{"path":"research/instant-entry-and-the-microtasks-it-needs.md","content":"# Instant entry, and the microtasks it needs\n\n**Status, 2026-11-21:** a session's assessment. Nothing here is ruled, nothing was switched on, and no model was\ncalled. His question that night: *\"Do we need the analogy-finding or can we get other microtasks reliable enough to\nturn them on (or well, the reliable ones!) and let a user have the kind of instant interactivity that Gemini talked\nup? I'd love that. Take your best stab at this as well.\"* The interactivity he means is UX principle P7b: a visitor\ntypes a short argument beside a map and sees it become cards at once (`docs/ux-principles.md`, **not built**: no path\nfor typing an argument into the map exists today).\n\n## The answer in short\n\n**Analogy finding is not needed.** Neither are the keeper (the microtask that reads two claims for sameness), the\nrelevance judge, priors or the independence reading, before a card shows. P7b already orders it so, in his words:\n*\"it should at least be split to the separability floor still, but I guess maybe sameness checking could be\ndeferred?\"*, and to *\"Do NOT fuse extraction … DO defer sameness checking\"*: *\"agreed\"*. A card waits for what P7b's\nfirst rule names: the argument split to the floor, each part passing the stranger test. In today's code that is four\nmicrotasks: extraction of the claims, reasons off, the splitter with its check, and a title.\n\n**Reliable enough to switch on: not yet, by his own condition.** A card lands *\"only if what we show in that card is\nsomething we're confident of is accurate and will not be transformed by later microtasks/whatever\"* (P7b, rule 2).\nThe record says *\"No pass is proven\"* (`proving-a-pass-reliable.md` § Where things stand), and his hold stands\n(24 October: *\"Microtasks should still be turned off in production\"*; theory first, then the instructions, then\nevals). That a landed card is a draft only its author sees limits who a wrong cut reaches. It does not meet the\ncondition.\n\n**The way there is short.** These four, on short typed arguments, are the smallest set anyone could read: about ten\ncases, read by him one at a time, which is the reading the Principles put before any proof (ledger Y9, ruled).\n\n## One example, whole\n\nA visitor types: *\"We should set the thermostat to 19, because heating costs a lot and Alex can just put on a\nsweater.\"* The third column is written by hand as an illustration; it is not a run.\n\n| Step | Microtask (its registered name) | What it would write for this sentence | Before a card shows? |\n|---|---|---|---|\n| 1 | Extraction of the claims only (`extraction`, with `DELIBERUS_EXTRACT_LINKS=0`) | Three claims, each naming what it is about: *The thermostat in the visitor's flat should be set to 19 degrees* · *Heating the visitor's flat costs a lot* · *Alex can put on a sweater* | Yes |\n| 2 | Reasons off (`reasons_off.read`, version 3; `reasons_off.check_conditional` for an *if*) | The *because* taken off the first claim: each reason a claim of its own with a support link to the conclusion, numbered by which argument it belongs to | Yes |\n| 3 | The splitter and its check (`split.propose`, `coverage.check`) | Nothing to cut here; from *cheap and safe* it would write two parts | Yes |\n| 4 | Titles (`titles`) | A short headline per card | Yes: a card needs a face |\n| 5 | Sameness against the graph (`keeper.pair`) | Whether *Alex can put on a sweater* says what a claim in the graph already says | No: marked afterwards as a merge candidate, the author told, never asked (P7b, rule 3) |\n| 6 | Relevance, priors and membership, independence, analogy, the scope reading | Each link's relevance, each claim's prior, kinships | No: standing sweeps, later |\n\nOne microtask is undecided between the two groups: `reported_speech.split`. A visitor who types *\"my landlord says\n19 is enough\"* has reported a position, and a card that shows it as the visitor's own claim would not be accurate.\n\n## What a wrong answer does, microtask by microtask\n\n- **A wrong cut moves no claim's strength.** Since the ruling that part links carry no strength (ledger BD16, ruled\n  2026-10-25) a part passes nothing to its whole. What a wrong cut does is put words under the author's name that\n  the author did not say: a written part is attributed to the claim's author\n  (`microtasks-in-broad-daylight.md` § 3e).\n- **Reasons off is the entry microtask whose error moves a claim's strength.** It writes the support links, and\n  since the ruling that questions prompt and claims carry (ledger Q51, ruled 2026-11-21) strength comes from reasons\n  only. The strength it moves is that of the visitor's own draft claims.\n- **The keeper and the relevance judge write between different people's claims**, which is why their errors reach\n  other readers and why P7b puts them after the landing.\n\n## What is measured about each, today\n\n- **None has an error rate from a person's reading** (`proving-a-pass-reliable.md`).\n- **The splitter.** Version 2 (30 September) asks of every part *if this part were false, would the claim be\n  weaker?* and caught all six harmful parts of version 1; its part test answered 45 of 45 parts the same way twice,\n  on Claude Sonnet 5, with splitter and checker of one model family. Known and open: an *or* cut into two required\n  parts, and a small gloss the claim did not state in nine of sixty fresh parts, as a model read them\n  (`microtasks-in-broad-daylight.md` § 3e, § 3g). One sentence was cut four ways by five models (§ 3h there). By the\n  commit record its part test was changed on 11 November (*false*, or only *less supported*) and proposals from\n  older versions are refused since 16 November. His review of the fresh parts was set for later (6 October: *\"We can\n  always review the cuts later after the reviews that's fine\"*); whether he has read parts since was not checked for\n  this note.\n- **Reasons off.** Ruled as ledger T9 (a claim that gives its own reasons is cut into a conclusion and its reasons);\n  version 3 groups reasons into arguments. No reading of its output by a person was found in the notes read here.\n- **Extraction.** Every source in the graph went through it, and code checks its quotes\n  (`scripts/measure_quote_fidelity.py`). The reviews so far read claims cut from long sources; a short typed\n  argument is a kind of input no review has covered.\n- **Titles.** Low stakes: a wrong title misleads no computation.\n- **The keeper** (not on the entry path): nine pairs set aside for testing wait for his reading, one at a time.\n- **Seconds per step: unmeasured where it matters.** The engine that runs on his plan answers one cut in minutes\n  (two to three for a call, `microtasks-in-broad-daylight.md` § 3g), so it cannot serve a visitor. The engine that\n  asks Anthropic directly on the API credit (`deliberus/extraction/anthropic_direct.py`) is the candidate, and its\n  time on a one-sentence input has not been measured. The figure in the November conversation with Gemini (cards in\n  about two seconds) was an estimate, removed from the docs as unmeasured on 16 November.\n\n## Where this already sits in the plans\n\n- **P7b**, three rules, reworded on his word on 15 November.\n- **`TODO.md`, order of work**: item 2 (*Split/Keeper decoupling*: interactive entry splits to the floor, sameness\n  deferred to the keeper's sweeps) and item 3 (the landing in the map).\n- **The order that cuts first exists in code behind a switch** (`deliberus/floor.py`, `DELIBERUS_FLOOR_FIRST=1`, off;\n  ledger AL9, ruled: the cut runs before linking, on every entry path).\n- **A claim's workup** (`deliberus/workup.py`, built 6 November, read by the keeper only) gathers the same\n  stages for each claim (`a-tree-per-claim.md` § 4).\n- **His order for opening microtasks** (`TODO.md`, the section of Saturday 2026-10-24 on revising the microtasks):\n  the theory that grounds the\n  prompts, then the instructions, then evals by the review protocol, the API credit going to those evals (*\"that\n  whole review by one example at a time protocol\"*). His proposed first test is the bench that hides a map's\n  sources (`asteroid-and-feeling-around-sources.md` § 6). And on call counts: *\"The whole thing about making repeated model\n  calls feels like a copout to me\"*.\n\n## Proposed, inside that order\n\n1. **Ten short typed arguments as the entry path's first eval set**, on his yes to the spend: under one dollar of\n   the month's API credit, an estimate until run (`whose-tokens-pay-for-mapping.md` § 3g, § 3h). One run answers two\n   things: how many seconds steps 1 to 4 take on the direct engine, and what the four write on input of this kind.\n   It writes nothing to the public graph.\n2. **He reads the ten, one case at a time.** Each reason he gives becomes a candidate rule for the instructions or\n   for code, tried on the rest.\n3. **A switch per microtask.** The live site has two blanket switches (`DELIBERUS_EXTRACTION_PAUSED`,\n   `DELIBERUS_MICROTASKS_OFF`); *the reliable ones* needs a list of which are on.\n4. **The landing itself**, as P7b words it: cards at stage `draft`, visible to their author, sameness afterwards.\n   The (+) on the preview front page is the door (`specs/landing-redesign/requirements.md` § 2026-11-21); today it\n   says extraction is paused.\n5. **Who pays, with a ceiling**: only visitors who have signed in, and a daily cap on the spend, as the router's cap works\n   today.\n\n**Considered and not proposed:** switching the whole pipeline back on (it writes links between sources with\nmicrotasks nobody has read); one fused call for short inputs (his answer of 12 November, above); switching the four\non unread because a draft is private (his condition is about the card's accuracy, not its audience); waiting for a\nproven rate before anything is on (the Principles put a person's reading before the proof, and a rate can be proven\nlater on real use).\n\n## What does not fit\n\n- A visitor who is not signed in gets nothing from this; whether the first taste needs an account is his call.\n- An author reading their own cut is not an outside check: they see what they meant.\n- A product that showed people the whole tree of sub-questions failed twice as overwhelming before a narrower form\n  worked (Elicit, in `listening-for-the-calculus-and-sturdy-microtasks.md` § 3 and § 4). The landing shows a\n  visitor a few cards, never the tree.\n- This is a cost of now: once the sweeps run as standing work the question does not arise.\n\n## A look back, 21 November: what the record says about the microtasks so far\n\nHis ask, with about twenty dollars of API credit granted for iterating on reliability: *\"Can we look back at how\nreliability of microtasks was when we filled up the AI safety institute graph/maps lately, to start with?\"* Read\nfrom the record of model calls (16,863 calls, 6 October to 9 November, read at 03:20 on 21 November;\n`scripts/judgment_facts.py`) and Codex's two notes of 13 November.\n\n**The institute maps were not filled by microtasks.** A session made them by hand: 250 claims and 334 links in each\nlanguage, built by a script and stored with `scripts/store_hand_extraction.py`. The record holds no call about any\nof their claims by id. One microtask read them, the relevance judge, on 24 October: 164 calls, of which 153\nanswered (134 by Gemini 3.1 Flash Lite, 19 by Gemini 3.5 Flash, the model the free tier falls back to) and 11\nfailed; 103 of those prompts name an institute. No person has compared its readings with their own.\n\n**What has been checked on that map's material** (Codex, 13 November,\n[his note on weighing two Swedish objections](how-independent-reasons-add/swedish-objections/README.md)): one figure in the hand-made map\nwas wrong against its source (four parties at or below 2.33, where the map said 2.30), an error of the session that\nmade the map; and a model asked what two of the map's objections establish passed 14 of 14 invented controls and\nthen 4 of 8 checks on the real sources. Fredrik had not assessed those judgments when that note was written.\n\n**Where microtasks did fill the graph: the night of 5 to 6 November** (27 more sources, $113.56 by the router's\ncount; `TODO.md` § THE NIGHT OF 5 → 6 NOVEMBER). From the record, over all its dates:\n\n| Microtask | Calls | Failed outright | The same question asked again: an answer of another kind |\n|---|---|---|---|\n| The splitter (`split.propose`) | 2,568 | 2 | 5 of 45 |\n| Its check (`coverage.check`) | 2,492 | 2 | 0 of 7 |\n| The scope reading (`scope.read`) | 2,447 | 75 | 14 of 98 |\n| The keeper (`keeper.pair`) | 1,852 | 53 | 27 of 91 |\n| Reasons off (`reasons_off.read`) | 446 | 0 | 0 of 9 |\n| Reported speech (`reported_speech.split`) | 362 | 0 | 0 of 9 |\n| The independence reading (`independence.pair`) | 135 | 1 | 0 of 25 |\n\nExtraction's focused step needed a second attempt on 283 of its 1,221 calls. ⚠ The last column counts a change in\nthe coarse outcome the record stores (*found*, *nothing*, *undecided*), and whether the instructions were the same\nboth times was not checked, so it mixes a microtask answering differently with its instructions having changed;\nthe keeper's instructions carry each pair's workup and changed often.\n\n**These are counts of failing and of answering alike. None is a rate of being right.** Nothing that night wrote has\nbeen read by him: the sittings that TODO section lists for his review are open. So the first thing an iteration\nneeds is cases with a known right answer, read by him one at a time, and Codex's result (14 of 14 on invented\ncontrols, 4 of 8 on real sources) says where to take them from: real inputs.\n\n## How this was done\n\nRead whole for this note: `microtasks-in-broad-daylight.md`, `listening-for-the-calculus-and-sturdy-microtasks.md`,\n`a-tree-per-claim.md`, `asteroid-and-feeling-around-sources.md`. Read in part: P7b in `docs/ux-principles.md`, the\norder of work and the plan for opening microtasks in `TODO.md`, the registered list in\n`deliberus/standing_work.py`, the head of `deliberus/floor.py`; and, before the first draft, the commit record of\n21 October to 20 November with `proving-a-pass-reliable.md`, `machine-work-as-a-standing-property.md`,\n`the-evals-loop-and-our-machine-passes.md` and `whose-tokens-pay-for-mapping.md` § 3g to 3h (how much of each was\nnot recorded). **Not read:** `writing-out-the-unwritten.md` and `one-challenge-one-home.md`, which the corpus graph\nalso names on this subject (the second searched only for the order that cuts first), and the review notes themselves\n(`scripts/read_review_notes.py`). No helpers, no outside check.\n\n**See also:** [microtasks in broad daylight](microtasks-in-broad-daylight.md) ·\n[a tree per claim](a-tree-per-claim.md) · [machine work as a standing property](machine-work-as-a-standing-property.md) ·\n[proving a microtask reliable](proving-a-pass-reliable.md) · [the evals loop](the-evals-loop-and-our-machine-passes.md) ·\n[UX principles](../ux-principles.md) (P7b)\n"}