{"path":"research/session34-open-loops-and-unweighed-answers.md","content":"# Session 34: open loops, and the answers he has not weighed\n\n**Date**: 2026-09-21 · **Type**: session record, written on his instruction — *\"Document all open loops\nand include any paragraphs/responses you've given me in this session that I've left uncommented\non\"* · **What the session was**: the map view's second day, the tail made addressable, the split\nmachinery built, the sameness sitting built, the reported-speech backfill run, and a long argument\nabout what a machine may write.\n\n**Why the second half of this document exists.** A chat answer he did not reply to is not an answer\nhe rejected and not one he accepted — it is one nobody has weighed, and it vanishes when the\ntranscript does. The standing rule is that cost of re-derivation decides what gets written down, not\nimportance; several of these cost hours to reach and would cost hours again.\n\n---\n\n## 1. Ruled and shipped this session\n\nRecorded compactly, because each has a home that carries the reasoning.\n\n| What | His words | Home |\n|---|---|---|\n| Kind is the box's own colour, not a dot | *\"I'd feel better if these dots were removed (from all maps) and instead the bg-color of all boxes was tinted according to type.\"* | `specs/map-view/design.md` |\n| The instruction line goes, everywhere | *\"Wasn't this line supposed to be removed?\"* | same |\n| The map gets the attention on a source page | *\"What's this empty box? … it's the map that should receive attention\"* | same |\n| Provenance in faint type under the link | *\"The provenance/author should perhaps be shown under the link or whatever, in faint/small type.\"* | same |\n| Titles vertically centred in their boxes | *\"The titles should be vertically centered in these boxes\"* | same |\n| All three tail proposals | *\"Please implement all of these. You can make the map larger vertically if you need more real estate for the tail stuff.\"* | same, § the map keeps one place |\n| The split machinery | *\"Make this happen!\"* | `the-jhu-decomposition-line.md` |\n| The sameness sitting, built the day it was asked for | *\"Let's do this sitting now!\"* | `the-tidying-line.md` § 3 |\n| The reported-speech backfill | *\"Please do the reported-speech backfill now.\"* | `earning-the-right-to-write.md` § 7 |\n| A machine-written split part caps like a person's | *\"Yes it should. We should expect good splits.\"* | same, § 2b |\n| A cut through a thick concept is an argument, not a split | *\"I would call that act backing the main claim with subclaims rather.\"* | `core-periphery-and-the-concept-layer.md` § 5a-ii |\n\n---\n\n## 2. Open loops — his to rule on\n\nOrdered by how much else waits behind them.\n\n0. **RULED 2026-09-21, both approved**: *\"Approved both candidates\"* — the **addressability test**\n   (does the output give a reader more places to disagree, or fewer) and **a cut through a concept\n   is an argument, not a split**. Both now in the canon principles; the first carries the bound he\n   supplied, that it is a gate for which passes may write and never an objective, since the\n   separability floor decides how far a cut goes.\n\n1. **The trust ladder — RULED IN PART the same evening.** Founder: *\"Wire and switch first, sitting\n   later.\"* The splitter is at rung two (writing-marked) by ruling rather than by record, which is\n   legitimate and is not the same thing as having one. **Still open**: the numbers — a measured\n   agreement rate over a stated minimum of human touches, read off the first sitting. And one\n   correction the build forced: the rung's `confirmed=false` was the wrong mark, since that field\n   already means *a person refused*; a written part leaves it unset. Home:\n   `earning-the-right-to-write.md` § 4a.\n2. **The addressability test as the committable tier's discriminator** — replacing *nothing that mints\n   a claim* with *nothing that lowers addressability*. Proposed as the sharper axis than\n   narrow-versus-deep. Unruled, and it is the piece the rest of the ladder rests on. Home: same, § 1.\n3. **A failed reconstruction re-files a split as `SUPPORTS`.** `CoverageReport.reconstructed` already\n   makes the model write what the parts jointly assert before it sees the original; nothing acts on\n   the comparison. Proposed, unruled. Home: `core-periphery-and-the-concept-layer.md` § 5a-ii.\n4. **Does a signed-out visitor see a split proposal that already exists, read-only?** The suggestion is\n   a machine's rather than a person's, and showing it would let a stranger see what the machinery does\n   without an account — which is the *show, don't tell* surface he has said matters. **Raised twice,\n   never answered.**\n5. **Should the hovered box on a source map carry its three tags** (kind, standing, strength)? The\n   source map never fetched them; the claim page does. **My open call, offered and unanswered.**\n6. **What a section should correspond to.** *\"I just noticed that sections started showing up… I'm not\n   sure what it should correspond to ideally, so it's worth philosophizing/strategizing about.\"* The\n   measurement exists (`how-a-map-is-cut-and-what-it-shows.md`); the ruling does not.\n7. **The five unruled map-view calls** carried from 2026-09-20, plus the new one: how loud a\n   provisional horizon reading may be on any surface.\n8. **Ledger rows still awaiting him** from earlier sessions: the delta-not-description proposal, the\n   audience-as-provenance 2×2, the sixth flattening mechanism.\n\n---\n\n## 3. Open loops — built but unexercised, or named but unbuilt\n\n| | State | What it needs |\n|---|---|---|\n| **The sameness sitting** | live at `/sameness/sitting-1`, 120 pairs, **unsat** | one hour of his keyboard; then the friends layer on the same salt |\n| **The looks-bundled review** | 1,536 claims, clusters measured, **unread** | the 167 conjunction-only rows first; `TODO.md` carries the reading order |\n| **The split proposer** | **writing-marked since 2026-09-21** — 7 of 11 stored proposals write, 4 hold for invention | its own sitting; `scripts/write_splits.py --propose N` is the sweep |\n| **The contested-word surface** | named 2026-09-21, unbuilt | *\"it should display a well-thought out UI/UX for dealing with contested claims\"* |\n| **A summary above the map** | deliberately deferred on his word | the tail work shipped today is its precondition |\n| **The stance field** | designed in `the-remaining-case-for-the-stance-guard.md` § 4, **0 claims carry it** | one enum at extraction; the root fix the backfill only patched around |\n| **The sweep surface and its ranking** | unbuilt | a flag is not a ranking; the proposer's own output is the natural ranker |\n| **`count_safe_summary`** | exists, **still no caller** | `/graph/stats` publishes undifferentiated asserted-plus-minted totals |\n\n---\n\n## 4. Answers he has not weighed\n\nEach of these was given in chat and drew no reply. None is a decision; all are findings or offers\nthat would otherwise evaporate.\n\n**On daemons, asked directly and answered.** *\"Have we in the past talked about recursive\ndecomposition and perhaps building a daemon that buzzes around now and then and proposes splits?\"* —\nyes: `graph-daemons-design-space.md` (2026-08-17) puts decomposition in the **propose-only** tier and\nnames six preconditions, of which the split proposer now satisfies five. The unmet one is **an undo\npath**, which that document says is *\"asserted in the tier table above and is not currently\nimplemented anywhere.\"* Staleness was named the strongest first candidate; the split proposer is in\nfact the first propose-tier instrument to ship.\n\n**The reported-speech base rate.** 153 read, 46 minted, 107 refused, 0 failed; `REPORTS` edges 1 → 53.\n**A 30% mint rate is the pre-filter working as designed**, since it errs toward reading: a false flag\ncosts one model call and a missed report is a position nobody can attack. The second run's rate was\nfar lower (2 of 39 against 44 of 114) because candidates are read longest-first, so the tail is thin\nby construction.\n\n**Three failure shapes in real splits**, answering *\"what would a wrong split even look like?\"*:\n**loss** (a dropped University-of-Washington attribution — the one he guessed, and the coverage call\ncatches it); **invention** (a proposal that came back coverage-*clean* while adding doctrinal\nspecificity the parent never carried — the shape that makes a split look *better* than its parent,\nwhich he did not name and which is the dangerous one); and **drift into commentary**, watched rather\nthan asserted.\n\n**Two corrections I made on my own claims.** The machine-provenance list held **by accident** — one\nentry, so the next pass to learn writing would have moved strengths silently; named positively now.\nAnd a sub-case was nearly got wrong in fixing it, caught by an existing test: a **missing**\nprovenance keeps the presumption, because the cap runs in the conservative direction and dropping it\nwould make wholes look *stronger*, which is the flattering direction.\n\n**The recruiting point, after he pushed back on science-versus-art.** Whether two things are separable\nis logic; how far down to go is a function of purpose. So **divergence between two careful reviewers\nis neither taste nor noise — it marks the places where nobody has said what the split is for.** That\nis a better reason to recruit for divergence than *find people with different intuitions*, and it\nreplaces what I first wrote.\n\n**The granularity number, narrowed.** 56% fine-grained against 91% coarse is about *sense*\ngranularity. It transfers to cuts **through** a concept, because those are the same problem wearing a\nclaim's clothes, and **not** to cuts between propositions. Citing it against claim splitting in\ngeneral overstates the ordinary case.\n\n**The truth-check caught a live 500.** `GET /claims/{id}/tail` was reported live on the strength of\nits source-scoped sibling working — describing a whole from having read a part. A variable-length\nmatch binds a single `Edge` rather than a list at one hop, so filtering the path with `all(x IN r\nWHERE type(x) IN …)` raised *Type mismatch: expected List or Null but was Edge*. The types belong in\nthe pattern. Fixed, with two tests that assert the queries **parse**, which is the part no unit test\nover the pure ranking function could reach.\n\n**Path collision, as a rule rather than a bug.** A page and an API route on one path is not a style\nquestion: the API wins and the visitor gets raw JSON. Singular for the page, plural for the API, and\na test now asserts that a client asking for HTML at a page path is never answered with JSON.\n\n---\n\n## 5. Raised in the last minutes, with no answer yet\n\n**The AI-doom argument diagram** (Andres Gomez Emilsson, published 2026-09-21) — *\"This can probably\nflesh out and connect to a lot of the AI alignment debate/extraction/subclaims.\"*\n\nThree things make it worth more than an illustration, and none of them has been acted on:\n\n1. **It is an independently invented ontology** — evidence, extrapolation, assumption, mechanism,\n   conditional inference — and those are **roles in an argument** rather than kinds of claim. Our\n   `epistemic_status` enum (`observed`, `causal_hypothesis`, `prediction`, `counterfactual`,\n   `feasibility`, `normative_assertion`) covers much of the same ground from a different angle, which\n   is convergent validation worth a paragraph in `academic-foundations.md` rather than a new schema.\n2. **The author declares his own crux**: the probability bridge is *\"the weakest part of classic doom\n   arguments\"*, and the real disagreement is *\"not whether a mechanism exists, but how much\n   probability to assign to each link and how correlated the links are.\"* That is a hand-computed\n   hinge, published before we touch it — **a rare source with a pre-registered crux, so extracting it\n   tests our hinge score against a human verdict.** Run 3F's finding was that the crux is usually\n   implicit; here it is stated, which is the complementary test case.\n3. **Correlation between links is exactly what our arithmetic does not handle.** The corroborative\n   channel inherits naive Bayes's correlation blindness, named as open in\n   `weakest-link-arithmetic-and-the-merge-hunch.md`. An outside author independently naming it as\n   *the* locus of disagreement is the strongest external evidence that the gap matters.\n\n**Proposed, unruled**: take it as a dogfood source, paired with an opposing-register source per the\ningest-by-cluster rule. It is a structured position by a named public author on the topic this\nproject's own positioning runs through. The social context around the post (who commented what) is\nnot corpus material and is not recorded here.\n\n---\n\n## Cross-references\n\n`specs/map-view/design.md` (everything shipped to the map) ·\n[earning-the-right-to-write.md](earning-the-right-to-write.md) (the ladder, the addressability test,\nthe reported-speech numbers) · [the-tidying-line.md](the-tidying-line.md) (the sweep that opened it) ·\n[core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md) § 5a, § 5a-ii (the\ntwo kinds of cut) · [how-a-map-is-cut-and-what-it-shows.md](how-a-map-is-cut-and-what-it-shows.md)\n(sections, and the tail's reserved place) · `TODO.md` (the looks-bundled review, with its clusters)\n"}