{"path":"research/testable-bets/README.md","content":"# Claims as testable bets: what prediction markets and the hedging literature say (2026-10-20)\n\n**His aim, verbatim** (2026-10-20, ruling the hedge widening): *\"On stance hedges: I want claims as close\nto testable as possible, testable bets that can be challenged and tried whenever possible/applicable.\nKinda similar to prediction markets like Polymarket I guess. How do they approach this and related\nquestions? Research this extensively online.\"*\n\nTwo research reports, written by helpers that searched the web, each in full with every source linked:\n\n- [prediction-markets-resolution-practice.md](prediction-markets-resolution-practice.md): how\n  Polymarket, Kalshi, Metaculus, Manifold, Good Judgment and the Forecasting Research Institute turn a\n  belief into a resolvable question, the documented disputes of 2025-2026, and AI-written questions.\n- [hedges-vagueness-and-testability-literature.md](hedges-vagueness-and-testability-literature.md):\n  words of estimative probability, Tetlock, the linguistics of hedging, falsifiability and bets, what is\n  lost when hedges go, and systems that keep a proposition apart from a speaker's confidence.\n\nAn earlier note covers market mechanisms, futarchy and Metaculus's conditional pairs:\n[../prediction-markets-argumentation.md](../prediction-markets-argumentation.md).\n\n## What both reports find\n\n1. **The ruling matches practice.** No platform puts *probably* in a question's text; the forecaster's\n   credence goes in a probability field. Toulmin's qualifier, the US intelligence standard (ICD 203), the\n   IPCC, FactBank and nanopublications all keep the proposition apart from the speaker's confidence.\n2. **A hedge word is a poor carrier of confidence.** Readers turned *may* in policy sentences into\n   anything from 0.08 to 0.59; with the US standard's table and no guidance, readers matched its ranges\n   32% of the time. So the stance field keeps the cue word, and any number beside it is marked inferred.\n3. **The danger runs one way.** Stripping a modal that is content writes a stronger claim than a named\n   author made, and LLM rewrites of science over-generalise often (26-73% in one 2025 study). Keeping a\n   stance hedge by mistake only leaves a vaguer claim. Human readers agree well that *might* is\n   epistemic and badly on *could* (kappa 0.20-0.26).\n4. **A testable bet is a separate object from the claim.** Every platform's question is a bundle: a\n   proposition in observable terms, which kind of event counts (occurred, announced, signed), a named\n   source with a rule for when it fails, a window, and a fallback. Nearly every famous 2025-2026 dispute\n   is one of those parts missing. The Forecasting Research Institute writes several operationalisations\n   of one expert's idea and sends them back to the expert.\n5. **Not every claim can be a bet.** Values cannot; causal and counterfactual claims need indicators,\n   conditional pairs or panels; claims about the present are verification, not forecasts; a *could*\n   claim is existential (one case verifies it, no finite search refutes it). Platforms have refusal\n   outcomes (Metaculus: ambiguous when reality is unclear, annulled when the question was flawed), and\n   these must count as no evidence.\n\n## Where the build of 2026-10-20 differs from the reports\n\n- **An ambiguous hedge is taken out** (ledger AM6, ruled *for now*, with his revisit open). Both reports,\n  for the asymmetry in point 3, would keep the cue in the wording and flag it. Proposed as AO1.\n- **A stated probability stays in the wording.** The hedging report notes that *probably E* moves out\n  while *70% chance of E* stays in, though the intelligence lexicons treat them as one thing at two\n  precisions; it proposes moving a credence about a one-off event to the stance field while a rate or\n  chance process stays as content. Proposed as AO2.\n- **Whose hedge it is.** The trial on 2026-10-20 dropped *Some have said* together with *may*\n  (`a-sentence-as-a-pointer-into-a-landscape.md` § 6e); the reports' remedy is an assertion record of\n  holder, stance and span, rendered as *holder, with stance, asserts P*. Proposed as AO4.\n\n## Proposals, filed as ledger Arc AO (unruled)\n\n- **AO1** An ambiguous hedge stays in the claim's wording, flagged, rather than going to the stance field. A fourth option, put in the walkthrough of 2026-10-20 (part four): mint both readings, *will* and *can*, each\n  marked as a reading of the author, and offer them to the author where reachable.\n- **AO2** A stated probability that is the author's credence about one event moves to the stance field;\n  one that describes a rate or a chance process stays.\n- **AO3** A testable bet is a proposal stored beside the claim, never the claim: named slots (the\n  proposition in observable terms, what counts as the event, source and its failure rule, window,\n  fallback), a clairvoyance-test gate, and *ambiguous* and *annulled* results that move no strength.\n- **AO4** Every place that attributes a claim shows the holder's stance with it (*holder, with stance,\n  asserts P*), and a quoted speaker's hedge belongs to the quoted speaker.\n- **AO5** A possibility or feasibility claim gets its own testable forms (a demonstration by a date; a\n  conditional), and a failure to find a case does not move its strength as a refutation would.\n\nQuestions the hedging report leaves for him: whether the stance field holds a band, the author's number\nwhen given, or both; and whether authors become scoreable on their stance bands when claims resolve.\n"}