{
 "bucket": "calibration-market",
 "note": "Every number is copied verbatim from a committed artifact (path in source_artifact). This is a calibrated predictor that publishes measured calibration and honest nulls.",
 "entries": [
  {
   "q": "Does the model actually beat the market comparison market?",
   "alt_phrasings": [
    "can you beat the market",
    "is the model better than the market",
    "does it outperform the odds"
   ],
   "a": {
    "status": "ok",
    "answer": "The three recomputed pregame comparisons show no measured improvement over the proportionally devigged reference price. MLB moneyline trails the close: model Brier 0.2429 versus 0.2390 across 13,992 held-out games, with a model-minus-close gap of +0.0039 and a 95% interval of [+0.0028, +0.0051]. Soccer Over/Under 2.5 also trails: 0.2465 versus 0.2390 across 7,558 matches, gap +0.0076, interval [+0.0059, +0.0092]. NBA moneyline is inconclusive at 372 held-out games: 0.1735 versus 0.1672, gap +0.0063, interval [-0.0042, +0.0168]. That NBA interval includes zero but is too wide to establish parity, and the reference price's capture time is unverified.",
    "source_artifact": "docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md",
    "as_of": "2026-09-24"
   },
   "tags": [
    "model-vs-market",
    "efficiency",
    "brier"
   ]
  },
  {
   "q": "Is the market efficient?",
   "alt_phrasings": [
    "are sports market comparison markets efficient",
    "is the closing line sharp",
    "can markets be beaten pregame"
   ],
   "a": {
    "status": "ok",
    "answer": "On every pregame slice measured, yes. A real-data candidate search across four sports and six independent corpora rejected every schedule/fatigue/form/h2h/totals signal, requiring BSS>0 AND Diebold-Mariano p<0.05 AND N>=200 on at least two corpora; nothing cleared the bar. Positive full-sample lifts reversed sign across calendar halves, which is the overfit signature.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "efficiency",
    "signal-rejection",
    "overfit"
   ]
  },
  {
   "q": "Is there anywhere the model out-sharpens the market at all?",
   "alt_phrasings": [
    "where does the model win",
    "does it beat the market anywhere",
    "any spot the model is sharper"
   ],
   "a": {
    "status": "ok",
    "answer": "In 3 of 17 head-to-head checkpoints, all of them MLB late-inning run-totals: end of inning 6 (paired delta 0.6392, n=54), inning 7 (0.7327, n=55) and inning 8 (1.4201, n=49). Every one is flagged MODEL_SHARPER_PROVISIONAL on small samples. Most other rows are underpowered or favor the market.",
    "source_artifact": "webapp/public/data/showcase/cross_sport_scoreboard.json",
    "as_of": "2026-07-22"
   },
   "tags": [
    "model-vs-market",
    "mlb",
    "provisional"
   ]
  },
  {
   "q": "In-game, does the win-probability model beat the market's win probability?",
   "alt_phrasings": [
    "is the live win prob sharper than the market",
    "in-play model vs market",
    "does the model beat live odds"
   ],
   "a": {
    "status": "ok",
    "answer": "No. Across 27,351 MLB rows, model Brier is 0.166759 versus market 0.151008 (BSS versus market -0.1043). Across 4,265 international soccer rows, it is 0.332522 versus 0.195157 (-0.7039). Lower Brier means less probability error. Every published checkpoint also has negative model-versus-market BSS. These are in-game comparisons, not evidence of a market advantage. This descriptive artifact was generated 2026-09-16; its observation window is not published. Counts are graded rows, not games.",
    "source_artifact": "webapp/public/data/showcase/brier_skill_scores.json",
    "as_of": "unknown"
   },
   "tags": [
    "model-vs-market",
    "in-game",
    "bss"
   ]
  },
  {
   "q": "How close is the model to the NBA closing line?",
   "alt_phrasings": [
    "nba model vs close",
    "how good is the nba prediction vs the market",
    "nba moneyline calibration"
   ],
   "a": {
    "status": "ok",
    "answer": "On 372 held-out NBA moneyline games, model Brier is 0.1735 versus 0.1672 for the proportionally devigged reference price. The model-minus-reference gap is +0.0063, with a 95% interval of [-0.0042, +0.0168]. The interval includes zero, but this sample is too small to establish parity or rule out a meaningful deficit. The price is labeled a close in the study, although its capture time is unverified.",
    "source_artifact": "docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md",
    "as_of": "2026-09-24"
   },
   "tags": [
    "nba",
    "model-vs-market",
    "closing-line"
   ]
  },
  {
   "q": "How does the model do against the MLB moneyline close?",
   "alt_phrasings": [
    "mlb model vs market",
    "mlb moneyline calibration",
    "baseball prediction vs the odds"
   ],
   "a": {
    "status": "ok",
    "answer": "The MLB moneyline model trails the proportionally devigged close on 13,992 held-out games. Model Brier is 0.2429 versus 0.2390 for the close; the model-minus-close gap is +0.0039, with a 95% interval of [+0.0028, +0.0051]. The interval stays above zero, so the model has higher measured Brier error on this sample.",
    "source_artifact": "docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md",
    "as_of": "2026-09-24"
   },
   "tags": [
    "mlb",
    "model-vs-market",
    "closing-line"
   ]
  },
  {
   "q": "Does the model match the market on soccer?",
   "alt_phrasings": [
    "soccer over under vs market",
    "soccer model calibration pregame",
    "football o/u vs the close"
   ],
   "a": {
    "status": "ok",
    "answer": "For soccer Over/Under 2.5, the model trails the proportionally devigged close on 7,558 held-out matches. Model Brier is 0.2465 versus 0.2390 for the close; the model-minus-close gap is +0.0076, with a 95% interval of [+0.0059, +0.0092]. The interval stays above zero, so the close is measurably sharper on this market and sample.",
    "source_artifact": "docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md",
    "as_of": "2026-09-24"
   },
   "tags": [
    "soccer",
    "model-vs-market",
    "totals"
   ]
  },
  {
   "q": "How does the model do on tennis pregame versus the market?",
   "alt_phrasings": [
    "tennis match-win vs the close",
    "atp prediction vs market",
    "tennis calibration pregame"
   ],
   "a": {
    "status": "ok",
    "answer": "It trails slightly: on ATP match-win the model's Brier is 0.2177 versus the close's 0.2028 over 7,374 matches, a gap of +0.0149, labeled BEHIND. The gap is attributed to same-day freshness (lineups, conditions) the public model cannot see, not a model defect.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "tennis",
    "model-vs-market",
    "freshness"
   ]
  },
  {
   "q": "Does the model keep up with the market on totals?",
   "alt_phrasings": [
    "over under model vs market",
    "totals calibration",
    "does the model match the total line"
   ],
   "a": {
    "status": "ok",
    "answer": "Totals trail the close by a freshness gap. NBA total RMSE is 19.17 for the model versus 18.11 for the close (n=372, gap +1.06); MLB total RMSE is 4.72 versus 4.44 (n=1,679, gap +0.28). Both are labeled BEHIND, attributed to injuries/lineups/pitcher/park/weather the market prices and the model cannot see.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "totals",
    "model-vs-market",
    "freshness"
   ]
  },
  {
   "q": "When the model and the market disagree, who tends to be right in MLB?",
   "alt_phrasings": [
    "who wins on disagreement",
    "is the model right when it differs from the market",
    "model vs market when they split mlb"
   ],
   "a": {
    "status": "ok",
    "answer": "The market. On the largest MLB disagreements (>=0.10 probability apart, n=33,402) the model is the closer forecaster only 37.7% of the time, with model Brier 0.2827 versus market Brier 0.2103. Bigger disagreement means the market is usually the one that is right.",
    "source_artifact": "webapp/public/data/showcase/market_disagreement_profile.json",
    "as_of": "unknown"
   },
   "tags": [
    "model-vs-market",
    "disagreement",
    "mlb"
   ]
  },
  {
   "q": "What happens on the biggest model-vs-market disagreements in soccer?",
   "alt_phrasings": [
    "soccer disagreement who is right",
    "does the model win when it differs on soccer",
    "big soccer disagreement outcome"
   ],
   "a": {
    "status": "ok",
    "answer": "The market wins even more decisively than in MLB. On the largest soccer disagreements (>=0.10 apart, n=4,406) the model is closer only 21.5% of the time, with model Brier 0.3306 versus market Brier 0.1720. Wide disagreement is a signal the model is off, not that it found something.",
    "source_artifact": "webapp/public/data/showcase/market_disagreement_profile.json",
    "as_of": "unknown"
   },
   "tags": [
    "model-vs-market",
    "disagreement",
    "soccer"
   ]
  },
  {
   "q": "Why doesn't being accurate mean beating the market?",
   "alt_phrasings": [
    "accuracy vs advantage",
    "why is calibration not enough",
    "does higher accuracy beat the odds"
   ],
   "a": {
    "status": "ok",
    "answer": "Accuracy alone does not establish that forecast probabilities beat market probabilities. In this published MLB snapshot, model/market Brier is 0.166759/0.151008. The 10-bin Murphy components show a reliability gap of +0.002194 and a resolution gap of -0.014875; the larger resolution difference describes discrimination in these bins, not a causal explanation or proof that recalibration cannot help. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/murphy_decomposition.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "resolution",
    "murphy"
   ]
  },
  {
   "q": "Are the predictions actually well calibrated?",
   "alt_phrasings": [
    "is the model calibrated",
    "how calibrated is it",
    "do the probabilities match reality"
   ],
   "a": {
    "status": "ok",
    "answer": "Calibration varies by sport and probability bin. In MLB (27,351 rows, 178 games), 3/10 eligible model bins have pointwise 95% game-cluster-bootstrap gap intervals excluding zero, versus 0/10 market bins. In soccer (4,265 rows, 27 games), the counts are 8/9 model and 7/9 market bins. These pointwise intervals are not a joint guarantee that all other bins are calibrated. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/calibration_stability.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "reliability",
    "model-vs-market"
   ]
  },
  {
   "q": "Is the model overconfident?",
   "alt_phrasings": [
    "does it overstate certainty",
    "is it too confident at the extremes",
    "overconfidence at high probabilities"
   ],
   "a": {
    "status": "ok",
    "answer": "The current MLB 0.9-1.0 bin does not establish high-end overconfidence: 1,437 rows across 61 games have mean predicted probability 0.9516, observed frequency 0.9798, and outcome-minus-prediction gap +0.0282. Its pointwise 95% game-cluster-bootstrap gap interval [-0.0063, 0.0488] includes zero, so this bin alone establishes neither overconfidence, underconfidence, nor calibration. Other bins differ; see the reliability curves. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/calibration_stability.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "overconfidence",
    "mlb"
   ]
  },
  {
   "q": "What is a Brier score and what is the model's?",
   "alt_phrasings": [
    "explain brier score",
    "what brier does the model get",
    "how is probability accuracy measured"
   ],
   "a": {
    "status": "ok",
    "answer": "A Brier score is the mean squared error of forecast probabilities against binary outcomes; lower means less probability error, but does not by itself isolate calibration. In this in-game snapshot, MLB model Brier is 0.166759 versus market 0.151008 across 27,351 rows. International soccer model Brier is 0.332522 versus market 0.195157 across 4,265 rows. A constant 0.5 forecast scores 0.25 for any binary outcome balance. This descriptive artifact was generated 2026-09-16; its observation window is not published. Counts are graded rows, not games.",
    "source_artifact": "webapp/public/data/showcase/brier_skill_scores.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "brier",
    "definition"
   ]
  },
  {
   "q": "How do the model's and market's ECE compare?",
   "alt_phrasings": [
    "expected calibration error model vs market",
    "ece comparison",
    "which is better calibrated by ece"
   ],
   "a": {
    "status": "ok",
    "answer": "For MLB in-game moneyline, model/market 10-bin ECE is 0.057367/0.049057 across 27,351 graded rows and 178 files. For soccer match result it is 0.400236/0.281866 across 4,265 rows and 27 files. The market has lower ECE in both published corpora; the soccer sample is smaller, and these descriptive scores do not establish lasting improvement. Files here are source files, not a count of games. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/calibration_by_market_type.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "ece",
    "model-vs-market"
   ]
  },
  {
   "q": "Is calibration improving month over month?",
   "alt_phrasings": [
    "is it getting better over time",
    "calibration trend",
    "does calibration improve across months"
   ],
   "a": {
    "status": "ok",
    "answer": "The exhibit reports only 2026-06 and 2026-07 groups, not a durable trend. MLB model/market Brier moves from 0.1968/0.1873 (5,030 rows) to 0.1600/0.1428 (22,321 rows); model/market ECE moves from 0.1265/0.1210 to 0.0750/0.0439. Soccer model/market Brier moves from 0.3273/0.1955 (1,213 rows) to 0.3346/0.1950 (3,052 rows); ECE moves from 0.3644/0.2596 to 0.4243/0.2952. Different month groups do not isolate model improvement. Published snapshot dated 2026-09-16; the month groups are reported, but exact first and last observation dates are not published.",
    "source_artifact": "webapp/public/data/showcase/calibration_over_time.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "trend",
    "over-time"
   ]
  },
  {
   "q": "How good is the soccer model's calibration?",
   "alt_phrasings": [
    "soccer calibration quality",
    "is the soccer model calibrated",
    "football in-game calibration"
   ],
   "a": {
    "status": "ok",
    "answer": "In the 4,265-row soccer pack, model/market Brier is 0.332522/0.195157. The pack's minute-bucket weighted ECE is 0.3748 over 3,383 minute-bucket rows, a subset of the 4,265 headline rows. This aggregation differs from the ECE in other exhibits. The pack does not state a game count; single-fold reads are descriptive and not durable. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/soccer_calibration_pack.json",
    "as_of": "unknown"
   },
   "tags": [
    "soccer",
    "calibration",
    "small-sample"
   ]
  },
  {
   "q": "Is the model-vs-market gap fixable by recalibrating?",
   "alt_phrasings": [
    "can recalibration close the gap",
    "would isotonic fixing help",
    "is the shortfall calibration or information"
   ],
   "a": {
    "status": "ok",
    "answer": "That is not established. In the 10-bin Murphy decomposition, MLB model-minus-market reliability is +0.002194 while resolution is -0.014875, so the observed MLB gap is mainly in resolution. Soccer reliability is +0.103894 and resolution is -0.032549, so its observed gap is mainly in reliability. Recalibration might address some reliability error, but this exhibit does not test a recalibration method or prove how much Brier would improve. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/murphy_decomposition.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "resolution",
    "recalibration"
   ]
  },
  {
   "q": "Does the model beat a simple base-rate baseline?",
   "alt_phrasings": [
    "is it better than climatology",
    "does it beat the base rate",
    "bss vs climatology"
   ],
   "a": {
    "status": "ok",
    "answer": "For MLB, yes: versus the published constant sport base-rate climatology, model BSS is +0.3327 and market BSS is +0.3957 across 27,351 rows. For international soccer, no: model BSS is -0.8569 and market BSS is -0.0898 across 4,265 rows; neither improves that published climatology reference. BSS above zero means lower Brier than the named reference, not superiority to a different reference. This descriptive artifact was generated 2026-09-16; its observation window is not published. Counts are graded rows, not games.",
    "source_artifact": "webapp/public/data/showcase/brier_skill_scores.json",
    "as_of": "unknown"
   },
   "tags": [
    "calibration",
    "bss",
    "baseline"
   ]
  },
  {
   "q": "Does conditioning on the in-game state actually sharpen the forecast?",
   "alt_phrasings": [
    "does in-game conditioning help",
    "is the live model sharper than pregame",
    "does mid-game state improve calibration"
   ],
   "a": {
    "status": "ok",
    "answer": "Yes, this is the one measured calibration win. Fusing the pregame rating prior with the realized mid-game state improves win-probability Brier from 0.209 to 0.159 in NBA and 0.241 to 0.126 in MLB on real out-of-sample corpora. It measures calibration: a live market source sees the same score, so no market-comparison test applies.",
    "source_artifact": "docs/INGAME_PROOF.md",
    "as_of": "unknown"
   },
   "tags": [
    "in-game",
    "calibration",
    "conditioning"
   ]
  },
  {
   "q": "How much of the in-game improvement is the model versus just the scoreboard?",
   "alt_phrasings": [
    "is the live lift real skill or mechanical",
    "how much does the model prior add in-game",
    "scoreboard vs model contribution"
   ],
   "a": {
    "status": "ok",
    "answer": "Most of it is the scoreboard. A three-arm split against a rating-blind score-only baseline shows NBA going 0.209 (prior only) to 0.172 (score-only) to 0.159 (combined), so the model's own prior adds the last ~0.014 Brier (~27%) over watching the score. In MLB the model prior adds only ~0.001 (~1%); by mid-game the scoreboard washes it out.",
    "source_artifact": "docs/INGAME_PROOF.md",
    "as_of": "unknown"
   },
   "tags": [
    "in-game",
    "attribution",
    "mechanical"
   ]
  },
  {
   "q": "What is the honest end-of-Q3 win-probability Brier?",
   "alt_phrasings": [
    "endq3 brier leak free",
    "third quarter win prob accuracy",
    "q3 brier number"
   ],
   "a": {
    "status": "ok",
    "answer": "The leak-free walk-forward end-of-Q3 win-probability Brier is about 0.141, measured after removing two 4th-quarter-derived features that had let an earlier model peek at the quarter it was predicting. A controlled A/B put the leak's inflation at about 4% relative. This is framed as a leak caught in the pipeline, not a competitive number.",
    "source_artifact": "docs/JOB_EVIDENCE_PACKET.md",
    "as_of": "2026-07-23"
   },
   "tags": [
    "in-game",
    "leak",
    "brier"
   ]
  },
  {
   "q": "What did the full-season win-probability backtest find versus the close?",
   "alt_phrasings": [
    "season backtest vs close",
    "walk forward win prob result",
    "full season market efficiency proof"
   ],
   "a": {
    "status": "ok",
    "answer": "Correction published 2026-09-14: the earlier full-season 0.208 model / 0.198 close pair could not be reproduced and has no supporting artifact in the public evidence-packet audit. It is withdrawn and should not be used as performance proof. See docs/JOB_EVIDENCE_PACKET.md.",
    "source_artifact": "docs/JOB_EVIDENCE_PACKET.md",
    "as_of": "2026-09-14"
   },
   "tags": [
    "backtest",
    "efficiency",
    "win-probability"
   ]
  },
  {
   "q": "How good is the win-probability model on its own?",
   "alt_phrasings": [
    "win prob model accuracy",
    "standalone win probability quality",
    "how accurate is the win prob stack"
   ],
   "a": {
    "status": "ok",
    "answer": "The 5-way NNLS win-probability stack scores 0.709 accuracy and 0.193 Brier in 3-fold walk-forward validation. That is a solid, well-calibrated forecaster; it just does not beat sharp closing lines, which is a separate and harder bar.",
    "source_artifact": "docs/JOB_EVIDENCE_PACKET.md",
    "as_of": "2026-07-23"
   },
   "tags": [
    "win-probability",
    "calibration",
    "accuracy"
   ]
  },
  {
   "q": "Which sport has the best-calibrated market probabilities?",
   "alt_phrasings": [
    "which market is sharpest",
    "lowest brier market",
    "most predictable sport by the market"
   ],
   "a": {
    "status": "ok",
    "answer": "This snapshot cannot rank calibration alone across sports: Brier measures total probability error and the outcomes, base rates, and row sets differ. MLB market Brier is 0.151008 on 27,351 rows; international soccer market Brier is 0.195157 on 4,265 rows. The MLB value is lower for these graded outcomes, but that does not establish better calibration. Market Brier is lower than model Brier in each sport. This descriptive artifact was generated 2026-09-16; its observation window is not published. Counts are graded rows, not games.",
    "source_artifact": "webapp/public/data/showcase/brier_skill_scores.json",
    "as_of": "unknown"
   },
   "tags": [
    "predictability",
    "brier",
    "cross-sport"
   ]
  },
  {
   "q": "Is the model tighter to the close in NBA or MLB?",
   "alt_phrasings": [
    "nba vs mlb which is closer to market",
    "which sport does the model track best",
    "closest match to the close"
   ],
   "a": {
    "status": "ok",
    "answer": "The measured moneyline Brier gap is +0.0039 for MLB (13,992 held-out games; 95% interval [+0.0028, +0.0051]) and +0.0063 for NBA (372 held-out games; interval [-0.0042, +0.0168]) against proportionally devigged reference prices. MLB trails its close; the NBA interval includes zero but is too wide to establish parity. These different sport samples do not establish which model is closer overall, and the NBA reference price capture time is unverified.",
    "source_artifact": "docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md",
    "as_of": "2026-09-24"
   },
   "tags": [
    "nba",
    "mlb",
    "closing-line"
   ]
  },
  {
   "q": "Does a tennis prior trained on one tour transfer to the other?",
   "alt_phrasings": [
    "atp to wta transfer",
    "does the tennis model generalize across tours",
    "cross corpus tennis calibration"
   ],
   "a": {
    "status": "ok",
    "answer": "The published receipt reports lower held-out Brier in both cross-tour tests: 0.1698 (state-time base) to 0.1597 (prior), and 0.1550 (state-time base) to 0.1432 (prior); lower is better. Its verdict is REPLICATED. However, its pregame direction labels and in-game tour counts conflict, so these pairs cannot be confidently assigned to ATP-trained versus WTA-trained transfer. The observation window and snapshot date are not published. These are Brier comparisons with a fitted state-time baseline; no in-play odds join exists.",
    "source_artifact": "webapp/public/data/showcase/tennis_showcase.json",
    "as_of": "unknown"
   },
   "tags": [
    "tennis",
    "cross-corpus",
    "calibration"
   ]
  },
  {
   "q": "Did the surface-specific tennis prior beat the surface-blind one?",
   "alt_phrasings": [
    "does court surface improve the tennis model",
    "surface specific hold prior result",
    "clay grass hard prior test"
   ],
   "a": {
    "status": "ok",
    "answer": "No, that one was rejected. A surface-specific hold prior did not beat the surface-blind baseline on either tour: ATP pooled Brier 0.165518 (surface) versus 0.165397 (blind), WTA 0.153053 versus 0.150561. The planted null died as designed, verdict REJECT.",
    "source_artifact": "webapp/public/data/showcase/tennis_showcase.json",
    "as_of": "unknown"
   },
   "tags": [
    "tennis",
    "signal-rejection",
    "surface"
   ]
  },
  {
   "q": "Is the NBA getting less competitive?",
   "alt_phrasings": [
    "nba parity trend",
    "are nba teams more lopsided",
    "league balance over seasons"
   ],
   "a": {
    "status": "ok",
    "answer": "Slightly less balanced across the three seasons on disk. The win-share Gini rose from 0.177 in 2023-24 (most balanced) to 0.2004 in 2025-26 (least balanced). This is descriptive of seasons already played, not a forecast, and it pools any playoff games that inflate win concentration.",
    "source_artifact": "webapp/public/data/showcase/league_parity_index.json",
    "as_of": "2026-04-12"
   },
   "tags": [
    "nba",
    "parity",
    "descriptive"
   ]
  },
  {
   "q": "How many NBA games are decided by five points or fewer?",
   "alt_phrasings": [
    "share of close nba games",
    "how often are nba games tight",
    "close game rate nba"
   ],
   "a": {
    "status": "ok",
    "answer": "Between 23.3% and 25.5% of games across the three seasons measured (23.3% in 2023-24, 25.2% in 2024-25, 25.5% in 2025-26), with mean absolute final margins of 12.58, 12.71 and 13.03 points. Roughly a quarter of games are one-possession-tight at the buzzer.",
    "source_artifact": "webapp/public/data/showcase/league_parity_index.json",
    "as_of": "2026-04-12"
   },
   "tags": [
    "nba",
    "margins",
    "descriptive"
   ]
  },
  {
   "q": "Which sport has the widest model-vs-market calibration gap?",
   "alt_phrasings": [
    "biggest gap to the market",
    "where is the model furthest behind",
    "worst sport for the model"
   ],
   "a": {
    "status": "ok",
    "answer": "Among the two sports with published 10-bin Murphy components, soccer has the larger reliability (calibration) gap: model-minus-market +0.103894 versus MLB +0.002194. The total Brier gaps are different quantities: model/market Brier is 0.332522/0.195157 in soccer and 0.166759/0.151008 in MLB. This is a ranking of these corpora, not a universal cross-sport ordering; soccer has only 4,265 rows in this decomposition. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/murphy_decomposition.json",
    "as_of": "unknown"
   },
   "tags": [
    "cross-sport",
    "calibration",
    "soccer"
   ]
  },
  {
   "q": "Do in-game prices overreact to score changes?",
   "alt_phrasings": [
    "do live markets overreact",
    "price overshoot after a move",
    "market overreaction test"
   ],
   "a": {
    "status": "ok",
    "answer": "It depends on the sport. After an in-game price move, MLB prices land a few points above the eventual outcome (moved-to-minus-outcome +0.0801 to +0.0518 across move sizes, a small overshoot), while soccer prices sit well below it (-0.1945 to -0.3648, an undershoot). This is a documented structural description labeled INCREMENTAL, and some soccer buckets are tiny (n=21).",
    "source_artifact": "webapp/public/data/showcase/market_overreaction.json",
    "as_of": "unknown"
   },
   "tags": [
    "overreaction",
    "in-game",
    "structure"
   ]
  },
  {
   "q": "Does the market get sharper as an MLB game goes on?",
   "alt_phrasings": [
    "does the market improve through the game",
    "in-game market sharpening mlb",
    "who pulls ahead late in a game"
   ],
   "a": {
    "status": "ok",
    "answer": "Yes, and its lead over the model widens. The market's Brier advantage over the model grows from -0.015 at the first checkpoint to -0.0902 by the ninth inning. As information arrives, the market absorbs it faster than the model does.",
    "source_artifact": "webapp/public/data/showcase/info_arrival_curve.json",
    "as_of": "unknown"
   },
   "tags": [
    "in-game",
    "information",
    "mlb"
   ]
  },
  {
   "q": "Does the model commit to an outcome faster than the market?",
   "alt_phrasings": [
    "does the model get confident faster",
    "entropy decay model vs market",
    "who hedges longer"
   ],
   "a": {
    "status": "ok",
    "answer": "In MLB, yes, and that is a warning sign. By the ninth checkpoint the model's forecast uncertainty has collapsed to 0.3656 bits while the market still sits at 0.7244, and their disagreement widens from 0.07 early to 0.2791 late. Paired with the overconfidence seen in the calibration exhibits, growing certain faster than the market is a flag, not a strength.",
    "source_artifact": "webapp/public/data/showcase/market_convergence.json",
    "as_of": "unknown"
   },
   "tags": [
    "in-game",
    "entropy",
    "overconfidence"
   ]
  },
  {
   "q": "Does the closing line get sharper than earlier lines?",
   "alt_phrasings": [
    "is the close sharper than earlier",
    "does the line tighten toward tip off",
    "closing line sharpening"
   ],
   "a": {
    "status": "ok",
    "answer": "In the WNBA sample it does: the devigged market Brier sharpens from 0.2454 at T-6h to 0.2280 at the close (delta +0.0174 over 30 games). This is the expected pattern of a market improving as game time approaches, though the window is a single capture and flagged provisional.",
    "source_artifact": "webapp/public/data/showcase/micro_closing_decay.json",
    "as_of": "2026-07-17"
   },
   "tags": [
    "closing-line",
    "wnba",
    "sharpening"
   ]
  },
  {
   "q": "Does the soccer line sharpen toward the close too?",
   "alt_phrasings": [
    "soccer closing line behavior",
    "does the football line tighten",
    "soccer close vs early"
   ],
   "a": {
    "status": "ok",
    "answer": "Not in this sample, but it is underpowered. Soccer Brier went from 0.1355 at T-24h to 0.1374 at the close (delta -0.0019 over just 13 games), so no sharpening is visible and the read is flagged UNDERPOWERED. Too few games to conclude either way.",
    "source_artifact": "webapp/public/data/showcase/micro_closing_decay.json",
    "as_of": "2026-07-17"
   },
   "tags": [
    "closing-line",
    "soccer",
    "underpowered"
   ]
  },
  {
   "q": "Is closing-line value real?",
   "alt_phrasings": [
    "is clv a real thing",
    "is the close sharper than the open",
    "does the line move toward truth"
   ],
   "a": {
    "status": "ok",
    "answer": "Yes, as a market phenomenon: on MLB the close beats the open on log-loss (Diebold-Mariano p=0.0010, N=27,975), with the close carrying a Brier skill of +0.43% over the open. The line does move toward the truth from open to close. That sharpening belongs to the market, not to a retrospective model.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "clv",
    "closing-line",
    "mlb"
   ]
  },
  {
   "q": "Can a model capture the open-to-close line move?",
   "alt_phrasings": [
    "can you predict the line move",
    "does the model see the close coming",
    "harvest clv with a model"
   ],
   "a": {
    "status": "ok",
    "answer": "No. A leak-free open-time model has essentially zero correlation with the realized open-to-close move: correlation of (model probability minus opener) with (close minus open) is +0.0038, 95% CI [-0.046, +0.055], i.e. indistinguishable from zero. The move is same-day information/speed, not a cleverer feature a retrospective model can extract.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "clv",
    "line-movement",
    "efficiency"
   ]
  },
  {
   "q": "Is the in-game overreaction finding novel?",
   "alt_phrasings": [
    "is this a new discovery",
    "has overreaction been shown before",
    "novelty of the market tests"
   ],
   "a": {
    "status": "ok",
    "answer": "No, and it says so. The in-game overreaction result is labeled INCREMENTAL because it replicates Moskowitz-style reversion tests, and the information-arrival instrument closely resembles a June 2026 prediction-market paper (arXiv 2606.07811, 409,512 rows, finding 0.64-for-one contemporaneous absorption). These are honest replications on the corpora on disk, not first-ever claims.",
    "source_artifact": "webapp/public/data/showcase/info_arrival_curve.json",
    "as_of": "unknown"
   },
   "tags": [
    "novelty",
    "overreaction",
    "literature"
   ]
  },
  {
   "q": "Do you publish your failed hypotheses?",
   "alt_phrasings": [
    "do you show your nulls",
    "how many null results",
    "do you hide the misses"
   ],
   "a": {
    "status": "ok",
    "answer": "Yes, and they outnumber the wins. Across five validation ledgers there are 351 nulls versus 168 confirms, a 2.1x ratio: NBA 45 confirmed / 31 null, MLB 32 / 37, soccer 29 / 24, tennis 24 / 20, and the interaction factory 38 confirmed / 239 null / 709 not-testable. A system that only ever confirmed would not be credible.",
    "source_artifact": "webapp/public/data/showcase/honesty_exhibit.json",
    "as_of": "2026-07-22"
   },
   "tags": [
    "honesty",
    "nulls",
    "methodology"
   ]
  },
  {
   "q": "How many candidate market signals were rejected?",
   "alt_phrasings": [
    "how many signals failed the gate",
    "reject rate of candidates",
    "did signals get thrown out"
   ],
   "a": {
    "status": "ok",
    "answer": "Every pregame candidate in the search was rejected across at least two corpora, and the two-way interaction factory alone logged 239 nulls, 709 not-testable and only 38 confirmed out of 1,003 candidates. A large null/not-testable majority is the gate doing its job, not a shortfall.",
    "source_artifact": "webapp/public/data/showcase/honesty_exhibit.json",
    "as_of": "2026-07-22"
   },
   "tags": [
    "signal-rejection",
    "gate",
    "honesty"
   ]
  },
  {
   "q": "Did any signal look good full-sample then fall apart?",
   "alt_phrasings": [
    "did a signal overfit",
    "sign flipping signals",
    "any signal reverse out of sample"
   ],
   "a": {
    "status": "ok",
    "answer": "Yes, and catching them is the point. NBA schedule signals (back-to-back, three-in-four, rest, travel) showed positive full-season lifts that reversed sign across the two calendar halves, the classic overfit signature. Since the schedule is known long before any line moves, the whole sample is a clean test bed, and the gate rejected them.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "overfit",
    "sign-flip",
    "nba"
   ]
  },
  {
   "q": "What is the single measured win in the whole system?",
   "alt_phrasings": [
    "what actually works",
    "the one real result",
    "biggest genuine finding"
   ],
   "a": {
    "status": "ok",
    "answer": "In-game conditioning. Fusing the pregame rating prior with the realized mid-game state sharpens win-probability calibration (NBA Brier 0.209 to 0.159, MLB 0.241 to 0.126) on real out-of-sample corpora. A live market source sees the same score, so this is reported as a calibration result.",
    "source_artifact": "docs/INGAME_PROOF.md",
    "as_of": "unknown"
   },
   "tags": [
    "in-game",
    "calibration",
    "headline"
   ]
  },
  {
   "q": "Does the in-game model beat the market's NBA win probability at the end of Q1?",
   "alt_phrasings": [
    "nba end q1 model vs market",
    "first quarter win prob comparison",
    "early nba in-game vs market"
   ],
   "a": {
    "status": "ok",
    "answer": "No, the market is sharper. At end-of-Q1 the paired model-minus-market Brier delta is -0.0084 (95% CI [-0.0161, -0.0008], n=1,592), verdict MARKET_SHARPER_PROVISIONAL. Later NBA checkpoints are underpowered with intervals spanning zero.",
    "source_artifact": "webapp/public/data/showcase/cross_sport_scoreboard.json",
    "as_of": "2026-07-22"
   },
   "tags": [
    "nba",
    "in-game",
    "model-vs-market"
   ]
  },
  {
   "q": "How big are the corpora behind these calibration numbers?",
   "alt_phrasings": [
    "sample size of the calibration data",
    "how many games and rows",
    "corpus size"
   ],
   "a": {
    "status": "ok",
    "answer": "The calibration-stability exhibit has 27,351 MLB graded rows across 178 games and 4,265 soccer graded rows across 27 games. The 27-game soccer sample is limited; read its 8/9 model and 7/9 market eligible-bin counts cautiously. The 95% intervals are pointwise game-cluster-bootstrap intervals, not joint guarantees. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/calibration_stability.json",
    "as_of": "unknown"
   },
   "tags": [
    "sample-size",
    "corpus",
    "methodology"
   ]
  },
  {
   "q": "Are single-fold results trustworthy?",
   "alt_phrasings": [
    "can you trust one split",
    "is a single window enough",
    "why require multiple corpora"
   ],
   "a": {
    "status": "ok",
    "answer": "The soccer pack reports 4,265 rows and explicitly warns that single-fold reads are not durable. It does not publish a game count or a cross-corpus replication test. A result from one fold can describe that fold, but these source values alone cannot establish stability or a general model improvement. Published snapshot dated 2026-09-16; the observation window is not published.",
    "source_artifact": "webapp/public/data/showcase/soccer_calibration_pack.json",
    "as_of": "unknown"
   },
   "tags": [
    "methodology",
    "replication",
    "multi-corpus"
   ]
  },
  {
   "q": "Does the market already price schedule facts like back-to-backs and rest?",
   "alt_phrasings": [
    "is rest already in the line",
    "does the market know about b2b",
    "schedule advantage in the odds"
   ],
   "a": {
    "status": "ok",
    "answer": "Yes. NBA back-to-back, three-in-four, rest, travel, altitude and home-court probes all rejected against the devigged close (home-court advantage was fully absorbed in the devig). The market already prices deterministic calendar physics, so these known facts are not exploitable information.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "schedule",
    "efficiency",
    "nba"
   ]
  },
  {
   "q": "How does the model's pregame MLB run-total forecast compare to the market?",
   "alt_phrasings": [
    "mlb total runs crps vs market",
    "pregame totals model vs market baseball",
    "run total sharpness pregame"
   ],
   "a": {
    "status": "ok",
    "answer": "Statistically tied on the data available. On MLB pregame total-runs the paired CRPS delta versus the market is -0.0523 (95% CI [-0.1657, 0.0482], n=300), verdict UNDERPOWERED, with 80% interval coverage of 0.75 for the model versus 0.81 for the market. No difference is established either way.",
    "source_artifact": "webapp/public/data/showcase/cross_sport_scoreboard.json",
    "as_of": "2026-07-22"
   },
   "tags": [
    "mlb",
    "totals",
    "crps"
   ]
  },
  {
   "q": "Why is the MLB run-total Brier left blank in the market-type breakdown?",
   "alt_phrasings": [
    "why no brier for mlb totals",
    "unscored totals market",
    "null brier by design"
   ],
   "a": {
    "status": "ok",
    "answer": "Because there is no resolved outcome to score against. The MLB run-total quote-tracks carry 2,113 rescued in-game ticks but no settlement join, so scoring them would mean inventing labels; the Brier is reported null by design and only the model-vs-market divergence (median 0.116128) is disclosed. Refusing to fabricate a score is the discipline being demonstrated.",
    "source_artifact": "webapp/public/data/showcase/calibration_by_market_type.json",
    "as_of": "unknown"
   },
   "tags": [
    "honesty",
    "totals",
    "no-fabrication"
   ]
  },
  {
   "q": "Does in-game conditioning let you beat a live book?",
   "alt_phrasings": [
    "can you beat live in-play odds",
    "is the in-game win an advantage",
    "does the live model beat the live market"
   ],
   "a": {
    "status": "ok",
    "answer": "No. The in-game Brier improvement (NBA 0.209 to 0.159, MLB 0.241 to 0.126) is forecaster sharpness on a real out-of-sample corpus, but a live market source sees the same score at the same time. It is a calibration measurement.",
    "source_artifact": "docs/INGAME_PROOF.md",
    "as_of": "unknown"
   },
   "tags": [
    "in-game",
    "calibration",
    "no-comparative-conclusion"
   ]
  },
  {
   "q": "Where is the in-game model most miscalibrated?",
   "alt_phrasings": [
    "worst calibrated in-game bucket",
    "where does the model fail most",
    "biggest live calibration error"
   ],
   "a": {
    "status": "ok",
    "answer": "In late soccer states at low probabilities. The worst-ranked bucket is soccer minute 75-90+, predicted-probability band 0.2-0.4, where the model's mean prediction was 0.2117 but the outcome rate was 1.0 (calibration error 0.7883, n=10). Late-game low-probability soccer is where the model is furthest from reality relative to the market.",
    "source_artifact": "webapp/public/data/showcase/state_conditioned_calibration.json",
    "as_of": "unknown"
   },
   "tags": [
    "in-game",
    "calibration",
    "soccer"
   ]
  },
  {
   "q": "Does the model or the market misfire more on late MLB favorites?",
   "alt_phrasings": [
    "late inning heavy favorites calibration",
    "mlb late favorite miscalibration",
    "who is off on strong late leads"
   ],
   "a": {
    "status": "ok",
    "answer": "The model. In late MLB innings (7+) for the 0.8-1.0 band the model predicted 0.9211 but teams only won 0.6853 of the time (calibration error 0.2357, n=5,037), while the market's same-band error was just 0.0206. The model badly overrates late favorites relative to the market.",
    "source_artifact": "webapp/public/data/showcase/state_conditioned_calibration.json",
    "as_of": "unknown"
   },
   "tags": [
    "mlb",
    "in-game",
    "overconfidence"
   ]
  },
  {
   "q": "Should I selection the model's pick when it disagrees with the market?",
   "alt_phrasings": [
    "is it a selection when model differs from market",
    "fade the market on disagreement",
    "act on model vs market gap"
   ],
   "a": {
    "status": "refused",
    "answer": "This is a calibrated predictor. On the largest model-vs-market disagreements, the market is the more accurate side: the model is closer only 37.7% of the time in MLB and 21.5% in soccer.",
    "source_artifact": "webapp/public/data/showcase/market_disagreement_profile.json",
    "as_of": "unknown"
   },
   "tags": [
    "refused",
    "no-market comparison-advice",
    "disagreement"
   ]
  },
  {
   "q": "What return measurement or return does this system produce?",
   "alt_phrasings": [
    "how much money does it make",
    "what is the return on investment",
    "is it successful"
   ],
   "a": {
    "status": "refused",
    "answer": "The cited recomputation measures Brier probability error, not realized returns. MLB moneyline and soccer Over/Under 2.5 trail their proportionally devigged close references; the NBA moneyline result is inconclusive at 372 held-out games. It provides no realized-return measurement.",
    "source_artifact": "docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md",
    "as_of": "2026-09-24"
   },
   "tags": [
    "refused",
    "no-return-measure",
    "efficiency"
   ]
  },
  {
   "q": "Give me a successful strategy based on the calibration findings.",
   "alt_phrasings": [
    "turn the calibration into a market comparison system",
    "how to return from this",
    "monetize the advantage"
   ],
   "a": {
    "status": "refused",
    "answer": "Every candidate pregame signal was rejected across at least two corpora and the model does not beat the market's Brier at any in-game grain, so the system's own evidence points to market efficiency.",
    "source_artifact": "docs/MARKET_EFFICIENCY_PROOF.md",
    "as_of": "2026-07-07"
   },
   "tags": [
    "refused",
    "no-market comparison-advice",
    "efficiency"
   ]
  },
  {
   "q": "Does the model beat Pinnacle's closing line specifically?",
   "alt_phrasings": [
    "clv versus pinnacle",
    "beat the sharpest book",
    "pinnacle close comparison"
   ],
   "a": {
    "status": "no_data",
    "answer": "The recomputation notes that its soccer Over/Under 2.5 model was scored against Pinnacle and trails on that measured slice. It does not provide Pinnacle-specific closing-line value or establish general outperformance against Pinnacle closes, so this question cannot be answered broadly from the cited evidence.",
    "source_artifact": "docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md",
    "as_of": "2026-09-24"
   },
   "tags": [
    "no-data",
    "clv",
    "closing-line"
   ]
  },
  {
   "q": "How does the model do on NFL or NHL?",
   "alt_phrasings": [
    "nfl calibration",
    "hockey predictions",
    "other sports beyond the four"
   ],
   "a": {
    "status": "no_data",
    "answer": "No data. The measured corpora are NBA, MLB, soccer and tennis; NFL and NHL are not covered, so there are no calibration or market-comparison numbers to report for them.",
    "source_artifact": "webapp/public/data/showcase/brier_skill_scores.json",
    "as_of": "2026-07-24"
   },
   "tags": [
    "no-data",
    "coverage",
    "cross-sport"
   ]
  },
  {
   "q": "Which reference source is slowest to move its lines?",
   "alt_phrasings": [
    "slowest book to update",
    "which reference source lags",
    "book-by-book line speed"
   ],
   "a": {
    "status": "no_data",
    "answer": "This is not measured. The market-comparison work uses devigged consensus/market probabilities, not per-source reaction speed, so there is no artifact ranking individual sources by how fast they move.",
    "source_artifact": "webapp/public/data/showcase/info_arrival_curve.json",
    "as_of": "unknown"
   },
   "tags": [
    "no-data",
    "book-speed",
    "market-microstructure"
   ]
  },
  {
   "q": "Which games tonight are the market getting wrong?",
   "alt_phrasings": [
    "what games is the market mispricing now",
    "tonight's mispriced games",
    "live market errors right now"
   ],
   "a": {
    "status": "no_data",
    "answer": "This corpus holds retrospective calibration and market-comparison measurements, not a live tonight feed, so there is nothing here that identifies mispriced games in real time. And the standing finding is that on large disagreements the market is usually the accurate side, so a gap is not evidence the market is wrong.",
    "source_artifact": "webapp/public/data/showcase/market_disagreement_profile.json",
    "as_of": "unknown"
   },
   "tags": [
    "no-data",
    "live",
    "disagreement"
   ]
  }
 ]
}