Methodology

ClearMarket ingests the full Kalshi + Polymarket universe and filters to the institutionally relevant set.

Three filter inputs

1. Category whitelist — 9 institutional categories IN: economics, financials, crypto, companies, technology, health, politics, geopolitics, climate. OUT: sports, entertainment, mentions.
2. Volume threshold — venue-specific minimum (tuned empirically).
3. Resolution horizon — resolves within 24 months.

Category derivation differs by venue. Kalshi ships a clean category field, so we map it directly. Polymarket has no category field, so we derive it from tags[]: a deterministic map keyed on top-level tags (weighting Polymarket's forceShow flag), with a constrained LLM fallback for events whose tags carry no category signal.

Every market's resolution source is captured from platform-provided data, never invented by a model. The mechanism differs by venue.

Kalshi — structured (direct)

The resolution source and URL live in a structured API field (series.settlement_sources). We read it directly; no extraction. Coverage across 703 unique series:

• 98% authoritative — Federal Reserve, BLS, BEA, NWS, Library of Congress, ICE, CF Benchmarks.
• 2% loose — a placeholder ("For example, Google Finance"), all in the equity-index family.
• 0% missing.

Polymarket — hybrid extraction with a verbatim gate

The source and URL are embedded in the market description free-text. We extract them without ever letting a model mint a URL:

1. Regex extracts the candidate URLs present in the description (the deterministic floor).
2. By count: 0 URLs goes to subjective/none; exactly 1 is used directly (no LLM); 2+ has an LLM select the primary by number among the candidates.
3. Verbatim gate: the selected URL must appear verbatim in the description, or it is rejected and flagged for review.

Result: 88% pure-deterministic (49% subjective-no-URL + 39% single-URL), 12% gated-LLM, and zero hallucinated URLs (6 borderline cases fell back to first-candidate and were flagged). The LLM is constrained to selecting among URLs that provably exist in platform text; it cannot introduce or mistype one.

Provenance tiers

Every value carries its origin. direct: platform API, structured field, or verbatim-extracted URL. editorial: LLM-drafted interpretation, e.g. underlying_reference. subjective: resolution by consensus, no named data source. The LLM is confined to the editorial tier; any verifiable identifier (source name, URL) traces to platform data.

Source commitment

Capturing a source is not the same as the venue committing to one. We classify every market's resolution source on a commitment axis, because a hedge or a placeholder is not a definitive source:

• named — a concrete, resolvable authority (Federal Reserve, BLS, the NYC Rent Guidelines Board, a deep-link .gov / official-electoral citation). No grade cap.
• uncommitted — a source is gestured at but not committed to. Two sub-cases: illustrative ("For example, Google Finance" — a candidate, not a commitment) and placeholder ("a consensus of credible reporting" — names no concrete authority).
• none — no source language at all.

The principle: "for example" means the venue did not commit to a definitive source. Source commitment is a separate axis from outcome objectivity — an objective outcome resolved against an uncommitted source still carries resolution risk, because which report decides is left unspecified. This is the gap the CFTC comment documents: regulated venues commit to a named authority; unregulated markets frequently do not. The commitment classification feeds a grade ceiling (see below), and the venue's literal source field is shown verbatim on each event page — including when it does not match the resolution rules — so the reader judges the source in the venue's own words, not ours.

A per-market A/B/C grade of how unambiguously a market will resolve. Per market, not per event: one event can hold a Kalshi market that resolves clearly and a Polymarket market on the same question that resolves subjectively. The grade is a weighted score (0–100) banded into A/B/C, then capped — not a single rule. It rolls seven factors grounded in documented resolution failures (Hall's a16z analysis + the CFTC ANPRM comment): the failures were not absent sources but specific structural defects, so the grade scores those defects directly.

Seven factors (weights sum to 100)

Scoring rules

• Always-on factors (trigger, contested, source clarity, arbiter, temporal, mutability — 92 of 100) score every market. The one situational factor (source-conflict rule) scores only when 2+ distinct sources are named and is excluded from the denominator otherwise (score re-normalized over applicable factors).
• Score = 100 × earned / applicable, banded A ≥ 80, B 55–79, C ≤ 54.
• Hard caps (ceilings, applied after banding): source uncommitted-illustrative goes to B; uncommitted-placeholder or none goes to C (§2 — a hedge is not a committed source; the binding constraint on most unregulated markets); 2+ genuinely competing sources with no precedence rule goes to C; discretionary trigger with no named source goes to C; permissionless oracle + discretionary trigger goes to C; adversarial ground truth (contested reality) goes to B. Caps only ever lower a grade, never raise it.
• The applicable-factor count is surfaced with the grade ("B — scored on 6 of 7 factors") for comparability across markets of different complexity.

Five LLM-judged factors — trigger objectivity, contested reality, source-conflict rule, temporal precision, source mutability — are scored in one per-event call (enhance.llm_rcg_factors). Source clarity is hybrid: it reads the §2 verified-source layer (deterministic for Kalshi, gated-LLM for Polymarket). Only arbiter incentive is purely deterministic (it reads the venue's arbitration model). Adding factors adds tokens, not calls. The full weighted model, scoring rules, and caps are in rcg-scoring-model.csv (repo root) — the canonical spec for this section.

The source-conflict rule was deterministic (URL-count + regex) through 2026-05-26; it moved to the LLM on 2026-05-27 because URL counting mis-read one source across multiple pages as "multiple sources" (false C) and missed prose-named sources entirely (false na). It is now a reading-comprehension judgment.

Validation

The model reproduces the four documented failures as C / C / B / C (Venezuela, Zelensky, OPM, Ukraine). Full-universe distribution is pending the per-event LLM rater's calibration run against a sample of ordinary markets — four tail failures are not a calibration set.

The cross-venue resolution-governance contrast (objective, named-authority Kalshi markets vs. the subjective-consensus share on Polymarket) is the signal the venues themselves do not surface.

On top of the verified-source layer, ClearMarket adds editorial fields (editorial_notes, underlying_reference, tags, canonical question) drafted by an LLM and tagged editorial. These interpret and gloss the verified data — e.g. naming the authoritative publisher (S&P Dow Jones Indices) where a venue cites only a placeholder (Google Finance) — but never serve as the source of record. Enrichment is per-event (shared across a market's strike ladder), not per-strike.