Every update, tracked in the open.

WTA 12-FACTOR IMPORTANCE WEIGHTS. The women's model now runs on a 12-factor weighted composite instead of the hold-%-is-the-engine stack — twelve individual factors, each weighted on its own: Return Points Won % 21, Break Points Converted % 17, Service Points Won % 10, Return Games Won % 9, Break Points Saved % 8, 1st Serve Points Won % 8, 2nd Serve Points Won % 8, Surface Win % 8, Service Games Won % 5, Double-Fault Rate 2%, H2H 2%, Ranking Differential 2% — 100% total. MECHANICS: wtaWeightedCompositeEdge computes one composite edge E per matchup (each factor a head-to-head differential normalized by 2x the 2026 tour standard deviation, clamped to +/-1, double faults inverted); E feeds the Monte Carlo loop directly — every simulated service game and tiebreak draws against edge-adjusted holds (+/- E x 0.05 per side), so the 12 factor weights bank against fresh noise game by game instead of pre-shifting the hold inputs — a max edge moves each hold 5pp in opposite directions, typical matchups 1-2pp per side. DATA: the WTA's own free stats API (api.wtatennis.com, no key) — one paginated sweep builds the full 2026 tour table, cached 6h in the Cache API, name-matched per player, minimum 5 matches to count. MISSING-DATA RULE: any missing input goes neutral (0), never a fabricated default — if either player's WTA row is missing, the composite stays off and the standard hold path runs untouched. NO-DATA GATE (all tennis): when neither player has verifiable data (no rank, no season W/L on either side), the matchup is declined as unable to process instead of simulating on pure estimates — the model either runs on real stats or it doesn't run. WTA and ITF: the composite runs for both tours (active only when both players have WTA tour rows); ATP never enters that path — its hold-% engine, surface logic, and writeups are unchanged. The WTA writeup now leads with the return-game framing. Untouched: 0.40 shrinkage, down-only band calibration, 50/50 rule + caps, best-of-3 for WTA, model 1.6.1. Covered by tests/wta-importance-weights.test.js and tests/tennis-no-data-gate.test.js. Full suite green.

CONSERVATISM SHRINKAGE SET TO 0.40 (0.55 → 0.40). The daily shrinkage watch's drift note showed the empirical optimum moved to 0.40 (90% noise band 0.35-0.47), with the 0.55 setting sitting outside the data's noise band — the sim's edges were being over-flattened into coin flips. Easing shrinkage keeps more of the sim's lean (a raw 70% edge reports ~58% instead of ~61%, a raw 80% edge ~62% instead of ~67%). All deflationary mechanisms stay in place: down-only band calibration, 50/50 rule + caps, grading, branding, model 1.6.1.

MANUAL ANNOUNCEMENT BLASTS. New "announcement" alert type plus an admin /admin/send-announcement endpoint (admin-key authed, title/body/url params): Drew can now push any message he wants to every push-subscribed user on demand. Every blast still flows through the single dispatchAlert choke point — per-user prefs (new "Announcements from ParlayKings" toggle in notification settings, default on), quiet hours, the 12/day cap, and 24h dedupe on identical messages (re-sending the same text within a day is a no-op, not a double blast). Announcements respect quiet hours — only game-day score alerts bypass them, by design. Covered by tests/alerts.test.js. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, grading, model 1.6.1.

FOOTBALL 15-FACTOR IMPORTANCE WEIGHTS. Drew's exact importance weights are now the football model — 15 factors, 100% total: EPA/efficiency 17%, QB factor 15%, Passing EPA/dropback 12%, Success rate 10%, Pressure/protection 9%, Turnover differential 8%, Explosive plays 7%, Points per drive 5%, Red-zone TD 4%, Early-down EPA 3%, 3rd/4th down 3%, Rush EPA 2%, Sack rate 2%, Special teams 1.5%, Penalty efficiency 1.5%. MECHANICS: footballWeightedCompositeEdge computes one composite edge E per matchup (each factor a head-to-head expected-points edge, clamped to +/-6 before weighting, normalized vs league average); bridgeToTeamSim applies it additively as +/- E/2 on the two Poisson lambdas before the realistic-range clamp. The old tiered-hierarchy multiplicative stack is bypassed for NFL/CFB — when the composite is on, the QB, RB, WR, O-line, pass-rush, kicker, efficiency, and turnover-margin multiplier slots all go neutral (exactly 1.0) inside calculateLambdaMultiplier so nothing is double-counted; rest, travel, form, weather, injuries, and SOS are untouched, and other sports never set the flag. GONE from the sim path: the standalone RB/WR multiplicative factors (not on Drew's list — team rushing lives in Rush EPA at 2%). DATA: NFL EPA table (nflverse play-by-play, all 32 teams) is embedded in the bundle and refreshed weekly by nfl_epa_refresh.py — EPA, success rate, passing/dropback EPA, rush EPA, early-down EPA, sack rates, special-teams EPA, points per drive, all shrunk by gp/8 early in the season. The QB factor (individual QB vs league average: CPOE/accuracy, decisions, sack avoidance, pressure play, recent form) and Passing EPA (the whole dropback ecosystem: QB, receivers, protection, scheme, opposing pass D) stay separate inputs at Drew's combined 27% — overlap acknowledged, not increased without backtesting. CFB runs the same 15-factor structure with the advanced EPA factors neutral until a real CFB EPA feed lands — the kicker factor is its special-teams proxy; nothing is faked. MISSING-DATA RULE: any missing input goes neutral (0), never a fabricated default — one-sided missing data produces no phantom edge. Untouched: 0.55 shrinkage, down-only band calibration, football always-picks-a-winner, neutral-site HFA behavior, 50/50 rule + caps for other sports, model 1.6.1. Covered by tests/football-importance-weights.test.js (102 checks: exact weights and 100% sum, 32-team EPA table sanity, neutral missing-data behavior, reversal symmetry, QB/passing separation, exact 4x early-season shrinkage, per-factor clamp, composite-on neutralization with legacy path intact). Full suite green.

UFC 15-FACTOR IMPORTANCE WEIGHTS. Drew's exact importance weights are now wired into the Monte Carlo loop — 15 factors, 100% total: Significant Strike Differential 16%, Opponent-Adjusted Performance 15%, Takedown Defense 10%, Recent Performance / Trajectory 10%, Significant Strike Defense 8%, Significant Strikes Absorbed/min 7%, Strength of Schedule 7%, Knockdown Rate 6%, Takedowns/15 5%, Takedown Accuracy 4%, Round-by-Round Performance 4%, Significant Strike Accuracy 3%, Submission Attempts/15 2%, Control-Time Differential 2%, Finish Rate 1%. Drew's five emphasized inputs carry 59%. MECHANICS: ufcWeightedCompositeEdge computes one composite edge E per matchup from the 15 normalized factors; each simulated round banks E * 0.75 (UFC_ROUND_EDGE_SCALE) against noise. GONE: the old strike-differential kicker (0.22 edge, x2.4 into the decision), the 1.15 takedown control credit, and the 0.98-1.02 form volume multiplier. KEPT: the KO/sub/TD event rolls at their old math (takedown prob multiplier 2.2, sub-finish coefficient 0.08, KO chance on landed strikes x0.035), the last-five exact-tie tiebreak (|raw-50| < 0.5), and the CANNOT_VERIFY gate. NEW DEEP STATS from ufcstats.com fight-detail pages (last 3 fights): knockdown rate per 15, control-time differential per round, round win % on significant strikes, and a trajectory metric (newest fight vs older average) blended 70/30 into the Recent Performance factor. FightMatrix factors (opponent-adjusted 15%, strength of schedule 7%) are neutral 0 until the FightMatrix feed lands — nothing is faked. MISSING-DATA RULE: no fabricated defaults — when SApM is missing on either side, both the SApM factor and the strike-differential factor go neutral instead of inventing a 95%-of-SLpM phantom. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, the tale-of-the-tape prose, model 1.6.1. Covered by tests/ufc-importance-weights.test.js (28 checks: all 15 keys wired, weights sum 100%, top-5 = 59%, FightMatrix neutral, SApM phantom fix, single-factor isolation, deep-parser fixtures, 3/5-round sims), tests/ufc-form-demoted.test.js (form as the real 10% factor), tests/ufc-tiered-hierarchy.test.js (composite structure, event rolls, physicals capped at +/-0.10), plus updated ufc-tiebreak-and-rounds and ufc-tale-of-the-tape. Full suite: 842 passed, 0 failed.

AUTO-GRADE DATE-MATCH HOLE CLOSED. findFinalResult's exact Event Date check previously only applied when multiple ESPN events matched a pick's matchup — a single candidate was graded without any date verification. This morning's run-all proved the hole: ten 9/24 MLB picks were graded against 9/23's finals (same matchups, series continuing) before any 9/24 game was played. Now a single candidate must also match the pick's Event Date exactly, otherwise the pick stays Pending — no guessing, single or multiple. The ten bad grades were wiped back to Pending in Airtable and the replica sheet. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, model 1.6.1.

BASEBALL TIERED HIERARCHY. Final baseball importance order, enforced by per-tier clamps with strictly decreasing max swing: TIER 1 = the starting pitcher as the single dominant weight — the QB of baseball — clamp widened 0.86-1.16 to 0.70-1.40 so an elite ace moves the number like the football QB; ERA regression toward 4.2, the sample gate (w<0.25 neutral), and starter-lock timing logic are untouched. TIER 2 = the bullpen with real weight: (a) FIXED the confirmed quirk where bullpen ERA was shrunk toward the 4.2 starter anchor so every elite bullpen flattened to 0.86 — relievers now regress toward their own 3.90 anchor with a tier-2 clamp of 0.80-1.25; (b) the full-lock pitching blend rebalanced 62/38 to 58/42 starter/bullpen (staff proxy 78/22 to 75/25), so the pen carries genuine tier-2 weight while the starter keeps the majority. TIER 3 = the bats — team runs scored stays the offensive engine, unchanged. TIER 4 = infielders sampled INDIVIDUALLY for the first time (the model previously knew zero batters — no lineups at all): each team's starting SS, 1B, and top 2B/3B (pooled, one combined factor) identified from the live roster by at-bats (min 100, regular-season only, OUT players skipped) and factored by OPS in Drew's positional order — SS 0.88-1.12 > 1B 0.90-1.10 > 2B/3B 0.92-1.08; Drew's caveat honored (bat over glove, OPS-driven). TIER 5 = home field (+0.15), rest, weather, travel, recent form — unchanged and smallest. Each tier rides its own lambda multiplier slot (starter+pen blend stays in the pitcherFactor slot suppressing the opponent; new ssFactor/firstBaseFactor/secondThirdFactor slots boost a team's own expected scoring). Untouched: 0.55 shrinkage, down-only band calibration, bivariate Poisson, NO 50/50 for MLB (Drew's standing call — baseball always picks a winner), neutral-site HFA behavior, model label SAM 1.6.1. Still open (flagged, not approved): neutral-site HFA gap, changelog-vs-code doc mismatches (BDL 50/50 vs 65% prose, rho 0.02 vs 0.28 note, recent-form exclusion note), park factors, platoon splits, umpires, bullpen workload. Covered by tests/baseball-tiered-hierarchy.test.js. Model 1.6.1.

FOOTBALL TIERED HIERARCHY. Final football importance order, enforced by per-tier clamps with strictly decreasing max swing: TIER 1 = the QB as the single dominant named-player weight, now clamped 0.70-1.40 (was 0.80-1.25 — like the pitcher in baseball, the QB gets the highest weight in the game); TIER 1 also = the lead back by rushing YPC (0.88-1.12) and the WR1 by receiving YPR (0.90-1.10), sampled individually from each roster by actual usage; TIER 2 = the offensive line via sacks allowed, rush YPC, and yards per play (0.92-1.08); TIER 3 = the pass rush via the OPPONENT's sacks created, TFL, and turnovers forced (0.93-1.06); TIER 4 = the kicker as a standalone meaningful input via the actual kicker's season FG% (0.95-1.04 — a good kicker can swing a coin-flip game); TIER 5 = the remaining team-efficiency bits (3rd/4th down, red zone, INT%, TOP, pace, punts, penalties), demoted from a 0.90-1.10 clamp to 0.96-1.03; TIER 6 = home field, rest, weather, travel, recent form — unchanged and smallest. QB CONSOLIDATION: the QB was triple-counted (team PPG, pass YPA/pass YPG inside the generic efficiency factor, standalone QB factor) — pass YPA, pass YPG, and yards per game are OUT of the efficiency factor, sacks allowed moved to the O-line factor, opponent sacks/TFL moved to the pass-rush factor, FG/XP bits moved to the kicker factor. Team PPG stays the scoring engine. Each tier rides its own lambda multiplier slot (qbFactor stays in the pitcherFactor slot; new rbFactor/wrFactor/olineFactor/passRushFactor/kickerFactor params on calculateLambdaMultiplier). The skill factors sample real athletes: QB by passing attempts (min 50, existing path), lead back by rushing yards (min 40 attempts), WR1 by receiving yards (min 20 receptions), kicker by kicking points; OUT players are skipped and never boost a team; missing data is neutral. Untouched: 0.55 shrinkage, down-only band calibration, football always-picks-a-winner, neutral-site HFA behavior, kickRetAvg sign quirk, CFB sosDiff (still 0), model label SAM 1.6.1. Covered by tests/football-tiered-hierarchy.test.js (50 checks: QB clamp widening, RB/WR/O-line/pass-rush/kicker factors, efficiency demotion cap, strict tier ladder, elite QB beats good form, elite WR beats max rest, O-line swing beats home field, kicker flips a seeded 50/50 sim, dead-even games still resolve). Tonight's UFC tiered hierarchy, tennis hierarchy, basketball fixes, wiring audits, NASCAR removal, and NHL restoration all verified still present. Model 1.6.1.

UFC TIERED HIERARCHY. Final UFC importance order: TIER 1 = striking trio (striking output SLpM > strike accuracy > strikes absorbed) AND grappling (takedowns/subs) at full strength — control score 1.15 per landed TD, takedown probability multiplier 2.2, sub-finish coefficient 0.08. TIER 2 = physicals: reach, height, stance, age. TIER 3 = everything else: recent form (+/-2% band), finish mix (+/-10%), home-country boost (0.015), weight-class finish biases (all demoted earlier tonight and still demoted). The earlier grappling demotion was REVERTED — grappling sits alongside the striking trio in tier 1, not below it. The striking trio's own math is untouched (SLpM x accuracy strike volume, x0.035 KO scaling, 0.22 strike-differential edge, x2.4 into the decision), as are the CANNOT_VERIFY gate, 0.55 shrinkage, down-only band calibration, UFC always-picks-a-winner behavior, and the prose. Covered by tests/ufc-tiered-hierarchy.test.js (12 checks: tier-1 fighter beats physicals-only fighter 73-27, even tier-1 with physical edge decides 72-28, tier-3 inputs with perfect form and maxed mix only nudge 53-47). Also folded into this entry: (1) LAST-FIVE EXACT-TIE TIEBREAK: last-five form is the very last thing — only when the raw sim split is within |raw-50| < 0.5 (inside Monte Carlo noise) does the better last-5 winPct get the pick; a raw 50.9-49.1 is a decision, not a tiebreak; last-5 tied or missing -> raw higher side stands; applied before shrinkage so confidence shrinks normally (lands 51-49); the prose explicitly states when last-five broke a dead-even sim. (2) EXPLICIT PRELIM ROUNDS: prelim fights are 3 rounds, period — the card-position rule is the authority (title fights and main events go 5; anything not identifiable as title/main-event is 3), ESPN's format.regulation.periods is the explicit scheduled-rounds field, the description '/N Rnd/' parse is only a cross-check now. (3) ROUNDS VERIFIED END TO END: traced fetchUFCFightRounds -> runMatchupSimulationCore UFC branch (roundsToUse = rounds === 5 ? 5 : 3) -> runUFCMultiStatSimulation (safeRounds = rounds === 5 ? 5 : 3, loop 'for (rnd = 0; rnd < safeRounds && !finished; rnd++)' iterates exactly rounds times, no hardcoded 3/5 in the loop path, no off-by-one; fatigue fade and per-round finish rolls scale with the actual round count). Covered by tests/ufc-tiebreak-and-rounds.test.js (21 checks: tiebreak fires on seeded dead-even sim with 5-0 vs 2-3 and picks the 5-0 fighter 51-49, 50.9-49.1 does not invoke form, prelim resolves to 3 even with missing description, title/main-event resolve to 5, 5-round sim shows materially higher finish rate than 3-round). Tonight's basketball fixes, wiring-audit fixes, NASCAR removal, tennis importance weights, UFC form demotion, and UFC tale-of-the-tape all verified still present. Model 1.6.1.

UFC TALE OF THE TAPE ON TOP. The physical tape and striking/grappling stat differentials (reach, height, age, stance, SLpM/accuracy/absorbed, takedown avg/accuracy, submission avg, strike and takedown defense) are now the top tier of the UFC model by construction: everything below them was demoted, nothing above them was boosted. DEMOTED: (1) ufcFinishMixScale compressed from clamp(0.72 + share*0.9, 0.78, 1.32) to clamp(0.90 + share*0.35, 0.90, 1.10) — a fighter with 100% KO history now gets at most a +10% finish-chance modifier instead of +32%. (2) The weight-class finish-bias table compressed halfway toward 1.0 across all classes (e.g. heavyweight KO 1.18 -> 1.09, flyweight KO 0.82 -> 0.91). (3) The home-country strike boost halved from +0.03 to +0.015. UNCHANGED AND FULL STRENGTH: the striking/grappling core math, frame edges (combined clamp +/-0.08), stance edges, the age speed edge and 35+ cliff, the 0.98-1.02 form band (form stays the bottom input), the CANNOT_VERIFY gate, 0.55 shrinkage, down-only band calibration, UFC always-picks-a-winner behavior, and the prose (the tale of the tape keeps printing in full — it is information AND the engine). Covered by tests/ufc-tale-of-the-tape.test.js (extracts the real bundle functions). Tonight's basketball fixes, wiring-audit fixes, NASCAR removal, tennis importance weights, and UFC form demotion all verified still present. Model 1.6.1.

UFC RECENT FORM DEMOTED. Recent form (raw last-5 W/L, no opponent-quality adjustment) is now the BOTTOM of the UFC importance hierarchy — the least important input in the sim. The form multiplier in runUFCMultiStatSimulation was compressed from 0.92 + 0.16 * (winPct/100) (a 0.92–1.08 swing) to 0.98 + 0.04 * (winPct/100) (a 0.98–1.02 swing): a 5-0 fighter now gains at most +2% strike volume and an 0-5 fighter loses at most 2%. Form can nudge a genuinely close matchup but can no longer flip a fight that the striking/grappling differentials decide. Prose is unchanged — the last-5 is still printed in the tale of the tape as information, just demoted in the math. Covered by tests/ufc-form-demoted.test.js (extracts the real bundle function). Untouched: all UFC sim math (striking, grappling, defenses, frame, stance, age, weight-class scale, finish mix, rounds, home-country boost), the CANNOT_VERIFY gate, 0.55 shrinkage, down-only band calibration, UFC always-picks-a-winner behavior, and tonight's basketball fixes, wiring-audit fixes, NASCAR removal, and tennis importance weights — all verified still present. Model 1.6.1.

TENNIS IMPORTANCE WEIGHTS. "Hold serve %, court surface, number of sets" are the three most important inputs — everything else subordinate. The tennis model now enforces that hierarchy in tennisServeHoldProb and the sim writeup. (1) HOLD % is the engine: real serve tape is used directly as the hold probability, untouched and dominant. (2) SURFACE is the co-primary driver, rules unchanged: grass +0.04 / clay -0.04 / hard 0 on hold probability plus the player's own surface W-L edge at 0.15 weight (5+ surface matches, delta capped at +/-25pp), total surface contribution clamped at +/-0.06, and NO modifier on top when real surface-specific hold tape already captures the surface effect (no double-counting). (3) SET FORMAT compounds structurally: verified that runTennisMatchSimulation already implements best-of-N as real set-by-set probability math (first to 2 sets in BO3, first to 3 in BO5) with no fudge factor — the stronger server's edge genuinely grows in BO5 (a 0.85-vs-0.78 hold edge gains ~5pp of match win probability going BO3 to BO5). ATP Slams = best-of-5, everything else = best-of-3, unchanged. DEMOTED to tiebreak-level nudges only: recent form's coefficient halved (0.015 -> 0.0075, so a perfect recent record moves hold by 0.00375 — it cannot overrule the serve/surface signal), and the no-tape fallback estimate's rank swing compressed (0.82 - log10(rank)*0.09 -> 0.84 - log10(rank)*0.045, so rank 1-vs-100 moves the estimate ~5pp instead of ~10pp) with the season W-L band halved — rank and generic W-L still differentiate when no serve tape exists, but they can no longer manufacture a favorite out of a ranking. H2H edge (max +/-0.008) was already tiebreak-level and stays. The sim writeup now states the hierarchy: "Hold % is the engine, then surface, then the set format. Rank, recent form and season W/L are tiebreak-level nudges only." Covered by tests/tennis-importance-weights.test.js (25 checks against the real bundle functions). Untouched: 0.55 shrinkage, tennis 0.72 both-taped carve-out, down-only band calibration, 50/50 rule + caps, per-league bias (removed, stays removed), tennis weather wiring, tonight's basketball fixes, wiring-audit fixes, and NASCAR removal — all verified still present. Model 1.6.1.

NASCAR REMOVED FROM THE DAILY SLATE. The daily board-run loop (buildDailySlate, "run all") does not touch NASCAR — it is not in SLATE_LEAGUES, so racing/nascar-premier is never fetched for the daily slate. Reason: there is no way to do NASCAR effectively — no real statistical data source exists, so predictions were running on model-estimated 0-100 power ratings with no live per-driver stats. The single-prediction path is unchanged: an explicit NASCAR matchup request still verifies via get_game_schedule and estimates driver ratings as before. Covered by tests/nascar-slate-removal.test.js. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, per-league bias (removed, stays removed), tonight's basketball fixes and wiring-audit fixes, model 1.6.1.

BASKETBALL FIXES. Five repairs across NBA/WNBA/CBB, all conservative and tested: (1) OUT STAR NO LONGER BOOSTS: fetchTeamRosterRaw never attached injury data, so isPlayerOut() was a silent no-op — an OUT leading scorer still boosted his team's lambda up to 1.25x while the injury haircut took at most ~6% (the 1.5.7 changelog claimed OUT suppression worked, but the branch never fired). Rosters now attach live injury statuses from ESPN's league injuries endpoint (same source getInjuryReport uses; a cached GET, one cheap call per league, fail-open to old behavior if down) matched by athlete id with display-name fallback — the existing suppression branch now actually fires, so an OUT star's factor reverts to neutral. This also repairs the same dead check for NFL/CFB/NHL key players. The general injury haircut is untouched. (2) NEUTRAL-SITE HOME COURT: a neutral venue previously fell through as "teamB is home" and handed +3/+4 to an arbitrary side — systematic in CBB conference tournaments and March Madness. New isNeutralVenueEvent() reads ESPN's neutralSite flag and homeAway "neutral" markers; neutral games now award 0 home-court points to either side and set both travel tiers to medium (neither side is "home"). Non-basketball leagues' behavior is unchanged. (3) CBB TRAVEL: basketball/mens-college-basketball added to SPORT_PATH_TO_TRAVEL_LEAGUE plus a new 72-program CBB arena-coordinate table (major conferences + Gonzaga/Memphis/SDSU/Dayton; newest arenas verified — Baylor's Foster Pavilion Waco 31.5567,-97.1216, Texas' Moody Center Austin). Same coarse tiers as every league (<200 mi short, >1500 mi long, else medium); unlisted programs fail open to the old medium default instead of the old flat 0.99x-for-everyone. (4) KEY-PLAYER SELECTION: basketball now picks the pool winner by max PPG, not max gamesPlayed — ESPN roster order is not depth-chart order, so a 12-PPG ironman was beating a 28-PPG star. NFL/CFB still sort by usage (passing attempts = the real starter, per the Dak Prescott verification) and NHL by games. (5) PER-LEAGUE SCORER SCALES: the single (18, 40) NBA-shaped scale compressed WNBA/CBB stars into a near-constant band — a 15-PPG CBB scorer even graded BELOW average (0.925). New keyPlayerScaleForLeague() mirrors the qbScaleForLeague precedent: NBA (18, 40) unchanged, WNBA (13, 29), CBB (11, 25), derived proportionally from each league's team-scoring environment (sim base rates ~115/83/72 PPG). Now an elite scorer in any league lands near the 1.25 clamp and an average starter sits at 1.00, exactly like the NBA. Covered by tests/basketball-fixes.test.js (51 checks against the real bundle functions). Deliberately NOT implemented (proposal-only): SOS builds, draw-sigma changes, 3PT-variance modeling, early-season-blend re-weighting, and the other 30 re-weight proposals from the loop audit. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, per-league bias (removed, stays removed), model 1.6.1.

WIRING-AUDIT CLEAN TARGETS. Five small fixes, no re-weighting. (1) SOCCER STAR RESOLUTION: getSoccerStarFactors returned after the first competition where EITHER team had a scorer, so when the two teams played in different competitions the second team's star silently dropped (no adjustment at all). New pure resolveSoccerStarsByCompetition() resolves each team's star independently in ITS OWN competition, matching how the base-rate fetch already works — e.g. Arsenal's scorer from the Premier League and Real Madrid's from La Liga in the same pick. Early fetch exit preserved (stops once both teams resolve). (2) SOCCER VENUE-SPLIT TRACE: the writeup built the 50/50 TOTAL-vs-HOME/AWAY venue-split note with += then immediately overwrote it with = — math was fine, the prose lied. The note now actually prints. (3) NHL GOALIE SAMPLE GATE: the probable-goalie path fetched games played but never used it — a 2-game .960 call-up got 1-(0.960-0.905)/0.1 = max 0.8 goal suppression from 2 games of noise. New nhlGoalieEffSavePct(): under 3 games the SV% is NOT used (goalie stays identified, factor 1, no fallback to a different goalie's numbers), 3-9 games shrink toward the .905 league average by games/10, 10+ games full weight. Exact magnitudes: min games 3, full sample 10, league-avg SV% .905; existing 0.8-1.25 factor clamp untouched. (4) UNUSED FOOTBALL STATS WIRED: the ESPN football fetch already pulled yardsPerGame/passYpg/rushYpg/fourthDownPct but footballProfileToFactor never read them — data fetched and discarded. Now wired conservatively: yardsPerGame 0.0002 coef clamped 0.98-1.02 and ONLY when passYpa is missing (no double-count with per-attempt efficiency), passYpg 0.00015 clamped 0.99-1.01, rushYpg 0.00015 clamped 0.99-1.01, fourthDownPct 0.0006 clamped 0.99-1.01. League anchors: NFL 335/218/117/52, CFB 390/245/145/48 (total YPG / pass YPG / rush YPG / 4th-down %). Existing early-season shrink and final factor clamp unchanged. (5) DEAD-CODE CLEANUP: removed unused tennis fromTape, computeTennisPowerRating and its ratingA/ratingB adjustment chain, the unused forecast from getTennisMatchupPowerRatings, and the unconsumed rawA from runTennisMatchSimulation; parsed.ufcData marked as an explicit dead field (parse shape preserved). computeUFCPowerRatings was flagged as orphaned but is live through getUFCMatchupPowerRatings — verified, kept. Covered by tests/wiring-audit-3.test.js (28 checks, all green). Untouched: 0.55 shrinkage, tennis 0.72 both-taped carve-out, down-only band calibration, 50/50 rule + caps, model 1.6.1.

CONTINUED AUDIT: CLEAN DATA WIRINGS ONLY. Empirical graded-pick review (30 days through 2026-09-23): UFC weakest at 51.4% against 63.9% published confidence, NHL weak on a tiny sample, NFL slightly overconfident; WTA/tennis/WNBA/NCAAF/soccer all working — nothing about their wiring was touched. Six activations, all pure connections of data the worker already fetches, zero new model factors: (1) UFC fallback merge now keeps real SApM/strikeDefense/takedownDefense/finishMix/country from the ufcstats.com chain instead of dropping them and forcing the sim onto league-average constants (mergeUfcFallbackStats). (2) UFC unknown stance is now neutral (null) instead of defaulting to orthodox — missing data can no longer manufacture a phantom +/-0.025 stance edge vs a known southpaw. (3) NFL/CFB cold weather: Open-Meteo tempF already arrived but never reached the sim — now <=32F = 0.99x and <=20F = 0.985x on both offenses, symmetric, conservative, no heat drag for football. (4) Soccer venue splits: the football-data.org standings response already contains HOME/AWAY tables in the same payload (zero new calls) — each side's home/away-table scoring rates now blend 50/50 with its TOTAL season rate (TOTAL stays the anchor); neutral-venue tournaments (WC/EC) and venue samples under 3 games fail open to TOTAL, and fallback paths (BALLDONTLIE/ESPN) pass through untouched. (5) Golf now prefers the current-season scoring-average split over career/all-time when ESPN labels splits; unlabeled responses fall back to prior behavior. (6) Bowling blends all available PBA season rows 60/30/10 (most recent first) instead of reading only the first 20xx row — identical behavior with a single season. Also a trace honesty fix: darts no longer claims "current-season" averages since the parser can't verify season context. Deliberately NOT implemented (proposal-only, flagged for Drew): NHL PP/PK (the deep-dive's claim was half-wrong — the data exists on NHL's own API as powerPlayPct/penaltyKillPct but NOT in any already-fetched response, so it's a new fetch, not a free wiring; hockey literature also rates PP/PK weak vs 5v5 goal differential), UFC layoff/ring-rust (fight dates not proven present in the already-fetched page), NBA 4-in-5 density drag, widening the basketball key-player candidate pool, tennis thin-tape rank shrinkage, and golf/bowling/darts modeling tweaks. Covered by tests/wiring-audit-2.test.js (49 checks against the real bundle functions). Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, NHL run-all exclusion, per-league bias (removed, stays removed), model 1.6.1.

QB KEY-PLAYER FACTOR: PER-SOURCE SCALES. The deep-dive audit flagged that NFL/CFB both ran their QB factor through one scale (avg 90, divisor 150) — but ESPN's athlete-overview "QBRating" field is a DIFFERENT stat per league, verified live against ESPN's real API 2026-09-23. NFL = classic passer rating: Mahomes 103.8, Allen 125.8, Fields career 84.7, top-25 mean ~104.8 (0-158.3 scale, starter average ~90) — the old params were CORRECT here, kept as-is. CFB = NCAA passer efficiency: Sayin 182.4, leaders mean ~134.6 / median 132, FBS average ~135 (scale runs ~100-200+, uncapped) — the old params were WRONG: an average 135-efficiency college QB computed 1 + (135-90)/150 = 1.30 and maxed the 1.25 clamp, and nearly every qualifying college QB sat at or near the ceiling, so the factor couldn't differentiate. The NCAA.com "Pass Eff" fallback is the same NCAA-efficiency metric, so one CFB scale fixes both the ESPN path and the fallback. Fix: new qbScaleForLeague() gives each league its own (avg, divisor) — NFL (90, 150), CFB (135, 225) — and new qbFactorFromValue() applies the shared 0.8-1.25 clamp; the factor now always measures "how far above/below average for THIS stat". The CFB divisor mirrors the NFL curve in standard-deviation terms: elite (182) -> ~1.21, average (135) -> 1.00 neutral, bad (95) -> ~0.82. The trace label is honest too: CFB now reads "Pass Eff" instead of "QB Rating". Backup-QB substitution behavior untouched (OUT QBs still revert to neutral); shrinkage, bands, 50/50 rule, caps untouched. Covered by tests/qb-factor-scale.test.js (18 checks: elite/avg/bad/terrible per scale, monotonicity, no floor-drag for good QBs, no max-out for average QBs, plus a regression check showing the old params maxed a 135-efficiency QB at 1.25). Untouched: everything else, model 1.6.1.

TENNIS SURFACE IS A FIRST-CLASS INPUT. Drew played college tennis at Eastern Illinois, and Two things were wrong. (1) DETECTION: the surface came from fuzzy venue-city matching against TennisMyLife tournament names (e.g. ESPN's "All England Club" or "London" never matching "Wimbledon"), and the tournament name ESPN already returns was ignored entirely — so surface detection silently failed and everything downstream of it (surface hold tape, surface win%, modifiers) went neutral. Now the tournament name is checked first: the Slams have FIXED surfaces (Australian Open = Hard, Roland-Garros/French Open = Clay, Wimbledon = Grass, US Open = Hard), with venue fuzzy matching only as a last resort. ESPN's tennis scoreboard was checked live and exposes no surface field, so there is nothing to read from ESPN. Unknown = neutral, never a guessed surface. (2) WEIGHT: surface was a tiny nudge (clay -0.03 / grass +0.025, and only when there was no surface hold tape at all). Now surface sits alongside hold-serve % as the two dominant inputs. Exact magnitudes: when real surface-specific hold tape exists (12+ service games on that surface from TennisMyLife), the tape IS the surface effect — no modifier on top, no double-counting. Otherwise the surface moves hold probability directly: grass +0.04, clay -0.04, hard 0 (roughly the real ATP/WTA surface differential off the hard baseline, symmetric), PLUS the player's own surface W-L edge at 0.15 weight (needs 5+ surface matches from TennisMyLife, edge delta capped at +/-25pp — real observed data, never fabricated). Total surface contribution clamped to +/-0.06 so it modulates the pick instead of single-handedly deciding it; overall hold stays clamped to [0.50, 0.94]. The trace now names how the surface was detected (tournament name vs venue match). Covered by tests/tennis-surface.test.js (24 checks). Untouched: 0.55 shrinkage (0.72 when both players have hold tape), down-only band calibration, 50/50 rule + caps, Grand Slam best-of-5 logic, tennis weather wiring, model 1.6.1.

MLB WEATHER IN THE SIM. The Stadium Intelligence weather machinery was gated to football only — MLB got the display line but zero run-scoring effect (the audit flagged Wrigley wind as the example). Now MLB games feed real Open-Meteo temp/wind/precip into the simulation through a new baseballFactor lambda input: heat 90F+ = 1.02x expected runs both sides (thinner air, more carry), 80s = 1.01x, cold 40F or below = 0.99x, 25mph+ wind = 1.015x, 15mph+ = 1.005x, 70%+ rain = 0.99x — symmetric on both teams, clamped to 0.97-1.03 total so most games move 0-1%. Deliberately NOT the football wind math (a passing/kicking suppression would be wrong for baseball): windMph/precipPct are never set for MLB. No park-orientation data exists in the stadium table, so wind stays direction-neutral (temperature does the heavy lifting) rather than faked. Indoor venues and missing forecasts skip gracefully to neutral. Tennis already had weather wired at the hold-probability level (wind/heat/humidity move the hold number) — verified, left untouched. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, model 1.6.1.

NHL CONFIRMED STARTING GOALIE. The goalie is the key player in hockey, but the sim was crediting the season usage-leader's save% — if the backup starts, the wrong goalie was modeled; if the usage-leader was out, neither was. ESPN publishes probableStartingGoalie (with a Confirmed status) on NHL scoreboard competitors — the same probables shape MLB uses for probable pitchers, verified live. New getNHLProbableGoalies() identifies the game matchup-first (same selection logic as the MLB pitcher fix: eventDate is only a tiebreaker) and the confirmed starter's real save% now drives the goalie factor ahead of the usage-leader. Same opponent-direction math and 0.8-1.25 clamp as before — only the selection got smarter, not bigger. When ESPN lists no probable goalie (or no A-vs-B game is found), the usage-leader path still runs but the trace says so explicitly instead of silently modeling the wrong goalie. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, model 1.6.1.

TENNIS GRAND SLAM FORMAT FIX. Men's Slams were simulating best-of-3: the slamHint detector was built from the surface string ("Hard"/"Clay"/"Grass"), which can never match a Slam name, so the best-of-5 branch could never fire — systematically overstating underdogs at the four biggest events of the year. Slam detection now reads the REAL tournament name from ESPN's scoreboard (Australian Open, Roland-Garros/French Open, Wimbledon, US Open, variants handled): ATP at those events simulates best-of-5, ATP everywhere else best-of-3, WTA best-of-3 everywhere (women's Slams are best-of-3), ITF best-of-3. Unknown tournament defaults to best-of-3, never a guessed best-of-5. The trace now prints the format line with the tournament name. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, tennis hold/weather wiring, model 1.6.1.

SAM 1.6.1. Model version bumped from SAM 1.6 to SAM 1.6.1: the label written to new Airtable pick logs (Model Version), the chat display label, and the test ping all read 1.6.1; historical changelog entries and graded rows keep their 1.6 labels. Rolls up tonight's data-integrity fixes: home-field advantage now enters MLB/NHL/soccer sims (+0.15 runs / +0.25 goals / +0.3 goals in expected-goals math), MLB probable-pitcher lookup identifies the game by matchup instead of trusting the model's date (wrong dates no longer pull the wrong starter; explicit uncertain result with 0% starter weight and a model-facing WARNING when no A-vs-B game is found), real great-circle travel distance tiers (159-venue table; long 0.97x / short 1.0x branches now fire), soccer star-power factor from football-data.org top scorers, and removal of the dead injurySeverity parameter (severity was already encoded per-player in the injury haircut — wiring it would have double-counted). Known exception: strength of schedule (sosDiff) is still neutral everywhere — honest SOS needs full-season schedules plus every opponent's win%, new fetching plumbing that is flagged for a follow-up build, not hacked in. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, NHL run-all exclusion, grading, branding.

SOCCER STAR-POWER FACTOR. Soccer was the only team league with no key-player input at all. New getSoccerStarFactors() pulls each competition's top scorers from football-data.org (same source already used for soccer scoring baselines, limit 50 so every club's top scorer is covered), takes each side's top scorer goals/game (minimum 5 appearances), and converts it to a "self" lambda multiplier analogous to the NBA leading-scorer PPG factor: league-average 0.35 goals/game is neutral, an elite ~0.75 goals/game striker lands near +20%, clamped to 0.8-1.25 like every other key-player factor. Applied per side through the same pitcherFactor slot the other leagues' key-player factors use. The model-facing trace names the scorer with goals and appearances. Fail-open: no scorer data means no adjustment, never a penalty. Also removed the dead if (key === "SOCCER") line inside getKeyPlayerFactor (unreachable since soccer never entered that function). Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, model 1.6.

REAL TRAVEL DISTANCE TIERS. The long (0.97x) and short (1.0x) travel branches in calculateLambdaMultiplier could never fire: getHomeAwayContext only ever returned "home" or "medium", so every road team got the same 0.99x whether it bussed across town or flew coast-to-coast. Now the away side's tier comes from the real great-circle distance between the two teams' home stadiums: under 200 miles = short (1.0x, e.g. Yankees at Mets, Jets at Giants sharing a stadium = 0 miles), over 1500 miles = long (0.97x, true coast-to-coast), everything else stays medium (0.99x). Built a verified 159-venue coordinate table covering NFL (32), NBA (30), WNBA (15 incl. 2026 expansion Portland Fire + Toronto Tempo), MLB (30, Athletics at Sutter Health Park Sacramento), NHL (32, Utah Mammoth), and EPL (20, current 2026-27 lineup) — 18 venues spot-checked against a second geocode source, all matched. Lookup reads the ESPN event's own team abbreviations with alias + tolerant matching (GSW/GS, CONN/CON, ATH/SAC) so variant spellings still resolve; anything unresolved falls back to the old home/medium behavior, never an error. CFB/CBB have no table yet (hundreds of venues — needs a bigger verified build or runtime ESPN team-venue lookups) and keep the previous behavior; flagged as a follow-up. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, model 1.6.

DEAD MULTIPLIER INPUTS RESOLVED. (1) injurySeverity REMOVED, not wired. calculateLambdaMultiplier accepted an injurySeverity input (OUT = 0.94x, Questionable = 0.97x) that no caller ever set. Checked for double-counting before wiring it: scoreInjuryReport already encodes severity per player (OUT = 1.0x, Doubtful = 0.7x, Questionable/day-to-day = 0.4x of position weight, capped at 0.82 floor). A flat most-severe-status multiplier on top would have counted the same signal twice, so the dead parameter and its branches were deleted instead of wired — the richer per-player encoding is the severity mechanism, now documented in a code comment. No live pick changes: the parameter was never populated. (2) sosDiff LEFT UNWIRED, flagged for follow-up. Strength of schedule is still 0 (neutral) everywhere: honest SOS needs full-season schedules plus every opponent's win%, which is new fetching plumbing the pipeline doesn't have (the 17-day rest/form window only covers recent games — a "recent SOS" from two weeks of data would be fake precision, so it was not hacked in). The sosDiff input stays in calculateLambdaMultiplier ready for that build. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, model 1.6.

MLB PROBABLE-PITCHER MATCHUP-FIRST LOOKUP. getMLBProbablePitchers picked its game by matching the model-passed eventDate against an 11-day scoreboard window — a wrong or missing date silently pulled the wrong game's starter. Now the game is identified by the MATCHUP: candidates are already filtered to games where the two teams play each other, and the pick is the date-matched game when eventDate hits (doubleheader-safe: soonest upcoming game that day), otherwise the nearest upcoming A-vs-B game in the window. A wrong model date no longer changes the answer. The uncertain fallback now fires only when no game between the two teams exists in the window at all — in that case the function returns an explicit uncertain result (instead of null) so the model gets a WARNING line and starter weight goes to 0 rather than silently running pitcher-neutral. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, model 1.6.

NOTIFICATION TAP DOUBLE-SEND FIXED (CROSS-TAB). Tapping "prediction ready" could render the reply twice: the original chat tab was still polling and the notification opened a second view, and the render guards were per-tab memory only. Seen/started flags are now also in localStorage so the second tab skips a job the first tab already rendered.

DUPE CHECK NOW KV-ONLY. The "is this game already logged" lookup was falling through to a slow Airtable call. Now it checks the KV replica only — instant. The replica is kept in sync on every log and grade.

PREDICTION WRITEUPS UPGRADED TO PROFESSIONAL ANALYST GRADE. Every prediction now reads like a professional analyst of that specific sport broke the game down: matchup setup, key numbers with named players, a substantive professional breakdown walking through the real edges (matchups, injuries, rest, form, weather, pitching, home/away), then the win-percentage split and definitive winner. No more short blurbs that hide the analysis.

HOME-FIELD + PITCHER DATA FIXES. (1) HOME-FIELD NOW ENTERS MLB/NHL/SOCCER SIMS. The sim told the model home advantage was "factored directly into this simulation" for these leagues, but no home-field number ever reached the trials — only NFL/CFB/NBA/WNBA/CBB got one. Now the home team gets a real edge in the expected-goals/runs math: MLB +0.15 runs, NHL +0.25 goals, soccer +0.3 goals (conservative, grounded in measured per-game home scoring edges). (2) MLB PROBABLE-PITCHER WRONG-GAME FALLBACK. When ESPN listed several games in a series and none matched the requested date, the sim silently used the first game's starter and the model got no warning. Now the pitcher data is flagged uncertain: a WARNING line goes to the model, and starter lock weight is cut in half so a possibly-wrong pitcher can't drive the pick — the blend shifts toward bullpen/staff instead. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, NHL run-all exclusion, grading, branding, model 1.6.

LEAGUE BIAS REMOVED. The per-league calibration bias is gone entirely: no more Supabase bias fetch/adjust on the sim text the model sees (runMatchupSimulation), the grader no longer trains it (updateCalibrationBias call removed), and the /admin/backfill-calibration endpoint is removed. Deleted dead code: applyCalibrationToResultText, getCalibrationBias, applyCalibrationBias, updateCalibrationBias, backfillCalibrationFromAirtable, CALIBRATION_LEARNING_RATE/MAX_BIAS, and tests/calibration-picked-side.test.js. The sim's numbers now reach the model unadjusted by league history. Untouched: 0.55 shrinkage, down-only band calibration, 50/50 rule + caps, NHL run-all exclusion, grading, branding, model 1.6.

CONSERVATISM SHRINKAGE SET TO 0.55 (0.50 → 0.55). The sim's raw edges were being flattened into coin flips: post-shrinkage sims average ~63% picked-side with only 10% landing 48-52, but published confidence lands at median 49% with 53% inside 48-52 — the band calibration fits on its own deflated outputs and compounds the pull-down. Easing shrinkage keeps more of the sim's lean (a raw 70% edge reports ~61% instead of 60%, a raw 80% edge ~67% instead of 65%) while all deflationary mechanisms stay in place: down-only band calibration, 50/50 rule + caps, NHL run-all exclusion, grading, branding, model 1.6.

DATA PIPELINE GAPS CLOSED. Deep audit found the model was missing data it should have had: (1) Rest days + recent form now factor into MLB, NHL, and Soccer sims — previously only NFL/NBA/WNBA/CFB/CBB got them, leaving NHL back-to-backs and MLB hot/cold streaks unadjusted. (2) CBB injury weights fixed — was silently using NFL position weights for college basketball; now uses proper basketball weights. (3) MLS/EPL/Premier League queries now normalize to SOCCER instead of silently degrading with NFL injury weights and no rest/form data.

SLATE TIMEOUT RAISED TO 5 MINUTES. The Run Today slate runner was aborting prediction requests at 120 seconds, causing "signal timed out" failures on slow games. Bumped to 300 seconds so slow signals can finish instead of failing.

NHL BACK IN RUN ALL. NHL is restored to the daily slate leagues after the 2026-09-20 removal. Single NHL predictions were never affected.

MARKET COMPARISON RESTORED, STAKING STAYS OUT. The model-vs-market commentary is back on every prediction: "SAM sees +X% edge over the market" / "SAM is X% below the market implied probability" / "SAM and the market are roughly aligned" now render on the Market Line block for new predictions and logged-pick displays. The edge-gate verdicts (BET/HOLD/PASS) and all stake sizing stay deleted — this is market-agreement numbers only, no imaginary-money recommendations.

DEFLATIONARY CALIBRATION RESTORED. The 2026-09-21 calibration redesign is reverted: (1) fitBandCalibration buckets on the LOGGED band-calibrated "Confidence %" again instead of the raw sim picked-side % — bands learn on their own past outputs; (2) the symmetric +/-10pt clamp is replaced by the down-only clamp (published confidence may only shrink toward the band's realized hit rate, never inflate); (3) the online league-bias learner trains on the logged confidence again instead of the raw sim split. Buckets are back to 50-59/60-69/70-79/80-89/90%+. Untouched: 0.5 shrinkage, 50/50 rule + caps, NHL run-all exclusion, the picked-side bias fix, grading, branding, model 1.6.

REFUSED PREDICTIONS NO LONGER EAT A FREE PREDICTION. A live test showed a guest asking for a matchup that isn't on the board ("Tigers vs Mariners" — the model correctly refused instead of guessing) still burned one of the 3 daily free predictions: the counter dropped from 3 to 2 with no pick delivered. The guest cookie was being incremented on every /api/chat reply, even when no pick came back. Now the incremented cookie is only set — and the remaining count only drops — when a real pick is produced. Logged-in users were already protected (their deduction only writes when a pick exists); guests are now covered too. Same fix applied to the tennis coin-flip decline path, which had the identical leak, and /api/chat now fails closed if the billing backend key ever goes missing instead of serving unmetered predictions.

MODEL SWITCHER REMOVED ENTIRELY. The "ParlayKings 1.6" tag by the chat input is gone completely — no button, no fold-out, no label. One model only (SAM 1.6 on Muse Spark 1.3), so there was nothing to switch or explain. The picker markup, all its styles, and its toggle script are deleted; the chat input now sits alone. Nothing else touched.

MODEL TAG FOLDS OUT AGAIN. Tapping the "ParlayKings 1.6" tag by the chat input now opens the info card again, showing the model name and its explanation line ("Sports Analysis Machine (SAM) 1.6 original model still most accurate."). Still one model only — SAM 1.6 on Muse Spark 1.3, nothing to switch between; the card is informational. Tapping anywhere else closes it.

MODEL LABEL NO LONGER LOOKS LIKE A BUTTON. The chat input's "ParlayKings 1.6" tag was still styled like the old dropdown (border, pill shape) so it looked tappable but did nothing when tapped. It's now plain text. Single-model setup is untouched: SAM 1.6 on Muse Spark 1.3, no picker, nothing to choose.

GAME-CARD FOLLOW-UPS REMOVED. The post-simulation gameCard feature from earlier tonight is fully removed: the five card functions, the /api/chat card wiring (request parsing, cached-response property, prompt grounding, main response property), the background-job card property, and all client-side card state (hold/send/receive/reset). Chat behavior is back to exactly what it was before the feature existed. Nothing else touched.

EDGE GATE REMOVED FROM OUTPUT. The edge-gate verdict block is gone from every prediction reply: no more BET/HOLD/PASS-by-edge lines, no edge-points numbers, no model "fair price", no stake sizing, and none of the "SAM sees +X% edge / X% below the market / roughly aligned" commentary. The Market Line block stays as plain market data (book, both sides' prices, implied %). The 50/50 note is unchanged. The gate's code (edgeGate/edgeGateVerdict, Kelly sizing, odds converters, thresholds) had zero remaining consumers after the rendering was cut, so it was deleted outright; the postmortem faded_market tag still works because it computes from the logged confidence vs the market-line implied % numerically. Regression test replaced: tests/edge-gate-removed.test.js (was edge-gate-demotion.test.js). Untouched: prediction flow, grading, calibration, 50/50 rule, 0.5 shrinkage, NHL run-all exclusion, alerts, recap, branding, model 1.6, /api/speak.

GAME-PRO FOLLOW-UP WORDING FIX. The follow-up grounding rule told the model to say the canned line "my sim notes don't cover that" when a question goes beyond the game card. Drew flagged it as robotic — replaced with an instruction to say so in its own words (e.g. that wasn't part of this simulation). No behavior change: still answers from the card, still never invents sim details. Nothing else touched.

POST-SIM "GAME PRO" FOLLOW-UPS (Drew's request). After a prediction, /api/chat now returns a compact gameCard built from the real sim outputs (matchup, league, both teams' sim percentages, picked winner, calibrated confidence, the sim's key-factors line, and the Market Line block — under 1.5KB). The client holds it per-conversation and sends it back with the next message; a new chat or opened thread starts clean so cards never leak across conversations. If the next message is a follow-up about that game, the card is injected as structured grounding with an anti-freelancing rule: answer FROM the card, say so in their own words when the card doesn't cover the question, never invent sim details — and if the message names a different matchup, run a fresh simulation for it. Follow-up vs new-game detection is deterministic: a message naming teams that match neither card team drops the card; everything else keeps it. Client-supplied cards are strictly sanitized (allowlisted fields, length caps, newlines collapsed) before they can reach the prompt. Background-job predictions also arm the card. Additive only: no card present means byte-identical behavior; prediction, grading, calibration, 50/50 rule, edge gate, shrinkage, staking, NHL exclusion, and all prompt rules untouched. Regression tests: tests/game-card.test.js (37/37).

BET VERDICTS DEMOTED TO HOLD-ONLY. The 10,000-run Monte Carlo staking backtest found the edge gate inverted: picks it classified BET (edge of 3+ pts) won only 39.1%, and half-Kelly sizing on them ruined the bankroll in 99% of simulated worlds — the gate bets exactly where SAM is overconfident vs the market. Per Drew's "only if it wins" rule no staking parameter was changed; instead the published verdict no longer prints a BET signal or any stake sizing. A BET-classified pick now reads as HOLD with the edge info preserved ("+X pts vs market … no stake recommended while the gate is under review"). Internal gate classification is kept for the eventual rework; HOLD/PASS verdicts unchanged. Regression test added (tests/edge-gate-demotion.test.js). Untouched: 0.5 shrinkage, 50/50 rule, NHL run-all exclusion, staking math itself, branding, model 1.6.

CALIBRATION REDESIGN: BANDS NOW FIT ON RAW SIM VALUES + SYMMETRIC CLAMP (Drew: "do whatever you think best"). The 2026-09-21 audit found the band calibration deflationary: 62% of picks logged below 50% confidence while that group hit 55%+, Brier 0.2434 vs 0.2444 for a constant base-rate predictor. Root causes fixed: (1) fitBandCalibration now buckets on the RAW band-stage input — the bias-adjusted sim picked-side % (max of logged "Team A %" / "Team B %") — instead of the already band-calibrated logged "Confidence %", which had created a feedback loop (bands fit on deflated values, then pulled new picks further down). New bands match the sim's honest structure: 50-60 / 60-70 / 70-80 / 80%+. (2) The down-only clamp is replaced by a symmetric +/-10pt clamp (BAND_CALIBRATION_MAX_SHIFT_PTS) inside applyBandCalibration, so legitimate pull-ups are possible again; exact-50 is never touched (the 50/50 rule owns it). (3) The online league-bias learner (updateCalibrationBias) now trains on the same raw picked-side % via new biasLearnerPredictedPct(), instead of the deflated logged confidence — it was learning on one scale and correcting another. Backtest (386 decided SAM 1.6 picks, chronological walk-forward, 21 re-graded excluded): Brier 0.2434 -> 0.2393, share logged below 50%: 62% -> 8%, new logged calibration is monotonic (50-55%: 52% hit, 70-75%: 79% hit), and the 70-80% raw sim segment is preserved (85 picks, mean logged 74.6%, 75.3% hit). Supabase sam_calibration values left to re-learn (not reset): the 21 re-graded picks' total contamination was only -10.5pts spread across 6 leagues (worst -4.5 NFL/WTA), the historical values are unreachable from here (worker-side secret), and the online updater self-corrects within a few graded picks per league now that inputs are honest. STAKING BACKTEST (Drew: "if it wins more bets... i am happy" — data-driven only): the edge-gate (BET at SAM-minus-market edge >= 3) is INVERTED on 243 decided picks with DraftKings lines — BET picks went 39.1% (69 BETs at 36.2% on old conf), and 10,000 bootstrap Monte Carlo runs show half-Kelly on gate BETs ruins the bankroll in 99.2% of worlds (median terminal 0.6u from 100u); quarter-Kelly, flat stakes, and a tighter +5 gate all lose too; even the diagnostic inverted gate (bet edge<0) only reaches median 97.9u. NO staking change shipped — nothing wins robustly, so per Drew's condition the staking layer is untouched. FLAG FOR DREW: the published "BET ... half-Kelly" verdicts are -EV as designed; recommend deciding whether to keep, demote, or rework the edge gate. Untouched by design: 0.5 shrinkage, 50/50 rule + caps, NHL run-all exclusion, staking params, branding, model number.

NHL MADE A FIRST-CLASS LEAGUE IN THE DATA PIPELINE. Airtable already carried "NHL" as a League select option and the worker was already wired for it (ESPN paths, sim distributions, grading margins, alert prefs) — the only gap was the system prompt's [PICK] rule-6 league list, which never told the model NHL was a valid League value, so NHL predictions would have been logged as "Other" and fragmented the history. Added NHL to that list. The Bets-replica chain (Airtable → KV → bets-replica.json → Google Sheet "Bets" tab) is league-agnostic with no per-league tabs or validation, so NHL rows flow cleanly; verified 19 league values already coexist there. The 2026-09-20 run-all exclusion is untouched: the slate loop still skips NHL unless Drew says otherwise.

CALIBRATION BIAS NOW APPLIES TO THE PICKED SIDE + 21 HISTORICAL 50/50 PICKS RE-GRADED TO PUSH. (1) Logic-bug fix in applyCalibrationToResultText: the Supabase league bias is computed from picked-side outcomes (stated confidence of the pick vs. actual), but it was being added to teamA's percentage unconditionally — teamA/teamB are just tool-arg order. When the model picked teamB in an overconfident league (NFL bias ≈ −12.6), e.g. teamA 35%/teamB 65% printed as teamA 22%/teamB 78% — the printed confidence got inflated instead of corrected, and feedback from inflated picks drove the bias further negative. The bias is now applied to the presumptive pick's percentage (higher sim side; teamA on an exact tie, matching the [PICK] rule) with the other side mirrored as 100−adjusted. ±15 clamp, 1–99 range, and the calibration note are unchanged. No Supabase calibration values were modified — this is a presentation/prompt-text correction, not a calibration value change. (2) Drew approved re-grading the 21 historical SAM 1.6 picks logged at exactly 50% confidence that the auto-grader had scored W/L (7–14) instead of Push, violating the standing 50/50 rule — each was verified live in Airtable before its Result was set to Push; nothing else on those rows was touched. Flag: those 21 picks previously fed the calibration updater as W/L, so their historical influence on the running Supabase bias values can't be cleanly unwound — the bias values keep that history by design. Untouched by design: 0.5 shrinkage, calibration math/values, confidence/staking/thresholds, risk settings, the SAM XAI 2.1 history classifiers, and /api/speak.

GRADER 50/50-RULE ENFORCEMENT + TENNIS/WTA LABEL STANDARDIZATION. (1) The auto-grader now enforces Drew's 50/50 rule on the confidence number itself: any pick logged at exactly 50% confidence grades as Push, even when the model named a side (previously only "50/50"/"no pick" winner labels were caught, so 21 historical conf-50 picks were auto-graded W/L). New isConfidenceFiftyFifty() helper; the graded Push is also excluded from calibration updates and tagged in Notes. Going-forward only — the 21 historical rows were left untouched per Drew. (2) Women's tennis is now a single label everywhere going forward: the [PICK] tag format and both tool league enums now state Tennis = men's ATP only and WTA = ALL women's matches (the model knew — e.g. calling Potapova/Andreeva "a WTA tour-level matchup" — but logged League=Tennis, fragmenting per-league calibration). Historical Airtable rows were not rewritten. Untouched by design: 0.5 shrinkage, all calibration math/values, confidence/staking/thresholds, risk settings, the SAM XAI 2.1 history classifiers, and /api/speak.

SPARK 1.3 SINGLE-MODEL CLEANUP. Drew's call after the Grok removal: with SAM 1.6 (Muse Spark 1.3) as the only model, strip everything that only existed for the old multi-model days. (1) Onboarding copy: the Workflow explainer still said “ParlayKings 1.6 or ParlayKings XAI 2.1, whichever you picked” — now just ParlayKings 1.6. (2) Deleted the entire dead Gemini LLM stack: callGemini/requestGemini, GEMINI_MODEL, TOOLS_GEMINI, geminiApiKeys (zero callers each, verified) plus dead provider===“gemini” branches in extractReplyText, fetchCacheCandidateRecords, fetchPendingSlateRecords, and fetchPredictionHistory; runMatchupSimulation's default provider is now “meta”. (3) buildDailySlate no longer fetches today's logged records twice — one KV-replica read; the alreadyLoggedGemini admin stat now aliases the meta count. (4) isPublicReplicaProvider() is now a constant true; the chat page's model-picker dropdown (one option) became a static “ParlayKings 1.6” label and the picker JS was removed. (5) fetchPredictionHistory's up-to-10-page Airtable scan is now cached 10 minutes in the SAM_BETS KV namespace (provider-scoped key); all downstream computation and postmortem side effects still run per message. (6) New single-model test asserts one provider target, no picker options, and no dead model stacks. Untouched by design: 0.5 shrinkage, calibration math, confidence/staking, the 50/50 rule, the NHL slate exclusion, the SAM XAI 2.1 history classifiers, and /api/speak (xAI orion voice).

SAM XAI 2.1 (GROK) REMOVED — SAM 1.6 IS NOW THE ONLY MODEL. Drew's call: the Grok picker option was barely used (one logged pick in the week before removal) and confused people more than it helped. Removed end to end: the "ParlayKings XAI 2.1" model-picker option, Grok from the auto-grade loop (gradeAllModelBets now only grades SAM 1.6 rows), the Meta/Grok cross-provider failover in /api/chat, the "get a second opinion" button (it only existed to ask the other model), and all xAI chat-model code (callGrok/requestGrok, GROK_MODEL, GROK_URL, the Grok key rotation + quota-skip machinery, allGrokKeysDown). The xAI text-to-speech endpoint (/api/speak, voice "orion") is untouched — it still reads GROK_API_KEY for voice only. Historical Airtable rows labeled "SAM XAI 2.1" are left alone; the Meta label matcher still excludes them so they never leak into SAM 1.6 track records or calibration.

NHL OUT OF RUN ALL. The Run All board run no longer fetches or runs NHL games — hockey is off the slate until Drew says otherwise. Single predictions can still pick NHL explicitly; this only stops the automatic whole-board run from touching it.

BOARD-RUN ALERTS GO TO EVERYONE, SCATTERED THROUGH THE DAY. When the admin runs the whole board, high-edge pick alerts now fan out to every push-subscribed user — each person only gets the picks matching their own leagues and confidence floor, with per-user dedupe so nobody's delivery blocks anyone else's. Previously the alert only went to whoever ran the prediction. To avoid one big burst, picks for games happening today go out immediately while future-game picks drip out at most one every 30 minutes per user (quiet hours still respected, same 12-a-day cap). Covered by tests/he-fanout.test.js.

GAME-DAY TEAM ALERTS. ParlayKings now pings you for your teams: Michigan football, Michigan State football, the Lions, and the Tigers — game start, halftime (football), and final score, straight from ESPN's scoreboard. Pick your teams in Profile under "My teams" (all off by default); team alerts go out live even during quiet hours, with 24-hour dedupe and the same 12-a-day cap as everything else. The worker checks on its every-minute cron but only calls ESPN when somebody actually has a team on, and only on days that team plays. Covered by tests/game-alerts.test.js.

RECAP STOPS COUNTING EVERYTHING TWICE. Both provider targets resolve to the same Airtable base, so the nightly recap fetched the same table twice and doubled every W-L line (MLB showed 14-4 for a real 7-2). fetchGradedRecords now dedupes by Airtable record ID — counts are identical whether one target or many resolve to the base. No other alert logic, thresholds, or timing touched. Covered by tests/alerts.test.js (59/59).

NAV DRAWER TRIMMED, GRIMALDI.TV UNDER TRIVIA. The hamburger menu carried a full podcast menu (live broadcast, photo gallery, blog, merch, clip vault) on top of every sports link. The podcast stack is back to HOME + EPISODES only, in both drawer copies (shared builder and main chat page), and the GRIMALDI.TV link sits under TRIVIA on every host instead of only parlaykings.net. Covered by tests/nav-links.test.js (29/29).

NOTIFICATIONS SAY WHAT THEY ARE. Every push ticket woke the phone with a fresh look at the latest finished job, so an alert ticket arriving later re-announced "Your prediction is ready" for a job long done — burying the real alert banners. The service worker now only announces jobs finished within the last 15 minutes, and when there is no fresh job and no pending alerts it shows nothing instead of a stale generic. Covered by tests/alerts.test.js wiring checks (53/53).

TAP A NOTIFICATION ONCE, SEE THE ANSWER ONCE. Tapping "prediction ready" could type the reply twice a few seconds later: the startup catch-up and the tap path raced each other (and a tap can arrive twice — service-worker message plus ?job= URL), and nothing marked a render as started until it finished, so an in-flight render never blocked a second one. Every render path is now single-flight (startedBgJobs is marked synchronously the moment a render begins; the startup sweep skips the tapped job; repeat taps are no-ops while failed taps stay retryable). Tapping while the run is still going also adopts the same 10-minute watch window as a fresh background send instead of risking eternal dots. Covered by tests/bg-resume.test.js (18/18) plus a bg-poll.test.js harness sync.

PICK ALERTS, NIGHTLY RECAPS, LINE-MOVE INTEL — ONE DISPATCHER. Every alert now flows through a single choke point (dispatchAlert): per-user prefs, quiet hours, and rate caps are enforced before any push goes out. Three alert types: high-edge pick alerts fire the moment SAM logs a pick above your minimum confidence; a grading recap lands nightly at 9 PM ET; line-movement intel fires when a logged pick's line moves 25+ cents (or 15+ cents within an hour), with a 6-hour cooldown per pick so it can't spam. Alerts are off until you enable them in settings — per-league toggles, 50-100 confidence floor, and quiet hours included. Capped at 12 alerts a day with 24-hour dedupe. Covered by tests/alerts.test.js (51/51).

PROFILE PHOTO SAVES WITH SAVE. Picking a photo used to strand you on a second Upload button nobody could find: preview, then Save does everything — photo goes up first, settings after, footer avatar and profile picture refresh together. Uploads also survive the storage lag that used to say "saved" while showing nothing (background retries plus self-reloading images), and iPhone HEIC shots get a plain-English nudge instead of a dead error. Covered by tests/profile.test.js (81/81).

SPORTSBOOK TAB TAPS FIRST TIME, EVERY TIME. The edge tab slid sideways on hover, so on touch screens the button moved out from under your finger and the tap landed behind it — flash, then nothing. The tab never moves now; it also carries an invisible wider tap halo and sits above every page layer. A well-meant 350ms close-blocker that ate quick follow-up taps is gone too. Covered by tests/sportsbooks.test.js (42/42).

NOTIFICATION TAPS LAND ON THE ANSWER, NOT THE DOTS. Tapping "prediction ready" could just foreground the frozen page you left — still showing the old thinking dots, with nothing re-running — until you clicked out and back in. The tap now wakes the open page and loads that prediction inline (adopts its chat, types the reply, clears the dots); only browsers that can't steer the open window fall back to a fresh load. Covered by tests/bg-resume.test.js (14/14).

USER PROFILES: NOTIFY PREFS, SETTINGS, AVATARS. The KV profile record (built for pins) is now a full user profile: notification prefs (master toggle, per-league toggles, 50-100 minimum confidence, quiet hours) plus settings with display name and avatar. New profile/settings UI in the chat drawer: prefs form plus picture upload with client-side 512px downscale. Avatars target R2 (bucket env.AVATARS) with MIME magic-byte check, 2 MB cap, and server-side dimension parsing for JPEG/PNG/GIF/WebP - and the whole flow degrades gracefully until R2 is enabled (upload control hidden, everything else works). To enable: create bucket parlaykings-avatars and add [[r2_buckets]] binding="AVATARS" to wrangler.jsonc. Covered by tests/profile.test.js (55/55). No notification sending logic - prefs only.

PIN CHATS AND PICKS. Chat history now supports pinning entire chats and individual games/picks, freely combined. Pins are per-user and persistent in a KV profile record (profile:) whose schema reserves notify/settings sections for later features - no notification logic in this change. Turn picks are persisted at save time so pinned games render without re-parsing text; pinned chats get their own history section with pin/unpin toggles, and each saved pick row carries a pin button. Deleting a chat prunes its pins. Caps: 100 chat pins, 200 pick pins. Covered by tests/pins.test.js (49/49).

TENNIS COIN FLIPS STAND, CAPPED AT 3 A DAY. Genuinely even tennis (model calls 50/50, or sim split within 2 points) now stands as an honest 50/50 — logged, grades Push — instead of being forced onto a side. Daily KV counter (ET day, 48h TTL) caps stands at TENNIS_FIFTYFIFTY_DAILY_CAP; past the cap the request is declined as too-close-to-call with nothing logged and no credit deducted. Other leagues untouched; counter failures fail open to standing. Covered by tests/tennis-fiftyfifty-cap.test.js.

BAYESIAN BAND CALIBRATION BACK ON. The per-band fit/apply helpers stayed in the bundle computing params, but the call-site block had been replaced with void finalCalibrationParams, so the correction was computed and thrown away. Restored after sim lock in the /api/chat handler (primary and failover paths share it): fitted bands may pull the sim-locked number DOWN toward the band's historical hit rate, never above the sim; the visible reply text is rewritten to match the logged value. Thin bands stay near raw via shrinkage (prior weight 10); the fit needs 20+ graded picks or it no-ops. Covered by tests/band-calibration.test.js (12/12: overconfident 60-69 band pulls 64 to 53, thin 80-89 band stays near raw); the sim-lock structural guard now allows the sanctioned down-only block.

SAVED CHATS KEEP THEIR BUTTONS. Opening a conversation from history rendered plain bubbles: no thumbs, no share, no second opinion. The restore path now runs every revived reply through the same action row as live messages (copy/up/down/share/speak/second-opinion; provider falls back when the turn predates it). One-line fix in openConvo. Covered by tests/saved-chat-actions.test.js (9/9).

SELF-SERVE PASSWORD RESET. Forgot-password link on the login page; emailed single-use reset links (1-hour expiry) via Email Sending from noreply@parlaykings.net; new-password page with min-8 + confirm. Same message whether or not the account exists (no enumeration), 3-per-hour throttle per identifier, tokens burned on use. Reuses PBKDF2 hashing + Supabase PATCH path. Covered by tests/password-reset.test.js (9/9).

POSTMORTEMS SAY WHY, NOT JUST "OOPS". The grader now reads the final score: league-aware margin bands (NFL/CFB 8/17, MLB 2/6, NBA/CBB 6/15, NHL 1, soccer 1/3) tag blowout losses rout (sim misevaluated strength) and within-one-score losses one_score (the short end of variance, process intact). Fading the market 5%+ and losing now also tags market_called_it (the market was right to be colder). coinflip_miss/favorite_miss demoted to fallback-only when no causal tag fires; high_confidence_miss untouched so the conservatism loop keeps its input. Autopsy lines state final margin plus a plain verdict. Covered by tests/postmortem-upgrade.test.js (54/54).

POSTMORTEMS CHECK THEIR OWN HOMEWORK. Two upgrades. (1) CALIBRATION OVERLAP GUARD: the sim-side conservatism pull (0.9 toward 50 after repeated high-confidence misses) and the per-band calibration correction could both fire on the same phenomenon and double-count. fetchPredictionHistory now measures both side by side: when global 70%+ bands are already correcting down 5+ points, the sim pull softens to 0.95 and says so in its note instead of stacking the full pull on top. (2) EFFICACY TRACKING WITH AUTO-RETIRE: every adjustment activation is recorded per league (KV-backed) with its tag baseline; after 8+ further losses the tag rate is re-measured — halved means working (baseline refreshed), flat means retired as ineffective, and high bands over-hitting while the pull is on means retired as overcorrecting. Retirements hold 30 days with the reason surfaced in the sim note, then the pattern can re-arm. Pattern faded means the record drops. Covered by tests/postmortem-upgrade.test.js (29/29).

TAPPING A NOTIFICATION NO LONGER LANDS ON ETERNAL DOTS. Opening ?job= on a fresh page meant a new chat ID, so the finished job never matched, and a stale storage read showing "pending" meant the page never started watching. The page now adopts the tapped job's chat, renders a finished job straight into it, and keeps polling when the first read is stale; every completed background render also clears the thinking dots. Covered by tests/bg-resume.test.js (11/11).

NOTIFY JOBS START ON SEND. Armed-bell sends used to sit pending until the minute cron noticed them (up to 60 seconds of dead wait): the enqueue handler swept the whole queue instead of running the job in hand, and storage-list lag could hide the fresh job until the next tick. The handler now runs that job object immediately via ctx.waitUntil(runBgJob(...)) and the push still fires the second it finishes. The minute cron keeps grading plus the replica dump and retries only stuck runs (a "running" job older than 6 minutes, down from 15). Bell, push, SW, jobs routes, and client polling untouched.

REASONING BUBBLE NO LONGER VANISHES FROM SAVED CHATS. Opening a chat from History showed only the user/answer bubbles -- the collapsible "ParlayKings Reasoning" bubble was gone every time. Root cause, three links: (1) the client saveConvoTurn posted only {convoId, message, reply}, never the thinking text; (2) /api/convos/save stored only {user, reply, at} per turn; (3) openConvo mapped thinking to '' and rendered only user/ai bubbles. Fix: saveConvoTurn now sends thinking (capped 8KB like replies), the server stores it on the turn, the background-job internal save threads job.result.thinking through too, and openConvo restores the bubble collapsed-by-default in its original spot plus keeps thinking in conversationHistory for follow-ups. Old saved chats without thinking render exactly as before. The bubble itself is unchanged behaviorally (collapsed by default, click the label to expand, chevron toggles) but renderReasoning is now built on DOM APIs instead of an innerHTML template, and whitespace tidying moved into a tiny tidyReasoningText helper with identical display rules -- both covered by tests/reasoning-history.test.js (26/26: collapsed default, toggle, literal-text safety, history restore, save round-trip). No notification, push, jobs, cron, or billing code touched.

NOTIFY-WHEN-DONE BACKGROUND PREDICTIONS. Bell button next to send (gold when armed, choice persisted per browser); armed sends go to POST /api/jobs (sign-in required, 3 pending cap) instead of the blocking /api/chat, and the every-minute cron runs them through the real pipeline via internal replay (same sim, same [PICK], same credit deduction), saves the turn server-side, and fires a payload-less Web Push ticket. New routes /sw.js, /api/push-public-key, /api/push-subscribe, /api/jobs, /api/jobs/mine, /api/job; new secrets VAPID_PUBLIC_KEY, VAPID_PRIVATE_KEY, VAPID_SUBJECT. Open page polls every 15s and renders finished jobs; tapped notifications deep-link /?job=. Follow-ups, same release train: jobs/push routes resolve their own session (fixed 500s on phones), live traffic rolled back to the message-queue build after the first notify builds broke core chat on mobile, bell moved to the History drawer footer with silent background runs, bell-off runs zero notification code (startup catch-up only for browsers that ever used background), and the thinking dots keep watching until an answer actually renders instead of clearing early. Verified live end-to-end (enqueue 200 through cron done in ~70s with reply saved to convo history).

TOOL-CALL TUNING + EDGE GATE + UFC FALLBACK. (1) LEAGUE ENUMS: every TOOLS_OPENAI league field now carries a JSON Schema enum (was prose-only lists the model had to parse), and the schedule/sim/injury lists gained the missing NCAAF/NCAAB values — all enum values verified against normalizeLeague. Legacy Gemini functionDeclarations left untouched. (2) SPARK SAMPLER: Muse Spark requests now send temperature 0.2 for more reliable tool arguments; reasoning_effort stays high. (3) PARALLEL HINT: system prompt rule 9 now says independent calls (schedule plus both sides' injury reports) may go out in one parallel block; schedule -> sim stays serial. (4) ARG-ERROR RETRIES: malformed tool arguments now return a retryable ERROR to the model instead of silently executing with empty {} (both Grok and Meta loops). run_matchup_simulation description intentionally NOT trimmed. (5) EDGE GATE: new deterministic bet/hold/pass layer (edgeGate/edgeGateVerdict + odds converters, pure functions). BET at +3 pts edge or better with a half-Kelly stake; HOLD within \u00b13 with a target price; PASS when the market sits 3+ above SAM or on coin flips. Verdict appended inside the Market Line block on fresh and cached paths so it persists in Notes with no Airtable schema change, worded to never false-fire the faded_market postmortem parser. (6) UFCSTATS VIA ANTIBOT IN FALLBACK: fetchUFCFallbackFighter now tries ufcstats.com first through the existing ZenRows-then-ScrapingAnt chain (was fightingdata/ultimatefightingstats only, env never threaded through), then fightingdata, then ultimatefightingstats — returns null fast when keys are absent so nothing else changes. (7) UFC CRASH FIX: the ESPN fallback path reassigned a const espnResult (TypeError whenever thin tape met a found fallback) — now let.

SIDE NAV IS DARK PURPLE WITH GOLD TYPE. The hamburger drawer background is a dark Parlay Kings purple; MENU, section labels, and links are gold. On parlaykings.net GRIMALDI.TV sits under Trivia instead of at the top of the podcast stack. ai.grimaldi.tv still has HOME in the podcast list.

UFC SIM GETS THE PREDICTIVE TAPE WE ALREADY FETCH. Significant strike differential (SLpM minus SApM) now swings decisions. Career finish mix (KO/sub/decision shares from the ufcstats fight log on the same page) scales KO vs submission chances. Home-country vs event country is a small cage boost when ESPN has both. Weight-class finish bias: heavyweights KO more, fly/straw less. Broadcast records (most wins, fastest KO) stay out of the 25k.

BASKETBALL FOUR FACTORS IN THE SCORE MODEL. NBA/WNBA/CBB now pull the ESPN team sheet (same one-call-per-side as football): eFG%, TOV% from estimated possessions, OREB vs the opponent's DREB, FT rate and FT%, 3P%, pace, AST/TO. Home court is +3 NBA/WNBA and +4 CBB. No PER/BPM/usage, no possession-by-possession rewrite.

FOOTBALL SCORE MODEL GETS THE ESPN SHEET ITEMS WE CAN ACTUALLY USE. Rush YPC and pass YPA (means — ESPN has no play-level stdev), sack % per dropback, INT % per pass, fumbles lost per play (still luck-regressed), FG% by <40 / 40-49 / 50+, XP%, punt net and inside-20, kickoff touchback % and return avg, seconds per play, 4th-down attempts/game. Home NFL side gets +2 expected points (+2.5 CFB). No drive sim, no EPA/DVOA, no player-level.

UFCSTATS.COM IS BACK AS A TAPE ENRICHMENT, VIA ZENROWS THEN SCRAPINGANT. ESPN is still the default UFC path. If SApM, strike defense, or takedown defense are missing, the Worker loads ufcstats.com with ZenRows js_render+antibot, then ScrapingAnt browser=true — not a raw fetch. Those three numbers now move the 25k (chin/defense, stuffing takedowns, KO chance). Challenge pages are rejected and cached fighter pages are reused.

FOOTBALL SIM USES THE ESPN TEAM-STAT SHEET, NOT JUST PPG. Same team-statistics payload now feeds yards/play, pass YPA, rush YPC, third/fourth down, red zone, time of possession, sacks taken vs created, tackles for loss, penalties, FG%, and explosive plays. Early-season samples are shrunk toward league average. Wind from the venue forecast now sets windMph on both sides (the old hook never fired). Combined football factor is capped 0.90-1.10 so scores stay in the real range.

FOOTBALL SCORES STAY IN A REAL RANGE + TURNOVERS IN THE SIM. CFB ESPN standings were feeding season totals (69-45 after two games) as if they were points per game, so the 25k sim printed 99-102 type scores. Football now converts totals to per-game, clamps NFL ~13-38 and CFB ~14-48 expected points, and caps a trial at 55 (NFL) / 70 (CFB). Takeaways and giveaways from ESPN team statistics now move expected scoring (~4.5 points per turnover, luck-regressed).

XAI 2.1 PICKER DESCRIPTION. ParlayKings XAI 2.1 subtitle is now "Sports Analysis Machine (SAM) XAI 2.1 Grok Version".

1.5.9 PICKER DESCRIPTION. ParlayKings 1.5.9 subtitle is now "Sports Analysis Machine (SAM) 1.5.9 original model still most accurate." XAI 2.1 subtitle unchanged.

PASSWORD IS ALWAYS REQUIRED TO SIGN IN. /login and /api/verify-login no longer let a sam_bettors username or email through if password_hash is empty. Blank password is rejected. Login form password field is required. Stripe can still create a row to attach a subscription; that row cannot sign in until a password is set.

HAMBURGER FITS ONE SCREEN. Nav letters are 13px with tighter row padding so the stack stays on one viewport. Lib News is out of the stack. No other menu items added or restored.

CRON IS EVERY MINUTE. Grade Pending bets and re-dump Airtable into the replica on * * * * *. No hourly trigger.

USER-FACING NAME IS PARLAYKINGS. Chat bubble, reasoning toggle, model picker, workflow, Run today status, failover notice, and connecting-error reply say ParlayKings / ParlayKings 1.5.9 / ParlayKings XAI 2.1. Airtable Model Version still writes SAM 1.5.9 / SAM XAI 2.1 so replica and auto-grade keep matching.

SPARK REASONING SET TO LEVEL 4 (high). Meta Muse Spark 1.3 reasoning_effort is now high (1 minimal, 2 low, 3 medium, 4 high, 5 xhigh). Was level 2 (low). Grok / SAM XAI 2.1 stays at medium.

REPLICA RE-SEEDS FROM AIRTABLE EVERY HOUR. Deletes in Airtable were leaving ghost tennis rows in KV (Michelsen vs Tiafoe still Pending in the replica after the Airtable row was gone, so tennis.grimaldi.tv showed a duplicate). Cron now grades then dumps the live table into the replica so the sheet ping and sport sites match Airtable. Google Sheet public CSV is still unpublished (401); the Drive file itself is updated via Apps Script from bets-replica.json.

MARKET LINE BACK ON THE REPLY. ESPN's scoreboard JSON no longer includes moneylines; they live on summary pickcenter (DraftKings homeTeamOdds.moneyLine). Flattening UFC/tennis cards was also dropping event id, so the summary fetch never ran. Odds lookup now keeps the event id, reads pickcenter first, then The Odds API. Cached replies keep the stored Market Line when stripping Auto-graded notes.

RUN ALL IS EVERY SPORT TODAY, NOT DWCS. Slate still pulls MLB, NFL, NCAAF, NBA, NCAAB, NHL, WNBA, UFC, Tennis, WTA, EPL from ESPN scoreboard?dates=YYYYMMDD. Dana White's Contender Series / DWCS is not a numbered UFC card and is skipped. Already-logged games (any Result, not just Pending) are leftovers-skipped so a second click doesn't look like "UFC only" after MLB already wrote. In-progress UFC fights are skipped.

CLOSE GAMES KEEP THE FULL WRITEUP. 50/50 is still gone — tennis/UFC always name a winner. The model was treating that as permission to skip stats and dump a short coin-flip blurb. Tool + prompt now require the four-part writeup even on a 51-49: game summary, cited hold%/tape/rank/W/L, why the higher side wins, then the 25k split that sums to 100.

DELETED AIRTABLE ROWS CAN BE RE-RUN. Cache/dupe checks were treating the KV replica as proof the pick still existed, so deleting Michelsen vs Tiafoe (or any row) and asking again served the old writeup and skipped the new log. Cache hits now GET the Airtable record; 404 drops it from the replica and logs fresh. Also strip Auto-graded / POSTMORTEM lines from cached user replies — those belong in Notes, not the chat.

/stats READS AIRTABLE LIVE. The public stats page and Recent Picks were reading the KV replica, so grades (including today's 50/50 Push rows) could lag the Ai Grimaldi Bets table. Both now paginate Airtable appnbWauDhdcM7mS9 / tbl9X3RoO9uY4jOhL. Sport sites still use the replica.

PARLAYKINGS.NET NAV TRIM. On parlaykings.net / parlaykings.app the hamburger drops Live Broadcast, Episodes, Photo Gallery, Podcast Merchandise, Clip Vault, Workflow, Blog, and the nav Run Today link. Launch Sports Analyzer is labeled ParlayKings.net. HOME is labeled GRIMALDI.TV. drewgrimaldi still gets the header Run today / Run All button on every host. ai.grimaldi.tv keeps the full Grimaldi network menu.

UFC WORKFLOW SHOWS TALE OF THE TAPE. /workflow now has a UFC path: ESPN age/height/weight/reach/stance/SLpM/TD/sub/last-5, those numbers run every round of the 25k sim, and the answer prints TALE OF THE TAPE plus WHY THE SIMULATION PICKED. If the model omits the tape, the Worker appends the sim's tape. ESPN physicals are kept when fight stats fall back to fightingdata.

CLOSE GAMES STILL PICK A SIDE. Tennis/WTA/ITF and UFC no longer lock Picked Winner to 50/50 when the sim is even or data is thin. Yesterday about a quarter of the slate logged 50/50 and graded Push. The 25k sim still prints the real split (51-49 stays 51-49). [PICK] must name a winner — higher sim side, first-listed on an exact 50-50. Already-logged 50/50 rows still grade as Push.

LOGIN CHECKS PASSWORD WHERE ONE IS SAVED. /login and /api/verify-login accepted any existing sam_bettors username or email with no password check. Both now verify the password against password_hash in Supabase sam_bettors with PBKDF2-SHA256 on any account that has one saved, before creating a session or returning ok:true. Accounts with no password saved keep the old behavior until one is set.

FOOTBALL SIM PLUMBING: OPPONENT ADJUSTMENT + TURNOVER REGRESSION + WIND. calculateLambdaMultiplier accepts three new optional inputs, all neutral by default so live picks are unchanged: sosDiff, turnoverMarginPerGame (regressed 70% to the mean at ~4.5pts per turnover), and windMph (0.985 at 15+mph, 0.97 at 20+mph, both sides). New tunables: FOOTBALL_WIND_15_FACTOR, FOOTBALL_WIND_20_FACTOR, SOS_POINTS_TO_FACTOR, TURNOVER_LUCK_DAMPEN, POINTS_PER_TURNOVER.

UFC TALE OF THE TAPE DEEP INTEGRATION. (1) SIMULATION ENGINE: The 25,000-trial UFC Monte Carlo now deeply factors the full Tale of the Tape into every round — physical frame leverage (height & reach combined distance control, center of gravity on takedowns), age dynamics (youth & athletic prime advantage, 35+ chin durability degradation and late-round cardio fatigue), and stance clash (Southpaw/Switch open-stance angles). (2) EXPLANATION & WRITEUP: Tale of the Tape advantages are explicitly calculated, reported, and analyzed in the prediction breakdown alongside striking and grappling metrics.

SAM 1.5.9 (META AI MUSE SPARK 1.3) IS THE MAIN FRONTIER MODEL. (1) FRONTIER SWITCH: SAM 1.5.9 (Meta AI Muse Spark 1.3 via META_API_KEY) is the default frontier model in the picker and for unspecified traffic on ai.grimaldi.tv and parlaykings.net. (2) REPLICA FEED: SAM 1.5.9 is the public replica provider. (3) SECOND OPINION: SAM XAI 2.1 (Grok 4.6) is the second opinion option in the picker. Cached lookups and auto-grade still skip the LLM.

SAM META 1.0 AUTOMATIC FAILOVER ON SAM XAI CREDIT RUNOUT. If SAM XAI 2.1 (Grok 4.6) encounters credit exhaustion, billing issues, or quota limits (HTTP 402/403/429), SAM META 1.0 (Meta Muse Spark 1.2) automatically jumps in first to finish the predictions and write them to the Ai Grimaldi Bets Airtable base and public replica.

SAM XAI 2.1 IS THE MAIN FRONTIER MODEL — ALL PREDICTIONS PULL OVER TO AI GRIMALDI BETS. (1) FRONTIER SWITCH: SAM XAI 2.1 (Grok 4.6) is the default frontier model in the picker and for unspecified traffic on ai.grimaldi.tv and parlaykings.net. (2) SINGLE AIRTABLE BASE: All models log into the single official "Ai Grimaldi Bets" base (appnbWauDhdcM7mS9), distinguished by Model Version. Other bases were deleted. (3) REPLICA FEED: SAM XAI 2.1 is the public replica provider, feeding sport sites (ufc, tennis, football, mlb, wnba, basketball). (4) SLATE: Run today runs SAM XAI 2.1 first to feed the public book and sport sites. (5) NAV CLEANUP: Removed Press & Upcoming Guests from all navigation menus.

RUN TODAY SIMS ALL THREE MODELS. Gemini (SAM 1.5.8) still goes first and is the only write that feeds sport sites (ufc / tennis / football / mlb / wnba / basketball.grimaldi.tv) via the Gemini replica. Then SAM XAI 2.1 and SAM META 1.0 each log the same slate into their own Airtable bases. Click Run today again for leftovers per model. Credits stay untouched. Cached lookups and auto-grade still skip the LLM.

THREE AIRTABLE BOOKS, ONE PER MODEL. SAM 1.5.8 (Gemini) writes AI Grimaldi Bets - Gemini — that stays the public replica / sport-site / Run today feed. SAM XAI 2.1 writes AI Grimaldi Bets - Grok. SAM META 1.0 writes AI Grimaldi Bets - Meta Ai. Asking Grok about a Gemini-logged game runs a fresh sim instead of serving the Gemini cache. Auto-grade walks all three bases. SAM 1.5.8 remains the frontier default and Run today always uses Gemini.

SAM 1.5.8 NOW RUNS gemini-3.8-flash (was gemini-3.7-flash) AND IS THE DEFAULT. Picker default stays SAM 1.5.8, labeled Gemini - Original Model. Thinking stays MEDIUM. Same intro rate as 3.7 ($0.75/$3.75 per 1M through end of 2026). SAM XAI 2.1 stays Grok, SAM META 1.0 stays Meta AI. Run today still uses 1.5.8. Cached lookups and auto-grade still skip the LLM.

PICKER SUBTITLES NAME THE ENGINES. SAM 1.5.8 now reads Gemini, SAM XAI 2.1 reads Grok, SAM META 1.0 already reads Meta AI.

THREE MODELS IN THE PICKER. SAM 1.5.8 is Gemini again (gemini-3.7-flash, native tool loop). SAM XAI 2.1 stays Grok 4.6. New third option is SAM META 1.0 (Meta Muse Spark 1.2 via META_API_KEY). Default and Run today stay SAM 1.5.8. Same picker on parlaykings.net and ai.grimaldi.tv (one Worker). Cached lookups and auto-grade still skip the LLM.

SAM 1.5.8 IS THE MAIN FRONTIER MODEL FOR ONE MORE DAY. Picker default is SAM 1.5.8 (Meta Muse Spark 1.2), labeled Frontier SAM Model. Unspecified traffic and Run today / Run All use 1.5.8. SAM XAI 2.1 (Grok 4.6) stays in the picker as the second-opinion option so Drew can compare another day of data. Cached lookups and auto-grade still skip the LLM.

SAM XAI 2.1 IS THE MAIN FRONTIER MODEL. Picker default is SAM XAI 2.1 (Grok 4.6), labeled Frontier SAM Model. Unspecified traffic and Run today / Run All use XAI 2.1. SAM 1.5.8 (Meta Muse Spark 1.2) stays in the picker as the second-opinion option. Cached lookups and auto-grade still skip the LLM.

PARLAYKINGS.NET SEO. Title, description, keywords, canonical, Open Graph, JSON-LD, robots.txt, and sitemap.xml so Google can find Parlay Kings for sports, sports betting, gambling, DraftKings, FanDuel, and MGM searches. Homepage copy: best sports gambling analysis on the web.

PREMIUM SPORTS ACCESS EXPIRES WITH THE MONTH. premium_until already gates SAM. Sport-site cookies now expire on that date, verify-login returns premiumUntil, and /api/sports-access lets each sport site revoke leftover localStorage the day after the month ends. Pack buyers still have no sports access. Lifetime / drewgrimaldi / Grimillionaire stay.

$5 / 5-PACK ON STRIPE. Clicking Predictions Remaining now offers 5 Predictions for $5 alongside the $10 / 10-pack and $25 Premium. Same Checkout flow as the 10-pack: paid session credits +5 to sam_bettors.credits_remaining, packs never expire.

PARLAYKINGS.NET GETS THE FULL GRIMALDI.TV HAMBURGER. Same podcast + sports list as the rest of the network (HOME, Live, Episodes, Gallery, Blog, Merch, Clip Vault, Press, sports sites). At the bottom: Parlay Kings links — Stats, Workflow, About, Version Notes, Sign In/Out, and Run Today for Drew.

UFC 50/50 LOCK (same rule as tennis). After the 3-round vs 5-round fix, the Aug 22 UFC card still went 5-5 because several fights were 53-54% coin flips that still got a crowned winner. UFC now matches tennis: if the sim is 54-46 or closer (or UFC_DATA: coin), log Picked Winner = 50/50, confidence 50, do not auto-grade Win/Loss (Push once final). Real edges still pick a side. 5-round title/main events keep using the ESPN scheduled round count.

SAM 1.5.8 SPARK REASONING STAYS AT LEVEL 2 (low). Drew tried level 3 (medium) then kept level 2. Meta Muse Spark 1.2 reasoning_effort is low (1 minimal, 2 low, 3 medium, 4 high, 5 xhigh).

WRITEUP OPENS WITH A GAME SUMMARY, THEN STATS, THEN WHY THE PICK. Every prediction now leads with a short 2-4 sentence summary of this game, then the stats, then a detailed explanation of why that team or player wins (the actual edges from the tools), then the win-percentage split. Run today / Run All always uses SAM 1.5.8 (Spark), even if the picker is on Grok.

SAM 1.5.8 SPARK REASONING SET TO LEVEL 2 (low). Meta Muse Spark 1.2 reasoning_effort dropped from medium to low (Meta's scale: 1 minimal, 2 low, 3 medium, 4 high, 5 xhigh). Cheaper and faster on the main model. Grok / SAM XAI 2.1 stays at medium.

SAM 1.5.8 IS THE MAIN MODEL. Picker default is now SAM 1.5.8 (Meta Muse Spark 1.2), described as Frontier SAM Model. SAM XAI 2.1 (Grok 4.6) stays in the picker as the second-opinion option. Cached lookups and auto-grade still skip the LLM.

SAM 1.5.8 NOW RUNS META MUSE SPARK 1.2 (was gemini-3.7-flash). Flagship stays SAM XAI 2.1 (Grok 4.6) and remains the picker default. Label stays SAM 1.5.8. The 1.5.8 slot uses Meta Model API (muse-spark-1.2 at /v1/chat/completions) with the same OpenAI-compatible tool loop as Grok (schedule → sim → [PICK]), reasoning_effort medium. Cached lookups and auto-grade still skip the LLM. Internal provider key stays "gemini" so second opinion, failover, and Airtable Model Version keep working without a UI change.

FOOTBALL ON /stats. NFL and NCAAF were already in the graded log (40 + 6 rows) but the By Sport list labeled them only "NFL" / "NCAAF", so the word Football never appeared. Those rows now read Football (NFL) and Football (NCAAF). Same records, same win rates.

LOGIN RESTORED ON sam_bettors. Sport sites and grimaldi.tv were POSTing /api/verify-login to Pages Functions that were not running (405 empty body, then the UI said Connection error). Login now checks Supabase sam_bettors on this Worker at /api/verify-login: username or email row exists = in. Password hashes on that table were never Grimaldi.tv's unlock, so membership is the check. Network sites call https://ai.grimaldi.tv/api/verify-login so they do not depend on Pages Functions.

GRIMALDI.TV → PARLAYKINGS.NET SSO. Signing in on grimaldi.tv now follows you to parlaykings.net. The old session cookies were Domain=.grimaldi.tv, so the browser never sent them to the .net host. Launch Sports Analyzer now goes through /sso/handoff (reads the grimaldi session) then /sso/consume (sets sam_session on parlaykings.net). Token is HMAC-signed and dies in 60 seconds. Logged-out clicks still land on parlaykings.net unsigned.

RHO IS FITTED, NOT 0.28. MLB/NHL/Soccer bivariate Poisson correlation was a leftover 0.28 placeholder. Fitted from two seasons of completed ESPN scores (not Airtable): MLB 2025-26 regular season n=4510 residual r=0.02; NHL 2024-26 regular n=2632 residual r negative so clamped to 0 (this construction cannot take negative correlation); EPL 2024-26 n=760 residual r=0.01. Airtable is a pick log, not two seasons, and has no NHL rows.

SHARE CARD. Sharing parlaykings.net now shows a large Parlay Kings logo (1200x630) on iMessage, X, Facebook, Slack, and the rest. Served at /og.jpg with Open Graph + twitter:summary_large_image.

iOS HOME-SCREEN TITLE. apple-mobile-web-app-title was still UNCLE SAM under the Parlay Kings icon; it now says Parlay Kings. PWA theme/background colors match the purple.

PARLAY KINGS PURPLE BACKGROUND + .NET DOMAIN. Chat/login/stats pages drop the Uncle Sam poster for the deep purple from the Parlay Kings shield. parlaykings.net and www.parlaykings.net are Worker custom domains alongside parlaykings.app.

GOLD BAR. The red-white-blue top stripe (and the matching strip under the logo) is now gold, matching the Parlay Kings crown.

PARLAY KINGS BRANDING ON THE APP + TENNIS AUTO-GRADE ACTUALLY GRADES. (1) Header no longer shows the Uncle Sam hat, "The Best Sports Analysis On The Web", or "Sports Analysis Machine · SAM XAI 2.1". Favicon, apple-touch icon, PWA icon, and the logo up top are the Parlay Kings mark Drew bought (the PNG in AIProject). (2) Tennis auto-grade was fetching ESPN scoreboards with a dates=RANGE, which returns zero events. Live ATP/WTA boards are tournament-scoped and already include finished earlier-round matches. Grader now fetches the live board (no dates), dedupes the same competition when it appears twice, and treats a walkover as Push. Older US Open qualifying rows that sat Pending were graded in Airtable.

SECURITY + ACCURACY PASS (version held at SAM 1.5.8 per Drew's call — label unchanged). (1) TWO OPEN ADMIN ENDPOINTS CLOSED: /admin/trigger-autograde and /admin/clear-pending-mlb (plus its /api/clear-pending-mlb alias) were sitting ABOVE the session/auth checks in the fetch handler and had no auth at all — anyone who knew the path could hit them with a plain GET, and clear-pending-mlb mutates data (wipes pending MLB bets). Both now require admin auth before doing anything. Behavior change: those two URLs now need ?key= or an X-Admin-Key header, same as the other /admin routes. (2) ADMIN AUTH HARDENED: new isAdminAuthed() helper replaces the three inline providedKey !== env.SESSION_SECRET checks (set-sam-password, feedback, backfill-calibration). It prefers a separate ADMIN_SECRET and only falls back to SESSION_SECRET if unset — so the session-cookie signing key no longer has to double as the admin key; accepts the key via X-Admin-Key header so it doesn't have to ride in a URL where CDN/proxy logs capture it; still accepts ?key= so nothing existing breaks; constant-time compare. Non-breaking until Drew opts in by setting ADMIN_SECRET. (3) TIE TRIALS SPLIT 50/50 (non-soccer): runTeamSportSimulation was folding every integer-rounded tie trial into side B via winPctB = 100 - winPctA, a small systematic tilt toward whichever side was B. Non-soccer branch now recomputes winPctA from (winsA + ties/2) before shrinkage. Soccer untouched — draws stay a real modeled outcome there. (4) RECENT FORM FINALLY WIRED (NFL/NBA/WNBA/CFB/CBB): calculateLambdaMultiplier's recentFormDiff input had been dead since it was written (always 0, formFactor always 1.0). getRestDaysForSides already fetches a 16-day scoreboard for both teams for the rest calc, so it now ALSO computes each team's last-5 scoring margin from that same already-fetched data — zero new ESPN calls, no added subrequest pressure. In the sim core, recentFormDiff = FORM_DAMPEN * (recent last-5 margin − season margin), i.e. how far a team is over/under-performing ITS OWN baseline lately (deviation, not raw win-streak momentum — deliberately sidesteps the hot-hand fallacy), bounded by the existing ±8% clamp, no-ops under 3 completed games, and printed in the tool trace for transparency. FORM_DAMPEN = 0.5 is a deliberately conservative starting value exposed as a top-of-file constant for tuning. Only the 5 leagues that already do the rest fetch get it; MLB/NHL/Soccer would each cost a new subrequest so they were left for a deliberate future call. (5) PER-LEAGUE CORRELATION MADE TUNABLE (no values invented): sampleBivariatePoisson's rho is now a per-league SCORING_MODEL field instead of a hardcoded 0.28, but MLB/NHL/Soccer were all left at 0.28 on purpose — no per-sport value was fabricated, since there's no fitted data to back one yet.

SAM 1.5.8 NOW RUNS gemini-3.7-flash (was gemini-3.6-flash). Same paid intro rate as 3.6 ($0.75/$3.75 per 1M through end of 2026) — not a cheaper swap, a better Flash for the tool loop (schedule → sim → [PICK]). Thinking stays MEDIUM. The 25k Monte Carlo is still Worker code (run_matchup_simulation), not the model. Cached lookups and auto-grade still skip Gemini entirely. Label stays SAM 1.5.8.

$10 / 10-PACK CHECKOUT NOW CREDITS THE ACCOUNT. Stripe Checkout for /buy?tier=10 was already creating a live $10 session, but paying it never wrote credits_remaining: the success URL dropped the session id, STRIPE_WEBHOOK_SECRET is not set on the Worker, and no sam_bettors row has ever received pack credits. Return from Checkout now retrieves the paid session from Stripe and applies the same unlock as the webhook (10 credits for the $10 pack, or 30-day premium for $25), with a KV lock so webhook + return cannot double-credit.

NETWORK HAMBURGER MATCHES GRIMALDI.TV. SAM's slide-out menu now uses the same starred order as the rest of the network: HOME, ⭐ LIVE BROADCAST, ⭐ LAUNCH SPORTS ANALYZER, Football, Baseball, Basketball, UFC, WNBA, Tennis, Lib News, Trivia, then SAM-only links (Stats, Workflow, About, Version Notes).

TENNIS AUTO-GRADE CATCH-UP. Older WTA/ATP rows were sitting Pending after the match was final because (1) ESPN tennis scoreboards return the whole tournament, (2) Event Date was often the tournament start or an earlier-round date, and (3) names like Lili/Lilli and Lin Zhu/Zhu Lin missed the exact-string match. Grader now range-fetches tennis ~16 days past Event Date, matches last names and reversed given/family order, and processes oldest Pending first. get_game_schedule prefers a still-scheduled match over a completed earlier-round result so the logged Event Date is the actual match.

TENNIS IS HOLD SERVE FIRST (Drew, college tennis at Eastern Illinois). If they cannot break you they cannot beat you — the sim now uses real service-game hold % from TennisMyLife match tapes (service games minus breaks against), not a rank-guessed hold. Surface-specific hold (clay/grass/hard) wins when there are enough games on that court. Weather actually moves the hold number: wind and rain make holding harder, heat speeds the bounce, humidity fluffs the ball. Indoor still plays faster. Rank/W/L are fallback only when hold sample is thin. Writeups must cite both players' hold %. MLB unchanged.

TENNIS THIN-DATA 50/50 (MLB LEFT ALONE). Drew: leave MLB for a few more days, it is hitting; the leak is weaker tennis where rank/W/L never come back and SAM still crowns a winner off a default #80 hold. Tennis/WTA/ITF only: if either player is missing rank or season W/L, or the hold/break sim is 50-51, lock the logged pick to 50/50, say why, still write the Airtable row, still serve the cache with Line movement. Do not auto-grade Win/Loss on those rows (Push once the match is final). MLB and every other league still pick a side.

BASKETBALL IN THE SAM MENU. The hamburger sports list now has Basketball under MLB, linking to https://basketball.grimaldi.tv (NBA + NCAAB). Same network link as football, WNBA, tennis, and UFC.

UFC ACCURACY PASS (post-release patch, version held per Drew's call). (1) 5-ROUND FIGHTS WERE BEING SIMULATED AS 3-ROUND FIGHTS: runUFCMultiStatSimulation had the round loop hardcoded to 3, with no exception for title fights or scheduled 5-rounders \u2014 a real bug affecting every main event and championship bout, since decision math, finish-rate odds, and cardio all play out differently across 5 rounds. Added fetchUFCFightRounds(), which looks up the matched fight on ESPN's UFC scoreboard and reads the real round count off the competition's own description field (\u201c5 Rnd (5-5-5-5-5)\u201d) plus its title-fight flag, run in parallel with the existing power-ratings fetch so it costs no added latency. Falls back to 3 rounds if the fight can't be matched on the board (same non-blocking spirit as the rest of the UFC pipeline). (2) REACH/STANCE EDGES WERE FLAT ACROSS ALL WEIGHT CLASSES: a 4-inch reach edge was scored identically for a heavyweight and a flyweight, which isn't physically right \u2014 reach differential matters proportionally more relative to a smaller frame. Added ufcWeightClassScale(), a lookup table (1.6x at strawweight down to 0.65x at heavyweight) now applied to the reach-edge calculation, keyed off the real weightClass field ESPN's athlete endpoint already returns (added to fetchUFCFighterStats's extracted fields \u2014 was being pulled by ESPN this whole time but not captured). (3) NOT FIXED, BY DESIGN: opponent-strength normalization (raw SLpM/TD stats aren't adjusted for the quality of who they were earned against) and ring-rust/layoff time are both real remaining gaps, but neither has a reliable current data source \u2014 ESPN doesn't expose opponent quality metrics, and pulling last-fight date would mean two more sequential ESPN calls per fighter on top of the already-tuned rate-limit-safe chain (the one that took a dedicated bugfix round to get right, see below). Faking either with a heuristic would make picks look more rigorous than they are, so left alone until there's a real data source worth wiring in.

MLB 64% DEFAULT LOCKED OUT. After five straight 8/19 losses, Airtable showed the Monte Carlo at 50-58% while Confidence % and the writeup were almost always 64% (40 of ~108 recent MLB picks). Grok was rounding coin flips up because the prompt said not to hedge and that favorites clear 60-65%. Fix is in code: logged confidence and the visible split are overwritten from the actual run_matchup_simulation numbers (picked side's sim %, both sides sum to 100). Band calibration may still pull a number DOWN, never above the sim. Dummy/tiny-sample starters (Stock 12.1 IP / 3 GS) no longer sneak through the sample-weight backdoor — under 15 IP and 4 GS is unused, matching the Friday rule. Starter lock restored: 0% weight more than 6 hours before first pitch, ramp 6h→3h, full inside 3h. Prompt no longer permits inflating a 52% sim to 64%.

CONSERVATISM SHRINKAGE SET BACK TO 0.50 (0.55 → 0.50). Per Drew's call after MLB looked off. A raw 70% edge reports 60% instead of ~61%, a raw 80% edge 65% instead of ~67%. Same constant as the original conservatism pass; applies to every sport including MLB.

POST-MORTEM TAGS ARE ACTUALLY ACCURATE. Grade-time tags were leaking starter names into LOSS PATTERNS (comma-split turned "actual starter(s): Pintaro" into fake tags) and faded_market was still looking for "SAM and the market disagree" — a phrase the Market Line never writes. Parser now keeps a whitelist (wrong_probable, dummy_era, injured_qb_used, preseason_stats, thin_data_signal, high_confidence_miss, favorite_miss, coinflip_miss, faded_market). faded_market fires from the real line: SAM vs implied % (>=5) or "**+X% edge** over the market" / "**X% below** the market", never "roughly aligned". Historical losses are re-read the same way when TRACK RECORD loads, so old cards count. Autopsy line is "POSTMORTEM: tags — Winner 62% — logged Stock, Pintaro started". If faded_market or thin_data_signal repeats enough in a league, the next sim gets a hard tool-result warning (do not widen a SAM-vs-market gap; sparse/zero stats are a red flag) instead of rewriting the whole engine after one bad night.

WNBA IN THE SAM MENU. The hamburger sports list now has WNBA under UFC/MMA, linking to https://wnba.grimaldi.tv. Same network link as football, tennis, and UFC.

STRIPE PAYMENT ACTUALLY UNLOCKS SAM. Paying the $25 VIP Payment Link or /buy never flipped sam_bettors: the live Stripe webhook omitted checkout.session.completed, the Worker only handled that one event, Payment Links have no client_reference_id, and STRIPE_WEBHOOK_SECRET was never set so every POST 400'd. Webhook now accepts checkout.session.completed and invoice.paid, matches the payer by session id OR email (creates a sam_bettors row on that email if needed), sets premium_until from the Stripe period end (or +30 days / +10 credits for /buy), and verifies events via the signing secret or by retrieving the event from Stripe when the secret is missing.

GOOGLE SHEET STAYS CURRENT WITH THE LIVE REPLICA. Airtable is still the official write. Sport sites already read https://ai.grimaldi.tv/api/bets-replica.json (not Airtable). After every replica save, SAM pings GOOGLE_SHEET_SYNC_URL if set so the Drive sheet Ai Grimaldi Bets replica refreshes immediately. A one-minute Apps Script fallback also pulls the same JSON.

ALL GROK KEYS DEAD → SAM 1.5.8. After every Grok key (primary + backups) fails or is skipped, a new sim goes to Gemini / SAM 1.5.8. Already-run games still serve the stored writeup. The reply notes that SAM XAI 2.1 was unavailable. Next request skips the dead Grok keys for 15 minutes instead of waiting on them again.

GOOGLE SHEET / BETS REPLICA. Airtable stays the official log. Reads (cache, slate leftover check, track record, /stats, sport-site pick APIs) now use a Cloudflare KV copy first so they stop burning the Free-plan 1,000 Web API calls. New picks and grades still write Airtable, then update the replica. Public feed: GET /api/bets-replica.json and GET /api/bets-replica.csv (newlines stripped so Google IMPORTDATA can parse Notes). drewgrimaldi can force a one-time Airtable dump at GET /api/bets-replica/refresh.

ODDS API BACKUP KEY. env.ODDS_API_KEY_BACKUP is tried when the primary Odds API key returns 401/402/429. Only after both keys fail do we skip The Odds API for 15 minutes and fall through to ESPN moneylines. Secret is set via wrangler, never hardcoded.

ESPN MONEYLINE FALLBACK WHEN THE ODDS API IS OUT. The Odds API returned 401 on cache hits (monthly credits gone — they reset on the 1st). Line movement went silent. fetchMarketOddsData now tries The Odds API first; on 401/402/429 or a missing key it reads moneylines off the ESPN scoreboard we already fetch (ESPN BET / DraftKings / FanDuel when present). No new paid key. A 15-minute skip avoids hammering a dead Odds API key. Failed odds still do not print a warning on the user-facing pick.

CACHED REPLY ALWAYS CHECKS LIVE ODDS FOR LAST-MINUTE LINE MOVEMENT. Person B asking "Tigers vs Nationals" after Person A already ran it still gets the identical stored pick/writeup (no second sim). SAM now always fetches the current market line on that cache hit and appends a movement note: same / toward / away from the pick, opening price → now. The previous 800ms skip dropped this too often. If the first log had no Market Line, Person B still sees the current line labeled as checked just now. Opening price is parsed from the stored Market Line and also saved as a ---ODDSLINE--- snapshot on new writes so later lookups do not depend on the regex.

AIRTABLE CACHE SHORTCUT ACTUALLY FIRES FOR ALREADY-RUN GAMES. The whole point of logging picks is to serve the stored writeup instead of re-running verify/stats/25k sim. That path existed but missed most real questions after a slate run: (1) only the first 100 Pending rows were searched, no pagination; (2) two Pending rows for the same MLB series (yesterday + today) counted as "ambiguous" and skipped the cache entirely; (3) Placed Date coming back as a datetime failed the exact === todayET() check so every hit was treated as stale and re-simulated; (4) lookup required the user to type "vs" or "at"; (5) a cache hit still waited on The Odds API (tennis can fan out many 4s calls) plus a non-peek credits check before answering. Fix: paginate today's + pending records, prefer today's unique matchup, match nicknames even without vs/at, compare dates by YYYY-MM-DD, and return the stored reply immediately (line movement has 800ms to append, otherwise skip).

SCRAPINGANT ESPN FALLBACK WHEN ZENROWS IS OUT. Drew has a ScrapingAnt API key to use while ZenRows is unpaid/capped (AUTH004). Added fetchEspnViaScrapingAnt() against https://api.scrapingant.com/v2/general (x-api-key header). tryEspnBypassFallbacks now tries ZenRows first, then ScrapingAnt, then the unused Browser binding. If ZenRows returns 402/AUTH004, a sticky flag skips it for 15 minutes so we don't keep burning failed ZenRows calls. fetchEspnCoreWithRetry (UFC/tennis athlete lookups) now uses the same fallback chain instead of ZenRows-only. Secret: SCRAPINGANT_API_KEY. No-ops until that secret is set.

SAM LOGIN IS sam_bettors MEMBERSHIP. Every row in Grimaldi.tv's sam_bettors table can sign in with that username or email. Password hashes on that table were never the same as Grimaldi.tv's screen-name unlock, so requiring a SAM hash rejected real members (credits and premium included). Lookup is username OR email; if the row exists, SAM starts a session on that bettor id so free/daily, credits_remaining, and premium_until apply as stored. Not on the list = incorrect username. Supabase errors surface as such. drewgrimaldi is still the only admin for Run today.

ADMIN SLATE RUNNER — RUN TODAY'S BOARD INTO AIRTABLE. drewgrimaldi-only. GET /api/slate pulls today's ESPN scoreboards (MLB, NFL, NCAAF, NBA, NCAAB, NHL, WNBA, UFC, Tennis, WTA, EPL), expands UFC/tennis cards into individual matchups, skips completed games and anything already Pending in Airtable, and returns the leftover list. The signed-in admin UI has a Run today button that walks that list one game at a time through the existing /api/chat path (same sim, same [PICK], same Airtable write as typing the matchup). slate:true on /api/chat is rejected unless the session username is drewgrimaldi; when it is, the prediction-credit meter is not touched. One game per Worker request so ZenRows/subrequest limits stay the same as a normal typed prediction.

NFL / NCAAF / NBA / CBB ACCURACY PASS. These are the volume sports. (1) NBA preseason is ignored the same way NFL August is — season_type pre and October games before the 20th do not count as this year's offense/defense. Under 8 real NBA/CBB games (6 WNBA), last year is blended in. (2) Rest was in the lambda math but never set. ESPN last-game lookback now sets restDays for NFL/CFB/NBA/WNBA/CBB. NBA/WNBA/CBB on a 1-day turnaround is treated as a back-to-back and scales offense down. (3) Basketball key-player skips OUT/IR so a hurt star's PPG is not the factor. Tool result prints rest and data vintage.

POST-MORTEM TAGS NOW CHANGE THE SIM, NOT JUST THE PROMPT. On auto-grade of a Loss, SAM compares the logged starter/QB to the actual ESPN box-score starter and writes structured tags (wrong_probable, dummy_era, injured_qb_used, preseason_stats, favorite_miss, plus the old thin_data / high_confidence / coinflip / faded_market). Last 20 losses per league are rolled up when TRACK RECORD loads. If a tag repeats enough: wrong_probable x3 in 6+ MLB losses halves starter-lock weight; dummy_era x2 in 4+ losses raises the IP/GS bar; high_confidence_miss x4 in 8+ losses pulls that league's win% toward 50. Friday-style Stock/Pintaro would tag wrong_probable and then actually tighten the next MLB card. One bad night does not rewrite the engine.

FOOTBALL WEEK-1 HARDENING (NFL + CFB). Season is close; do not let August preseason or a 0-game 2026 slate pretend to be this year. (1) NFL BALLDONTLIE games tagged preseason, week<=0, or dated August/late July are dropped. If fewer than 4 real regular-season games exist, last year is blended in (or used outright if this year is empty). (2) ESPN NFL/CFB standings do the same: under 4 games played, blend last year instead of treating a Week 0/1 sample or empty 2026 table as a full season. (3) QB key-player: injured QBs are skipped so a backup is the input; Preseason splits are never used for QB rating. Tool result now prints the data vintage so a Week 1 pick says "2025 regular season / preseason ignored" instead of looking like live 2026 form.

MLB BULLPEN + DON'T LOCK A STALE PROBABLE. Follow-up to Friday's card. (1) BULLPEN: ESPN team pitching stats are pulled for both clubs. A real relief/bullpen ERA is preferred; if ESPN only exposes overall staff ERA it is used as a weaker staff proxy. That factor is blended into expected runs so the 8th/9th inning is not just "whoever started." (2) STARTER LOCK: listed probable ERA is not applied more than 6 hours before first pitch (0% of the starter weight), only partially from 6h to 3h, and fully inside 3 hours or after first pitch. A Stock-style 10.13 probable the morning of cannot manufacture a 62% pick anymore. Tool result states STARTER LOCKED / NOT LOCKED and the hours to first pitch. Re-run inside 3 hours to lock the actual starter.

MLB PITCHER INPUTS HARDENED AFTER A BAD FRIDAY CARD. Friday's graded MLB slate went 2-5 (likely 2-6): dummy 0.00 ERAs (Jobe), a listed probable who did not start (Stock 10.13 / Pintaro actually threw), and ERA-only weight turning 62% favorites out of mid-rotation matchups. Fix is in the sim inputs, not the prompt. Probable-pitcher lookup now pulls IP/GS/G with ERA. A 0.00, ERA over 12, under 15 IP and 4 GS, or a reliever/opener profile (starts are a small share of appearances) is NOT used to scale expected runs, and the tool result says so. Usable ERAs are sample-shrunk toward 4.20 and capped 0.86-1.16 so one ugly or tiny-sample number cannot manufacture a 62% pick. Star-bat IL weight raised (OF/DH/SS) so a Judge/Stanton/Bellinger-style pile-up actually moves the Yankees-type number instead of getting ignored as three 1.5% OF dings.

UFC TALE OF THE TAPE IS NOW PART OF THE SIMULATION OUTPUT. The fight sim returns the actual tape it used (age, height, weight, reach, stance, SLpM, strike acc, TD/15, TD acc, sub/15, last-5) plus the 25,000-trial win% split by KO/sub/decision and a why line computed from those same inputs (which tape edges moved the number, most common finish path). Not extra summary color — those rows are the simulation's inputs and results.

PREDICTION ENGINES: POSITION-WEIGHTED INJURIES + UFC MULTI-STAT + TENNIS HOLD/BREAK. Three code-level upgrades so winner accuracy comes from better inputs and better sims, not more prompt text. (1) TEAM INJURIES: ESPN reports are now scored by position and status (QB/ace/star OUT haircuts harder than a depth-chart Questionable) and applied automatically to each side's lambda, cap 18%. Simulation writeup names the notable absences. (2) UFC: fights no longer mash tale-of-the-tape into one power number plus N(0,15). New per-round model uses SLpM, strike accuracy, takedown avg/accuracy, submissions, reach, stance, and last-5 form from fightingdata.com (KO/sub finish paths + decision). (3) TENNIS/WTA/ITF: matches no longer use the same mashed fight Gaussian. New hold/break set sim (best-of-3, best-of-5 for ATP slams) with hold probabilities from rank, season W/L, TennisMyLife surface record, indoor/outdoor, recent form, and H2H. Conservatism shrinkage still applied to the raw win% on all three paths.

CONSERVATISM SHRINKAGE SET TO 0.55 (0.50 → 0.55). Per Drew's call after the fresh deploy landed back on 0.50. Same small nudge as the earlier 0.50 → 0.55 pass: a raw 70% edge reports ~61% instead of 60%, a raw 80% edge ~67% instead of 65%. Holding here.

GROK MODEL BUMPED TO 4.6, REASONING EFFORT LOWERED TO MEDIUM. Swapped GROK_MODEL from "grok-4.5" to "grok-4.6" (Drew is getting a fresh API key for it) and dropped reasoning_effort from "high" to "medium" per Drew's request. Also discovered while fixing the base-consolidation bug above that Grok's calibration/track record has zero real history to date — every prior write silently 404'd against a base that never existed — so this reasoning-effort change starts from a clean slate anyway.

REAL ROOT CAUSE FOUND VIA WRANGLER TAIL: ZENROWS CONCURRENCY LIMIT, NOT ESPN BLOCKING. Deployed the ZenRows-first change, ran wrangler tail live against a real WNBA prediction per Drew's request, and got a definitive answer instead of another round of static-code guessing: 8 athlete-overview calls fired, 5 succeeded, 3 came back AUTH006 (Too many concurrent requests) directly from ZenRows' own API -- not ESPN, not a subrequest-ceiling error, a distinct problem from everything chased earlier this session, and one the Cloudflare Workers Paid upgrade does nothing for since it's a ZenRows-side limit, not a Cloudflare-side one. Traced the exact mechanism: getKeyPlayerFactor's candidate-stats loop used Promise.all across 4-6 players, AND the outer call site ran both teams' getKeyPlayerFactor through Promise.all too -- a nested fan-out that, for a WNBA matchup, meant up to 8 simultaneous ZenRows sessions competing for the same concurrency slot the instant today's earlier change (routing every ESPN call through ZenRows first, no direct attempt) went live. That earlier change fixed the original latency problem but directly caused this one -- more calls landing on ZenRows at once, more collisions. Fixed both nesting levels to sequential (await one candidate/team fully before starting the next) rather than parallel. Also converted tennis's two-player Promise.all to sequential pre-emptively, same mechanism at smaller scale (2 concurrent vs. up to 8) -- Drew confirmed tennis is also still broken and this is a cheap, low-risk fix for the same class of bug rather than waiting for a second tail session to prove it independently. NOT YET touched: getInjuryReport's both-teams Promise.all (2 concurrent) -- not implicated in the log, left alone for now; worth revisiting if injury-report-specific errors show up in a future tail session.

ALL ESPN CALLS NOW GO STRAIGHT TO ZENROWS, EVERY SPORT, NO DIRECT ATTEMPT FIRST. Drew confirmed the error was \u201cSAM's having trouble connecting right now\u201d \u2014 the generic message shown when the whole model call fails, not an ESPN-specific error surfacing cleanly. That pointed at latency/timeout rather than a clean 403: every blocked direct fetch was paying a full connect-attempt-then-fail cost before ever falling back to ZenRows, and with WNBA/tennis routing several ESPN calls per prediction, that failed-attempt tax could plausibly stack up into the whole request timing out \u2014 which would explain a generic connection failure instead of a traceable ESPN error. Drew's ask: skip the 403 entirely, wire it all up with ZenRows first. Removed the conditional gating in both fetchEspnWithRetry() and fetchEspnCoreWithRetry() (isTightMarginEspnUrl, isEspnDirectLikelyBlocked's sticky flag, and the forceZenRows flag added earlier today for tennis) \u2014 now both functions try ZenRows immediately whenever env.ZENROWS_API_KEY exists, for every sport, every endpoint, with the direct fetch demoted to a last-resort fallback only if ZenRows itself comes back empty or erroring. isTightMarginEspnUrl() and isEspnDirectLikelyBlocked() are now unused (left in place rather than risk over-editing; harmless dead code, not wired into anything). Cleaned up the now-meaningless forceZenRows argument from all four tennis call sites since every call gets that behavior unconditionally now. This trades a real increase in ZenRows usage (every ESPN call, not just the previously-flagged tight-margin sports) for the fastest possible path per call \u2014 Drew is already on the 45k/month paid tier anticipating this kind of volume.

TENNIS ZENROWS COVERAGE GAP FOUND AND FIXED; WNBA CONFIRMED ALREADY FULLY COVERED. Drew reported still getting errors on both. Pulled the actual live deployed worker directly from Cloudflare first (not just the local file) to rule out a stale-deploy mismatch \u2014 it wasn't one, the live code already has every WNBA/tight-margin fix from this session. Traced every WNBA data path (team resolve, roster, standings, athlete search, athlete stats, schedule, injuries, grading) and confirmed all of them go through fetchEspnWithRetry(), which does have the guaranteed-ZenRows check (isTightMarginEspnUrl) \u2014 no gap found there. Tennis is a different story: its entire pipeline (fetchTennisPlayerStats \u2014 athlete lookup, rank list, rank detail, statistics, four separate calls per player) runs exclusively through fetchEspnCoreWithRetry(), which never got the WNBA-style guaranteed-first-try fix at all \u2014 it only fell back to ZenRows reactively, after a 403 had already been hit once this invocation (the sticky flag), or inline on that same call after a failed direct attempt. Since three of the four tennis calls use ESPN-returned $ref URLs rather than predictable literal paths, pattern-matching them the way isTightMarginEspnUrl does for WNBA would be fragile, so added an explicit forceZenRows parameter to fetchEspnCoreWithRetry() instead and set it true at all four tennis call sites \u2014 tennis has zero BALLDONTLIE coverage and its whole stat pipeline sits on this one function, so it needs the same guarantee WNBA already has, not URL-guessing. NOT YET CONFIRMED: whether this was the actual cause of the WNBA errors specifically, since that path checked out clean \u2014 need the real error text/message from Drew to pin that one down rather than guessing further.

CHAT BACKGROUND SWAPPED TO THE NEW UNCLE SAM POSTER. Replaced BACKGROUND_BASE64 with the newly uploaded promotional artwork. The original PNG upload was 2.5MB \u2014 too large to embed directly without pushing the worker script size up significantly, so it was resized to roughly match the previous background's dimensions (1100px wide, same aspect ratio) and re-encoded as JPEG at quality 78, landing at about the same final size as what was already in place (~316KB vs. the previous ~280KB) rather than ballooning the deployed script. Verified the new base64 decodes back into a valid JPEG before shipping.

NCAA.COM ADDED AS A GENERAL-PURPOSE STAT SOURCE FOR CFB, WIRED IN AS A QB-RATING FALLBACK. Drew asked for \u201call of the stats\u201d from NCAA.com's individual leaderboards, used as a fallback only when ESPN's college football data is thin. Checked the page directly first: it's a plain, simple server-rendered HTML table (no JS, no bot-check) with a consistent structure across every category, and found all 41 category IDs from the site's own dropdown (Rank/Name/Team/Cl/Position/G plus category-specific columns) \u2014 confirmed category 453 in Drew's link is specifically Passing Yards, not Passing Efficiency (the real QB Rating equivalent, which is category 8, \u201cPass Eff\u201d column). Built this as reusable infrastructure rather than a one-off scrape: parseNcaaStatsTable() generically parses any category's table by reading the header row and mapping columns dynamically, so adding coverage for a different stat later is a one-line addition to NCAA_FOOTBALL_STAT_CATEGORIES, not a new parser. fetchNcaaFootballStatCategory() fetches one category by name; fetchNcaaFootballPlayerStat() searches a category's full leaderboard for a specific player by name, disambiguating by team when multiple players share a name. Wired the Passing Efficiency category specifically into getKeyPlayerFactor() for CFB: when ESPN's per-QB stat fetch comes back empty for every candidate on the roster (common for less-marquee college programs ESPN doesn't track closely), falls back to looking up the top-of-roster QB by name and team on NCAA.com's real Passing Efficiency leaderboard, using \u201cPass Eff\u201d as the QBRating-equivalent value and \u201cPass Att\u201d as the usage stat \u2014 same downstream math, just a real number instead of a missing one. Verified the full path (table parsing, name lookup, team disambiguation, and a real not-found case) against live NCAA.com data before shipping.

TWO NEW UFC FALLBACK DATA SOURCES ADDED, BOTH VERIFIED SCRAPABLE WITHOUT ZENROWS. Drew's original ask was whether ufcstats.com could be added now that ZenRows exists \u2014 checked it directly first: it's not a simple IP block like ESPN, it's a client-side JS proof-of-work challenge (computed SHA-256 in-browser), a fundamentally harder bot-check that a plain fetch() can never solve regardless of where it runs from. Left that one alone pending real verification against ZenRows' antibot+js_render mode. Drew then suggested two alternatives instead: ultimatefightingstats.com and fightingdata.com. Confirmed ultimatefightingstats.com's fighter stats are on its free tier (not paywalled) before touching anything \u2014 it's a Next.js app whose fighter pages are server-rendered but the actual data sits inside a React Server Component payload with inconsistent backslash-escaping depth, which made SLpM/Str. Acc./TD Avg. reliably extractable via a backslash-tolerant regex but left reach/age/takedown-accuracy/submission-avg in a more deeply nested structure not worth force-parsing. fightingdata.com turned out to be the stronger source: plain server-rendered HTML with a genuinely simple table structure (/ pairs,