Do the verdicts work?

How it's measured. Each snapshot day, each card's verdict is a pick. Entry is that day's raw price; exit is the first price on or after 30, 60 or 90 days, within 7 days. The market is the median return of every single in the same era and price band (under $50, $50–250, $250 and up) over the same dates. Rates show only from 20 settled picks. Grade and slab picks are measured on the raw price move, not your grading result (see Grading results). Model 2026-10-09.1: 4 snapshot days; snapshots use the default settings, never your budget.

What the replay can't rebuild. As of each monthly anchor it rebuilds the Setups board from prices up to then and applies today's raw buy rule. It can't rebuild: PokemonPriceTracker confirmation (price_mismatch blocker, tracker-mismatch penalty, raw_moved penalty) (No tracker history exists, so every card is treated as confirmed: no mismatch blocker, no -7 / -5 penalties (optimistic)); Raw price freshness (stale_raw_price blocker, aging / stale penalties) and the stale-feed blocker (The entry must be a real trade on the anchor date, so the price is always 'fresh' (live, it can be days old)); Daily raw price and listing low (The anchor's sample (the month's last trade on that date): market price and its low for the spread (extreme spread > 2× is the one raw blocker applied)); Series length (the -8 short-history penalty) (Count of the printing's raw price dates on or before the anchor (the archive starts 2024-03-08, so early anchors get the penalty, as live would)); Which mode the card's verdict would have led with (Only the raw rule is replayed: a card counts when raw mode would buy, even if live its grade or slab mode would have ranked first); The wide pool's other sources (top EV, scarcity, saved cards) (Left out: those cards have no momentum signals, so raw mode can't buy them (neutral setup fields)); Graded comps, gem rates and population (grade and slab verdicts) (Not replayed: there are no historical PSA comps. Grade and slab picks can't be backtested). Grade and slab verdicts can't be backtested: there are no historical PSA comps.

How calibration works. Per era, once 5 grades are back, the miss split becomes (prior × 20 + your counts) / (20 + your misses), and the PSA 10 odds are multiplied by your PSA 10 share over the model's predicted rate, blended the same way and held to ×0.5–×1.5. Inactive everywhere today, so every verdict uses the starting assumptions.

The rule. A model change ships only if the raw buy backtest's hit rate drops by at most 5 points, its edge's confidence interval doesn't fall below zero, the Breeder, Articuno and Duskull regression cards keep their actions, and (once there is data) the live scorecard and the grading-ROI rank correlation don't get worse by more than their noise. Decisions are written up in docs/MODEL_CHANGES.md.