{"path":"research/the-ratification-record-and-the-triage-feed.md","content":"# The Ratification Record and the Triage Feed\n\n**Date**: 2026-08-30 · **Type**: design direction from a founder side-chat, captured before it\nevaporates (the side-chat flagged itself as unfiled twice). **Status: on the table, unruled** —\nthis doc merges two items that were sitting separately on the founder's desk (the anti-pattern-9\nrewrite and the confession-must-pay rule) because the reasoning showed they are one rule seen from\ntwo sides, and adds the founder's own synthesis: the triage feed.\n\n---\n\n## 1. What the harness conviction killed in anti-pattern 9, honestly\n\nThree planks of *\"Scaffold, never replace\"* as written do not survive the founder's delegation\nclarification and the harness conviction:\n\n1. *\"Never a substitute for judgment\"* — contradicted directly: **offloading is the purpose.** The\n   machine substituting for large amounts of human cognitive work is the point, not the hazard.\n2. *\"The confirmation step must provoke engagement\"* — this **conscripts** engagement, against the\n   standing formulation *review always invited, never conscripted*. Friction designed to force\n   thinking is imposed on someone who may legitimately, wisely delegate.\n3. The premise that the human's engagement is the quality-bearing element — relocated by the\n   harness conviction into harness plus grounding. If the harnessed machine reasons better,\n   forcing re-derivation can make outcomes worse.\n\n## 2. What survives: a record-rule, not an engagement-rule\n\nThe real failure in the corpus was never that people clicked without thinking — it is that **the\nclick was recorded as a check**. `confirmed=true` means *a human wrote this* while every reader\ntakes it as *a human examined this*: a provenance lie, the same shape as the badge incident (not\nwrong to compute on people's behalf; wrong to have no way of saying *\"I haven't looked\"*). So the\nrule migrates **from regulating the person's cognition to regulating the record**:\n\n- **Ratification states carry their weight honestly** — *examined* and *accepted-without-review*\n  as distinct first-class states, and (per the § 3b resolution, founder-agreed) that weight is\n  **telemetric, never settling**: an examined-state raises confidence the way a reviewed-by line\n  does, and settles nothing on contested content.\n- **The honest state is inferred, never clicked.** Nobody would press an \"I didn't look\" button —\n  so there is no confession button. Accepting records exactly what happened: *accepted*.\n  *Examined* is the state that needs evidence, earned by observable behavior: opened the premises,\n  answered a critical question, edited something. A default that does not lie, plus a stronger\n  state you earn — the same move as the badge fix, computed from provenance already present rather\n  than asked for.\n- **Display follows**: a claim whose only human touch is an unexamined click does not render like\n  an examined one.\n- **Engagement gets elicited — by stakes, by a visible gap with a move offered — never\n  manufactured by friction.** Make *yes* and *didn't look* equally cheap, so honesty costs nothing.\n\n**The merger**: this record-rule and the confession-must-pay candidate\n([reflection-and-the-graph.md](reflection-and-the-graph.md) § 4) are one rule from two sides —\n*never record agreement as examination* fails if the honest label penalizes you (people learn to\nlie), and costless labels fail if nothing records them (decoration). **They should be ruled\ntogether.** Candidate headline family, stranger test with Malin pending: *\"a click is recorded as\na click, a check as a check.\"* Every dead headline before this (*scaffold, never replace*;\n*confirming is not checking*) was trying to regulate cognition; a record-rule self-decodes more\neasily.\n\n## 3. The founder's synthesis: the triage feed\n\nHis words, from the side-chat (verbatim):\n\n> \"a human surfing around Deliberus should be subtly encouraged (but not conscripted/required) to\n> evaluate different kinds of discernments made by LLMs in whatever circumstances we can't be 105%\n> confident in the reliability of the targeted prompting of the LLM-step... but ideally humans\n> would PRIMARILY be nudged (perhaps a feed could be created) to sift through suspected candidates\n> for human-reserved classes of error, with incredibly clearly articulated yet short 1-sentence\n> prompts for the human user. It could make optimal use of the humans perusing Deliberus, instead\n> of having them re-check judgments and epistemic glue/inferences/connections we're already quite\n> confident in.\"\n\n**The primary human surface becomes a triage feed**: machine-uncertain discernments plus suspected\ncandidates for the human-reserved error classes (coherent absence, frame lock-in), each with a\none-sentence prompt — and nothing the instruments already hold with measured confidence. Four\ncorpus threads have been circling this without saying it: **materiality** from auditing (effort\ngoes where misstatement risk is highest; most entries are never examined, by design),\n**`worth_asking`** (hinge-ordered questioning already ranks attention within a descent), the\n**feed-as-dashboard principle** (Bayesian surprise; *\"this needs evidence and you have\nexpertise\"*), and the human-reserved classes themselves.\n\n**The supply side already exists.** Every instrument produces an uncertainty ordering whose tail\nhas had no consumer: cross-family jury disagreement · low terminus-classifier confidence · the\ngray band · stale claims · `does_not_fit` flags · the completeness oracle's unsupported value\npremises — which is the standing violation of *no gap displayed without a move offered*, fixed by\nexactly this feed (the gap arrives WITH its one-sentence move). And the **`needs-help` feed already\nships** (`deliberus/feed.py`, sorry-density ranking): the triage feed is that feed generalized from\none signal to the full instrument-uncertainty inventory, with the prompts made one-sentence-clear.\n\n## 3b. What the feed invites, and the voting principle it must honor (founder question, 2026-08-30)\n\nThe founder, verbatim: *\"I think in the past I've expressed serious skepticism about humans getting\nto vote on anything where they could lie based on motivated cognition or manipulation/corruption\netc, so they should only be able to vote on whether they agree/disagree with a claim... but the\ntriage-feed and the actions/judgments that it would invite from humans seem to be something else?\nThe invitation/incentive here would be to contribute with something outside of the mental\nbox/framing etc... to submit new related claims or some such?\"*\n\n**The record confirms both halves.** The skepticism is the shipped Option-B ruling (2026-03-31,\n`ux-principles.md` P2): humans vote on stance ONLY — agree/disagree, the one thing each person is\nsole authority on — while quality is never voted, because LessWrong measured that >95% of quality\nvotes just track agreement (motivated cognition, quantified); quality is instead **computed** (the\nbadge, from CQ evidence) or **argued** (a claim that must defend itself). And the 2026-08-27\nadditivity refinement ([what-human-judgment-is-for.md](what-human-judgment-is-for.md) § 4d) gives\nthe underlying criterion: **does the contribution mean something on its own, or only relative to a\ntally?** An absent premise means what it means at n=1; a vote means nothing alone — and tallies are\nwhere numbers can win.\n\n**So yes — the triage feed's invitations sit in the tally-immune class by design.** A\ncoherent-absence prompt invites *the missing claim*; a frame prompt invites *a reframe arriving as\nclaims*; a CQ prompt invites *an answer that argues for itself*; a `does_not_fit` review invites\n*a reasoned note*. Each is a standalone-meaning contribution that no brigade can amplify by\nrepetition, because adding it twice adds nothing. The feed routes humans TOWARD the act class the\nadditivity principle protects and AWAY from tally-shaped acts — the founder's instinct and the\nratified principle are the same design.\n\n**The seam, and the founder's catch (2026-08-31) that closed it properly.** The feed also surfaces\nmachine-uncertain DISCERNMENTS (terminus type, scheme fit), and checking those is a judgment. This\ndoc first said its defenses were \"of a different kind\" — seeded probes, inferred states, bounded\nweight. The founder: *\"doesn't the skeptic-refusal (3b) of voting/ratification apply to this\nremaining seam as well, unfortunately?\"* **Yes — and the listed defenses do not cover it.** Worked\ncase: a claim in an abortion debate is classified `empirical`; a motivated user asked *\"is this\nright?\"* can drift the discernment toward their side exactly as quality-votes drifted toward\nagreement — the judgment is about contested content, it affects standing (scheme → base weight,\nterminus → the wager's accounting), and generic seeded probes miss it, because **a motivated\nratifier passes neutral probes and lies selectively where it counts**. Nor does\none-ratification-not-a-tally save it: single-ratifier authority just swaps brigade-risk for\nfirst-mover capture.\n\n**The resolution is the same demotion votes received — one principle now covers both**:\n**human force enters the graph only through contributions that argue for themselves; bare\njudgments are telemetry.** Concretely: *assent* to a machine discernment never settles it — it is\nrecorded as \"examined, no objection filed,\" a weak honest confidence signal like a view-count,\nnever a verdict; *dissent* gets full force but must arrive as a reasoned reclassification — a\nrecord-object that argues, is attributable, and is challengeable, which is the tally-immune class\nby construction. So the examined/accepted states (§ 2) keep their record-honesty job but carry\n**zero settling authority on contested content**. Two riders: **motivated-shaped probes** — seed\nwrong proposals that would FAVOR a ratifier's apparent side, since only those detect selective\nhonesty; and a **phase split** — operator-phase calibration (the founder and friends, dogfooding,\nno stakes) may treat review as authoritative, while at scale, with motivated stakeholders real,\nthe demotion above is the design. **AGREED by founder 2026-08-31** (*\"Agreed on everything\"*); § 4d's ledger-OPEN row (strength-layer defenses\nuntested against a motivated stakeholder) still stands behind all of it.\n\n## 4. Why ratification is worth routing carefully — CORRECTED after the § 3b resolution\n\nThe first version of this section called ratification \"the audit signature,\" the one act that\n*transfers warrant* without changing the object. **That over-promised**: an audit signature carries\ninstitutional authority, and the founder-agreed demotion says bare assent never settles anything.\nWhat survives, stated honestly: attention is still the only signal humans alone can mint — *a\nspecific person attended to this specific structure* — but its force is **telemetric** (a weak,\nper-person, attributable confidence signal, like a reviewed-by line), while **the strong form of\nhuman force is the argued objection**, which is a contribution and already tally-immune.\n\n**The sole-authority criterion turns out to unify all three human channels**:\n\n| Channel | What it judges | Sole authority? | Force |\n|---|---|---|---|\n| Stance vote (agree/disagree) | your own position | yes | full, as self-report — but demoted to instrument reading in aggregate |\n| **Consent** | a rendering of YOUR OWN position | yes | **full and settling** — the one verdict no majority can win |\n| Discernment review (scheme, terminus, merge) | the shared world | **no** | assent = telemetry; dissent must argue |\n\nOne criterion — *are you the sole authority over the thing judged?* — decides where a bare human\njudgment may settle. This is the additivity principle and the Option-B voting ruling arriving at\nratification, and it is why the consent channel (already split from ratification in\n[what-human-judgment-is-for.md](what-human-judgment-is-for.md)) keeps full force while\ndiscernment-confirmation does not. (Current state, from the ratification audit: five\npropose-mechanisms, zero ratify actions; `confirmed=true` set by authorship — so the states in § 2\nare the missing half being designed, not a tweak to an existing one.)\n\n## 4b. The founder's two deflating questions (2026-08-31), and the simpler end-state they force — **BLESSED same day** (*\"Anyhow yes I agree, blessing this whole proposal you gave now, go ahead.\"*)\n\nVerbatim: *\"To me the most fundamental question here is, should there even be a voting/ratification\nbeyond the stance (agree/disagree) vote on a claim. I don't think the seeded-probe thing sounds\nuseful and elegant and worthwhile, do you? ... OTOH: if we don't have discernment-ratification,\nmaybe it'll unbalance or fuck up the ontology somehow?\"* And on consent: *\"Either you authorize an\nextracted claim from a text/URL/PDF of your own I guess (presumably mostly in the verbatim-copy\ncases) or you want to correct/tweak it, in which case you'll have to propose an edit/split of a\nclaim/concept anyway. Right?\"*\n\nBoth instincts survive deep inspection, and together they dissolve what remained of the act:\n\n**There is no ratify act.** The § 3b demotion, carried to its end: assent was already telemetry —\nand telemetry needs no button (the record-rule's own logic: *examined* is inferred from observable\nbehavior — opened the premises, dwelled, moved on without objecting). Dissent was already a\ncontribution (a reasoned reclassification through the correction pipeline, which exists). What was\nleft for a \"ratification act\" to do? Nothing. The act inventory collapses to two categories:\n\n> **The graph has contributions and telemetry, nothing else. Some contributions have a sole\n> possible author — your stance vote, your endorsement or disavowal of a rendering of your\n> position — which is all \"consent\" ever meant. Verdict-acts do not exist.**\n\n**Consent, deflated correctly (the founder's \"Right?\" is right).** Authorizing an extraction of\nyour own text is telemetry-plus-submission (you sent it and moved on) with one genuinely settling\nsliver — the authored-vs-source provenance fact, where you are the only witness. Objecting to a\nrendering is an edit/split proposal — a contribution. And disavowing someone ELSE's rendering of\nyour position mints a new FACT-contribution only you can author: *\"X does not endorse Y as their\nposition\"* — displayed beside the rendering with full standing, deleting nothing\n(add-never-overwrite). \"Settling\" was never a special verdict power; it was **sole-mintership**:\nnobody can outvote your disavowal because nobody else can mint it. The § 4 three-channel table\nreduces to this.\n\n**The seeded probe drops from live design** (founder lean, Claude concurs): planting wrong content\nin the live graph pollutes the record, carries trust cost, and sits badly with the confession\nethos — and its measurement is available deception-free: (a) **natural mistakes are free probes** —\nthe historical record already contains machine discernments later corrected; time-to-first-objection\non those measures whether anyone is looking, continuously, with no deception; (b) **offline\nevals** (the founder's lazy merge-review with friends) measure matcher accuracy and judge\ndivergence in a sandbox — and THAT is where motivated-shaped test pairs belong, as lab material,\nnever as live plants.\n\n**Does dropping discernment-ratification unbalance the ontology? No — it forces honest\nstratification instead.** Machine discernments carry machine provenance forever, displayed as such,\nchallengeable forever; \"untouched\" becomes the permanent honest default rather than a debt awaiting\na gate that scale could never staff (nobody was ever going to review 3,666 scheme labels).\nSame-hand bias is carried by the cross-family jury (machines checking machines), not by human\nreview; the strength layer's discipline is that unobjected machine structure must never be\nTREATED as human-grade — a display and provenance rule, already the standing principle. The one\nreal cost, owned: the ratification-tether calibration (measuring whether humans retain validation\ncapacity) loses its per-ratifier instrument — replaced, more weakly, by natural-mistake\ntime-to-objection and by what co-present live sessions observe directly.\n\n**Held for later thought, founder direction (2026-08-31): endorsement claims are factual claims —\nand so are identity claims.** Verbatim: *\"This whole thing deserves some thought later. Claims\nabout who endorses what should probably be factual claims just like anything else. That way ppl\nwon't fight over which username has a tick/is the real Person Z... I don't wanna have to deal with\nthat mess if I can help it...\"* The elegant closure this points at: the no-verdict principle then\ncovers identity too — the platform never issues verification ticks, because *\"account @z is Person\nZ\"* is itself a claim with evidence, supportable and attackable like any other, and a disavowal\ndisplays with its provenance chain (account → identity-claims about the account) weighed\naccordingly. Sole-mintership becomes: only the account claiming to be Z can mint Z's disavowal, and\nwhether that account IS Z is contested-and-mappable rather than platform-adjudicated. Deliberately\nNOT designed further — the founder wants to think; this note is the held direction.\n\n**Consequence for the `confirmed` → `ratified` rename (ledger C8)**: under this end-state there is\nno \"ratified\" state to rename INTO — the field decomposes into what actually happened:\nauthorship provenance plus examination telemetry. C8 wants re-ruling in that light.\n\n## 5. The seeded-probe, named properly\n\nThe side-chat used the coined compound *seed-flawed-proposal probe*; the concept is canonical (the\nratification-tether entry's candidate calibration instrument, from the cognitive-commons read):\n**occasionally slip a known-flawed machine proposal into the ratification stream and measure the\ncatch rate** — the audit profession's test transaction, the smoke detector's test button, the only\nway to know whether the human gate filters anything. Status: candidate, unruled, with a real trust\ncost — deliberately showing users wrong content wants a disclosure policy, probably operator-only\nduring dogfooding. **Upgrade required by § 3b (founder-agreed): probes must include\nmotivated-shaped seeds** — wrong proposals that would FAVOR the reviewer's apparent side — since\nonly those detect selective honesty; neutral seeds measure sloppiness, not motivation.\n\n---\n\nCross-references: [ux-principles.md](../ux-principles.md) (anti-pattern 9 and the naming note this\nsupersedes in direction) · [reflection-and-the-graph.md](reflection-and-the-graph.md) § 4\n(confession-must-pay, now merged here) · [what-human-judgment-is-for.md](what-human-judgment-is-for.md)\n(the human-reserved classes; ratification-as-failable) ·\n[cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md) (the ratification tether;\ndetect-but-defer) · [the-harness-conviction.md](the-harness-conviction.md) (§ 7, the failable\ncheck) · [active-inference-context-acquisition-and-deliberus.md](active-inference-context-acquisition-and-deliberus.md)\n(`worth_asking`) · [feed-algorithm-design.md](feed-algorithm-design.md) (the dashboard principle)\n"}