{"path":"research/corpus-graph/readers/2026-10-10-bias-descent-and-game-feel-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on bias, the descent and game feel (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 11 notes:\n\n- `docs/research/cognitive-bias-codex-and-human-contribution.md`\n- `docs/research/does-descending-change-the-descender.md`\n- `docs/research/game-feel-and-the-first-screen.md`\n- `docs/research/coupling-the-poles.md`\n- `docs/research/what-the-walkthroughs-and-the-reviews-are-for.md`\n- `docs/research/game-feel-mechanics-sketchbook.md`\n- `docs/research/philosophical-foundations-stress-test.md`\n- `docs/research/wild-weighing-dialects-and-the-question-set.md`\n- `docs/research/taxonomy-gaps-and-the-closed-enum.md`\n- `docs/research/reflection-and-the-graph.md`\n- `docs/research/active-inference-context-acquisition-and-deliberus.md`\n\nIt made 119 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\n11 of 11 notes read in full.\n\n### F1\ntarget: docs/research/what-the-walkthroughs-and-the-reviews-are-for.md:125\ntarget_words: \"principles. So the common ground the wager bets on may not be at the bottom. Or the bottom may\"\nolder: docs/research/what-a-rule-rests-on.md:596\nolder_words: \"publics the measured pattern is the reverse. And Toulmin's commission is weaker evidence than it\"\nkind: later-correction\nwhy: Checked the same night, this note withdraws the reading the thought leans on. Sunstein's first kind of agreement is the reverse (principle agreed, cases disputed). Toulmin's commission was mandated and read to work on principles. In large publics the measured pattern is the reverse. Only a deciding body with a shared tradition agrees on a case without a theory. Anyone using \"common ground may sit in the middle\" should keep only that narrower claim.\nstrength: strong\n\n### F2\ntarget: docs/research/game-feel-mechanics-sketchbook.md:105\ntarget_words: \"This is the one mechanic on the list that directly targets the measured asymmetry (supports are\"\nolder: docs/research/lowering-the-cost.md:184\nolder_words: \"People have written 9 supports and 21 attacks, too few to read a tendency from.\"\nkind: contradicts\nwhy: The sketchbook builds idea 8 on a \"measured\" asymmetry between supports and attacks. This correction says it is only a prediction about human contributors. The corpus ratio is near 1.17 to 1 because machines mint both directions, and people have written 9 supports against 21 attacks, too few to read. Idea 8 therefore aims at an untested hazard. The ratio of human-contributed edges is the thing to watch.\nstrength: strong\n\n### F3\ntarget: docs/research/cognitive-bias-codex-and-human-contribution.md:272\ntarget_words: \"any new instrument: for each rater, compare their rigor ratings on claims they agreed with against\"\nolder: docs/ux-principles.md:62\nolder_words: \"Single-axis human voting (agree/disagree only). The QBAF badge IS the quality signal — computed from CQ evidence, not subjective votes.\"\nkind: contradicts\nwhy: Section 4.5 proposes measuring belief bias from each rater's rigor ratings. P2 records that on 31 March 2026 the project chose single-axis voting: only agree/disagree is collected, and quality is the badge computed from critical-question answers. LessWrong's finding that over 95% of quality votes matched agreement is cited there. bridging.md says the two-axis vote was decided against. No human rigor rating exists, so the test cannot run as written. The nearest version uses critical-question answers per rater.\nstrength: strong\n\n### F4\ntarget: docs/research/does-descending-change-the-descender.md:240\ntarget_words: \"discomfort is **correct** and designing it away would silence a working alarm; the answer is an exit\"\nolder: docs/research/session26-open-rulings.md:208\nolder_words: \"a frame gap also deserves attention and to be exposed as a closable gap to be engaging instead of anxiety-inducing.\"\nkind: later-correction\nwhy: The note keeps frame-gap discomfort as a working alarm. On 9 September the founder ruled on the same question (ledger K2 to K4): a frame gap deserves to be exposed as a closable gap, and with a way to contribute the anxiety is short-lived and productive. K3 records this as softened: the anxiety is a signal, the answer is a contribution path, never a soothing. K4 withdraws \"you cannot make a frame gap feel closable\".\nstrength: strong\n\n### F5\ntarget: docs/research/philosophical-foundations-stress-test.md:51\ntarget_words: \"the platform can rebut but cannot undercut. Formats that reify inference steps as nodes have carried undercutting since\"\nolder: docs/research/the-squirrel-and-the-strength-calculus/strength-rulings-and-code-as-read-2026-10-08.md:351\nolder_words: \"an undercutter has no separate edge type. It acts as an answered connection question\"\nkind: already-answers\nwhy: The stress test says the graph cannot undercut and asks for edge-attacks or reified inference nodes. Reading the code in October, this report finds an undercutter has no edge type of its own but acts as an answered connection question, or as an attack on the link's relevance claim. Ledger AD22 and philosophy-of-the-unwritten-step.md:476 say that slot has existed since 28 September. What remains is typing it (ledger Q9, open) and use: no link carries such a claim yet.\nstrength: strong\n\n### F6\ntarget: docs/research/does-descending-change-the-descender.md:74\ntarget_words: \"| `dissolved` | moderation is the *point*; the disagreement was verbal |\"\nolder: docs/research/session26-open-rulings.md:282\nolder_words: \"names an outcome, not a remainder type** — the honest type is *definitional* (the disagreement was about the word;\"\nkind: later-correction\nwhy: The table predicts moderation per terminus type and treats `dissolved` as the case where moderation is the point. On 16 September the founder agreed that `dissolved` names an outcome, not a remainder type, and chose \"definitional\". Ledger O2, O4 and O5 add that terminus became horizon kind, and kinds are read from the arrangement as dated, defeasible readings, never set. A per-type test must now be per dated reading, and for `dissolved` the prediction is circular.\nstrength: strong\n\n### F7\ntarget: docs/research/what-the-walkthroughs-and-the-reviews-are-for.md:113\ntarget_words: \"two neighbours and an oak; Toulmin's commission | two opponents shown where they already agree\"\nolder: docs/research/legibility-under-power-red-team.md:72\nolder_words: \"A system whose founding wager is that decomposition reveals shared care has no category for the case where decomposition destroys a cooperative equilibrium.\"\nkind: same-idea-earlier\nwhy: Sunstein's incompletely theorized agreements, and the Good Friday and Resolution 242 fudges, are the older form of \"agreeing without agreeing why\". Hole 2 argues such agreements can hold only because the divergence stays unspoken. Publishing it can detonate them, \"anyone can detonate the fudge\", and the ontology has no category for load-bearing illegibility. Showing opponents their agreement (AD28 shows different rules as a state) needs a check that it does not expose a fudge doing work.\nstrength: strong\n\n### F8\ntarget: docs/research/does-descending-change-the-descender.md:233\ntarget_words: \"The standing design rule that answers it is already ratified — **no gap displayed without a move\"\nolder: docs/research/scaffolding-versus-difficulty.md:37\nolder_words: \"the platform may always show the door and must never walk through it for you\"\nkind: narrows\nwhy: The note answers Mutz's demobilisation with the ratified rule \"no gap without a move\". The August scaffolding note found the opposite pull where the experiment had run: supplying the next step on every gap trained learners to ask what is wanted. Its resolution separates moves that say where to look from moves that say what to conclude, and the rule needs only the first. Read \"move offered\" as closability, not closure. That also keeps the self-generation section 5 requires.\nstrength: strong\n\n### F9\ntarget: docs/research/philosophical-foundations-stress-test.md:268\ntarget_words: \"(1) A fifth residue type: hinge commitment / bedrock practice. Unchosen (so not axiom choice), not a merit relation (so not\"\nolder: docs/research/the-load-bearing-unsaid.md:67\nolder_words: \"Renaming fixes the confusion; it does not fix the underlying question, which is whether hinges can be Claim nodes\"\nkind: narrows\nwhy: The stress test asks for a fifth residue type and a renamed hinge score. The later note says renaming does not touch the real question: whether a hinge can be a Claim node at all. If hinges are non-propositional and enacted (Moyal-Sharrock) they cannot. If they are a reasons-unresponsive attitude (Pritchard) they can, but strength computed over them is meaningless. The fifth type is therefore a representational fork, not another slot. Ledger O7 notes the shipped `is_bedrock` was born Deutsch-flavoured, not Wittgensteinian.\nstrength: strong\n\n### F10\ntarget: docs/research/coupling-the-poles.md:19\ntarget_words: \"after a descent decomposes a user's position, render *their own decomposed argument through the other side's lens*.\"\nolder: docs/research/storytelling-and-the-story-layer.md:90\nolder_words: \"(a tempting feature), the steering result applies at full force — rendering IS restating, so any\"\nkind: narrows\nwhy: The return edge has the machine render a decomposed argument through the other side's lens. The storytelling note names this as a watched hazard. A machine narrative rendering of a region is restating, so it inherits the facilitation rule (structure and questions, never their words) or manufactures convergence. The coupling note's own safeguard is to keep the quote and the person attached. Ledger R10 flags machine-rendered other-side lenses as a borderline case of simulating an end.\nstrength: medium\n\n### F11\ntarget: docs/research/reflection-and-the-graph.md:71\ntarget_words: \"machine pass gets scored so that a confessed *does-not-fit* strictly beats a wrong nearest-fit.\"\nolder: docs/research/what-counts-as-evidence/testimony-experience-and-normative-evidence.md:306\nolder_words: \"they make honest reporting of uncertainty the optimal strategy.\"\nkind: same-mechanism-elsewhere\nwhy: The candidate rule makes a confessed does-not-fit beat a wrong nearest-fit. The forecasting literature has the formal family: strictly proper scoring rules, under which honest reporting of uncertainty is the optimal strategy. The evidence note calls them \"the formal version of confession must pay\". Such a rule bounds both sides, so a pass cannot win by always confessing either. That is a ready shape for scoring machine passes, with no animal analogy needed.\nstrength: medium\n\n### F12\ntarget: docs/research/game-feel-and-the-first-screen.md:306\ntarget_words: \"1,000 words, a fixed floor around 76 s** — so a 1,500-word source is roughly **100–165 seconds**.\"\nolder: docs/research/live-election-test-design.md:91\nolder_words: \"paid Gemini is roughly 4–6× faster running all eight passes than the Claude backend is running two of them.\"\nkind: narrows\nwhy: The 100 to 165 second \"map assembles\" wait is the paid-Gemini rate. The live-session note measured the Claude backend at 2.45 words/s, 4 to 6 times slower while running only two of eight passes. The founder ruled density, quality, attribution and addressability over latency (ledger A7; CLAUDE.md topic index), which is why a mandated break exists. If the Claude backend is what runs, the demo wait is several times longer, and \"shown, it is the demo\" needs re-pricing.\nstrength: medium\n\n### F13\ntarget: docs/research/cognitive-bias-codex-and-human-contribution.md:208\ntarget_words: \"a persistent minted claim is exactly an anti-hidden-profile device (private information becomes an\"\nolder: docs/research/the-group-premise-evidence.md:30\nolder_words: \"reasons exchanged inside one frame move the group further out\"\nkind: narrows\nwhy: The note treats minted claims as an anti-hidden-profile device and asks only for a pre/post attitude reading. The evidence pass adds that arguments exchanged inside one frame polarize a like-minded group more than comparison does (Isenberg), and that hidden-profile groups are eight times less likely to find the answer. The graph supplies three of five conditions, not \"minds that differ\". The founder re-cut the tile to \"Minds that disagree beat one mind, when every reason is on the table.\" The attitude instrument should record whether participants share a frame.\nstrength: medium\n\n### F14\ntarget: docs/research/wild-weighing-dialects-and-the-question-set.md:7\ntarget_words: \"one register (campaign politics) induced these; expect the next register to break the set again.\"\nolder: docs/research/what-human-judgment-is-for.md:113\nolder_words: \"will carry the distinctions of whatever register produced it and break on the next — it needs a\"\nkind: supplies-a-missing-piece\nwhy: This note warns that the next register will break the question set but gives the set no way of saying so. The human-judgment note, on the same fourteen questions, says the set is itself a frame and needs a does-not-fit value from day one. It also says questions must be harvested from real cases, not invented. The four new ones were harvested. The set as a whole has only the typing question's closing \"or none of these?\" (deliberus/weighing.py:230).\nstrength: medium\n\n### F15\ntarget: docs/research/active-inference-context-acquisition-and-deliberus.md:17\ntarget_words: \"Questions are ranked but never *priced*: nothing decides when the best remaining question is no longer worth a human's next tap\"\nolder: docs/research/kartpaus-auto-enrichment-design.md:26\nolder_words: \"and stops where nothing it could add would change what a reader sees.\"\nkind: same-mechanism-elsewhere\nwhy: The note prices the human's next tap and says a descent has no stopping rule. The kartpaus enrichment design (direction ruled) applies the same display-band price to the machine's own spending. It takes one more decomposition level or fetch where the expected badge movement is highest, and stops where nothing it could add would change the display. One price with no new constant would then govern both the person's attention and the daemon's budget.\nstrength: medium\n\n### F16\ntarget: docs/research/taxonomy-gaps-and-the-closed-enum.md:90\ntarget_words: \"degree that confessing beats a forced nearest-fit for whoever presses it, human or model. Candidate\"\nolder: docs/research/proving-a-pass-reliable.md:304\nolder_words: \"The strongest tool in the literature is to let the machine decline, and to prove the bound only on\"\nkind: same-mechanism-elsewhere\nwhy: The note makes the confession channel pay through scoring. The reliability note shows the other classic route: abstention with coverage as the price. A judge declines below a threshold set on a human-labelled set with an exact bound, and the declined items were the ones people disagreed on (agreement 0.815 against 0.902). A does-not-fit rate can then be read as coverage that tracks contestedness. A threshold from a point estimate is no guarantee.\nstrength: medium\n\n### F17\ntarget: docs/research/game-feel-and-the-first-screen.md:100\ntarget_words: \"this is *no gap displayed without a move offered* given a concrete form, and it answers the\"\nolder: docs/research/scaffolding-versus-difficulty.md:39\nolder_words: \"The decomposition guide seeds a non-blank starting structure, which is closer to the line.\"\nkind: narrows\nwhy: The note offers a fill-in slot with grammar as the concrete form of \"a move offered\". It also calls the oracle's bare report a violation. The scaffolding note's test is that a move should say where to look, not what to conclude. By that test the oracle passes (line 39). A seeded non-blank structure sits closer to the line and is worth an audit. The grammar slot is safe only if it frames the question without hinting the answer.\nstrength: medium\n\n### F18\ntarget: docs/research/coupling-the-poles.md:73\ntarget_words: \"state their position in a way they would endorse, and do you feel why it is held?\"\nolder: docs/research/incentives-analysis.md:166\nolder_words: \"Community Notes has already operationalized the ideological Turing test at scale, and solved the incentive.\"\nkind: same-idea-earlier\nwhy: The note's test is Rapoport's rule, the ideological Turing test. The incentives note records that Community Notes already runs it in production: a note is published only when raters who usually disagree both accept it, so the contributor is paid for writing what the other side signs off. That gives the return edge's check a scoring protocol and a precedent. The founder's 9 September \"Listening\" tile ruling uses the same test.\nstrength: medium\n\nNoticed: Three more pairs were real but cut for the cap. (1) cognitive-bias §4.1's open question whether person-claims should exist is answered earlier by scheme-bounded-decomposition-and-evidence-as-subgraph.md:231 (\"Decompose into empirical claims about persons + inferential claims about relevance\") and by what-counts-as-evidence/philosophy-of-evidence.md:266 (bias evidence belongs on the source or inference). (2) does-descending §6's pre/post instrument was proposed earlier as revote vote-delta in feed-algorithm-design.md:153, and the-motte-and-the-periphery.md:225 reuses the slider as the entrenchment detector. (3) reflection §1's \"reconstruction presented as description\" has a measured two-round instance in four-cuts-of-the-ai-debate.md § 1. The walkthroughs note names session26-open-rulings.md, so no ledger rows could be used against it. coupling-the-poles names lowering-the-cost.md and specs/landing-redesign/constellation-map.md, so I used incentives-analysis for the Turing-test pairing. CLAUDE.md topic-index entries still point at §§8–9 of the cognitive-bias note, which do not exist (TODO.md:1025).\n"}