{"path":"research/corpus-graph/readers/2026-10-10-strength-red-team-and-descent-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on strength, the red team and the descent (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/strength-layer-audit.md`\n- `docs/research/what-the-necessary-default-changes.md`\n- `docs/research/heylighen-value-complexity-and-deliberus.md`\n- `docs/research/evidence-division-and-the-foundation.md`\n- `docs/research/deep-descent-experiment.md`\n- `docs/research/weighing-scheme-and-terminus-classifier-design.md`\n- `docs/research/slicing-stability-first-run.md`\n- `docs/research/red-team-synthesis-2026-07.md`\n- `docs/research/peer-review-and-the-reasoning-layer.md`\n- `docs/research/frontier-extraction-experiment.md`\n- `docs/research/ai-slop-and-the-reasoning-layer.md`\n\nIt made 114 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-necessary-default-changes.md:146\ntarget_words: \"3. **Mislabelled** — this was never a part; it is evidence. Propose changing the relationship type\"\nolder: docs/research/session39-what-the-graph-stopped-claiming.md:83\nolder_words: \"**The census reliably detects *not a part*; it says little about *what instead*.**\"\nkind: narrows\nwhy: The third bucket proposes re-typing a mislabelled part as evidence. A hand read of the census (70 edges labelled miscast as evidence) found that re-filing the 31 machine-written ones as support would have made about 18 false: some ran the other way, some were reported speech, some were not support at all. The note concludes that a not-a-part verdict licenses retiring the link, never writing a substitute. The founder ruled the wider rule, store the blank, not a plausible stand-in (Q30).\nstrength: strong\n\n### F2\ntarget: docs/research/what-the-necessary-default-changes.md:164\ntarget_words: \"Those two signals are countable without judgment. A model handles what is left, which is the genuine\"\nolder: docs/research/the-parts-review.md:48\nolder_words: \"**The 77% silence is not a defect of the structural reader.** It is the grey zone Yu & Zenker\"\nkind: narrows\nwhy: Section 8 says the shared-subject and carved-fragment signals need no model and a model handles the rest. The parts review ran two readers over all 197 decomposition edges. The structural reader was silent on 152 (77%), the two readers disagreed on 18 and agreed on 27, and no person has ever set a support_semantics tag by hand. The countable tier covers under a quarter of the graph, and the contested remainder needs a person's tie-break.\nstrength: strong\n\n### F3\ntarget: docs/research/evidence-division-and-the-foundation.md:62\ntarget_words: \"route the same corpus under k reasonable policies, compare badge distributions), and nobody has measured it\"\nolder: docs/research/slicing-stability-first-run.md:24\nolder_words: \"for 79 of 87 mothers, every filing that produces a verdict produces the *same* verdict\"\nkind: already-answers\nwhy: The falsifier this sentence calls unmeasured was run the next day. Slicing stability recomputed 87 decomposed mothers under four reasonable filings: verdict stable for 79 of 87 (90.8%), median spread 0.000, and the two sensitive claims trace to one unrecorded join interpretation. Its own limits: critical questions are mostly unanswered, and no motivated stakeholder has filed adversarially. A reader should cite this result and those limits, not write that nobody has measured.\nstrength: strong\n\n### F4\ntarget: docs/research/evidence-division-and-the-foundation.md:55\ntarget_words: \"it changes every number in the graph, which is why it waits\"\nolder: docs/research/slicing-stability-first-run.md:28\nolder_words: \"**Fair-share changed nothing anywhere measured** — expected, and itself evidence\"\nkind: narrows\nwhy: This note defers the fair-share adoption decision because it would change every number. The first slicing run applied fair-share to all 87 decomposed mothers and found it changed nothing. The one known shared-evidence case, the fasting meta-analysis, had already been dissolved by evidence decomposition plus edge supersession. Someone putting the adoption question to the founder can say it is currently inert on decomposed claims, and that it matters only once real sharing appears.\nstrength: strong\n\n### F5\ntarget: docs/research/peer-review-and-the-reasoning-layer.md:164\ntarget_words: \"work by pre-empting the objection that he is being pedantic. The four-type taxonomy\"\nolder: docs/research/taxonomy-gaps-and-the-closed-enum.md:17\nolder_words: \"Adding it as a fifth type would make `type` mean two incompatible things at once\"\nkind: narrows\nwhy: Finding R1 files the speaker-attitude hedge as a gap in the claim-type taxonomy. The next day's note, built from this run plus two others, says adding it as a fifth type is a category error: types classify subject matter, while a hedge is classified by its role in the exchange, so it wants its own field or nothing. It also prescribes a does-not-fit value on every enum, with a required reason. Anyone designing the fix should start there.\nstrength: strong\n\n### F6\ntarget: docs/research/deep-descent-experiment.md:27\ntarget_words: \"Either way, the current silent behavior is a mis-typing: without a decision, the classifier will force\"\nolder: docs/research/synthetic-stress-suite-and-ontology-reflection.md:45\nolder_words: \"typed on the what-is-it-like leaf via the production path, verdict-claim ground minted\"\nkind: later-correction\nwhy: The passage treats the is-bedrock type as an undecided founder call, with mis-typing meanwhile. A later note records that the is_bedrock kind exists and fired end to end on a phenomenal-floor leaf (22 Aug), and that the residue fraction was left untouched by the is-side floor, by the accounting shipped that morning. What stayed open afterwards was the name (ledger O7 proposes is_open), not whether the kind exists.\nstrength: strong\n\n### F7\ntarget: docs/research/weighing-scheme-and-terminus-classifier-design.md:39\ntarget_words: \"Source-provenance weighings auto-open; authored input and sacred-marker weighings get invitations\"\nolder: docs/research/the-load-bearing-unsaid.md:109\nolder_words: \"The brake reads **claims**, and against all 1,678 authored and extracted claims in the graph its shipped lexicon matched **zero**\"\nkind: narrows\nwhy: This records the sacredness brake as shipped. A later measurement found its lexicon (absolutist phrases such as never acceptable) matched zero of 1,678 claims in the graph: inert, not miscalibrated, because argued prose uses sanctity vocabulary instead. It was rebuilt on 13 Aug and afterwards matched one case. So the sacred-marker invitation path almost never ran, and the July provenance split with brake behaved as a provenance split alone until the rebuild.\nstrength: strong\n\n### F8\ntarget: docs/research/red-team-synthesis-2026-07.md:62\ntarget_words: \"1. **Brake-as-graph-state** (convergence 1): make the sacredness brake a first-class challengeable node marker so the oracle stops contradicting it\"\nolder: docs/research/parity-and-the-zone-interior.md:47\nolder_words: \"I don't see any fundamental need to put up stop signs for ourselves in facilitating either/both however, just because it's sensitive and sacred.\"\nkind: narrows\nwhy: The synthesis ranks a challengeable protected-value marker as the top founder decision. The founder's 25 Aug sacred-domain statement, woven into the parity note, rules that nothing is protected against analysis. Sacredness changes the attunement owed and the humility required, and what is withheld is premature verdicts and advice. The note itself calls that the substance the July brake re-scope was waiting for (consent and timing, never epistemic exemption). A marker meaning declined for decomposition runs against it.\nstrength: strong\n\n### F9\ntarget: docs/research/ai-slop-and-the-reasoning-layer.md:13\ntarget_words: \"nothing in the pipeline strips or flags instruction-shaped content before it reaches Gemini\"\nolder: docs/research/jev-and-the-judgement-layer.md:427\nolder_words: \"Injection guard (`extraction/injection_guard.py`) | six English regexes, propose-only flags,\"\nkind: later-correction\nwhy: The note calls ingestion-time prompt injection zero-instrumented with no defense. A deterministic first tier was built the next day (instruction patterns, hidden-HTML text and zero-width scans, flags only, never blocking). A later note tabulates it as six English regexes and proposes a model-based second tier, because a published pattern list arms attackers. Registering this threat as unmitigated is now wrong; the open gap is the second tier.\nstrength: strong\n\n### F10\ntarget: docs/research/ai-slop-and-the-reasoning-layer.md:19\ntarget_words: \"Flooding a graph with modest unchallenged claims buys nothing under honest absence-reporting\"\nolder: docs/research/relevance-and-the-two-factors.md:208\nolder_words: \"nothing, while a decomposition child with no standing still enters its parent at a count-based\"\nkind: contradicts\nwhy: The note says absence-is-a-state means a flood of modest unchallenged claims buys nothing. That holds on the link channel only. The 28 Sept relevance note found the parts channel differs: an unexamined decomposition child still enters its parent at a count-based prior, and two unscored supports lift it to 0.632, so structure alone can manufacture strength. Changing this would move every decomposed claim, so it was left to the founder. The mill defence is weaker than stated.\nstrength: strong\n\n### F11\ntarget: docs/research/frontier-extraction-experiment.md:60\ntarget_words: \"Implicit-premise recall is therefore the single most important axis of the whole comparison\"\nolder: docs/research/cross-source-premise-run.md:76\nolder_words: \"**And the pattern across all ten is sharper than any one of them: every premise is frame-level, not first-order.**\"\nkind: already-answers\nwhy: The comparison protocol was never run, but the axis it names was tested with a dedicated cross-source pass on four opposed pairs, pre-registered with a positive control. Result: ten premises, none generic, restated or common ground, and the control exceeded. All were frame-level (unit of analysis, reference class, burden of proof, governing law, market model), not missing facts. Anyone scoring implicit-premise recall has a pre-registered key and a class list, with n=4 in registers the corpus already held.\nstrength: strong\n\n### F12\ntarget: docs/research/what-the-necessary-default-changes.md:23\ntarget_words: \"confirmed untagged part now caps its whole at a modest, rebuttable weight (half of the full\"\nolder: docs/research/what-confirmed-means.md:21\nolder_words: \"whether a part caps its whole. No page shows it to a reader.\"\nkind: narrows\nwhy: The 17 Sept completion decides the cap by whether an untagged part is confirmed. A later history found the field was never defined, carries five meanings in one bit, and is fully determined by who wrote the link; extraction's links are confirmed by default (3,748 true), so confirmed mostly marks machine-extracted parts. The founder ruled to retire the field (Q32, Yes). Anyone tuning the cap weights should know its discriminator is being replaced.\nstrength: strong\n\n### F13\ntarget: docs/research/peer-review-and-the-reasoning-layer.md:21\ntarget_words: \"Researchers routinely coin terms that piggyback on prestigious ideas\"\nolder: docs/research/prime-real-estate-words.md:152\nolder_words: \"**The standing passes to whatever the word is applied to.**\"\nkind: already-answers\nwhy: The reviewer's remedy is to replace the coined term. The wanted-words note, blessed 7 Oct (AC1 to AC4), says some words carry standing that passes to whatever they name, that standing cannot be moved to a new word by definition, that offering two new words is refused on a wanted word, and that the word should stay on both senses with the holders' names. A terminology-objection artifact should record standing, not only senses.\nstrength: strong\n\n### F14\ntarget: docs/research/peer-review-and-the-reasoning-layer.md:89\ntarget_words: \"a falling foundational claim weakens everything leaning on it through the same published gradual semantics used everywhere else in the graph\"\nolder: docs/research/relevance-and-the-two-factors.md:312\nolder_words: \"**The principle: a reason that fails drops out; failure counts against a claim only through a\"\nkind: narrows\nwhy: Two later facts narrow retraction propagation. Until the 2026-09-28 build a supporter's own strength never reached its target (line 25 there), so a falling foundation moved nothing through support edges. And the 2026-09-30 ruling, for now, says a failed reason drops out and subtracts nothing; failure counts against a claim only through an objection, an argued need, or a thorough empty search. Dependents lose support but are not attacked by the retraction.\nstrength: medium\n\n### F15\ntarget: docs/research/strength-layer-audit.md:63\ntarget_words: \"summing raw strengths would let three unexamined supporters manufacture a strong claim out of silence\"\nolder: docs/research/relevance-and-the-two-factors.md:209\nolder_words: \"prior (two unscored supports lift it to 0.632), so structure alone can still manufacture strength\"\nkind: narrows\nwhy: The audit calls the zero contribution of unexamined supporters right, since summing would build strength from silence. The 28 Sept relevance note found the two channels disagree: a supporter with no standing lends nothing, but a decomposition child with no standing enters its parent at a count-based prior (two unscored supports reach 0.632). The route from silence to strength the audit rules out still exists through parts, and the note left it unchanged because it moves every decomposed claim.\nstrength: medium\n\n### F16\ntarget: docs/research/strength-layer-audit.md:167\ntarget_words: \"An ontology addition before it is a formula change: the graph cannot currently say which kind a decomposition is.\"\nolder: docs/research/session20-the-strength-layer-and-the-unspecified-middle.md:91\nolder_words: \"A `necessary` reading **eliminates the splitting incentive** — three parts and twelve parts at 0.70 both yield 0.502\"\nkind: later-correction\nwhy: Item 2 says the join kind cannot be recorded. The same day's session note shows a support_interpretation field (necessary, corroborative, undecided) added to DECOMPOSES_INTO, and a necessary reading removes the splitting incentive: three and twelve parts both yield 0.502. Items 2 and 3 are not open as written. What remained open was who records the tag, which the 31 Aug ruling then answered by changing the default.\nstrength: medium\n\n### F17\ntarget: docs/research/slicing-stability-first-run.md:26\ntarget_words: \"so the additive and the jointly-required readings are both live, and they disagree substantially\"\nolder: docs/research/what-the-necessary-default-changes.md:18\nolder_words: \"cannot raise a parent, and a weak part pulls it down a little), while the hard **cap** (a whole can be\"\nkind: later-correction\nwhy: The study ran on the 22 Aug code. On 31 Aug untagged parts began to aggregate by their weakest part, and on 17 Sept a confirmed untagged part also began to cap its whole at half the full weakest-link cap. So the 0.491 versus 0.060 spread on the two sensitive claims describes the old default, and untagged no longer means additive. The instrument is re-runnable and should be re-run before those numbers are quoted.\nstrength: medium\n\n### F18\ntarget: docs/research/strength-layer-audit.md:65\ntarget_words: \"So the correction-UX promise is a **two-step loop presented as one**: attach evidence, then answer the critical questions it raises.\"\nolder: docs/research/synthetic-stress-suite-and-ontology-reflection.md:32\nolder_words: \"This is the badge-boundary defect (a two-step contribution presented as one step) surfacing on the note text itself.\"\nkind: later-correction\nwhy: The audit leaves the two-step problem open as a copy or interaction decision. Five days later the synthetic stress suite met the same defect in the badge note, which told readers no evidence bears on the claim while an attack was attached. It was fixed that day: the note now says N edges are attached, none scored yet, and that answering their critical questions is what moves the claim. The copy half is done and regression-pinned; the interaction half remains.\nstrength: medium\n\nNoticed: pairs read at both ends but left out for the cap: docs/open-problems.md:31 (public page) says shared evidence \"is discounted so it isn't counted twice\" while evidence-division line 19 and CLAUDE.md say fair-share is built but NOT adopted, so the public page overstates; heylighen-value-complexity-and-deliberus.md:9 (anti-orthogonality, danger boundary is selection) bears on frontier-extraction-experiment.md:40 (orthogonality and aliveness are the unstated premises of the 3F batch-1 crux), and neither note names the other; red-team-synthesis-2026-07.md:17 (oracle flags braked nodes, \"ships today\") is also hit by the-load-bearing-unsaid.md:109 (brake matched 0 of 1,678 claims), and its :13 immunity gradient reappears in four-cuts-of-the-ai-debate.md:222 for machine moral patienthood, its :49 cycle point is measured in strength-layer-audit.md:31 (3 two-cycles, 1 three-cycle) and its :46 uptake metric has a computable form in peer-review-and-the-reasoning-layer.md:69; peer-review :97 (anonymity, chilling) has legibility-under-power-red-team.md:186 (Chatham House rule, measured chilling); what-the-necessary-default-changes.md:41 (remove the reward, do not police) has twins at ai-slop-and-the-reasoning-layer.md:37 and relevance-and-the-two-factors.md:189; strength-layer-audit.md:103 (sentinel rule) has data-layer twins in session39-what-the-graph-stopped-claiming.md:105 and the-residual-error-taxonomy.md:97 (dead instruments believed alive); weighing-scheme-and-terminus-classifier-design.md:109 (auto-open leans yes) is narrowed by the personal-domain carve in kartpaus-auto-enrichment-design.md:34, and its :65 (frontier-grade classifier) lacks the ask-twice steadiness test in pass-quality-and-the-choice-of-model.md:150. The helper report the-squirrel-and-the-strength-calculus/strength-rulings-and-code-as-read-2026-10-08.md is a compact map of every strength ruling with founder words.\n"}