{"path":"research/session33-the-horizon-vocabulary.md","content":"# Session 33: The horizon vocabulary — premature finality, named and ruled\n\n**Date**: 2026-09-14 → 2026-09-15 · **Type**: glossary blessing pass (loop 2) that opened into an ontology arc; one ruling shipped into canon, one into the design, eleven ledger rows, nothing built.\n\nThe session began as riff-mode glossary blessing, one term per turn, and the second term broke it open. Asked what a *terminus* is, the answer given was wrong in a way the founder caught immediately, and the correction ran through the whole horizon vocabulary, into `CLAUDE.md`, and out the far side as a design ruling about when a kind comes to exist at all.\n\n---\n\n## 1. The error that opened the arc\n\nAsked to gloss *terminus*, the session said that at an empirical terminus **the map is done and evidence takes over** — a handover of move-set. The founder: *\"'terminus' does feel like a misnomer for the factual case, since evidence breaks down into a subgraph, and all of the subgraph claims can/should be dealt with in depth and are amenable to all sorts of various analysis/breakdowns, right?\"*\n\nHe is right, and the corpus already said so in its own words: *a horizon is a fallible fixed point, not proven exhaustion* (`convergence.md`, Jul 9, Deutsch's point that proving no further decomposition exists would mean proving nobody will ever have a creative idea); *a terminus is dated, and the map does not say so* (`convergence.md`, Aug); *no residue is a dead end* (Aug 20/24). The claim layer had already chosen process-relative names (*currently undecomposed*, the *sorry frontier*) and the concept layer names lifecycle positions (underdefined · emerging · bifurcated · fragmented · locally stable). **The horizon enum was the one place end-state names survived.**\n\nThe move-set framing was withdrawn the same turn: the available moves are the same on both sides of the line (split a claim, ask a checklist question, attach evidence, weigh, clarify a term, add a missing premise, say what it costs you, reframe). What differs is only **what could close the remainder**, and even that continues below the line in a new register.\n\n## 2. Dissolution is not what happens to a definitional fight\n\nFounder, verbatim: *\"if I know humans right, often a conflict with all its associated conflict energy might just move to the level of fighting over whose definition of a term is somehow most obviously correct (or attempting to win the fight by default by simply implying their definition/use of it is obviously correct) and they will unfortunately often rationalize whatever defense and evidence they can to support their cause. But this is motivated reasoning in debate/politics and it is only one facet of the human condition, there is also honesty and curiosity and depth and empathy etc. So... clarifying the different meanings/senses of a word/term could absolutely help, but 'dissolved' is maybe too strong.\"*\n\nAnd on his second pass: *\"whenever there are competing definitions of a word, in human society and reality people will not have dissolved the conflict even if they all are clear on the fact that there is widespread semantic confusion/conflict. So 'dissolved' is a misnomer until idk either the contested term is split into two new words or people somehow magically merge the multiple senses ... or one sense is somehow proven to be most accurate or they come up with new words that serve the relevant purposes better.\"*\n\nThe canon already predicts this: the founding conviction says clarification **dissolves** a confusion-conflict and only **clarifies** an interest-conflict. So *dissolved* names an outcome, and when the fight relocates to the word, the outcome did not happen.\n\n**Outside names** (verify before citing outward): Chalmers, *Verbal Disputes* (2011) gives the **method of elimination** — bar the contested word, restate both positions without it, see whether the dispute survives; the corpus's own false-dissolution test (*after the dissolution, does anything about what to actually do change?*) is that method arrived at independently. Plunkett & Sundell (2013) name the founder's case exactly: a **metalinguistic negotiation**, a dispute that looks like it is about the world but is a negotiation over how a word should be used — and they argue it is **substantive**, not a confusion. His four ways such a fight actually ends (split the word · merge the senses · prove one · coin better words) is conceptual engineering (Cappelen 2018), and each is a concept-lifecycle transition the concept layer can record. Gallie names the case where it never ends.\n\n**Correction inside the session**: the first proposed rename was *semantic*, which fails the collision check — the extraction pipeline already uses *semantic* for **core**-level contestedness against *criterion* for **periphery**-level, and a fight over whose definition is right is usually criterion-level. Settled candidate: **definitional**.\n\n## 3. The founder's diagnosis of why the names came out final\n\nVerbatim: *\"all in all, the finality implied in 'terminus' and 'dissolved' etc is premature. ... They feel to me like they are based on coming from the origin and thrust of extraction (from some source/text/claim/mother-claim) and breaking it down and wanting to name the parts but naming them prematurely, because the work is nearly neverending or we cannot tell clearly whether we are nearing complete resolution and clarity and peace on earth or whatever. Do you understand what I am getting at?\"*\n\nConfirmed against the origin: the four value kinds were induced on 2026-07-10 from two hand-walked descents (`deep-descent-experiment.md`), one person descending in one sitting, stopping, and naming what was at the stop. **From inside one sitting the stop feels like a bottom; from the map's side the same point is a horizon that moves.** The names photographed the sitting.\n\nMeasured footprint of the surviving finality words at the time of the arc: `terminus` 184 lines in 18 files · `verdict` 152 in 29 · `residue` 45 in 8 (despite the Aug 25 rename ruling) · `is_bedrock` 25 in 7 · `dissolved` 14 in 5.\n\n**The principle, blessed hedged** (*\"Otherwise I guess I am game for the naming principle.\"*): every name in this vocabulary names a **position and a direction** — where shared mapping currently ends, and what kind of work continues past it — never an end state.\n\n## 4. The permissive zone, examined claim by claim\n\nThe founder asked for concrete candidates, diligently assessed, and for what the kind is worth. Measured first: the live graph holds **three typed claims** (one fittingness, one empirical, one is-bedrock) and **no permissive_zone** — the kind has never fired on real data, and the corpus's one found instance is the white-lie descent leaf, a fittingness/parity blend the taxonomy could not hold cleanly.\n\nSeven candidates were assessed in `parity-and-the-zone-interior.md` § Addendum 2026-09-14, including two negatives that show the kind is earned rather than declared (the minimum wage inside a band whose edges are still empirically contested; *Bach is greater than the Beatles*, which usually reads permissive and is not). The boundary in one line: **a permissive zone is a fittingness dispute in which both sides acknowledge more than one fitting answer.**\n\n## 5. The grant, proposed and withdrawn the same day\n\nThe first computed-reading design made one human move load-bearing — *I accept your landing as permissible*. The founder rejected it: *\"Relying on humans stepping in and saying something like that feels rare and should not be counted on. I would rather let the graph score of how well-supported any and all 'options' claims are determine what shows up as a 'permissive zone'. It feels more realistic and simpler rather than requiring a special type of move?\"*\n\nHe is right by his own 2026-08-31 ruling, which the grant violated: *human force enters the graph only through contributions that argue for themselves; bare judgments are telemetry.* A grant is a bare judgment. Permissibility is read instead from what survives attack. The resulting template is the badge's own three honest states lifted from one claim to a set of sibling options (`what-belongs-in-the-ontology.md` § 6c.2).\n\n## 6. The label is never a home for content\n\nThe founder's test case, verbatim: *\"someone could post a claim saying 'It is worth up to 15% risk of catastrophe to achieve automated AI-driven healthcare in the Nordics if it were to triple effectiveness/reach/QUALYs saved' or whatever (also needs a definition of 'catastrophe' ofc...) and then ppl could post competing claims, and then ppl could claim something about how large the risk % ACTUALLY is for the initiatives that are ongoing etc... so I am not sure how/when this label is a better home for this content rather than the text inside the claims. But do not take this as me having made up my mind.\"*\n\nWalked through, every piece has a claim home and none needs the label: posture claims are the options, actual-risk claims an evidence branch, *catastrophe means…* a definitional branch, *your 15% contradicts your posture elsewhere* a cross-attack. The word is what the map uses when it **counts** (the wager scoreboard) and when it **routes** (which move to offer at this leaf). Recurrence across debates is served by claim sameness on the option claims, not by the label.\n\n## 7. Loop 1, the root, ruled: read, never set\n\nEstablished first, by reading the code rather than recalling it: **nothing sets a horizon kind automatically.** A signed-in person runs the classifier (`POST /claims/{id}/terminus/classify`, stored as `terminus_llm_*` proposal fields that never touch the cache) and then confirms a kind (`POST /claims/{id}/terminus` → `set_terminus_with_ground`), which writes it once, undated, and mints the verdict-claim. Nothing re-reads it when the claims beneath change. The three live typed claims were set by hand.\n\nThree options were put; the founder ruled **\"The second.\"** — never set, compute the reading from the arrangement, refresh on change, classifier proposes structure only.\n\n**His caveat, which is the honest residual**: *\"But it could be premature, if the refreshed re-read subtree maybe is not final (in reality, in hindsight after a day or a year) but LOOKS final to the LLM or whatever is judging this at the time of the read/re-read, which there is probably some vulnerability to. We cannot foresee magically all future moves carried out on the subtree... Am I right?\"*\n\nYes. Two things could look final, the judge and the arrangement. The ruling removes the judge. What remains is that the arrangement looks settled because the move that would unsettle it has not been contributed — **coherent absence** by name, one of the two human-reserved error classes, invisible to any computation over the graph because it lives in what is not in the graph. It is also the ordinary condition of every defeasible conclusion: the strength badge has the same exposure and nobody calls it premature, because it claims *given what is here*. The reading claims the same. The design constraint that follows (dated · cross-attack record shown · open-world states · slicing stability · novelty rate as liveness · no gate) is in `what-belongs-in-the-ontology.md` § 6c.3.\n\n## 8. What was ruled, and what is still open\n\n**Ruled this session**: the naming principle (hedged); *read, never set*; the canon line **Nothing is permanently terminal**, blessed into `CLAUDE.md` beside its atom-side twin and reworded to carry his conservatism (*\"We should also be a bit conservative about what we classify as terminal, even when knowing it is not permanently terminal, obviously\"*). **Ruled the next morning (2026-09-16)**: the auto-minting of the July verdict-claim goes — *\"Right. So we ditch the automatic minting? Does make sense yes, sure, do it.\"* — with the code change sequenced together with the computed reading, since removing the minting while the kind is still hand-set would leave the bare stamp July rejected.\n\n**Ruled and built the next morning (2026-09-16)**: the reading and the end of auto-minting landed together (*\"Ok but then just make both changes please.\"* — `deliberus/horizon.py`, `GET /claims/{id}/horizon`, the setter answering 410); the rename table's first word, *dissolved* → *definitional*, with *terminus → horizon kind* and *verdict → reading* agreed as following from loop 1 (*\"ah, yes, agreed, and choose \\\"definitional\\\"\"*). Live: 201 candidates read, 0 value horizons readable because no option set has been cross-attacked, so the wager fraction reports *not yet measurable* rather than 0.0 (`what-belongs-in-the-ontology.md` § 6c.4).\n\n**Open, filed as nine loops in `TODO.md` § OPEN REVIEWS** (loops 1, 6 and 7 closed; loop 2 has its first word): the rename table's remaining words · the buckets sorted by who could settle a question · the zone template · whether the 15%-case walkthrough settles the label question · ledger E2 · the seven candidates · then glossary loop 2 resumes at *fittingness*.\n\n**Received 2026-09-19, and deployed the same day**: the leaf rule (a kind is read only at the last claim of a descent, never partway down) was accepted as revisable, and his answer opened a loop above it — *\"Since this \\\"stop\\\" is only ever a provisional stop, right, we shouldn't maybe put too much emphasis on what the current \\\"guess\\\" is\"* — which is now loop 10: how loud a provisional reading may be on any surface. Proposed and unruled: lead with the record of what was tried, trail with the reading (`what-belongs-in-the-ontology.md` § 6c.5, ledger O12). The reading itself went live on deliberus.com that afternoon with the first deploy since 2026-08-28.\n\n**Corrections made inside the session, both on the founder's challenge**: *semantic* → *definitional* (collision with the concept layer); and the claim that the computed reading *supersedes* the July hybrid and *retires* the verdict-claim, softened to what is true — the cache half survives upgraded, the auto-minting becomes unnecessary, and the verdict-claim remains available as ordinary content.\n\n## Cross-references\n\n- `docs/research/what-belongs-in-the-ontology.md` § 6c (enum or claim), § 6c.1 (premature finality, the rename table), § 6c.2 (the grant withdrawn, the zone template, the 15%-risk walkthrough, why `CLAUDE.md` did not carry it), § 6c.3 (loop 1 ruled, the residual, the honesty kit)\n- `docs/research/parity-and-the-zone-interior.md` § Addendum 2026-09-14 — the permissive zone claim by claim\n- `docs/research/weighing-scheme-and-terminus-classifier-design.md` § Addendum 2026-09-14 — pointer from the design home of the July ruling\n- `docs/research/session26-open-rulings.md` Arc O — thirteen rows, four ruled (count by `rg '^\\| O' | rg -c RULED`; the number here decays)\n- `docs/research/the-canon-distillation-pass.md` § 6 candidate 11 — blessed\n- `CLAUDE.md` — the canon line; `TODO.md` — the nine loops\n"}