the\nmachine may write under a link: *\"It may and should write deeper steps. But the microtasks in charge of\ndoing so need to be quality-controlled, just like the other microtasks that are in play in Deliberus. Ground\ndeeply on this. Use the corpus graph to the best of your ability to make sure we keep track of all\nconnections both known and unknown to the rest of the ideas in the corpus.\"* (ledger AG6).\n\nThis reader's part: who checks machine-written steps and whether that checking can be relied on, what being shown a machine's reasoning does to a reader, which errors no check by a machine or a sample can see, and when writing the unsaid is unwanted. It was told to read\n[the-chain-inside-a-link.md](../the-chain-inside-a-link.md) in full first, to ask the corpus graph about\nevery note before opening it, to change nothing and to call no model. It made 116 tool calls. The three\nbriefs are in [the-three-briefs.md](the-three-briefs.md). Its report is below, unchanged, except that\nthe machine's own path to the repository is cut to the path inside it.\n\nWhat was done with it: [../quality-control-for-deeper-steps.md](../quality-control-for-deeper-steps.md)\n§ 11 says which of its findings were checked against the files and what each changed. A finding is a\nreader's, and stays a hypothesis until both passages have been opened. What was kept is in\n`../corpus-graph/found.jsonl` (lines whose `where` names this file).\n\n---\n\nAll paths are under docs/research/ unless marked. \"Note\" means the-chain-inside-a-link.md. I wrote and changed nothing.\n\n## 1. Bearings (not related in the note first, then partly)\n\n| # | Passage (file:line, quote) | Bears on | In the note? | What it would change |\n|---|---|---|---|---|\n| 1 | cognitive-commons-and-deliberus.md:12, :33 \"followed the advice anyway\" (80.7% had detected the error); \"rubber-stamping well-formatted proposals\" | Note :402 \"No person has read a written-out step\"; AG6 | no | A person's check decays to surface validation while counts stay healthy. Never record \"shown\" or \"not contested\" as \"checked\". Ledger R11 (session26-open-rulings.md:431): \"A tidy graph looks examined.\" |\n| 2 | split-review-and-the-founders-intentions.md:563-564 \"the panel did not render for a week, not that nobody objected\" | AG6 QC | no | The instrument that shows whether anyone looks needs its own liveness check. CLAUDE.md:559: a guard that never fires looks like one that works. |\n| 3 | outside-reading-dependable-small-tasks.md:130, :111, :84 \"anchoring gaps of 19–26 points\"; \"wrongly admits a mistake on 98% of questions\"; \"Leniency moved 28.9 points with a detailed prompt\" | (d); AG5, AG6 | no | The checker gets one call, minimal context, no writer's argument, no earlier verdict, no \"are you sure\". Acceptance rate is measured across wordings and models (rules 5, 6, 11, :347-353). |\n| 4 | the-motte-and-the-periphery.md:261-262 \"instructional prompts raise every model (a task that carries a false assumption, rather than a question about it)\" | AG5 \"write the first level\" on a link already drawn | no | \"Write the step under this link\" is the instruction form. Ask \"no relation\" first (how-far-a-premise-is-from-its-claim.md:301). Write the step for the link and for its opposite, then diff (:287-292; outside-reading-machines-naming-premises.md:490). The test design is inferred. |\n| 5 | ux-principles.md:143 \"knowing you *could* check increases trust even when you don't\"; :141 \"'One of six used' is a score wearing a disclosure's clothes\" | AG1 ratio (Note :154-155); hidden by default | no | Hidden steps raise trust unopened (inferred). Mark kinds (\"nobody has opened this\"), not ratios. AG1's \"1,568 of 1,569\" needs a ruling against :141 (also specs/map-view/design.md:233). |\n| 6 | does-descending-change-the-descender.md:120-121 \"a reader of a beautifully decomposed graph gets none of this\"; :171-175 entrenchment | AG3 \"how\" question; Note § 4.4 | no | Mechanism humbles the person who generates it, not the reader of a machine's. A value claim split into reasons is \"reason-enumeration\" and may harden; no instrument sees it. |\n| 7 | scaffolding-versus-difficulty.md:37 \"the platform may always show the door and must never walk through it for you\" | AG6; Note § 4.5 | no | A written step (\"what ties the case to a rule, and the rule\", Note :289) says what to conclude. Show it after the reader's own try (cognitive-commons-and-deliberus.md:29). Tension with CLAUDE.md:556 \"No gap displayed without a move\" is my inference. |\n| 8 | the-tidying-line.md:161-163 \"provided it arrives as an invitation rather than a verdict\"; :218-219 signed-out visibility unruled; TODO.md:1669 \"Raised three times, never ruled\" | AG1 \"Every link shows its state\" | no (Note cites kartpaus-auto-enrichment-design.md, not this) | Rule on form, not volume. specs/map-view/design.md:15-16 \"as few boxes and as little text as possible\"; :342 \"machine-made versus person-made are not on the box\". A link state should be a mark on the line, as the dotted path is (:123-124). That is inferred. |\n| 9 | decomposition-axes.md:98 \"gating on contestation gates on the rarest behaviour in the system\"; engagement-gradient-priors.md:27 decomposers \"fraction of 1%\" | Note :271 \"Shown on contest\" | partly (B2 used for what to write, :58, not for who checks) | If people check only what they contest, the checkers are the rarest tier. QC needs a sampled audit counting steps \"eligible, sampled and actually examined\" (machine-work-as-a-standing-property.md:325-328). |\n| 10 | what-human-judgment-is-for.md:33 \"a human anchored on a machine proposal is not an independent generator; they are a second sample of the machine's framing\" | Note :302-304 (person asked first), § 4.5 | partly (asked-first yes; the checker's anchoring no) | A \"question with a move\" shows the machine's sentence, so a click on it is not independent. Counted checks must be blind (proving-a-pass-reliable.md:242-247; the-parts-review.md:59-61 withholds the proposal in the API). |\n| 11 | Ledger :741 (AD7) \"The verdict follows the question, the order, or the wording of the named claim\"; :739 (AD5) | Note § 4.1 row 3 (:136) | partly (only the 5-of-30 figure, :127) | A step's check cannot be a machine's verdict that it holds. Use checks needing no verdict (AF2, :783: five tests; the fourth failed on 14 of 44 steps) and blind, order-swapped reads. |\n| 12 | the-residual-error-taxonomy.md:71-73 \"because it was written to fit an answer already given\"; :13 \"cannot catch coherent tilt ..., coherent absence ..., or coherent drift\" | Note § 5.2 first level; § 8 :394-399 | partly (§ 8 states the principle) | A step written after the link is drawn can only rationalise (inferred application). Add chain-level checks: stability under re-decomposition (CLAUDE.md:707), proving-a-pass-reliable.md:283-287. |\n| 13 | proving-a-pass-reliable.md:269-279 \"A machine checking a machine is independent only if it fails differently\"; \"the three of the writer's own family all passed the fault and one model of another family caught it\"; writing-out-the-unwritten.md:436-437 \"A reader of the writer's own family takes for granted what the writer takes for granted.\" | Note § 4.1 row 3; § 8 :400 | partly (named, not measured) | Give the checker its own seeded test and measure joint failure with a second family; it \"helps without guaranteeing independence\". Nine judges carried \"about 2.0–2.5 independent votes\" (outside-reading-dependable-small-tasks.md:290). The blessed design leans on the jury (the-ratification-record-and-the-triage-feed.md:212). |\n| 14 | what-a-rule-rests-on.md:809-812 \"a pass writes what it found and holds what it added (13 of 15 written rules stood in neither text, so they were added)\"; earning-the-right-to-write.md:371-377 | AG6; Note § 4.3 \"Made\" row | partly (§ 5.2 :308-310 \"one witness asked four times\") | AG6 lifts \"holds what it added\" for deeper steps. QC must supply what it supplied: cases and counter-cases found in a text or by another family (ledger :761). |\n| 15 | legibility-under-power-red-team.md:64, :72 \"Anyone can detonate the fudge; the platform prices the detonation at one URL submission.\"; \"No concept of load-bearing illegibility.\" | Note § 5.2 :304-306 carve; § 8 :404 | partly (carve covers a person's body, spirit, sacred only) | Collective texts (treaties, coalition texts) are uncovered, and \"a text's own step\" runs eagerly on them. I found no founder ruling on constructive ambiguity. |\n| 16 | the-load-bearing-unsaid.md:47, :93, :120 \"the wrong tool for a hinge\"; sacred lexicon \"zero of ~62,000 words\"; \"0 → 2 firings\" | Note :305 \"nothing uninvited\" | partly | The carve rests on a detector that fires on 2 claims. Ask first for any value step. Give hinge-like steps \"cannot be written\" (older-corpus-bearings-people-and-agreement-2026-10-08.md:27). |\n| 17 | store-the-uses-read-the-meaning.md:296 \"Nothing is guessed into the use.\"; parity-and-the-zone-interior.md:55 \"humans may be wrong about the causes of their own experiences\" | Note :302-304, :404-405 | partly (AA7 is open) | A made step in a person's name guesses a mind. Carry who wrote it (AG5 record number). Their refusal today stands; their account of earlier meaning gets \"no extra weight\" (:304). |\n\n## 2. What a person's check of a machine-written step can and cannot be relied on for\n\nCan:\n- Refuse, today, a rendering of their own position: \"What cannot be wrong is their refusal, today, to stand behind a rewrite\" (store-the-uses-read-the-meaning.md:305-306; what-human-judgment-is-for.md:170).\n- Supply absence, frame and stake no machine pass holds (what-human-judgment-is-for.md:15-16). But a model raised 47 of 47 considerations on 13 arguments, 0 corpus-absent, and cannot see its corpus's absences (:55).\n- Break ties: 18 edges where two machine readers disagreed (the-parts-review.md:43-46).\n- Bound a rate: about 60 blind items none wrong give under 5% at 95%; about 150, under 2% (earning-the-right-to-write.md:337-340).\n- Find kinds of fault before any rate is proven (Y9, ledger :598).\n\nCannot:\n- Stay independent once shown the step: 80.7% (cognitive-commons-and-deliberus.md:12). One item in five is served blind to measure it (the-parts-review.md:59).\n- Stay in practice unused: \"that ability decays when it is not used\" (understanding-after-the-proof-machines.md:117-118).\n- Be many: decomposers \"fraction of 1%\" (engagement-gradient-priors.md:27).\n- Be seeded live: dropped 2026-08-31 (the-ratification-record-and-the-triage-feed.md:199). A seed \"tests attention to known mistakes, not whether reviewer and model share a blind spot\" (proving-a-pass-reliable.md:252-253).\n- Agree on sufficiency: alpha .27 to .47 (outside-reading-machines-naming-premises.md:29). Granularity 56% fine vs 91% coarse (earning-the-right-to-write.md:283-285). No answer key.\n- Stay independent as a pool: recruited reviewers retire rubber-stamping \"by selection\", which \"expires the moment the population widens\", leaving \"same-hand bias with people\" (earning-the-right-to-write.md:262-274).\n- Be helped by a wrong explanation: \"worse off than none\" (writing-out-the-unwritten.md:507-508).\n- Know their own reasons: fallible (parity-and-the-zone-interior.md:55).\n\n## 3. What no check sees, and the only remedy the corpus gives\n\n- What the corpus is a sample of: \"A model cannot know what the corpus it was given is a *sample of*\" (what-human-judgment-is-for.md:15). Remedy: a person other than the author (CLAUDE.md:561), plus a counting instrument and disclosure (the-residual-error-taxonomy.md:38, :388-391).\n- The frame the graph was built inside (the-residual-error-taxonomy.md:109). Remedy: a person positioned differently on the thing, via the reframe invitation (K5/K6 ruled).\n- Tilt and drift of the whole (:13). Remedy: population measures such as disagreement preservation, head/tail balance and re-decomposition stability, not step checks (proving-a-pass-reliable.md:283-287).\n- One family's shared blind spot (writing-out-the-unwritten.md:436-439). what-counts-as-evidence.md:394-397 says \"another agent is not another observation\". Remedy: a case found in a text or by another family (ledger :761); no guarantee.\n- A step's rightness: \"A complete step is not a right one\" (writing-out-the-unwritten.md:541). alien-proofs-and-the-library.md:157-158: \"nothing outside a reader can settle the matter\". Remedy: a located contact (Q2, ledger :338; Note :178-182) or a person's examination.\n- Added rhetorical force: \"Persuasive drift is invisible to it\" (the-scrutiny-gap.md:91). Remedy: none built (:131).\n- Illegibility that does work: none (legibility-under-power-red-team.md:72).\n- Silence: \"may be uncontroversial or may be unexamined\" (steadying-an-outside-judge.md:25). Remedy: the gray band.\n\n## 4. The founder's words, most recent first\n\n- 2026-10-10, ledger :799 (AG6): \"It may and should write deeper steps. But the microtasks in charge of doing so need to be quality-controlled, just like the other microtasks that are in play in Deliberus.\"\n- 2026-10-10, ledger :455 (S8, \"Yes!\"): members \"keep their own position, the ability to check, and the freedom to leave.\"\n- 2026-10-10, Note :11-13: \"We must not make too-large leaps/steps, that would be irresponsible for any scientist. We must examine and make it possible to review each step of the reasoning tree/chain/graph.\"\n- 2026-10-10, outside-checks/README.md:35-36: \"Don't think it's gonna fall for that, if it arrives JUST after coming up with the same content itself? Haha.\"\n- 2026-10-07, Note :62: \"explicated and exposed and scaffolded automagically. To the extent that process can be trusted of course. This depends if we can split it up into microtasks that we can rely on deeply.\"\n- 2026-10-06, CLAUDE.md:554: \"Sure\", to \"A person reads a machine pass before anyone measures it. Rates are measured only on the kinds of fault that reading found.\"\n- 2026-09-29, proving-a-pass-reliable.md:412: \"(or to check the machines' mistakes without being overwhelmed)\". Ledger :432: \"can probably be trusted since (yet if and only if) they are reliable enough on the microtasks in question\".\n- 2026-09-21, earning-the-right-to-write.md:229-231: \"'review it later' is worse than simply weaving in the reviewing naturally in the system\". :355: \"Wire and switch first, sitting later.\"\n- 2026-09-21, the-tidying-line.md:13-16: \"tidying up regularly where it makes the most difference to reach and show examples of more clarity\". Showing to signed-out visitors: unruled.\n- 2026-09-20/21, specs/map-view/design.md:15-16, :174-176, :253-254: \"as few boxes and as little text as possible on each view\"; \"let's avoid all unnecessary text in the UI in general\"; a summary \"needs careful deep thought until it's usable and accurately captures the tail\".\n- 2026-09-15, parity-and-the-zone-interior.md:91: \"Relying on humans stepping in and saying something like that feels rare and shouldn't be counted on.\"\n- 2026-09-13, the-tidying-line.md:36-37: \"Yes, replace by itself, keep the old link visible.\"\n- 2026-08-31, the-ratification-record-and-the-triage-feed.md:169-171: \"I don't think the seeded-probe thing sounds useful and elegant and worthwhile, do you?\" He blessed the no-ratify-act end-state at :167.\n- 2026-08-30, same file :61-62: \"subtly encouraged (but not conscripted/required)\".\n- 2026-08-27, what-human-judgment-is-for.md:4: \"Let's not install meaningless buttons for users to press\".\n- 2026-08-25, kartpaus-auto-enrichment-design.md:7: \"All unstated premises need to be clarified and scaffolded.\" Ruled: do eagerly, \"invisible on the default map\" (:17). parity-and-the-zone-interior.md:47: \"even more risk averse and epistemically humble in the personal/spiritual domain ... But even this should not be artificially protected against analysis ultimately.\"\n- Not found: a founder sentence on who speaks first for a step. \"Ask the author first\" is Claude's (what-a-rule-rests-on.md:804-806). I also did not find his verbatim for \"never decomposed uninvited\"; the brake was ratified 2026-07-06 (weighing-scheme-and-terminus-classifier-design.md:39).\n\n## 5. Contradictions\n\n1. The note vs AG6. Note :38-40 \"The machine writes one level in its own words. Below that it finds ... and it asks\"; :171-172 \"today that is one made level\"; :271, :292 \"made only within the cap\". Ledger :799 says \"deeper steps\", and the AG2 one-level sentence is withdrawn.\n2. AG6 vs \"writes what it found, holds what it added\": earning-the-right-to-write.md:371-377; what-a-rule-rests-on.md:809-812. Not reconciled.\n3. T1 vs AG6: CLAUDE.md:552 \"the measurement caps how deep\"; proving-a-pass-reliable.md:96-98 \"With no measured rate, it may not build ... at all\". Ledger :471 says the no-rate sentence is \"still proposals\". It fits only if QC is the measurement.\n4. Q33 (ledger :369: a second-family machine proposes relevance, which weighs the link) vs AD7 (:741: no machine reading may weigh) vs AD17 (:751: \"weighs nothing\"); what-a-rule-rests-on.md:772-780. A step-checker is such a reading.\n5. Jury as the remedy (the-ratification-record-and-the-triage-feed.md:212-213) vs writing-out-the-unwritten.md:543-546 (\"a second family would buy less than it seems\"), outside-reading-dependable-small-tasks.md:290-293, game-feel-and-the-first-screen.md:283-286.\n6. Forcing engagement: outside-reading-dependable-small-tasks.md:360 (\"make them act on an assumption\") and ux-principles.md:268 (\"must be designed to provoke engagement\") vs the-ratification-record-and-the-triage-feed.md:17-19 (\"conscripts\"), ux-principles.md:439 \"Never force depth\", the-harness-conviction.md:68-70.\n7. Who speaks first: incentives-analysis.md:120 \"let opponents propose your premises\" and what-human-judgment-is-for.md:180-183 (consent, \"Ship it\") vs what-a-rule-rests-on.md:804-806 and Note :302-304 (author first), against what-human-judgment-is-for.md:33.\n8. Hiding depth: curiosity-as-growth-fuel.md:165 \"The mitigation is not to hide depth\" vs kartpaus-auto-enrichment-design.md:17 and Note :271 (hidden by default). Ratio: ux-principles.md:141 vs Note :154-155.\n9. Seeds: the-harness-conviction.md:186-188 (seed, measure the catch rate) is not annotated against the-ratification-record-and-the-triage-feed.md:199 (dropped); the same note keeps it at :241-245 as \"candidate\".\n\n## 6. What the corpus graph held (`around`)\n\n- the-chain-inside-a-link (read in full): AG6 [RULED] names it home; it names T1 and B2 as ruled.\n- what-human-judgment-is-for (full): K5, N8. A 2026-10-08 reader on AD28: anchoring means \"I pay\" may be an echo; ask the author's reason first.\n- the-ratification-record-and-the-triage-feed (full): C8 moot (Q32). Reader F6: the feed never samples the confident region. F7: the silent panel needs a liveness check.\n- cognitive-commons-and-deliberus (full): nothing.\n- fragile-checkers-and-the-verification-bottleneck (full): reader: an \"opponent told to resist\" is a pressured judge.\n- the-scrutiny-gap (full): session: the harness conviction's limit is measured in § 3.\n- steadying-an-outside-judge (full): nothing.\n- the-residual-error-taxonomy (full): J15. Reader F7: a missed merge is not inert.\n- the-harness-conviction (full): session: scrutiny gap § 3.\n- does-descending-change-the-descender (full): it supersedes bias-codex § 4.4. Reader: recognition arriving as structure is \"unmeasured\".\n- the-load-bearing-unsaid (full): readers F5, F2; AD29 needs \"cannot be written\".\n- analysis-and-attunement (full): J9 superseded; F12.\n- parity-and-the-zone-interior (full): O4 struck the grant.\n- engagement-gradient-priors (full): a model \"stranger\" is advisory.\n- curiosity-as-growth-fuel (full): nothing.\n- incentives-analysis (full): F6, F9, AD22.\n- the-tidying-line (full): F5, F9.\n- earning-the-right-to-write (full): F7, F12, F14.\n- understanding-after-the-proof-machines (full): F8, F14.\n- game-feel-and-the-first-screen (full): nothing judged.\n- specs/map-view/design.md (full): `around` says it is not a note or ledger row.\n- ux-principles.md (full): J9, M18.\n- store-the-uses-read-the-meaning (plain-language section, §1, §3d only): F9, F7.\n- what-counts-as-evidence (§0, 2c, 4a, 8c-8d, 8l-8m): AD23 home; F6, F11.\n- the-motte-and-the-periphery (§3-3c): F9, F10.\n- grounded-critics-and-quarantined-imagination (full): nothing.\n- bengio-safety-from-honesty-and-deliberus (full): P5 superseded.\n- what-a-rule-rests-on (§13, §14 full): 51 older notes bear on AD22 to AD29.\n- The three older-corpus-bearings reports (full): nothing.\n- CLAUDE.md: lines 501-575 and 697-711 only.\n- Opened extra: kartpaus-auto-enrichment-design (full; reader F14: T1 vs invisible draft work). scaffolding-versus-difficulty (full; F9, F13, F14). proving-a-pass-reliable (§1, §4, §9). machine-work-as-a-standing-property (§3d). writing-out-the-unwritten (§8b, §9, §10) and the two outside-reading reports (parts): nothing judged.\n- Opened extra, with `around` run after opening: decomposition-axes (§4; B1), legibility-under-power-red-team (Hole 2), the-parts-review, split-review-and-the-founders-intentions, weighing-scheme-and-terminus-classifier-design, pass-quality-and-the-choice-of-model, threads-synthesis-and-gaps (two lines).\n- Opened with no `around` run: alien-proofs-and-the-library (§4.1-4.2, §4.8-4.9; the hook says the founder joined on 2026-10-10).\n"}