{"path":"research/corpus-graph/readers/2026-10-10-small-tasks-and-live-tests-as-a-reader-returned-it.md","content":"# What a reader found to bear on eight notes on small tasks, live tests and the first screen (2026-10-10)\n\nA read-only reader (Sonnet), one of those sent through the notes no reader had yet read, after the founder\nasked for the graph's coverage to be widened (README § 12). It was given the brief in\n[../reader_brief.md](../reader_brief.md) and these 8 notes:\n\n- `docs/research/quality-control-for-deeper-steps.md`\n- `docs/research/kartpaus-auto-enrichment-design.md`\n- `docs/research/live-election-test-design.md`\n- `docs/research/claude-code-as-extraction-engine.md`\n- `docs/research/weighing-detection-semantic-tier.md`\n- `docs/research/islands-of-coherence.md`\n- `docs/research/deeper-democracy-and-the-oscillation-cost.md`\n- `docs/ux-principles.md`\n\nIt made 129 tool calls. Its report is below, unchanged. What was kept from it, after the quoted words were\nchecked by code and both passages of each finding were opened, is in `../found.jsonl` (lines whose `where`\nnames this file).\n\n---\n\n8 of 8 notes read in full.\n\n### F1\ntarget: docs/ux-principles.md:181\ntarget_words: \"The more the community deliberates, the more the system can auto-connect, auto-deduplicate, and auto-suggest.\"\nolder: docs/research/the-reuse-flywheel-prior.md:189\nolder_words: \"The flywheel's shape as measured on the OLD linker**: per-source cross-source connection yield does not rise with\"\nkind: contradicts\nwhy: P10 states the flywheel as fact. The flywheel note measured it on 31 sources and found another shape. Cross-source yield per source does not rise with corpus size. It tracks whether an adversarial counterpart is already mapped (0.44 to 0.94 edges per claim for debate pairs, 0.02 to 0.04 for lone sources). Only 4 concepts span more than one source. The linker was later superseded, so the figures are suggestive. The supported sentence is \"paired, not cumulative\".\nstrength: strong\n\n### F2\ntarget: docs/ux-principles.md:258\ntarget_words: \"you don't need a dedup algorithm if the user never creates the duplicate in the first place\"\nolder: docs/research/claim-sameness-philosophical-readings.md:7\nolder_words: \"rejected as friction at the wrong moment, on the grounds that the person at the door is a poor sameness judge\"\nkind: contradicts\nwhy: The founder ruled on 26 August against nothing-merges-without-explicit-action. The ruled shape is \"anyone types anything, careful auto-merge runs afterward, anyone can split\". He called write-time suggestion too much friction. So the page's claim that user-side prevention removes the need for a dedup algorithm is overruled: the ruled design runs a dedup algorithm after the fact. Retrieval while typing can stay as context, but not as the dedup mechanism.\nstrength: strong\n\n### F3\ntarget: docs/research/islands-of-coherence.md:51\ntarget_words: \"make each island genuinely coherent and let adjacency do the propagation\"\nolder: docs/research/the-reuse-flywheel-prior.md:189\nolder_words: \"The flywheel's shape as measured on the OLD linker**: per-source cross-source connection yield does not rise with\"\nkind: narrows\nwhy: The note says adjacent coherence making the next act cheaper is the measurable mechanism, and that adjacency propagates it. The 29 August measurement on 31 sources shows yield per source does not rise with corpus size. It tracks whether an adversarial counterpart is already present, and only 4 concepts span more than one source. Islands propagate in pairs, not by nearness in general. Suggestive: old linker, small corpus. It changes what \"let adjacency do the propagation\" can promise a funder.\nstrength: strong\n\n### F4\ntarget: docs/ux-principles.md:99\ntarget_words: \"The loop written above is the **answer-a-CQ** loop, and it works exactly as described.\"\nolder: docs/research/what-counts-as-evidence/corpus-map-code.md:405\nolder_words: \"Eighteen confirming answers move C to 0.581. A single edge can never reach\"\nkind: later-correction\nwhy: The page says answering a critical question works exactly as described, ending in a badge colour change. The computed example in this code map (ledger Q27, measured 27 September) shows one answer leaves a lone-edge claim at 0.500, UNSETTLED. A saturated one-question edge stays UNSETTLED for good. Eighteen answers reach only 0.581. Answers are counted, not weighed. The colour changes once, from gray to neutral. \"Exactly as described\" is stronger than the record supports.\nstrength: strong\n\n### F5\ntarget: docs/research/islands-of-coherence.md:175\ntarget_words: \"which renders as a confident blue badge, so at the exact moment two participants have just disagreed, the instrument\"\nolder: docs/research/strength-layer-audit.md:83\nolder_words: \"So the grey `UNEXAMINED` state added today is a restoration, not an invention.\"\nkind: later-correction\nwhy: The sentence predicts that a confident blue badge will hide a fresh disagreement. The strength-layer audit, written the same day (17 August), added a gray UNEXAMINED state for claims with nothing answered, restoring the founder's four-state sketch. The code now returns it for that case. A claim with edges and no answers no longer renders as confident blue. The sentence is out of date. The live gap is now that answers barely move a badge (ledger Q27).\nstrength: strong\n\n### F6\ntarget: docs/research/live-election-test-design.md:400\ntarget_words: \"the badge display for fresh disagreements (confident blue at 0.500 — still unruled, and it ships as-is if it stays so)\"\nolder: docs/research/strength-layer-audit.md:83\nolder_words: \"So the grey `UNEXAMINED` state added today is a restoration, not an invention.\"\nkind: later-correction\nwhy: The decision list, dated 9 September, carries the blue-at-0.500 display as unruled and shipping as-is. The audit added a gray UNEXAMINED state on 17 August, and UNSETTLED followed on 23 August. The code returns UNEXAMINED when nothing is answered. So a fresh disagreement with unanswered questions renders gray. The blocker-table row (line 250) and the TODO item are stale too. What remains open is that a few answers barely move the badge.\nstrength: strong\n\n### F7\ntarget: docs/research/weighing-detection-semantic-tier.md:73\ntarget_words: \"The session then produces, as a by-product, **the corpus's first Swedish weighing dataset**\"\nolder: docs/research/weighing-eval/ground-truth-public.md:3\nolder_words: \"labels written 2026-08-24 BEFORE any detector (lexical or LLM) ran on\"\nkind: later-correction\nwhy: That dataset exists already. Seventy-two wild Swedish election sentences were labelled blind on 24 August, before any detector ran, and adjudicated to 6 weighings and 66 non-weighings. So precision, the half the 4-of-4 demonstration cannot show, can be scored now without waiting for a session. The key also shows that wild weighings are rare (about 8%). It pre-registers where the LLM tier's false alarms should fall. This changes what the election session is needed for.\nstrength: strong\n\n### F8\ntarget: docs/research/deeper-democracy-and-the-oscillation-cost.md:142\ntarget_words: \"Deliberus proposes no rival form of democracy; it adds a layer any form can stand on — the record in which the\"\nolder: docs/research/blobs-community-sandboxes-and-the-advice-process.md:14\nolder_words: \"The advice process is a way of deciding that has no counting step, so its whole quality rests on\"\nkind: same-mechanism-elsewhere\nwhy: The blobs note is a worked design for exactly this layer inside a living non-majoritarian form, the advice process, where nobody votes. It names the same cost at community scale: \"The same fight every year. Reasons live in old threads and in heads\" (line 40). The oscillation cost is that mechanism at polity scale. It also answers the commenter's question about better models with a concrete form already in use, and with a test plan for the layer.\nstrength: strong\n\n### F9\ntarget: docs/research/claude-code-as-extraction-engine.md:112\ntarget_words: \"it falls through to the untouched Gemini chain unless `DELIBERUS_CLAUDE_STRICT=1`. The swap can degrade to today's behaviour, never below it.\"\nolder: docs/research/proving-a-pass-reliable.md:265\nolder_words: \"A proof is a snapshot of one version.** A new model, a new prompt or a new backend resets it:\"\nkind: supplies-a-missing-piece\nwhy: The proving note says the Claude and Gemini backends are different passes, and a measured rate belongs to one version. Fallthrough on error, plus a default alias (line 136) that tracks the current model, means one run can mix passes. The \"never below\" claim holds for availability, not for attributable quality. For section 7.9's strict-versus-fallthrough and model-default decisions, strict mode and a pinned model are what keep a rate attachable to the output.\nstrength: medium\n\n### F10\ntarget: docs/research/kartpaus-auto-enrichment-design.md:34\ntarget_words: \"Normal-use enrichment is the founder's recorded lean, capped by the action price.\"\nolder: docs/research/claude-code-as-extraction-engine.md:26\nolder_words: \"Public extraction on the operator's subscription is the banned shape.\"\nkind: supplies-a-missing-piece\nwhy: The note prices the loop at $6–16 per pause on subscription window limits and says it generalizes to normal use. The engine note records that subscription-run extraction is operator-only, and production refuses the Claude backend in code (line 95). So for ordinary users the loop has no engine and no payer under current rulings. Whose key pays is the unruled question of the 4 October note on whose tokens pay. A builder of piece 1 should read this before treating cost as only a stopping rule.\nstrength: medium\n\n### F11\ntarget: docs/research/weighing-detection-semantic-tier.md:7\ntarget_words: \"The founder's instinct converges with what the corpus already concluded twice in its own words\"\nolder: docs/research/dogfood-run-7-swedish-election-structured-vs-unstructured.md:319\nolder_words: \"Neither gap is fixed by the other, and a semantic tier addresses both where a second word-list addresses one.\"\nkind: same-idea-earlier\nwhy: This is a third earlier arrival, from 17 August. It shows the lexicon also misses 10 of 10 type-eligible claims in English. The same addendum records a founder call that claims enter the graph in English \"this time\" (lines 301–303), with the permanent policy open. Whether a Swedish lexicon is needed for stored claims depends on that policy. Option A's half-day build should be weighed against it.\nstrength: medium\n\n### F12\ntarget: docs/research/weighing-detection-semantic-tier.md:51\ntarget_words: \"Score against the founder's question: the LLM tier found 4/4 with mechanism explanations\"\nolder: docs/research/quality-control-for-deeper-steps.md:345\nolder_words: \"The hand that wrote a rule, imagined the case for it and\"\nkind: same-mechanism-elsewhere\nwhy: The 4-of-4 comes from a synthetic paragraph made for the demonstration and judged by a Sonnet-class model. The case and the judge share a model class. The quality-control note's rule is that a made case counts for nothing: the writer, the case and the check are one witness asked three times. The result shows the model can find weighings in text built to contain them. It is no estimate of recall or false alarms on wild text. It cannot carry the choice among options B, C and D.\nstrength: medium\n\n### F13\ntarget: docs/research/kartpaus-auto-enrichment-design.md:22\ntarget_words: \"Auto-fetched evidence carries a NEW flattening mechanism: selection bias\"\nolder: docs/research/what-counts-as-evidence.md:131\nolder_words: \"Missing entirely**: source reliability separate from the claim; source independence; likelihood\"\nkind: supplies-a-missing-piece\nwhy: Both-sides fetching answers selection bias, not repetition. Bulk fetched items that echo one origin would each count as separate support, because the evidence layer is listed as having no notion of source independence. Fair-share routing is built but unadopted. The sibling note on law and intelligence (section 3.5) documents the same failure, \"many reports, few sources\", and the Iraq case. Builders of piece 3 need an origin check, not only a quality tier and dedup.\nstrength: medium\n\n### F14\ntarget: docs/research/quality-control-for-deeper-steps.md:267\ntarget_words: \"**The worst error in this design is a made sentence marked as found.** It lends a text's standing\"\nolder: docs/research/mapping-the-machine-consciousness-debate.md:411\nolder_words: \"00:43) a public page said a named research group held a claim the source says they did not make.\"\nkind: supplies-a-missing-piece\nwhy: A real instance of that error class, on a public page for about an hour: a position the source says the researchers did not hold was stored under their name. The step that separates who said it from what was said had no way to say \"reported as not holding\". The two rates proposed for pointing (found, wrongly pointed) and the 1,789-claim answer key have no case where the text denies the sentence. A third case belongs in the key.\nstrength: medium\n\n### F15\ntarget: docs/research/live-election-test-design.md:223\ntarget_words: \"**Consent consequence**: the existing line tells participants everything from start to stop is recorded, which covers this.\"\nolder: docs/research/legibility-under-power-red-team.md:202\nolder_words: \"**No public deletion policy, no right-to-recant statement, no data-minimization commitment**, and no answer to jurisdictional erasure rights for reasoning records.\"\nkind: supplies-a-missing-piece\nwhy: The consent line covers recording only. The 9 September plan has the extraction page appear on deliberus.com the moment it finishes, with speakers attributed to named guests (lines 14, 379). The red-team note's open list says the attributed permanent record has no deletion policy, no right to recant and no erasure answer. A person drafting the script would need to say what happens to the guests' words after the session, not only that they are recorded.\nstrength: medium\n\n### F16\ntarget: docs/research/islands-of-coherence.md:93\ntarget_words: \"Polarization has no structural answer**, and pretending otherwise would be exactly the kind of\"\nolder: docs/research/trust-economics-and-false-polarization.md:186\nolder_words: \"Heterogeneous groups** — polarization depends crucially on group homogeneity; include people from different sides\"\nkind: narrows\nwhy: The literature the corpus holds lists design conditions that prevent group polarization: heterogeneous groups, balanced briefing, trained facilitators, random selection. In America in One Room, 22 of 26 extreme proposals moved closer (line 179), and the effect lasted at least a year. A matched, facilitated pair meets some of these. So \"no structural answer\" overstates it. The ledger row E17 is the open flagged call. The pre/post reading stays right, but the claim should be that composition and facilitation are the known answers, not that none exists.\nstrength: medium\n\n### F17\ntarget: docs/research/quality-control-for-deeper-steps.md:742\ntarget_words: \"One family of models all the way down.** A second family for the reader was not used, since it\"\nolder: docs/research/pass-quality-and-the-choice-of-model.md:429\nolder_words: \"Two cheap checkers of another family** on each of 49 claims add about 5 to 9 cents.\"\nkind: supplies-a-missing-piece\nwhy: The note leaves out a second family because it would spend his money. The pass-quality note puts the price at 5 to 9 cents for two cheap checkers on 49 claims, about a fifth of a cent a claim, through one router key that reaches every family. It also records the one fault a checker of another family caught that the writer's family missed. The cost argument against Y4 and against AG10's checkers is therefore small, and he should rule on it knowing that.\nstrength: medium\n\n### F18\ntarget: docs/ux-principles.md:541\ntarget_words: \"One historical question, checked and unresolved**: whether Deutsch and Feynman ever publicly\"\nolder: docs/research/session26-open-rulings.md:201\nolder_words: \"CLOSED 2026-09-09 as unrulable: a claim that cannot be verified to exist has nothing to rule on; the anecdote stays out of every doc.\"\nkind: already-answers\nwhy: Ledger row J16 closed this question on 9 September as unrulable and said the anecdote stays out of every doc. The public page still carries a paragraph on it, with a \"not found\" result presented as an open item. The row records no founder words, so this is a ledger closure, not a quoted ruling. Even so, the paragraph is out of line with it and could be cut to match the closure.\nstrength: medium\n\nNoticed: (1) A stale TODO.md item still lists the blue-at-0.500 badge as open (lines 2243 and 3352) though the code returns UNEXAMINED. (2) The QC note grew from 810 to 867 lines while I read; I used the current line numbers. (3) I avoided pairs already kept in found.jsonl (islands with the-group-premise-evidence; ux-principles with the QC note). (4) Leads left out for space: kartpaus with grounded-critics-and-quarantined-imagination section 3 (machine pre-answers and the strength layer) and with the-tidying-line section 4a (question-framed showing); deeper-democracy with the-post-that-states-the-problem-from-inside section 6 (such posts are read as centrism); QC with law-intelligence-history-journalism-forensics section 3.5 (retraction not propagated); live-election \"ten-question descent\" is now fourteen (wild-weighing-dialects-and-the-question-set line 3); islands section 5 with parity-and-the-zone-interior line 59 (experience reports are exempt from challenge). (5) Process slip: I created and then deleted a helper file /tmp/ms.sh before I recalled the read-only rule; no repository file was touched.\n"}