How the score works.

Before any card is scored, Bet Data Labs pulls a sharp-market anchor for the underlying game. We try three sources in order. The first one with a posted quote wins.

1. Kalshi. US-regulated event exchange. Per-game markets KXMLBGAME, KXWNBAGAME, KXNBAGAME, KXNHLGAME. Typically quoted by midday ET; dark in the morning.
2. Polymarket. Decentralized prediction market. Per-game moneyline markets when available; player K markets surfaced separately. Coverage is thinner than Kalshi but lights up earlier.
3. Bet Data Labs Log5 baseline. Our internal model: Log5 from season records plus league-typical home-field advantage. Always available, but tagged as “Model” so cards that rely only on it cap at WATCH tier until an external sharp confirms.

Betfair Exchange and Smarkets are on the roadmap as Tier-1 sources, sitting above Kalshi in the waterfall.

Alpha Score is additive, capped at 100. Four components:

Per-sport line and diff thresholds. Sourced from the Bet Data Labs playbook (Rule 2). Examples:

• MLB Pitcher Ks: line greater than 3, diff% greater than 17, raw diff greater than 1.0.
• NBA / WNBA Points: line greater than 15, diff% greater than 20.
• 3-PT Made: line greater than 1.5, diff% greater than 50.
• Rebounds / Assists: line greater than 5 (reb) or 4 (ast), raw diff greater than 1.5.

Low-line guard: when line is below 1.5, we require raw diff greater than 1.0 to clear the gate. Otherwise small lines look like 200% edges when they are really 1-stat gaps.

Every card publishes a Kelly fraction at full Kelly, capped at 5% of bankroll. The recommended exposure is quarter-Kelly. Multiply the published Kelly by 0.25 for the recommended unit size.

Every card is graded the morning after it tips, using ESPN final box scores. Win, loss, or push is computed per the published line and side. Units are tracked at flat 1u per card with profit at the published price. ROI is decided units / decided plays. Pushes and ungraded cards are excluded from ROI. The stat-key map covers PITCHER_KS, PITCHER_OUTS (from IP × 3), POINTS, REBOUNDS, ASSISTS, THREES_MADE (parses compound "made-attempted" strings), STEALS, BLOCKS, HOME_RUNS, TOTAL_BASES, HRRBI (h+r+rbi), SHOTS_ON_GOAL, SAVES, RECEIVING/PASSING/RUSHING_YARDS, RECEPTIONS, plus combo props PRA/PR/PA/RA (summed from box-score parts) and ANYTIME_TD (binary from rush+rec TDs). Team-side markets (moneyline, spread, total) grade from the final score.

Graded history feeds back into scoring. Every settled card is bucketed by sport and market, and the bucket's realized hit rate is compared against the win probability the model published. Once a bucket has 30+ settled cards, the gap (shrunk by sample size, capped at 5 points) adjusts the win probability of new cards in that bucket before EV, Kelly, and Alpha Score are computed. A market the model has over-called gets marked down automatically, and the correction itself is printed on the card. Team markets are display-only: their probabilities come from external sharp anchors, not our projection model. The full predicted-vs-actual table and reliability curve are public on the track record page.

That feedback loop had an assumption buried in it: that a card graded three weeks ago was produced by the same model as a card built this morning. It was not. The prop centre changed on July 31, the lean allowed off the market changed on August 1, and the de-vig that sets fair value changed on July 30. Each of those moved the published win probability, which is exactly the number calibration measures. Pooling all of it corrects today's model for a bias that a fix has already removed.

So every graded card now carries a weight: how much it still says about the model running right now. A card built before a change that moved the published probability counts as a fraction of a card, halving for each such change it predates, with a floor so distant history whispers instead of disappearing. Corrections are sized against that effective sample rather than the raw count, and a bucket that cannot muster 30 effective cards stops correcting scoring at all until fresh results arrive. Changes that do not touch the probability, like tier thresholds or page design, do not age anything.

On the day this shipped the effect was blunt: across every prop market, about 13 percent of the graded history described the model actually running. Total bases had 2,953 settled cards behind its correction and an effective 386 of them. Corrections shrank toward zero accordingly, one market lost its correction entirely, and one changed sign. That last part matters. This is a reweighting, not a discount. A market whose recent results disagree with its distant ones can end up corrected harder, which is the whole point of asking the question this way.

The reliability curve on the track record page stays unweighted. It is the record of what we said and what happened, and reweighting it by our own opinion of which cards still count would be marking our own homework.

Edge is stated in percentage points of win probability: our number for the side we took, minus what the de-vigged market price says. An edge of +3.2pp means we make it 64.4 percent where the market makes it 61.2. It is not a percentage change, and it is not the distance between our projection and the line.

That distinction used to be wrong on prop cards, which published how far the projection sat from the line as a share of the line. On a 0.5 total-bases line a projection of 0.55 read as a 10 percent edge off a gap of one twentieth of a base, so the loudest cards on the board were mechanically the lowest-lined ones. Across 7,144 graded cards that number carried no signal: win rate sat flat near 50 percent from 0 through 20 percent and above. Every card now states the same quantity the team markets always did, so the Edge column compares like with like and the board sorts on one scale.

A prop card's edge is a deviation: our number sits some distance off the posted line. How much that distance is worth is not the same in every market, so we measure it. For each prop type we regress the realized result on the claimed deviation across every graded card, which returns the fraction of our claimed edge that actually shows up in the box score.

Over 5,673 settled props (July 2026), only total bases came back clean, and even there the slope was 0.145: the market absorbs roughly 85 percent of what our model thinks it sees. Threes and rebounds came back positive but thin. The WNBA possession-sum composites (points, assists, PRA, PA, PR) regressed negative, meaning a bigger claimed edge has been a worse bet. That is the measured version of a rule the desk already ran on instinct: the loudest card on the board is usually the noisiest one.

So the lean our model is allowed off the de-vigged market price is set per market, scaled by how well that market's edges have held up, and it can only ever tighten. A market whose deviations cannot beat their own standard error is priced at the market, gets no EV credit and no Kelly stake, and says so on the card. A market with no graded history yet keeps the full lean rather than being silenced on no evidence.

The Board is the browsing surface. The Slate is the product. Every night the model selects a small set of plays to press, sized in units, with a hard 8u daily exposure cap. Selection is deterministic: given the same day bundle and rules, we emit the same Slate.

• Eligibility. PRESS or STRONG tier only. Full Kelly ≥ 0.75u. No active kill signals. No top-edge fade or heed-your-own-flag flags.
• Caps. Max 3 plays per sport. Max 8 plays total. Deduped by (game, player, prop).
• Sizing. Quarter-Kelly per play, floor 0.25u, ceiling 3u. If the total would breach 8u, scale down proportionally. 1u = $50 default.
• Near-misses. Cards that just missed appear on the Slate page under "What almost made the cut," grouped by the bucket that blocked them.

Two rules from the playbook are always-on gates that downgrade cards after they score:

• Top-edge fade. The board's single biggest edge historically went 0-5. That one card is capped to LEAN (score ≤ 74), marked topEdgeCaution, and can never anchor a parlay.
• Same-game cap. Across July, cards ranked 3rd or deeper inside their own game went 22-43 (33.8%, -13.3u) while the 1st and 2nd reads of a game went 127-114 (52.7%, +9.7u). The gap survives controlling for tier, so it is correlation and model artifact, not rating. Every published card carries its rank in sameGameCluster; reads past the 2nd decay 10 to 18 points, floored at WATCH so nothing is deleted, and are barred from the parlay.
• Heed your own flag. When a favored moneyline card carries its own kill signal (weather, injury, opener risk), it's capped to LEAN. Caveated favorites went 0-3 in the training sample.

Every card also carries a per-market CONSENSUS gate check (Section 04) and per-sport kill signals: MLB weather + park factors, WNBA rebound cannibalization, injury-out player filters, MLB starter scratch detection.

MLB weather drives three things: a rain kill, a hot/cold/wind flag, and the only lean the game-total lane is allowed to take (thin hot air carries a batted ball). All three assume the reading describes the air the game is actually played in.

Until 2026-08-01 they did not. The indoor test read a field ESPN does not populate. Measured across 409 games and all 30 venues of the season to date, it was absent on 100% of them, Tropicana Field included, so every roof guard behind it was dead code. Roof state now comes from a table keyed on ESPN's venue id, which survives sponsor renames (Minute Maid Park became Daikin Park, same id).

• Rain. No roofed park can be rained out. A precipitation kill is only possible at the 22 open parks.
• Air. A temperature or gust reading is a playing condition at an open park, and at T-Mobile Park, whose umbrella roof covers the field without enclosing it. Under a sealed retractable roof it is not, because the readings extreme enough to fire a signal are close to the same conditions that get a roof shut.
• Totals. The lane only clears its edge gate at or above 96.7°F, or at or below 43.3°F. Outdoor baseball is almost never that hot. Retractable parks are, and they sit in the hot markets by design, so nearly every card this lane could publish was a game played indoors with the air conditioning on.

The correction is subtractive only. It withdraws readings we were not entitled to and never adds a lean, so it can cost a card its edge but can never manufacture one. An unlisted venue is treated as open, which changes nothing. On the 2026-08-01 board it cleared a rain kill and a 95°F heat flag off the nine cards in Texas at Houston, a game under a shut roof with a 45% chance of rain outside.

Card detail pages carry a "Recent form" strip: L5 / L10 / SZN hit-rate cells scored against the card's line. The stat is pulled from ESPN's player gamelog and cached 4 hours per player. Green ≥ 60%, red ≤ 40%. When L5 and SZN disagree, the line is fighting the trend and worth extra caution.

Team-form is a separate strip: L10 record, streak, home/road split, run/point differential. Sourced from the ESPN standings hydrator (12h cache).

Soccer moneylines and totals run through a Poisson-goals model per team: λ = league avg × attacking strength × opposition defensive strength × home-field advantage. Three-way market probabilities (win, draw, loss) come from convolving both team distributions. Confidence scales with games played (min 20, max 85).

Anytime and first-goalscorer projections distribute team lambda to individual players using position weight (FW 5.0 / MF 2.0 / DF 0.5 / GK 0.0), presumed-starter status, and a curated elite-scorer map with S / A / B tier multipliers (4× / 2× / 1.3×). Known penalty-taker duty adds a fixed weight bonus so the PK anchor appears at the top of the team's share.

The /edgespage ranks every card by the gap between our fair probability and DK's implied. Distinct from the Board (Alpha Score composite) and Movers (line movement over time). Same underlying data, framed for prediction-market bettors: where the book is most wrong, in either direction.

Every surface has a machine-readable variant. Integrations pull without HTML scraping:

• /api/slate - tonight's Slate as JSON. ?compact=1 lightens the payload. ?history=7 attaches the receipts strip.
• /api/slate/[date] - historical Slate joined to graded outcomes.
• /api/board - filtered board (sport, tier, minScore, sortBy).
• /slate/feed.xml - RSS for slate-only subscribers.
• /api/slate.ics - iCalendar subscribe; plays and kick times land in your calendar app.
• Telegram push - the cron endpoint fires a formatted digest when the Slate publishes.

Followers can tail cards locally (the ☆ button on any list surface). The tail log lives in localStorage - nothing about your bets touches our server. /me joins your tails to graded outcomes to compute personal record.

• We do not pick after games tip.
• We do not change historical cards.
• We do not hide losses.
• We do not run a Discord DM.
• We do not size beyond quarter-Kelly. Bankroll first.