{"path":"research/corpus-graph/readers/2026-10-11-four-top-level-pages-kinship-as-a-haiku-reader-returned-it.md","content":"# What a Haiku 5.5 reader found to bear on four top-level pages, asked only for kinship (model test, second run) (2026-10-11)\n\nA read-only reader on another model than the readers before it: Haiku 5.5 at high effort, started through\nthe command-line tool with no settings loaded, with the tools to read and to search files and nothing\nelse. It is part of the test in [../haiku-5-5-for-the-filling.md](../haiku-5-5-for-the-filling.md), where\nthe prediction written before the run and the scoring stand. It was given the brief in\n[../model-tests/reader-brief.md](../model-tests/reader-brief.md) and these 4 notes:\n\n- `docs/convergence.md`\n- `docs/analysis-and-attunement.md`\n- `docs/the-missing-layer.md`\n- `docs/open-problems.md`\n\nIt took 60 steps and six minutes. Its report is below, unchanged. What was kept from it, after the quoted\nwords were checked by code and both passages of each finding were opened, is in `../found.jsonl` (lines\nwhose `where` names this file).\n\n---\n\nRead in full: 4 of 4 notes. Subject-free sentences written for 44 passages.\n\n### F1\ntarget: docs/convergence.md:7\ntarget_words: \"Cross-cultural research supports this: machine-reading analysis of 256 societies finds the same seven moral structures present across all of them\"\nolder: docs/research/contingency-and-convergence-across-levels.md:118\nolder_words: \"same solution by a constraint stronger than the space is large). Only the second is evidence about the\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: agreement counts as evidence about a shared basis only when the arriving routes are independent and one constraint funnelled them; shared lineage or shared source is copy number. The target counts Curry's 256 societies and Parfit's three frameworks as convergence without asking whether they share descent or source. The contingency note states the test that the target lacks. A person working on the target would check lineage before counting agreement as support for bedrock, and would expect shared-descent agreement to read as inheritance, not discovery.\"\nstrength: strong\n\n### F2\ntarget: docs/convergence.md:140\ntarget_words: \"and the verdict is itself a challengeable claim that new moves can reopen.\"\nolder: docs/research/what-counts-as-evidence/law-intelligence-history-journalism-forensics.md:725\nolder_words: \"from you, or because you or the intelligence collectors have not looked for it? Heuer also\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: a null result is informative only if the record shows what was looked for; absence cannot be read without its cause. The target proposes that terminus verdicts carry moves-tried metadata so a horizon can be reopened. Heuer's question asks why an absence is absent: not happened, not observable, concealed, or not looked for. Only the last can be fixed by searching, so a verdict with no search record cannot distinguish a real horizon from an unsearched region. A person building moves-tried metadata would separate not-tried from tried-and-failed and ask Heuer's four-way question of each null.\"\nstrength: medium\n\n### F3\ntarget: docs/convergence.md:45\ntarget_words: \"the wager has survived by reformulation, and that series needs a ledger\"\nolder: docs/research/the-analogy-daemon.md:233\nolder_words: \"Telling a refinement from a retreat to a safer definition, and Lakatos on a counterexample: the good\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: a reformulation after a counterexample is progress only if it narrows the claim's condition rather than changing what its terms mean. The target says the wager's reformulation series needs a ledger, but gives no test for entries. The daemon kinship supplies one: a good response states the hidden condition as a condition on the claim, while a poor one changes the term's meaning, as a retreat to a safer definition does. A person extending the ledger would classify each reformulation by that test and record retreats as retreats.\"\nstrength: medium\n\n### F4\ntarget: docs/open-problems.md:27\ntarget_words: \"How strands add, and when two strands are one strand counted twice, is not settled.\"\nolder: docs/research/what-a-record-linkage-project-teaches-deliberus.md:164\nolder_words: \"The founding model of record linkage adds up agreement across fields, and it assumes the fields are\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: an additive score over parts is sound only when the parts draw on independent sources; shared sources get counted once per part. The target leaves open when two strands are one strand counted twice. The record-linkage note says the founding additive model assumes independent fields, and where they are correlated agreement is counted twice; one study measured the overstatement, 16.87 ignored against 13.55 modelled. A person working on strand addition would test pairwise dependence before summing, and would treat a correlation term as part of the sum rather than an optional refinement.\"\nstrength: strong\n\n### F5\ntarget: docs/the-missing-layer.md:45\ntarget_words: \"a machine reader has no unfamiliar interface to blame, so if the structure does not improve its answers over the raw source texts\"\nolder: docs/research/dogfood-run-7-swedish-election-structured-vs-unstructured.md:15\nolder_words: \"which is the harder test for structure. The reverse order would have flattered it.\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: a candidate earns credit only for what it adds over a baseline the rival already has, and the baseline must run first or its knowledge gets counted as the candidate's. The target says structure must beat raw source texts to earn its cost, but does not fix the order or the credit rule. The dogfood design runs the unstructured arm first and credits the structured pass only for what the reader did not already know, because the reverse order would flatter it. A person testing the layer's claim against machine readers would adopt that ordering and credit rule.\"\nstrength: medium\n\n### F6\ntarget: docs/open-problems.md:21\ntarget_words: \"both genuinely sensitive claims turned out to hinge on one unrecorded choice\"\nolder: docs/research/starting-values-and-the-problem-of-the-prior.md:528\nolder_words: \"the fraction of well-examined claims that are still prior-dependent measures how far they\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: a verdict must be recomputed under every reasonable configuration, and the cases that flip are reported as sensitive rather than averaged into the score. The target re-runs the score under four filings and flags two claims that hinge on one unrecorded choice. The starting-values note applies the same test to starting beliefs: robust if the conclusion holds under every reasonable start, prior-dependent if it flips, with the still-prior-dependent share as a measure of how far beliefs have washed out. A person extending the filing sweep would adopt the robust and prior-dependent display state as its summary.\"\nstrength: medium\n\n### F7\ntarget: docs/open-problems.md:47\ntarget_words: \"Peace treaties work the same way: write down exactly what each side means by the key sentence, and the treaty collapses.\"\nolder: docs/research/the-load-bearing-unsaid.md:11\nolder_words: \"trying to recover it as a premise destroys the thing you were recovering.\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: a joint that holds only while unstated is broken by making it explicit, so the explicit form must be handled as a refusal to recover rather than a premise to write. The target asks for a mark on vagueness that is doing work, and a consent rule before dissection. The load-bearing-unsaid note separates two cases: the crux was unwritten and can be recovered, while the sacred cannot be said and recovering it destroys it. A person designing the handle-with-care mark would need that split to know which unsaid items to extract and which to leave unwritten.\"\nstrength: medium\n\n### F8\ntarget: docs/open-problems.md:63\ntarget_words: \"but the same person who builds the instruments decides what they measure.\"\nolder: docs/research/the-analogy-daemon.md:49\nolder_words: \"separate calls (ideally separate model families, per the two-family jury), or the prior-setter\"\nkind: same-mechanism-elsewhere\nwhy: \"Pattern: a measurement is credible only when the party that sets the number is not also the party that audits the pattern behind it. The target says the instrument's builder decides what is measured, with no rulebook or amendment procedure. The analogy-daemon note makes the same move inside its pipeline: the pass that proposes a number and the pass that audits its pattern must be separate calls, or the prior-setter grades its own homework. A person drafting the rulebook would take that separation as the minimum, and make the rule-setter a different party from the instrument's author.\"\nstrength: strong\n\nNoticed: the mechanism and vocabulary-search fields led to the finds; searches on boundary objects, pharmacovigilance, value of information, Hirschman exit, Goodhart and sycophancy gave nothing usable.\n"}