{"path":"research/session26-open-rulings.md","content":"# Session 26 — the ruling ledger\n\n*2026-08-27. Founder: \"each paragraph or even sentence that I've not commented on is an open loop\nto me. I wanna make up my mind about EVERYTHING you've spit out in this session.\"*\n\n**This document has a defined lifetime.** It is a decision surface, not a record. When every item\nbelow is ruled, it is deleted and the rulings live in their home docs. Do not let it become a\nparallel store.\n\n**Status key** — `RULED` he settled it · `OPEN` asserted by me, unremarked · `SUPERSEDED` I\ncorrected it myself mid-session and the dead version may still be in your memory.\n\n---\n\n## Arc A — Measurement: extraction speed and the live-session budget\n\n| # | Claim | Status |\n|---|---|---|\n| A1 | Paid Gemini, full 8-pass: **9.4–15.6 words/s, ~40 claims per 1,000 words**, fixed floor ~76 s | OPEN |\n| A2 | Claude backend: **~2.5–6 words/s on two of eight passes, 88.5 claims per 1,000** — order of magnitude slower for ~2× density | OPEN |\n| A3 | **The model question is closed**: backend already defaults to Sonnet 5; Haiku measured *slower*; Opus ~30% slower still | OPEN |\n| A4 | Five runs of one fixture spread **77–122 s**, so single-run figures from this backend carry no precision | OPEN |\n| A5 | `DELIBERUS_CLAUDE_EFFORT` is default-off and **not a recommendation** — the signal sits inside run-to-run spread | OPEN |\n| A6 | The free-tier Gemini A/B is **void** as a comparison (fallback models cannot do the full job) | RULED (yours) |\n| A7 | You ruled **density/quality/attribution/addressability over latency** | RULED |\n| A8 | Therefore a **mandated break**; break audio is a separate meta-commentary artifact, never fed to the map | RULED |\n| A9 | **Cadence does not change the ratio** — only break duration and one cycle of lag | OPEN |\n| A10 | The honest cost: a bad draw either runs the break long or shows a thinner map | OPEN |\n| A11 | Break-recording **consent awareness** is owed in the session script | OPEN |\n| A12 | The focused-pass wall clock is **n=1** and the larger half of the break budget rests on it | OPEN |\n\n## Arc B — Decomposition: the misattribution, the JHU line, the builds\n\n| # | Claim | Status |\n|---|---|---|\n| B1 | `decomposition-axes.md` **misattributed** your Session-9 sentence as warrant for contest-gating; it sets a *completeness* criterion | RULED (you caught it) |\n| B2 | **The floor is separability** — decompose until nothing separable is still bundled; contest-gating demoted to a budget heuristic | OPEN |\n| B3 | *Molecular-not-atomic* is about **decontextuality**, a different axis — it never capped separation | OPEN |\n| B4 | The JHU group win with **theory in the prompt** (Russell's atomism + neo-Davidsonian event semantics); their pure-syntactic arm scored **worst** | OPEN |\n| B5 | Three quality dimensions the corpus lacked: **coverage · coherence · atomicity**; low atomicity is a named failure mode | OPEN |\n| B6 | **Coverage fails silently** and has no instrument here | OPEN |\n| B7 | `decompose_claim.py` built, 13 tests, propose-only, **not wired** — machine-proposed decomposition entering the correction UX is your call | OPEN — decision |\n| B8 | `check_coverage` is a **separate call** because a model must not grade its own omissions; caught a real invention on its first run | OPEN |\n| B9 | **CORE's gaming finding**: decompose-then-verify metrics are gameable by adding repetitive subclaims — the additive-energy defect as a *strategy-class* attack | OPEN |\n| B10 | `support_semantics.necessary` is the shipped payoff-removal defence and is **nearly inert** — 132 of 149 decomposition edges untagged on 2026-08-27, the stress suite having tagged 17 (12 `necessary`, 5 `corroborative`). **Corrected during the Phase-5 truth-check; the earlier \"all 131 untagged\" was stale.** Live: `uv run python scripts/graph_facts.py` | OPEN |\n| B11 | Only Pass 2b's output is stored — **the raw atom has no field**, so *which part of a claim is being attacked* is undecidable from the record | OPEN |\n| B12 | Cross-source premise run: **10–11 premises, 4 adversarial pairs, $0.64, zero junk**, positive control exceeded | OPEN |\n| B13 | The cross-source pass's premises are **all frame-level, none within-sentence** — because it reads stored claims, not raw sentences | OPEN |\n| B14 | Reading the text itself and reading around it are **structurally blind to each other** | OPEN |\n| B15 | The implicit-premise **ratio is not a meaningful figure**; corrected to ~10 pipeline-produced premises across 107 arguments the pass actually saw | OPEN |\n| B16 | That is **consistent with a prompt saying \"prefer fewer\"** — so it is NOT by itself evidence of under-extraction | OPEN |\n| B17 | The only real signal is hand extraction finding **~1.7× more per claim (n=2)** | OPEN |\n\n## Arc C — What human judgment is for\n\n| # | Claim | Status |\n|---|---|---|\n| C1 | Only **two** of eight residual-error classes are human-fixable: coherent absence and frame lock-in | OPEN |\n| C2 | **Ratification is the worst available use of a human**, on three independent grounds | OPEN |\n| C3 | Four acts: absence · reframe · refusal-to-endorse · stake testimony | OPEN |\n| C4 | Act three: *\"that is not what I meant\"* is a **self-interpretation** with no elevated credence; only the **present refusal to endorse** is incorrigible | RULED (you corrected me) |\n| C5 | **CONSENT vs RATIFICATION** — one word doing two jobs; consent ships, ratification stays operator practice and must be failable | OPEN |\n| C6 | **There is no ratify action** — `confirmed=true` is the default on human-authored edges, so the flag means *a human wrote this*, not *a human checked this* | OPEN |\n| C7 | Propose-then-ratify has a propose half built **five times over** and a ratify half built **once** | OPEN |\n| C8 | Rename `confirmed` → `ratified` (42 refs, 15 files) | RULED in principle, **not done** |\n| C9 | Votes survive **demoted** to instrument readings | OPEN |\n| C10 | An unratified human claim is **not a lesser claim** — the gray band says *unexamined*, never *unpermitted* | OPEN |\n\n## Arc D — Canon and the two principles\n\n| # | Claim | Status |\n|---|---|---|\n| D1 | Conviction: *what is invisible from inside a position is visible from outside it* | RULED — live |\n| D2 | **Positionality** is the resource; **visibility is not reducibility** | RULED (your correction) |\n| D3 | Principle: *never ask a mind to audit its own frame* | RULED — live |\n| D4 | Principle: *an act's value is not its rung* | RULED — live |\n| D5 | *Humans contribute objects, machines compute over them* | **SUPERSEDED** — a vote is an object too |\n| D6 | Narrowed: *contributions are added, never overwritten — nothing needs permission to exist* | RULED — live |\n| D7 | The real distinction is **tally-semantics**, not object-hood | OPEN |\n| D8 | **Existence is protected; weight is not** — the strength layer IS an aggregate and IS brigadeable | OPEN |\n| D9 | Weight is defended separately by provenance, attribution, sameness-merging, the `necessary` ceiling — **none tested against an interested party** | OPEN |\n| D10 | Three later-review questions: do the aggregate defences hold · does *nothing needs permission* survive at scale · is existence-without-weight honest to a contributor | OPEN |\n| D11 | Principle: *a guard that never fires looks exactly like a guard that works* | RULED — live |\n| D12 | Global directive: *a guard that holds only by accident is not a guard* | RULED |\n| D13 | Global directive: *a correction is not done until every surface that repeats it is corrected* | RULED |\n\n## Arc E — The cognitive bias codex\n\n| # | Claim | Status |\n|---|---|---|\n| E1 | The codex is **a map of what a mind does when it must decide under constraint** | OPEN |\n| E2 | **The graph never has to decide now** — permissive zone + `undecided` are the anti-act-fast affordance | OPEN |\n| E3 | Four pressures map to four structural reliefs (filtered material has somewhere to go · gap-filling becomes writable · no forced conclusion · persistence) | OPEN |\n| E4 | Honest claim: **Deliberus does not debias anyone; it removes the constraint that produces bias** | OPEN |\n| E5 | Supported by the one durable result: **technological interventions beat cognitive strategies** | OPEN |\n| E6 | Myside bias is **independent of intelligence** → no expert gate; your instinct is the literature's position | OPEN |\n| E7 | **No general bias-proneness factor**; test-retest 0–.82 → contributor-rationality scoring is measuring noise. **An empirical kill, more durable than a values argument** | OPEN |\n| E8 | Bias blind spot uncured by awareness, slightly *larger* in the sophisticated, **not mediated by actual susceptibility** | OPEN |\n| E9 | Therefore **self-report about one's own reasoning gets no elevated credence** | OPEN |\n| E10 | Correlated errors **accumulate rather than cancel** → register diversity from preference to **precondition** | OPEN |\n| E11 | Risk: a bias vocabulary is an **ad-hominem weapon** | OPEN |\n| E12 | **The ontology already answers it** — `bias` is a shipped Walton scheme (`schemes.py:73`) with `bias_exists`/`bias_influences` | OPEN |\n| E13 | **Never build a bias-tagging affordance** | OPEN — design refusal |\n| E14 | The reflexive trap: belief bias makes your side's ad hominem look relevant and theirs fallacious | OPEN |\n| E15 | **Open ontology question: should person-claims be first-class here at all?** | OPEN — decision |\n| E16 | Sunstein's four failures apply to the co-present session; minted claims are anti-hidden-profile, *nobody has said X* is anti-cascade | OPEN |\n| E17 | **Polarization has no structural answer** → I added a pre/post attitude reading to the workshop design **without asking** | OPEN — flagged call |\n| E18 | Do not import the taxonomy · do not claim we debias · **never let the LLM assess its own debiasing** | OPEN |\n| E19 | *\"A dropdown of 188 biases… is an ad-hominem machine with an academic finish\"* → recorded-phrasings list opened | RULED |\n| E20 | Threshold: **at three or four entries a phrasebook becomes a decision** | OPEN |\n\n## Arc F — Borrowed taxonomies\n\n| # | Claim | Status |\n|---|---|---|\n| F1 | Assembly theory and the codex fail in **opposite** directions | OPEN |\n| F2 | AT **overclaims a mechanism**; the codex **has no theory to contest** | OPEN |\n| F3 | Shared cost: importing a taxonomy imports its **induction corpus plus its unresolved disputes** | OPEN |\n| F4 | *\"You inherit the scar tissue without the scars\"* | OPEN |\n| F5 | The Walton set is the **controlled case**; `does_not_fit` is the general recipe | OPEN |\n| F6 | **The discriminator: a borrowed category earns its place by yielding a MOVE** | OPEN |\n| F7 | In both rejected cases the **abstraction one level up survives** | OPEN |\n| F8 | Default is **no**, since an imported taxonomy is opinionatedness at its most concentrated | OPEN |\n| F9 | The three-object correction: taxonomy-as-schema (no) · findings-as-design-knowledge (yes, already load-bearing) · arrangement-as-evidence (no) | RULED (you caught the collapse) |\n\n## Arc G — Belief bias and the weakest assumption\n\n| # | Claim | Status |\n|---|---|---|\n| G1 | **Belief bias is the named mechanism by which two-axis separability fails** — and appeared **once** in the whole corpus before today | OPEN |\n| G2 | Effect is **stronger on invalid arguments** — accepting bad reasoning you like is what a rigor axis exists to prevent | OPEN |\n| G3 | ROC refinement: **response bias, not discriminability** → calibratable rather than fatal | OPEN |\n| G4 | Measurable with no new instrument: **per rater, rigor ratings on agreed-vs-disagreed claims** | OPEN |\n| G5 | The persuasion literature establishes \"argument quality\" by **pretest** and is attacked for circularity | OPEN |\n| G6 | **Critical questions are already the independently-motivated criterion** that answers it | OPEN |\n\n## Arc H — Does descending change the descender\n\n| # | Claim | Status |\n|---|---|---|\n| H1 | *Decomposition is the debiasing act, endorsement the entrenching one; the graph can check* | **SUPERSEDED** — wrong in three ways |\n| H2 | The literature has **three** states: works · nulls · **reverses** (N=5,139) | OPEN |\n| H3 | **The dissociation**: understanding-collapse replicated, attitude moderation did not | OPEN |\n| H4 | ***Descent humbles*** stands on firmer ground than ***descent moderates*** | OPEN |\n| H5 | The moderator is **what the claim bottoms out in** → the terminus enum already encodes it | OPEN |\n| H6 | Sharp form: **descent moderates in proportion to how close the branch bottoms out to a checkable fact** | OPEN |\n| H7 | A flat result across terminus types would be **evidence against the enum's own cut** | OPEN |\n| H8 | The sacred re-scope is **empirically correct, not merely humane** | OPEN |\n| H9 | **Value-claim decomposition IS the reason-enumeration condition** — the arm that did not moderate | OPEN |\n| H10 | So the graph may hold a moderating and an entrenching act **behind the same button** | OPEN |\n| H11 | An interface asking *how does this work* differs from one asking *what is this resting on* → the causal axis may need **prompting differently** | OPEN |\n| H12 | The effect needs **self-generation** → readers get none of it; the benefit belongs to the contributor tier only | OPEN |\n| H13 | Audit correction: **not checkable here** — vote is binary, no attitude scale exists | OPEN |\n| H14 | Minimal instrument: **position slider (−3…+3) re-asked after a descent** — the same instrument the workshop pre/post wants | OPEN — decision |\n| H15 | Threat-model entry: the entrenchment hazard is **invisible to the entire instrument suite by construction** | OPEN |\n\n## Arc I — Sustained exposure and sacred values\n\n| # | Claim | Status |\n|---|---|---|\n| I1 | Your timescale objection is **correct** — everything else measures one sitting | RULED (yours) |\n| I2 | Bail: a month of opposing **conclusions** increased polarization | OPEN |\n| I3 | But it exposed **flags, not reasons** — and nobody has run that design with reasons | OPEN |\n| I4 | Kalla & Broockman: **arguments alone NULL**, arguments+narrative **durable at four months** | OPEN |\n| I5 | That is this project's dialectic **as a field experiment, with our arm as the null one** | OPEN |\n| I6 | Wood & Porter: **backfire is far weaker than folklore** (>10,100 subjects, 52 issues, none) | OPEN |\n| I7 | **Synthesis: the difference is neither dose nor duration but flag-versus-reasons** | OPEN |\n| I8 | Sacred values: **material incentives backfire, symbolic concessions work** | OPEN |\n| I9 | ***More reasoning is a material incentive in argumentative clothing*** | OPEN |\n| I10 | **Stake testimony is the graph's concession channel**, not decoration | OPEN |\n| I11 | The **position fingerprint** is structurally a symbolic concession | OPEN |\n| I12 | Honest limit: **a graph is not a party and cannot apologise** — whether recognition works as structure is unmeasured | OPEN |\n| I13 | Mutz: cross-cutting exposure lowers participation, and we may be industrialising it | **SUPERSEDED** — see I14 |\n| I14 | **Matthes meta-analysis: no overall relationship in either direction**; contingent on mediators | RULED (you challenged, correctly) |\n| I15 | **PACE is the mediator**; *no gap displayed without a move offered* is the answer | OPEN |\n| I16 | **Diffuse ambivalence vs located uncertainty** — the second is not less certainty | OPEN |\n| I17 | **Hinge + `worth_asking` already ships as the anti-Mutz instrument** and the corpus had not connected it | OPEN |\n| I18 | Narrowed risk: any surface showing a gap with no move; the **completeness oracle** is the one violation | OPEN |\n\n## Arc J — The over-broad refusal\n\n| # | Claim | Status |\n|---|---|---|\n| J1 | Four instances, all re-scoped **by you**, none caught by any instrument | OPEN |\n| J2 | **Under-broad fails loudly, over-broad fails silently** → coherent absence with an internal cause | RULED (you agreed) |\n| J3 | The tell: over-broad names a **noun**, correct names a **use** | RULED |\n| J4 | Two existing instruments aimed elsewhere: **stranger test** and **separability floor** | RULED |\n| J5 | **Rules should err narrow**, except where the hazard is irreversible | OPEN |\n| J6 | Relation to leveling dissolution: **argues a distinction away vs legislates it away** | OPEN |\n| J7 | Audit: 38 bolded refusals, **one live instance** — not widespread rot | OPEN |\n| J8 | **The failure lives in the HEADLINE, not the rule** — compression drops the scope clause first | OPEN |\n| J9 | Principle 9 rewritten to *\"Scaffold, never replace\"* | **SUPERSEDED** — fails the stranger test |\n| J10 | Proposed instead: **\"Confirming is not checking.\"** | OPEN — decision |\n| J11 | Deutsch's objection: **\"scaffold\" imports justificationism**, wrong for a graph built on attack edges | OPEN |\n| J12 | Feynman's relocation: **you can't tell having thought from having agreed**; rubber-stamping is cargo-cult ratification | OPEN |\n| J13 | A headline needs **two** tests: over-broad **and** stranger | OPEN |\n| J14 | Headline test as a **per-turn hook line** — costs tokens every turn, fleet-wide | OPEN — decision |\n| J15 | Killed as over-building: a re-scope register, and an LLM pass running the stranger test over rules | OPEN |\n| J16 | Deutsch–Feynman fight: **could not verify one exists** | OPEN |\n\n## Arc K — The two gap-types and the frame gap\n\n| # | Claim | Status |\n|---|---|---|\n| K1 | Two gap-types: **visible** and **frame** | RULED (yours) |\n| K2 | *No gap displayed without a move* **does not reach** frame gaps | OPEN |\n| K3 | **There the anxiety is correct and must not be designed away** | OPEN |\n| K4 | *You cannot make a frame gap feel closable, and you should not try. You give it an exit.* | OPEN |\n| K5 | `REFRAMES` exists; **the path from felt unease to a statable claim does not** | OPEN |\n| K6 | Design: **articulation help**, not a contentless flag | OPEN — decision |\n| K7 | The 14-question weighing descent is its **working sibling** | OPEN |\n| K8 | Two cautions: the question set **is itself a frame** (needs `does_not_fit`); and it must be **harvested, never invented** | OPEN |\n\n## Arc L — Population, coupling, funding\n\n| # | Claim | Status |\n|---|---|---|\n| L1 | **The population IS the horizon map** — exact, not figurative | OPEN |\n| L2 | **Co-presence is not engagement** — run 6: 402 candidate pairs, zero at threshold | OPEN |\n| L3 | **N users is not N frames** — a thousand from one reading community is one frame sampled a thousand times | OPEN |\n| L4 | Frame diversity is something to **select for, not await** | RULED (you agreed) |\n| L5 | The cross-source premise pass **is** the frame-engagement machinery, and is unwired | OPEN |\n| L6 | **Zoom is not attunement; lens is** | OPEN |\n| L7 | The stranger test is **subtractive of attunement's raw material** | OPEN |\n| L8 | But `source_span` + `attributed_to` + `raw_text` retain it → **missing from the read path, not the data** | OPEN |\n| L9 | **Build the return edge first, the lens second** | OPEN — decision |\n| L10 | Two-phase session; **the attunement phase forbids rebuttal**, and that prohibition is the active ingredient | OPEN |\n| L11 | Three clocks: NLnet **7 days** to open (68 to deadline) · Gemini ~11 · election **17** | OPEN |\n| L12 | **Funding and readiness compose; the election and readiness do not** | OPEN |\n| L13 | The Claude backend **unblocks readiness without money** | OPEN |\n| L14 | Suggested next: **the NLnet draft** | OPEN — decision |\n\n---\n\n## Arc M — asserted AFTER this ledger was written (the session outran the snapshot)\n\n*The ledger closed at commit `90e2d10`; seven commits followed. Everything below is post-ledger, so\nnone of it went through the count above. Added 2026-08-27 on the founder's completeness question.*\n\n| # | Claim | Status |\n|---|---|---|\n| M1 | **The meta layer is the gate** — the philosophy must be right before launch or it falls | RULED (yours, verbatim in `vision.md`) |\n| M2 | The three structural reasons I gave for it (no layer beneath the epistemology · a corpus cannot be un-thought · the credibility failure is an inversion) | OPEN — mine, not yours |\n| M3 | *\"No second first impression\"* | **SUPERSEDED** — you softened it; costly, not terminal |\n| M4 | **The discovery rate falling** as the checkable criterion for when the gate lifts | OPEN — entirely mine |\n| M5 | **Commit vocabulary**: seven harvested prefixes, `docs:` retired, no-prefix as the confession value | RULED (\"Sure\") |\n| M6 | Thinking precedes building **the same day, doc first, 4 of 5** | measured |\n| M7 | ...therefore *\"one motion, and the prefix cut it at the seam\"* — the diagnosis of the corpus/codebase tearing | OPEN — my interpretation of M6 |\n| M8 | `scripts/log_arc.py`'s path→kind classification scheme | OPEN — my design |\n| M9 | **Zero corrections in a window prints an alarm, not a pass** — built into that instrument | OPEN |\n| M10 | `scripts/graph_facts.py` + the convention *cite the command, never copy the number* | OPEN |\n| M11 | **A filter is an argument** — global directive, discharge by printing the complement | RULED (you ruled all three moves) |\n| M12 | **The success prior as the third always-loaded block** | RULED |\n| M13 | My compression of the four conditions into CLAUDE.md is faithful to the doc | OPEN — unverified by you |\n| M14 | The five consequences read as **standing mandates** rather than observations | OPEN — my framing |\n| M15 | *\"whether systems LIKE THIS get adopted\"*, not what the world does to us | RULED (your catch) |\n| M16 | The **why-now section in `vision.md`** — the two halves as one condition, two of four conditions arriving within three years | OPEN — my composition, unread |\n| M17 | **A headline needs TWO tests** — over-broad *and* stranger | OPEN (you ruled the first before the second was found) |\n| M18 | The inversion principle's **status section** in `ux-principles.md` | OPEN — my composition |\n| M19 | The **reference-implementation posture** in `competitive-landscape.md` | OPEN — my composition |\n| M20 | §16.1's Esperanto warning **composes with** today's *borrowed category yields a move* test | OPEN — my claim |\n| M21 | *\"Sweeps check presence, not truth\"* as a class-5 sibling measured on ourselves | OPEN |\n\n**Running total: 179 numbered claims, 145 open.** The ledger is a snapshot and will keep needing\nthis treatment while the session continues — which is itself the argument for ruling it down rather\nthan letting it grow.\n"}