Methodology change log

Instrument fitness diagnostics block shipped (experimental, opt-in)

An opt-in component table on the v0.2 card and packet APIs (include=fitness) and a collapsed packet-page section: depth durability, spread and price impact, time to resolution, resolution reliability, concentration exposure, and a structural basis-risk note. Each component states the observable on the stored card it derives from and reports an honest insufficient_data state when the card lacks the input; time to resolution is insufficient_data on every current card because the v0.2 schema stores no end date. Deliberately no combined number; the not-advice sentence, experimental label, and methodology version are fixed payload fields. Default API responses are unchanged.

Market Trust

Deviation note 1 for #295174: UMA terminal-dispute classification

The resolution-quality classifier was corrected so a UMA dispute counts as a bad outcome only when it is the terminal state; a disputed-then-re-resolved-clean market is clean. Recorded as a dated deviation note rather than a re-filing: zero resolution rows existed when the correction shipped, so the change preceded any forward evidence. The freeze hash is unchanged because it pins the frozen formula and test design, not the classifier implementation.

Deviation note 1 for #294147: SELL-trade win/loss correctness fix

A correctness bug in the shared position aggregation resolved SELL fills against the raw outcome index while keying positions on the normalized index, inverting win/loss for SELL-derived positions. Fixed in the data-generation path with a forward-archive reset; recorded as a dated deviation note because the forward collector imports that aggregation. The fix landed before forward evidence accumulated under the buggy rule.

Market Trust v1 calibration freeze filed (AsPredicted #295174)

Pillar weights, verdict cutpoints (80/65/45), evidence-confidence caps, the unknown-flow ladder, and the use gates were frozen as-is (operator-set, deliberately not tuned before filing). The single confirmatory primary outcome is resolution quality; the forward window starts 2026-06-07 and the verdict matures on resolved outcomes alone. A pre-registered null keeps Market Trust in canary_preview and is a publishable outcome. The freeze hash is pinned by an agreement test that fails CI if the frozen object drifts.

Pre-registration: AsPredicted.org #295174 (private filing; receipt held internally)

Market Trust v0.2 pillar merge: liquidity and depth become executability

The liquidity and depth pillars merged into a single executability pillar (weight 20); the freed weight moved to orderbook-independent pillars (coherence 20 to 25, resolution_reliability to 20) so the weights sum to exactly 100. The weights remain operator-set candidates, not calibrated, pending the outcome-ledger calibration track.

Wallet-skill FDR forward-persistence test filed (AsPredicted #294147)

Strictly prospective 90-day test (window 2026-06-02 through 2026-08-30) of whether FDR-cleared wallet-skill candidates persist out of sample. The frozen candidate registry and methodology hash are pinned in code; the daily verdict reports an honest insufficient_sample until at least 40 candidate wallets clear the 10-position floor and at least 1,000 qualifying forward positions resolve.

Pre-registration: AsPredicted.org #294147 (private filing; receipt held internally)

CME realized-vs-control forward-only validation filed (AsPredicted #294035)

Strictly prospective matched-control test of realized CME signal performance against controls, maturing on data and time alone with no human in the loop. The methodology hash is pinned; the daily verdict artifact reports an honest insufficient_sample until at least 30 K=20 pairs resolve. A null is a publishable pre-registered outcome.

Pre-registration: AsPredicted.org #294035 (private filing; receipt held internally)

Market Trust v0.2: evidence-status-gated cards under canary_preview

Versioned card substrate (migration 083, dated 2026-05-07) with SHA-256 row-hash receipts and previous_row_hash chaining; the rating predicate gates a use verdict on measured-pillar count and a pending-pillar ceiling. Every card carries contract_status canary_preview, fixed not_claiming fields, and the not-calibrated caveats for operator-set cutpoints.

Market Trust | Market Trust card research page

Market Trust v0.1 scorecards shipped (experimental)

Daily artifact-first scorecards with four ratings (use, use_with_caveats, discount, do_not_cite) and operator-set cutpoints at 80/65/45. Explicitly experimental; superseded by v0.2 once the evidence-status-gated rating predicate landed.

Market Trust card research page

Coherent Markets Engine V0.1 frozen

Constraint-projection detector for structural inefficiencies across linked Polymarket events, operating on market prices only (no skill weighting). Frozen configuration committed in the repo. The same day, the V0.2 walk-forward backtest was pre-registered on AsPredicted before running on historical data. Track record accrues on resolutions.

Pre-registration: AsPredicted.org #288046 (V0.2 backtest)

CME V1 research page

V2.8.2 skill-weighted aggregation sweep: no aggregator beat the baselines

A 24-aggregator sweep on the V1-M cohort tested whether skill-weighted aggregation improves per-market price priors. The primary hypothesis did not reject the null and none of the 24 aggregators beat the market-implied baseline; the primary aggregator underperformed both the equal-weighted and market-implied baselines. Published 2026-04-27 as a negative result that closed skill-weighted aggregation as a price prior.

Model-selection writeup

V1.5 pre-registered follow-up: both primary tests failed; published openly

Two ex-ante primary tests (E2 per-wallet temporal holdout, E7 per-quarter IC stability) were filed on AsPredicted before analysis and both failed at their ex-ante thresholds. Published 2026-04-27 as a negative result: the composite is a cross-sectional ranker, not a per-wallet temporal predictor.

Pre-registration: AsPredicted.org #287368

V1.5 paper | Negative results

Edge Score V1-M cross-venue extension shipped

V1-M extends the V1 frozen-coefficient ranker across Polymarket, Kalshi, and Manifold using the same composite formula. Active as the production wallet ranking, published with a public data bundle.

V1-M methodology paper

Edge Score V1 frozen

Frozen-coefficient three-pillar composite anchored on 8,656 Polymarket wallets with at least 5 resolved positions. Coefficients were committed to git before the validation script ran; V1 is internally ex-ante, not externally pre-registered. Paper published 2026-04-18.

V1 methodology paper