Track record

What the rule patterns actually did — hits and misses

A retrospective study over 1,156 historically resolved Polymarket markets: for each rule pattern, how often markets carrying it actually resolved disputed, voided, or ambiguous, versus the 17.6% base rate. Every figure below is recomputed from the engine's own grading data (last recomputed July 16, 2026) — including the patterns that don't predict risk.

Retrospective backtest on resolved markets — not a live trading recommendation, not prospective performance

Rule pattern	Fired on	Observed problem rate	Wilson 95% interval	Lift vs base
Legal-challenge / recount / appeal language	22	77.3%	57–90%	4.4×
Multiple qualifying outcomes ("first of", "either")	48	60.4%	46–73%	3.4×
Conflicting resolution sources ("X or Y")	11	54.5%	28–79%	3.1×
Title / rules mismatch	56	26.8%	17–40%	1.5×
Explicit cancellation clause (assumed risky — measured protective)	388	3.9%	2–6%	0.2×
Generic "ambiguous language" (assumed risky — measured protective)	224	8.5%	5–13%	0.5×

Read the misses too. Two patterns most traders assume are dangerous measured neutral-to-protective here: explicit cancellation clause (0.2× on 388 markets) and generic "ambiguous language" (0.5× on 224 markets) — an explicit clause tends to remove ambiguity rather than add it. A tool that only reported the scary findings would be selling fear; we report what the data says in both directions. A tool that only reported the scary findings would be selling fear; we report what the data says in both directions.

Sample, exclusions and what "inconclusive" means

The dataset contains 1,235 closed markets pulled from the public Polymarket record; 79 were excluded as unobservable (their final resolution quality could not be established either way — the honest "inconclusive" bucket), leaving 1,156 graded markets. The base problem rate across them is 17.6%. Small per-pattern samples carry wide intervals — that is why every row shows its n and a Wilson 95% interval instead of a bare percentage. The engine grades 11 rule patterns in total; the 5 not shown above measured no reliable signal in either direction on current samples — we publish the count so the table above can't be mistaken for a curated highlight reel.

Does the composite model beat guessing?

Yes, out-of-sample. Graded on a held-out split, the fitted model reaches a Brier score of 0.119 vs 0.149 for a constant base-rate baseline, and beats the category, segment and rules-only baselines as well — verdict MODEL_BEATS_BASELINES. That's a modest, honest edge over guessing the base rate, not a crystal ball. Method: Methodology → the model.

Retrospective, not prospective

Everything on this page is measured on markets that already resolved — it shows the patterns were real in the past, not that any future market obeys them. We publish no forward performance claims, no avoided-loss figures, and no cherry-picked wins. The honest use of this page: it tells you which rule patterns have historically deserved your attention before you commit to a thesis.

How to verify this yourself

Run a live Resolution Watch on any market you hold: it quotes the exact rule fragment behind every flag, so a deterministic finding is checkable against the market's own text in seconds. The associations above are one historical dataset (dated, with intervals) — association, not proof of causation.