The model is frozen. Now we grade it.

The current board's realized performance holds on picks made after the freeze.

That's the whole claim. Numbers earned on past picks are easy to flatter after the fact — the only honest way to test them is on picks that don't exist yet, measured by rules fixed in advance. This page fixes those rules.

The model was frozen at 2026-07-14 03:00 UTC, after one final change: a gate demoting MLB player-prop overs (PR #260). A test of a moving target proves nothing, so until this test resolves, nothing changes.

The freeze is an artifact, not a promise. The frozen configuration is pinned to git commit adc979a5 (2026-07-14) — a fixed point the running code can be checked against.

A record is only as honest as its denominators, so every cohort is defined here, before a single forward pick exists:

An ISSUED pick is a row whose Whet score is ≥ 50 at first scan — the customer-visible board.

A DEMOTED pick is tracked but below the floor due to a pre-registered gate.

A VOID pick is one that could not be graded, published with a reason code (player-dnp, stats-unavailable, event-cancelled, grading-unsupported).

PRIMARY ENDPOINT: flat-stake win rate and ROI of ISSUED picks made after the freeze, with event-clustered 95% confidence intervals.

Two primary metrics: forward win rate and flat-stake ROI on picks made after the freeze, reported with event-clustered 95% confidence intervals. Picks from the same game move together, so clustering by event is what keeps the intervals honest instead of flattering.

And no claim before the sample floors are met: at least 14 independent game dates and 200 graded picks per segment. Until both floors clear, the only honest label for any forward number is “too early” — and that's the label it gets.

The full segment grid is pre-declared: sport × market type × side. Every segment that reaches its floor is reported with win rate AND ROI AND confidence interval together — never one metric alone, never a subset of segments. Cherry-picking which slice to show, or which number to lead with, is the oldest trick in this industry; pre-declaring the grid takes the option away.

The secondary metric is closing-line value measured against Pinnacle's two-sided closing quote, accruing since 2026-07-14.

We retired our old consensus-anchored CLV because it measured our own book mix churning, not the market. A CLV number is only worth reporting against an anchor we don't control — so the forward test grades against Pinnacle's close and nothing else.

One honest limit: Pinnacle quotes only a subset of our markets (~32% of recent props, and unevenly across segments), so sharp-anchored CLV is reported with its coverage rate and is never the primary endpoint.

The freeze's one final change was a gate demoting MLB player-prop overs. A pre-registration owes that gate a written exit condition, decided now rather than argued later. The bar is comparative — the gate's job is to separate worse picks from better ones, not to predict which side of break-even a cohort lands on:

Demoted prop overs continue to be scanned, tracked, graded, and published identically to issued picks, flagged as demoted. The revert evaluation happens exactly once, at the verdict date. The gate reverts only if the demoted-overs cohort's forward flat-stake ROI is within 2 points of (or above) the issued board's forward ROI over the same window, with ≥ 200 graded demoted overs — that is, only if the gate demonstrably failed to separate worse picks from better ones.

No judgment call, no quiet extension, no early or repeated evaluation. One check, one rule.

Prices move fast. A surfaced pick is typically bettable for about one refresh window (~5 minutes); after that the market has often reconverged. Our record grades picks at their listed entry price — always verify the live price at your book, and if it is worse than listed, pass.

Every pick is locked and timestamped before first pitch. The record is append-only — picks can be added, never edited or removed. 100% of issued picks are graded; there is no “didn't count” pile. And when the test resolves, we publish the outcome either way — pass or fail.

The running numbers publish automatically every day from day one — we have no discretion over timing. The pass/fail verdict is computed exactly once, on the earlier of 2026-10-01 or the date all pre-registered sample floors (≥ 14 independent game dates AND ≥ 200 graded picks per segment) are met, and is published either way.