Every number, defined before you read it.

2·1precision@N: the headline

At each sample time we reconstruct the model's inputs exactly as they stood one hour earlier: open interest, long/short ratio, entry-price distribution, strictly as-of that moment, with no look-ahead. The model ranks all price buckets by estimated liquidation density and the top ~10% are marked "bright" (a small number of symbols use a wider fraction, disclosed per symbol). precision@N is the dollar-weighted fraction of the liquidation events that actually printed in the following window which landed inside a bright bucket. Every published precision@N figure carries its sample size (n) and its Wilson 95% lower bound.

Picking 10% of buckets at random would score roughly 0.10. That baseline is shown for context but never headlined; see changelog entry #1 for why the multiple-of-baseline framing was retired.

The ruler mix, disclosed. Most symbols are graded on the top-10% ruler described above; ETH, SOL and XRP are graded on a wider top-20% ruler (a volatility-aware override of 24 July 2026, disclosed per symbol on /proof). The pooled out-of-sample precision@N therefore mixes two ruler difficulties by design: a random pick scores ≈0.10 on the narrow ruler but ≈0.20 on the wide one, so the honest baseline for the mixed pool is the dollar-weighted mix (≈0.146 at the current composition), not the flat 0.10. Every pooled baseline comparison on this site uses the mixed figure; the flat 0.10 is kept for reference only. Alongside the pooled number we publish a same-ruler shadow: the identical served configurations re-graded solely on the uniform top-10% ruler over the same held-out days (n = 13 independent day-clusters, with its own Wilson 95% lower bound), labeled descriptive and never headlined. A walk-forward placement decomposition (BTC + ETH, n = 13 held-out days) measured the ruler mix contributing ≈+19.75pp to the pooled point. The mix is stated here so the number survives scrutiny; the shadow shows the same evidence on one ruler.

2·2The near-spot tautology: why capture metrics exclude the zone at spot

Roughly 88% of liquidation dollars land within about 1.5% of the current price. That means any model, including a trivial guess that paints the bands just off spot, scores well on a metric that counts near-spot events. A magnet sitting next to price gets "reached" almost by definition. Reading a near-spot-inclusive score as model skill would flatter us, so we publish companion metrics that exclude that zone entirely and grade only the harder, off-spot calls a trader actually cares about.

2·3Off-spot magnet capture (>1.5% from spot)

Of the liquidation dollars landing more than 1.5% from spot, the fraction caught by a top-ranked predicted band, scored walk-forward (out-of-sample), counted only on the (symbol, horizon) cells that cleared their own validation gate, averaged with equal weight across cells.

2·4Deep-magnet capture (>3% from spot): the strictest cut

The number we consider most defensible under diligence. Of the liquidation events landing more than 3% from spot (deep enough that a trivial near-boundary guess collapses), the fraction caught by a top-ranked predicted band. Computed walk-forward, pooled across every (symbol, horizon) cell with an adequate sample (at least 100 deep events), event-weighted, and deliberately not filtered to the model's winning cells: filtering to winners would cherry-pick the model up and collapse the comparison baseline, overstating the edge. Losing cells drag the number down, truthfully.

It is always published side-by-side with the trivial boundary baseline computed over the same unselected events, plus a dollar-at-risk-weighted variant, so you can see exactly what a no-skill guess achieves on the identical event set.

2·5Wilson 95% lower bound: why we headline it

Wherever a per-symbol rate is published, the number we treat as load-bearing is not the raw rate but the lower edge of its 95% Wilson score interval, computed on the de-overlapped effective sample count (2·6 below). A rate on a small or dependent sample overstates certainty; the Wilson lower bound is the version of the number that remains defensible if the sample was lucky. A symbol grades "validated" only when its Wilson lower bound clears the random baseline on at least three independent windows. A strong-looking point estimate with a weak lower bound is labeled low-confidence instead, and is barred from carrying strong claims anywhere downstream.

2·6De-overlapped effective n: overlapping windows are not independent samples

Calibration runs are logged far more often than their evaluation windows elapse: a run lands every hour or few hours, but each one scores a multi-day trailing window, so consecutive measurements share most of their underlying events. Counting every logged run as an independent trial would inflate n and dishonestly shrink the confidence interval. For the published interval we count a measurement only when its own evaluation window has fully elapsed since the previously counted one, and the effective sample count scales the raw observation total by that non-overlapping fraction. Effective n is always ≤ raw n, and both are published per symbol so nothing is hidden in the aggregation.

2·7Walk-forward / out-of-sample protocol

Parameter selection and scoring never touch the same data. Any tunable parameter is chosen on the earlier slice of a chronological split (roughly the first 60%) and scored on the later, held-out slice: a genuine forward test, never a shuffle. Two additional guards: a split where one slice contains essentially no scored liquidation mass is refused the out-of-sample label rather than reported as one; and changes to the served model are gated on the held-out score, never the in-sample one. Every scored window reconstructs inputs strictly as-of their timestamps; no future information reaches any input.

2·8Per-magnet labels: n / reach% / horizon

Every magnet rendered in the product carries three labels: n, the sample count behind its statistics; reach%, the empirical share of past cases in which price reached a magnet at that distance within the stated window; and horizon, the window that reach% refers to. Reach rates are descriptive history, n-weighted across tracked symbols, and refreshed continuously. For scale, at a ~1% distance from spot, as of July 2026:

The position-alert surface grades its "trim" (partial de-risk) alerts under a separate, pre-registered rule, distinct from the headline alert accuracy: a trim is vindicated if price moves adversely for the position beyond ±0.3% within 4 hours of the alert (same horizon and same price source as the headline grading); a favorable move beyond that line counts against it as given-up gains; smaller moves are neutral and skipped, and every skip is accounted under its own reason, never silent. Trim results publish with their own sample size (n) and Wilson 95% lower bound once the sample matures.

Separate from headline accuracy by construction: trim outcomes are recorded under their own distinct event type, so the headline alert-accuracy roll-up structurally cannot ingest them; the separation is enforced by the data model, not by policy. And because the rule was registered before results accumulated, it cannot be re-fit after the fact to flatter the outcome.