Model Percentile Scorecard v10.4-alpha-mlb

Fully deterministic JS pipeline computes all projections, edges, and Kelly sizing before the AI layer. Claude validates and writes narrative only — structurally cannot override math, add picks outside the candidate table, or increase units. Zero bypass paths.

v10.3 shrinkage calibration: every cover probability starts from the no-vig market price and earns only a fitted fraction of the model's claimed deviation (totals 0.30, spreads 0.35, moneylines 0.50 — fitted on 428 graded real picks, Mar–Jun 2026). Displayed EV is honest 3-15%, not marketing numbers. Quarter-Kelly sizing with 0.5u steps, EV > 3% calibrated floor, never worse than -250, moneyline dogs capped at +160. Every rule is enforced in code, not by prompt instruction.

4u daily exposure cap. 0.5u-3.0u range per pick. Drawdown detection triggers 25% Kelly reduction after 3+ consecutive losses. v10.4 per-market caps enforce measured discipline: totals (the segment with positive closing-line value) carry conviction units; full-game moneylines and F5 are capped at 0.5u until they earn a positive-CLV track record. Parlay sized at 0.5u (0.25u when a lean is on the card). Institutional bankroll management adapted for retail scale.

Cross-sport edge ranking on calibrated expected value (v10.4 — z-scores were not comparable across markets, so they systematically floated moneylines and F5 above totals; z is now a display diagnostic only). MLB cover probabilities come from a discrete negative-binomial run-distribution (analytic convolution of each team's projected run rate) rather than a normal CDF on a point estimate — baseball scoring is discrete and right-skewed, so this correctly prices the 1-run / half-run mass on run lines and the push region on totals. NBA and NHL still use sport-specific total standard deviations (NBA 18.5, NHL 2.3 goals) pending their own discrete models. Edge discount curve compresses outlier edges (log compression above 1 std dev). Consensus-tightness and line-movement discounts shrink edges the market has already priced. The 3% calibrated EV floor filters noise indistinguishable from variance at -110.

Three-method ensemble: Elo-based power ratings, sport-specific efficiency models (pace-adjusted for basketball, shot-based for hockey), and PPG averaging. Injury impact model adjusts projections for star player absences. Back-to-back rest penalty (3% scoring reduction) applied for NBA/NHL. Home-court adjustment baked in. All computed deterministically in JS.

Automated daily at 9am ET via Netlify scheduled function; First-Five lines fold into the same card from a 10am ET pass. 15-minute background worker, now hardened with per-call timeouts + exponential backoff and a deterministic fallback so a slow or rate-limited API can never stall or block a run. Early generation maximizes closing-line-value runway. Picks are final once generated — no mid-day overrides. No human in the loop for generation or delivery.

Five independent data sources: real-time odds (13 sportsbooks), ESPN stats + injuries, daily-updated Elo ratings, prediction market cross-reference (Kalshi + Polymarket), and real-time breaking news (X/Grok). Each source fails independently without taking down the pipeline.

NBA, NHL, and MLB active, plus MLB First-Five-innings (F5) folded in as a built-in market — the one genuinely soft corner, EV-discounted ×0.5 and slot/size-capped until closing-line value validates it. NCAAB disabled (sub-50% real accuracy) and all 8 soccer leagues disabled — 38.5% real daily-run record, and the normal goal model is structurally wrong for draw-heavy soccer (a discrete distribution is the prerequisite to re-enable). Quality over quantity: the cross-sport mandate evaluates every game with available lines and selects the highest calibrated EV regardless of sport.

Spreads, totals, moneyline, and puck lines evaluated across all sports. Market-specific edge thresholds per sport. Alternate spreads and totals attempted with automatic fallback to core markets based on API plan. System considers every available market type per game before selecting highest-edge option.

Type-aligned validation against Kalshi and Polymarket. Parses market titles to classify bet type (moneyline, spread, total) and only compares matching types — totals to totals, spreads to spreads. Multi-market aggregation weighted by volume and bid-ask quality. Proportional unit adjustments in 0.5u steps. Prevents false flags like comparing a moneyline market against a totals pick.

Closing-line value is tracked on every pick 3x daily — Pinnacle-anchored, two-sided no-vig — including a 3am ET overnight run that settles night games (doubleheaders, postponements, and F5 all disambiguated) before the morning pipeline reads bankroll state. v10.4 changed the correction philosophy: the old loop that let trailing performance mutate live probabilities and Kelly multipliers was removed — it chased variance and inflated whatever segment ran hot. Self-optimization is now an observer: each Sunday it aggregates 60 days of CLV by market and by parlay leg, computes real-dollar ROI, and checks the grade ladder, then posts the verdict to Discord. Sizing levers only move on deliberate, human-approved config changes gated by measured CLV — never on a hot streak. Daily 10:30am QA pass audits every published pick's math and narrative.