r244 · Game-level Marcel snapshot infrastructure — validation-first, NO live behavior change. Pure capture ship to enable backtest validation BEFORE flipping any live Marcel-related changes. Every game bet now carries gameStatsAtAnalysis stamped at recommendation time: season + L15 + 3-component-blended RPG/OPS + RA9 for both teams, sample sizes (gpSeason / gpL15), raw blend weights actually used (currWt/priorWt/lgWt), trajectory labels (consistent_strong, trending_better, etc.), and prior-year-used flag. ~30 fields, ~600-800 bytes per snapshot. Threaded through all three card-add paths (sendBetToCard, addAllToCard, addBestBetToCard). Why infrastructure first: live analyzeGame already uses rpgBlend (the 3-component blend) for game total/F5 prediction since long before this ship — the actual gap was that game bets didn't capture the inputs that fed those predictions, so backtesting was impossible. Pre-r244 entries lack the snapshot and are silently skipped by the diagnostic harness. Also ships window.runGameMarcelDiagnostic() — segments graded post-r244 game bets by Marcel-shift magnitude (max(|awayRpgBlend−awayRpgSeason|, |homeRpgBlend−homeRpgSeason|)) into four cohorts: minimal (<0.10), moderate (0.1-0.3), strong (0.3-0.5), extreme (≥0.5). Compares high-shift vs minimal-shift WR + ROI. Inline verdict: ✅ if high-shift outperforms by ≥3pp WR or ≥5pp ROI (Marcel adds signal — keep weights, fine-tune in r245); ❌ if high-shift underperforms (Marcel hurts — r245 reduces/reverts); ⚠️ if similar (inconclusive); ⏳ if <30 bets per cohort (wait). r245's decision tree fully driven by this output. Run after ~30 days of post-r244 graded game bets accumulate.

r276 · Fade Away Dog +100-149 — paired-direction signal, default ON. Companion ship to r275. Where r275 stops us from betting the losing side (away dog), r276 captures the empirical inverse: the same 23 games where the away dog lost 70% of the time mean the HOME favorite WON 70% of the time. Edge-boost pattern fires when iterating side === 'home' AND the opposite team's ML odds (oppOdds) are in the +100 to +149 band. Adds +0.06 to the home-side probability (6pp model edge boost) and pushes the "fade_away_dog" tag into firedPatterns for backtest segmentation. Calibration: observed 69.6% home WR is ~12pp above typical -130 breakeven (57.4%), so 6pp boost is half the empirical edge — matches the standard discount-for-sample-size convention applied to other patterns. The boost combines with the model's existing edge derivation for home ML; on this cohort it typically promotes borderline plays to FIRE/MAX tier with appropriate sizing. Empirical replay: 23 cohort games saw home win 16-7. At flat 1u sizing on home ML at average -130 odds: ~+5u net. At elevated tier sizing (3u-5u), expected gain ~+15-25u. Combined with r275's -50.23u suppression, the paired r275+r276 ships convert what was a -50.23u bleed into a positive-EV signal on the same 23 games. Console kill switch: window._disableFadeAwayDog = true. Why this pattern matters beyond bankroll: it's the first explicitly empirically-derived "fade the model" pattern. Surfaces a new investigative approach — when a cohort's WR is significantly below 50%, the inverse direction may be a positive-EV signal. r277+ candidates: scan all bet types for similar inverted cohorts (e.g., are there Total Over scenarios where the model's PASS-tier under is the actual play?), per-pattern home/away asymmetry regression once 50+ post-r276 plays accumulate.

r275 · Away ML +100 to +149 floor gate — backtest-validated, default ON. First game-bet floor gate (prior gates were prop-side: r265 Outs, r267 K, r269 HR/TB). Surfaced by r275 audit of 371 graded game plays in r257 backup: away ML at small-dog odds (+100 to +149) lost 30.4% WR / -44.8% ROI / -50.23u sized over 23 plays — EVERY tier was unprofitable (FIRE 1/5 at -15.4u, MAX 0/3 at -21.5u, PLAY 4/9, SOLID 2/6). Same odds at HOME side were profitable (50%+ WR / +10-74% ROI), confirming this is specifically an away-side miscalibration of the ML edge calculation, not a "small-dog ML is bad" general phenomenon. Heavy away dogs (≥+150) were already filtered to TRACK in pre-r275 logic; r275 completes the away-dog suppression with the sibling +100-149 band. Empirical lift on the 118-play ML cohort: -56.52u → -6.29u (+50.23u improvement, ~2.85× the next-largest single ship r265 Outs at +17.6u). Mechanism: matches existing heavy-dog filter pattern — pushes a TRACK rec (units:0, tier:'MARG', visible in analyzer output marked SUPPRESSED) so data continues accumulating for re-evaluation, but no bankroll impact. Filter triggers when side === 'away' && mlOdds >= 100 && mlOdds < 150. Console kill switch: window._awayMlAllowDog = true reverts immediately. What this implies about the model: the ML edge calc treats away vs home symmetrically but historical reality has an asymmetric "home-field" effect bigger than the model accounts for. The away dog band is where the asymmetry shows up most sharply (book is already pricing the away team to lose; if the model thinks it has edge there, the model is overconfident). Future ship: r276+ regression analysis on home/away asymmetry across all bet types could surface whether ML at neutral odds (~-110 to +100) also has a smaller version of the same effect, or if it's only the dog band.

r269 · HR + TB floor gates — architectural consistency, default ON. Completes the prop-floor gate sequence started by r265 (Outs) and r267 (K). HR analyzer now mirrors the K pattern: tier=PASS plays are not auto-tracked or shown as recommended. Empirical sample for HR is currently too small for a backtest verdict (only 6 graded HR plays in r257 backup) so the gate ships as architectural insurance — when HR data accumulates over the next ~30 days, the gate prevents the same floor-bleed pattern that hit Outs (-12.59u over 37 plays) and K (-2.95u over 23). TB analyzer adds a floor gate layered on top of the existing r228 suppression: while suppression is active (default), tier is already forced to PASS in the suppression block so the new gate evaluates false and recommended is false anyway — no behavior change. When TB suppression is eventually lifted (post ~30 days of post-r228+r241 NegBin data), the floor gate stays in place so micro-edge plays do not flood back into recommendations. Mechanism mirrors r267: _hrRecAllowed = _hrRecAllowPass || tier !== 'PASS'; same shape for TB with extra suppression-AND clause. Console kill switches: window._hrRecAllowPass = true and window._tbRecAllowPass = true revert respective gates immediately. With r269 all four prop analyzers (Outs / K / HR / TB) now have consistent floor behavior — tier=PASS plays universally produce recommended:false. Future ships transition to data-accumulation phase: r270+ candidates are TB suppression review (post-r228+r241 cohort), HR floor empirical validation (post-r269 cohort, ~30 days), bullpen per-weight ROI harness (handoff r261, fully unblocked).

r267 · K props floor gate — backtest-validated, default ON. Same pattern as r265 (Outs floor) but for K props. Backtest of 23 graded PASS-tier K plays in r257 backup: 47.8% WR, -12.8% ROI, -2.95u total — bleeding the otherwise break-even K cohort. Direction split is striking but on small samples: PASS Over (n=13) at 15.4% WR / -8.76u, PASS Under (n=10) at 90% WR / +5.81u. The asymmetry suggests model Over-lean bias on micro-edges, but neither sample is large enough to ship a direction-aware fix (95% CIs span 2-45% and 56-99%). Conservative gate filters all PASS; counterfactual lift on the same 23-play cohort is +2.95u (forfeits the +5.81u PASS Under wins which are likely partly noise). Mechanism: hoists tier to a local _kTier for both bet objects, then derives _kRecAllowed = _kTier !== 'PASS'. recommended is gated on _kRecAllowed && (direction === 'Over'|'Under'). Because _propUnitsFor already returns 0u for prob-edge <2pp and count-edge <0.5, and tier is already PASS for those edges, the only behavior change is recommended:false (true triple-strip in spirit). Console kill switch: window._kRecAllowPass = true reverts to pre-r267 behavior immediately without reload (gate consults window. on each call). Future ships: r268+ candidates: TB suppression review (needs ~30 days post-r228+r241 data; r257 backup pre-dates that window so 0 post-r228 plays available); HR analyzer audit (only 6 graded HR plays in backup, sample too small for diagnostic); generic prop tier=PASS rule unification once all four prop analyzers have empirical floor data.

r265 · Outs threshold tightening — backtest-validated, default ON. Per r242 the live Outs model lost money on both sides (Overs 46.5% / Unders 43.9% WR, both below 52.4% break-even). r264 ran runOutsDiagnostic() which surfaced two distinct broken cohorts via segmentation rather than a single uniform inversion. Floor: 37 of 113 graded Outs plays sit at |edge| < 1.0 with 35.1% WR, -34.0% ROI, -12.59u total — by far the dominant cause of the bleeding. Bucket WR is 12pp below the 47.6% no-skill floor so noise alone cannot explain it. The model emits direction (and recommended:true) for any non-zero edge, so micro-edges below noise floated through to user recommendations and to auto-tracking. ACE cap: 5 of 5 plays where pq='ace' AND |edge| ≥ 2.0 lost (0% WR, -100% ROI). Smaller sample but extreme outcome with theoretical justification — the model overadjusts ace projections (recent IP boost + OP3 wide-ump boost + base PQ default 6.3 IP can compound), and modern pitcher management (pitch counts, third-time-through avoidance, bullpen-first decisions) overrides those depth predictions for the very pitchers the model is most confident about. Cross-check: SOLID × |edge|≥2.0 × Over in the same backtest runs +18.9% ROI, so the issue is specifically with ACE pq class, not all high-edge predictions. Combined backtest: 71 of 113 plays survive (63% retention), 54.9% WR, +1.0% ROI, +0.71u total — flips the cohort from -16.88u to +0.71u (+17.6u swing on a 113-play sample). Mechanism mirrors r228 TB suppression: failed gate → recommended:false + tier:PASS + units:0 (clean three-field strip). Bet object still emitted so analyzer + props page can list it; snapshot still captured so calibration data continues to accumulate. Tunables: window._outsMinEdgeFloor (default 1.0; set 0 to disable floor), window._outsAceEdgeCap (default 2.0; set Infinity to disable cap). Setting both to no-op values reverts to pre-r265 behavior immediately without reload (gate consults window. on each call).

r243 · NegBin-tail K model — backtest-validated, default ON. Replaced the count-difference K analyzer (which decided direction by expectedK − line) with a probability-tail framework using ptbNegBinSurvivalGE. Default VMR=2.5; tunable via window._kDispersionVMR; rollback via window._kUseNegBinTail = false. The empirical case: backtest on 107 graded K entries from r242-era data showed count-diff (current) +3.53% ROI vs NegBin VMR=2.5 +12.14% — a +9pp improvement. Critically, Poisson-tail alone produced −1.13%, so the win is specifically from NegBin's overdispersion correction (real K outcomes have variance > mean because pitchers have correlated good/bad days; Poisson can't capture this, NegBin can). Tier thresholds calibrated from per-bucket data: <2pp = junk (42.9% WR, −27.66% ROI) → PASS; 2-5pp = strong (68.4% WR, +30.55% ROI) → MARG; ≥5pp = strong (60.6% WR, +15.06% ROI) → PLAY. Unit sizing buckets in _propUnitsFor extended with parallel prob-edge scale. Snapshot fields added: kDistribution, kDispersionVmr, kNegBinN, kNegBinP, kModelOverProb, kEdgeProb, overFair. New edgeUnits field on bet objects disambiguates downstream (legacy count-edge consumers fall through correctly via p.edgeUnits === undefined check). No suppression layer — current K model is profitable, not inverted; r243 is a quality lift not a hedge against a known failure (unlike r228's TB suppression). Full live simulation against the backup data showed 90 bets surviving tier, 62.2% WR, +14.97% ROI per unit, +38.46u sized profit (vs +3.78u flat at current count-diff system).

r242 · runDistributionBacktest console utility — Poisson vs NegBin sweep across all four prop models. Companion to r241. Re-scores every graded propsTracking entry under candidate distributions {Poisson, NegBin VMR=1.3/1.5/1.8/2.0} and reports recommendations / wins / WR / ROI per (propType, distribution). Console-only utility (matches runBlendBacktest pattern). Methodology: derive expectedX from p.expectedValue (count mean), recompute modelOverProb under each candidate, derive actualOver from original W/L × original direction, then compute hypothetical newWin = (newDirection matches actualOver). ROI uses the actual stored book odds for whichever side the candidate distribution recommends. Outputs console.table with ★ on best-VMR row. Answers two empirical questions: (1) does NegBin recover TB agree-cohort to ≥47% — i.e., is r228 suppression now safe to lift; and (2) does NegBin help K/Outs/HR too — i.e., should r241 extend beyond TB? Usage: window.runDistributionBacktest() for the full sweep across all four props.

r241 · TB Poisson → Negative Binomial. The actual recalibration of the model that r228 suppressed. Replaced the inline poissonTailGE call inside evaluateBatterTB with ptbNegBinSurvivalGE (already used for live game totals since r104). Parameterization: VMR (variance-to-mean ratio) controls dispersion, default 1.5 from baseball-counts literature. Math: p = 1/VMR, n = expectedTB/(VMR-1) — preserves mean exactly. Pre-ship validation table comparing Poisson vs NegBin showed: at expectedTB=1.5 / line=1.5, Poisson tail = 0.4422 but NegBin(VMR=1.5) tail = 0.4074, NegBin(VMR=2.0) tail = 0.3813 — NegBin pulls 3-7pp of probability mass off the right tail at typical TB scenarios, exactly the inversion-correction mechanism the 35.7% WR cohort needed. Suppression remains default ON — empirical data drove the suppression, theoretical math drove the fix; we wait for ~30 days of post-r241 propsTracking before lifting. Tunable: window._tbDispersionVMR in console (range ~1.01-2.5). Override: window._tbModelSuppressed = false unlocks the model. Snapshot extended with tbDistribution / tbDispersionVmr / tbNegBinN / tbNegBinP for full reproducibility — pre-r241 entries treated as Poisson if backtest re-scores them.

r240 · Closing Line Value tracking for prop bets. Mirrors the existing game-bet CLV pattern: openingOdds captured at track time (both manual and auto-track paths), closingOdds + closing book/Pinnacle pair stamped at autoGrade time via _capturePropClosingOdds helper. Resolves odds per propType from g.propOdds or g.batterPropOdds (matches by line for multi-line TB). Display: inline ↑/↓ CLV pp badge after odds in tracked-prop rows + Avg CLV stat tile in calibration summary card + per-type CLV breakdown. Why CLV matters: it's the leading indicator of edge — a 53% W/L model that beats CLV consistently is finding real price inefficiencies; a 60% W/L model bleeding CLV is just lucky. CLV gives a meaningful signal in ~30 days vs ~300 days for raw W/L. Capture is lazy-at-grade-time (matches game-bet shape), so it works as long as odds were refreshed at some point before game start; misses if user only opens the app post-game once API has dropped the market. Pre-r240 entries gracefully show "—" until they get re-graded.

r238 · Marcel-blend HRPG on the HR analyzer. analyzeBatterHRBets now calls marcelBlend() for HRPG. r236 fetch extended to compute recentHrpg from the existing lastGames[].hr data (sum_HR / recentGames). Uses BLEND_CONFIG.batter_hr (default 40/60 recent/season — HR is rarer than TB so the default weights season more, 30-PA gate). Snapshot adds hrpgSeason, hrpgRecent, hrpgBlended, hrpgSource, recentPA, recentGames, plus weatherMult/pitcherMult/disciplineMult/overFair on batterStatsAtAnalysis for backtest reconstruction. runBlendBacktest() extended with a batter_hr branch — recomputes expectedHR through the full multiplier chain, then probHR = 1−exp(−expectedHR) compared to captured overFair (HR-specific direction because line is always 0.5 but probHR > 0.5 only when expectedHR > ln 2). Live behavior changes — unlike TB which is suppressed, HR is active, so blended HRPG flows into recommendations and tier assignments going forward. barrelPctRecent deferred — needs a separate Statcast Savant fetch path.

r237 · Marcel-blend SLG on the TB analyzer. analyzeBatterTBBets now calls marcelBlend() for SLG, mirroring the K analyzer pattern from r229/r233. Uses recent batter form from r236 (last 15 games SLG) blended with season SLG via BLEND_CONFIG.batter_total_bases (default 50/50, 30-PA gate). Snapshot adds slgSeason, slgRecent, slgBlended, slgSource, recentPA, recentGames on batterStatsAtAnalysis. runBlendBacktest() extended with a batter_total_bases branch — reads the new fields, recomputes expectedTB by multiplying blended SLG through the cached multiplier chain (park × pitcher × weather × BvP × xwOBA × hard-hit × GB) with the live floor=0.3, sweeps the same [0/20/40/60/80/100% recent] ratios. Live tier still PASS — TB suppression (r228) intact. The 35.7%-WR inversion is a Poisson tail problem, not a SLG problem; r237 lays the Marcel groundwork so the queued Poisson→NegBin fix slots in cleanly without snapshot migration.

r236 · Recent batter form fetch (gameLog hydration). Mirrors the pitcher lastStarts pattern for batters — same MLB Stats API endpoint, group=hitting instead of pitching, batched 50 IDs at a time. Fetches each batter's last 15 game splits (~57 PA at 3.8 PA/game, comfortably above the 30-PA gate in BLEND_CONFIG). Computes recentSlg from totalBases/AB sums (weighted by AB so a hot 6-AB game outweighs a quiet 2-AB one), recentHrpg from sum_HR / recentGames (added in r238), plus recentPA, recentAB, recentGames, and a lastGames array. Attached to enriched g.{away,home}BatterStats[] entries. Cache key bumped v1→v2. Consumed live by r237 (TB analyzer) and r238 (HR analyzer).

r233 · BLEND_CONFIG + marcelBlend() utility + runBlendBacktest() harness. Centralized blending utility for season ↔ recent-form weighting across every model. Default config: pitcher_k 60/40, pitcher_outs 60/40, batter_total_bases 50/50, batter_hr 40/60, team_rpg 40/60. Outs snapshot extended with k9Season+k9Recent+ipSeason+ipRecent+ipBlended for backtest. window.runBlendBacktest() sweeps blend ratios [0/20/40/60/80/100% recent], reports leaderboard sorted by hypothetical -110 ROI. Going-forward only — needs ~30 days of post-r233 graded plays for reliable signal.

r229 · Marcel-blend K/9 (60% recent / 40% season). Pre-r229 K model used pure season K/9. Backup analysis showed K is the strongest prop model (57.9% WR on agree cohort, n=79) but season-only weighting lags pitcher form swings. Sample gate: ≥3 starts AND ≥5 IP across recent. Snapshot captures k9Season+k9Recent+k9Blended+k9Source for calibration regression. r231 surfaced PROP MODEL CALIBRATION card on Model page (was hidden on Props page).

r228 · TB model suppression based on empirical analysis. Backup analysis on 305 propsTracking entries showed: P K agrees 57.9% (n=79, working ✓), P Outs agrees 46.2% (n=81, marginal), B TB agrees 35.7% (n=31) — INVERTED vs neutral 47.4%. Pattern is consistent with cross-sport literature on Poisson-tail player props (under-predicts variance, fires "agree on Over" too liberally, reality regresses to under). r228 forces TB tier=PASS, recommended:false, units=0 when window._tbModelSuppressed !== false (default true). Tracking continues so propsTracking accumulates calibration data. UI banner with research-mode toggle. Real fix queued: Poisson → Negative Binomial (r235+).

r222 · Structural opposite-side conflict resolver + same-game concentration cap. Pre-r222 the analyzer could show OVER 57.6% AND UNDER 57.2% on the same line — mathematically impossible. doTotal/doF5Over/doF5Under each computed probabilities independently from book-implied; nothing forced p(over) + p(under) ≤ 1. Fix: after all bet types compute, scan for OVER+UNDER and F5 OVER+F5 UNDER on same line, drop lower-edge side, tag survivor with structural-conflict factor. Plus same-game concentration cap: when 3+ aligned bets remain, scale units by edge rank (1× / 0.75× / 0.5× / floor 0.5u).

r227 · CRITICAL bug — renderHistory filter parity. Discovered that r212/r218/r221 filter work all landed in renderResults() (orphaned page) instead of renderHistory() (the page the user actually navigates to). Two completely separate History UIs existed. r227 ports tier/odds/divergence dropdowns + matchesFilter logic + search box to renderHistory. The "Divergence" filter (✓ Model agrees / ~ Neutral / ⚠ Model diverges) is finally accessible.

r115 · Audit-driven re-grade migration. Found a TB @ CLE play stored as W when CLE actually won 3-2 (TB -1.5 RL = L). Fixed via doubleheader-safe migration that walks days/oddsResearch/patternResearch and corrects miscounts without touching ambiguous DH cases. Net P&L correction -$334.

r113 · RL grader fav/dog detection fix. Pre-r113 the grader had a catch-all that classified ALL post-r106 RL bets as -1.5 fav, silently misgrading +1.5 dog bets. Replaced with proper spread-resolution chain (p.line → game string → betType raw → odds-sign heuristic).

r109 · Against-pattern attribution. Six firing sites (B8, P11, S13, S14, S18, S19) used to push their "X against" branch to firedPatterns, polluting pattern stats. Now only the FOR-bet branch contributes to W/L tally. Migration cleaned legacy data; preship audit Class E permanently flags ternary regressions.

S-TIER 80%+Full edge applied. Min 5u when firing. Historical data is deep and consistent. Bet aggressively.

P6 S-70%GB/Sinker arm → Under modifier. 55%+ career GB rate → -5% over / +5% under. Applies to both starters. Does NOT override in domes or at Coors — still applies. Current confirmed arms: Abbott, Kirby, Pallante, Soriano, Holmes. Fires in: detectPatterns → doTotal + doF5Under

S-70%Pattern 6 — Sinker/GB suppression → under modifier. 5/5 — 100%. 55%+ career GB rate → -5% over / +5% under. Stacks on tier. Applies even at Coors (Sanchez confirmed — Coors went under both games). Confirmed arms: Holmes (60%+), Abbott (58%), Kirby (55%+), Pallante (67%), Soriano (MLB GB leader), Sanchez (PHI). ↳ recalibrated S-70% from observed S-88% — sample stabilization.

S-90%Pattern 3 — ACE vs ACE → UNDER default. Both K/9 10+. Full 100% park+weather cap. 15% min edge for over. NEVER play over below 15% edge. Biggest loss ($400) came from violating this rule. ↳ promoted to S-90% from A-72% — high-conviction signal.

runBlendBacktest(opts) — Re-scores graded prop plays under multiple Marcel blend ratios. Default sweeps [0/20/40/60/80/100% recent]. Reports leaderboard sorted by hypothetical -110 ROI. Skips entries missing season+recent fields (pre-r229 data unusable). Wait ~30 days for ≥30 post-r233 graded plays for reliable signal.
window.runBlendBacktest() // both K and Outs
window.runBlendBacktest({ modelKey: 'pitcher_k' }) // K only
window.runBlendBacktest({ ratios: [0.5, 0.6, 0.7] }) // custom sweep

BLEND_CONFIG — Per-model blend ratios. Read or override at runtime; persists for the session only.
window.BLEND_CONFIG // inspect all 6 entries
window.BLEND_CONFIG.pitcher_k.recent = 0.7 // tune K to 70/30

marcelBlend(input) — Standalone blend utility used by every model. Input: { seasonValue, recentValue, modelKey, sampleSize, denomSize }. Returns { blended, source, weights, gateReason? }.
window.marcelBlend({ seasonValue: 9, recentValue: 12, modelKey: 'pitcher_k', sampleSize: 3, denomSize: 18 })

./run_all.sh — Runs every harness and reports a unified pass/fail tally. Split into READ_ONLY_HARNESSES (parallel-safe) and MUTATING_HARNESSES (smoke alone, runs after) since r220 to eliminate a race where multiple harnesses read polluted index.html during smoke's inject windows. ~38 seconds wall time. Exits non-zero on any failure.
./run_all.sh # standard run
./run_all.sh --verbose # show each harness output

./preship.sh — Hard gate before packaging. Runs every check that could fail silently at deploy time: syntax, all tests, version sync (APP_BUILD vs CACHE_NAME), structural integrity, critical runtime sanity, bundle size drift. 31 of 31 gates green = safe to package. Failure = DO NOT SHIP.
./preship.sh # full run (~50s)
./preship.sh --skip-tests # syntax + structural only (fast)
./preship.sh --verbose # show output of each step