{"path":"research/the-tidying-line.md","content":"# The tidying line: what the machine may clean up, and what a visitor may see it having cleaned\n\n**Date**: 2026-09-21 · **Status**: corpus sweep and advice; **the tidying half is already ruled and\ndesigned, the showing half is not ruled** · **Homes of the decisions it gathers**:\n`graph-daemons-design-space.md` (tiers, preconditions) ·\n`soft-canonical-clustering-and-reversible-merge-semantics.md` (the housekeeping ruling's decision home) ·\n`kartpaus-auto-enrichment-design.md` (the design that already does this)\n\n**Origin.** The founder, after being told that splitting sits in the propose-only tier: *\"I'm curious\nabout maybe redrawing this line… to (in general, as often as is wise) spare users from having to\nencounter chaos/clutter [meaning maps/claims that are unsplit when they could confidently, accurately\nbe split by machine/daemon/pre-runs, or perhaps irrelevant spam-claims that could be marked irrelevant\nby machine, or unconnected/similar/sameness-claims that could have been marked same/similar/connected\nin some way] and instead to the extent possible without making glaring mistakes tidying up regularly\nwhere it makes the most difference to reach and show examples of more clarity and illustrate the power\nof separability/decomposition/addressability etc. Read up on everything that's remotely relevant for\nthis in the existing corpus pls and advise.\"*\n\n---\n\n## 1. The headline: the line he wants to redraw, he already redrew\n\nThree rulings of his own, none of which the propose-only tier table reflects on its face:\n\n**2026-08-25, the eager-decompose ask.** *\"I want Deliberus to automatically recursively decompose all\nclaims in each mapping-break… automatic recursive decomposition and deepening exploration and relevant\nresearch/googling should fire/engage… and I'm leaning towards that this should also occur (to some\ndegree/cap at least) in normal use.\"* Graduated into `kartpaus-auto-enrichment-design.md`, whose status\nline reads **direction RULED, build sequenced behind the live-test readiness suite**.\n\n**2026-09-13, the housekeeping default.** *\"Any regular housekeeping should be done that could be done\nwith high confidence in safety/accuracy and could spare human users from having to later clean up and\norganize the graph manually etc, to instead free up their mental cycles for the uniquely human-needing\ntasks.\"*\n\n**2026-09-13, the same day, on a machine-written link that fails a confident re-check**: *\"Yes, replace\nby itself, keep the old link visible.\"*\n\nSo the answer to *may the machine tidy* is **yes, ruled twice**, and the propose-only tier's line about\ndecomposition is the pre-2026-08-25 statement of a boundary two later rulings moved. **The tier table\nwas never re-stated after them**, which is why it reads today as a prohibition on exactly the thing he\nasked for — a documentation lag, not a live disagreement.\n\n### 1a. And the principle it seemed to violate was never ratified\n\nThe collision was found and resolved in the Aug-25 design: *\"decompose along an axis only when someone\ncontests along it\"* was **corrected 2026-08-27 as never ratified** (it misattributed a completeness\ncriterion to the founder). Contest-gating survives only as a **budget heuristic for where to spend\nmachine effort**, so eager decomposition *\"does not violate a principle, it spends a budget.\"*\n\n---\n\n## 2. His three examples are the corpus's named core work, not a new workstream\n\n`drowning-in-claims.md` § 4 reframes the whole flood question and lands exactly on his list: **in a\ngraph, volume only costs when unconnected or unmerged.** Two pathological forms, both his:\n\n| His words | The corpus's name | Where it already lives |\n|---|---|---|\n| unsplit when it could confidently be split | bundledness below the **separability floor** | the split machinery, built 2026-09-21 |\n| similar / sameness claims unmarked | **identity scattering** — evidence diluted, false disagreement manufactured | the merge ladder, designed, eval-gated |\n| unconnected claims | the **orphan** | auto-connect, live |\n| irrelevant spam-claims | *(the only one with no home)* | see § 5 |\n\n> *\"the anti-drowning programme is not a new workstream — it is claim-sameness plus auto-connect plus\n> honest strength, which are already the named core activities.\"*\n\n**And two of the three are already one machine.** The 2026-09-13 ruling's stated first instance: a\n*contains* or *overlaps* label in the pair review **raises a decomposition proposal by itself**, after\nwhich the pair re-labels at leaf grain. The sameness review doubles as a separability audit, and the\nordering was ruled the same day (*\"Confirmed the proposed ordering, thx\"*): the seven-key sameness check\nruns **before** any relation is written.\n\n---\n\n## 3. What actually blocks it, and it is one sitting rather than a decision\n\n`graph-daemons-design-space.md` § 5 lists six preconditions. Measured against today:\n\n| Precondition | State |\n|---|---|\n| A proposal layer | **exists for splitting** (`split_proposal`, one property per claim, 2026-09-21); no general one |\n| The ontology constraint layer | approved, partly built (the badge query names its types since 2026-09-17) |\n| Provenance for non-human authors | `count_safe_summary` exists and had no caller when this was written; since 2026-10-10 `/graph/stats` calls it, and a machine's sentence under a link has its own kind and count. A daemon's writes have neither yet |\n| An undo path | **partly real**: originals-never-destroyed, edge supersession live since 2026-08-22 |\n| A spend gate | designed; billing-off is a normal mode here |\n| **The strength layer settled first** | the August audit's four defects are fixed; this one has cleared |\n\nThe single live gate is the **evidence gate he himself revised on 2026-08-31**: a few hundred\nmachine-proposed pairs reviewed smoothly by keyboard, with **seeded wrongs** mixed in (so agreement is\ndistinguishable from rubber-stamping) and a **blind subset** (so the anchoring effect is measured rather\nthan assumed away), then friends taking the same test so per-kind divergence sets the bars empirically.\n\n**BUILT 2026-09-21, the same day the question was asked** (founder: *\"Let's do this sitting now!\"*).\nIt is at `/sameness/sitting-1`, signed in, one keystroke per pair, backspace to change an answer.\n120 pairs built from the live graph in 19 seconds: 33 near (cosine ≥ 0.90), 34 close, 33 loose, plus\n**20 seeded wrongs** and **24 shown blind**. `deliberus/sameness.py`, `deliberus/database.py`\n(`sameness_answers`), `GET|POST /sameness/{sitting}` and `GET /sameness/{sitting}/results`.\n\nThree build findings worth keeping:\n\n- **The seeded wrongs need no fabrication.** A pair joined by an `ATTACKS` edge is in conflict, so\n  *the same* is definitely the wrong answer — and the negation literature says those are exactly the\n  pairs a cosine matcher scores as near-identical, so they look like merge candidates and are not.\n  1,852 of them exist in the graph; the sitting draws its seeds from there.\n- **The candidates could not come from the stored `SIMILAR_TO` edges.** Measured: **56** of them\n  across 4,852 claims, every one carrying `similarity = 0.8` — the linker's own threshold written\n  down rather than a measured score. A sitting drawn from those would have calibrated a bar against\n  a constant and looked like it worked. The pairs are computed instead, by asking the database for\n  each sampled claim's nearest neighbours.\n- **The blind subset is withheld by the API, not hidden by the page.** An anchoring measurement is\n  worthless if the number is sitting in the response for a reader to notice.\n\nWhat it does not do, deliberately: nothing is written to the graph, and no bar is set anywhere. The\nscorer reports the catch rate on seeds, the same-rate with and without the machine's proposal, and\nthe per-band and per-kind breakdowns — the material a bar is read off once more than one person has\nsat it.\n\n---\n\n## 4. The genuinely new half of his message — and the corpus argues against it as stated\n\nHis ask has a second clause the earlier rulings do not cover: tidy *\"where it makes the most difference\nto reach and **show** examples of more clarity and **illustrate the power** of separability/\ndecomposition/addressability.\"*\n\nThe Aug-25 design resolves the tidying question **by hiding the tidying**: *do eagerly, show on contest.*\nEverything the machine works out lands at **draft lifecycle, machine provenance, invisible on the default\nmap**; what a participant meets stays contest-gated, so the descent is already there the instant someone\nattacks a claim. That is *capture immediately, publish deliberately* applied to the machine's own work,\nand it is what keeps the enzyme constraint intact — the machine changed the **rate** of a move a human\nwas about to make, never the **equilibrium**.\n\n**So the design achieves his tidiness and refuses his showing.** Three independent warnings say why:\n\n1. **Interpassivity / the T-cell warning.** The corpus's own T-cell passage is a *warning*, not an\n   endorsement: an intellectual who pre-digests challenges spares the in-group from engaging with full\n   complexity. *\"A daemon that pre-digests does exactly that to Deliberus's own readers.\"*\n2. **The experimental backing, added 2026-08-25.** Question-framed AI assistance is **developmental**;\n   finished-output assistance **attenuates** (Danry, Everett, Wiles, Rind) — converging with the DeepMind\n   facilitation result, where question-asking was clean and restatement steered. Hence the sharpened rule:\n   **the machine's pre-computed descent must arrive as questions requiring the participant's own\n   attempt — ratification, polarity answers, counterweight-naming — never as finished analysis.**\n3. **The ratification tether.** Every propose-only instrument is safe *because* a ratifier can tell a good\n   proposal from a plausible-wrong one. Tidying at scale grows the pile of ratified-but-unvalidated\n   structure, whose failure shape is that volume metrics stay healthy while validation goes surface.\n\n### 4a. The resolution: it is the FORM of the showing, not the amount\n\nThe three warnings all fire on **finished analysis**, and none of them fires on **structure that still\ncarries its question**. So the two halves of his ask are compatible if the distinction is drawn at the\nform rather than the volume:\n\n- A split shown as **done** — *here are the six parts* — is finished analysis, and every warning above\n  applies at full strength.\n- A split shown as **parts with the question still open** — *this bundles six things; which one do you\n  actually disagree with?* — is question-framed, and it illustrates separability **better**, because the\n  reader performs the separation instead of reading about it.\n\nThis is the same move the product already makes with a sorry marker, which displays a gap *and* a move\nrather than a finding. It is also what makes the showing safe for a **signed-out** visitor, which was the\nopen question that started this: a stranger may see the machine's structure, provided it arrives as an\ninvitation rather than a verdict.\n\n### 4b. Two display constraints already ruled, which the showing must respect\n\n**A map is a summary-shaped surface** (`saturation-and-the-long-tail.md` § 3b, his own *\"super important\nfinding\"*), and **since 2026-09-21 it is the main door.** Measured on the AI-governance source: a first\nview of 12 boxes held **1** thinly-linked claim and **1 of 3** unspoken premises. A tidy-up that raises\nthe best-connected material further compounds exactly that. So any visible tidying ships with the map's\nown tail rule: one reserved and marked place for the rare item, and an omissions line.\n\n**And what a surface shows shapes what arrives.** Showing the common breeds restatement; showing the\nrare-and-strong breeds new contribution. A tidied map that displays only well-formed, well-connected\nmaterial would train its own contributors to produce more of it.\n\n---\n\n## 5. The one item with no home: *irrelevant spam-claims*\n\nNothing in the corpus covers marking a claim irrelevant. The nearest things are all different:\n\n- The **orphan** (`drowning-in-claims.md`) is *unconnected*, which is a fixable state rather than a\n  judgment about worth — and auto-connect is its answer.\n- **Lifecycle quarantine** holds `raw` / `draft` / `superseded` material out of default retrieval, which\n  handles immaturity, not irrelevance.\n- An early sketch (`epistemic-gamification.md`) proposed an *important / unimportant* binary vote, from a\n  design generation the project has since moved away from.\n\n**And it collides head-on with a ratified principle**: *contributions are added, never overwritten —\nnothing needs permission to exist.* An irrelevance mark that hides a claim is deletion-by-judgment\nwearing a filter's clothes, and it is precisely the move a motivated stakeholder would want. Adversary\nclass matters here: bundledness and orphaning are **entropy-class** (no agent, so a detector built once\nworks forever and publishing it is free), while *irrelevant* is **strategy-class** the moment anyone has\na stake — someone will argue their opponent's premise is off-topic.\n\n**Recommended shape if it is wanted**: never an *irrelevant* flag, but the existing machinery pointed at\nthe same discomfort — an orphan stays visible and unlinked, a duplicate merges, and material below the\nattention price displays **as a state** rather than as a prompt (the `worth_asking` inversion, already\nbuilt: below the line a gap shows as a state line rather than a question).\n\n---\n\n## 6. Advice, in one place\n\n1. **Re-state the tier table** so it reflects the two 2026-09-13 rulings instead of contradicting them.\n   Its decomposition line is a pre-ruling artefact and will keep misleading sessions — it misled this one.\n2. **Run the evidence sitting.** It is the only live blocker, it is one lazy hour by his own design, and\n   it calibrates merge + containment + novelty together. Everything downstream is gated behind it.\n3. **Build the tidy-up on the existing loop, not a new one**: the shared post-store layer already runs\n   passes after every extraction, the split proposer already mounts there, and the stopping rule is\n   already mechanical (`worth_asking` — spend the next action where the expected badge movement across a\n   display band is highest; stop where nothing you could add would change what a reader sees).\n4. **Keep the showing question separate from the tidying question**, and rule it on form: structure that\n   carries its question may be shown to anyone; finished analysis may not.\n5. **Do not build an irrelevance mark.** Point the existing instruments at the same discomfort instead.\n\n**Unruled, and his to settle**: (a) whether machine-tidied structure is visible to a signed-out visitor\nat all; (b) whether the *shown as a question* form above is the right resolution of § 4; (c) whether a\ntidy-up daemon may run continuously or only on ingest.\n\n---\n\n## Cross-references\n\n[kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md) (the design this is the\ndisplay-side sequel to) · [graph-daemons-design-space.md](graph-daemons-design-space.md) (tiers, enzyme\nconstraint, preconditions, the rejection-rate pause) ·\n[soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md)\n(the housekeeping ruling's decision home; the evidence gate he revised) ·\n[drowning-in-claims.md](drowning-in-claims.md) (duplicate and orphan as the only pathological forms) ·\n[cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md) (question-framed warrant; the\nratification tether) · [saturation-and-the-long-tail.md](saturation-and-the-long-tail.md) § 3b (the map\nas a summary-shaped surface) ·\n[how-a-map-is-cut-and-what-it-shows.md](how-a-map-is-cut-and-what-it-shows.md) § 10 (the tail's reserved\nplace) · [the-jhu-decomposition-line.md](the-jhu-decomposition-line.md) (the split machinery and CORE's\ngaming finding) · [what-human-judgment-is-for.md](what-human-judgment-is-for.md) (contributions added,\nnever overwritten; the two reserved classes)\n"}