Every number on this site was checked this morning.

Warnings (do not close the gate):

• range-soft: shots.type 30/112091, contracts.entryLevelLabel 8/245, contracts.signingDate 10/894, contracts.termVsSeasons 9/927i
• control-alert: 2025-26.shotsPerGame 2026-04-08 downi
• anomaly-defects: 120 older games carry record defects (see /integrity)i

The deploy job reads this gate before it stages a build. Closed means yesterday's build stays up. A stale gate (over 30 h) is a warning in the deploy log, not a refusal, so an assay that failed to run cannot silently stop the approved daily deploy.

• goalsAgree: Goals agree across the shot file, the goal record and the events
• shiftSums: Shift charts sum to a real number of players and never double up
• shooterOnIce: Every shooter and goalie was on the chart at the shot's second
• teamAgree: The shooter's club on the chart is the shooting club in the shot file
• assistsOnIce: Every assist went to a player on the ice
• emptyNetGoalie: The empty-net flag agrees with the chart
• timeBounds: Every time sits inside its period

Every game due this season is complete.

Source: Game passports · as of 2026-09-29

• shots.type: 30 of 112091i rows; example {"game":"2025020184","type":"unknown"}
• contracts.entryLevelLabel: 8 of 245i rows; example {"club":"boston-bruins","player":"JJ Peterka","termYears":5,"capHit":7700000,"why":"an entry-level contract runs three years at most"}
• contracts.signingDate: 10 of 894i rows; example {"club":"carolina-hurricanes","player":"Andrei Svechnikov","signedOn":"2021-08-26","firstSeason":"2024–25","why":"signed 3 summers before its first season"}
• contracts.termVsSeasons: 9 of 927i rows; example {"club":"boston-bruins","player":"Fraser Minten","termYears":3,"seasons":"2022–23 to 2026–27","span":5}

Contract ledgers: contracts.capHit 0 of 1068i; contracts.term 0 of 1068i; contracts.tier 181 of 1068i.

• 2025-26.shotsPerGame 2026-04-08i: reading down (71, z -3.2i)

No stale builds.

at

—

generatedAt

2026-09-29T15:08:07.322Zi

spine

players: 2304, byStatus: (prospect-pool: 1456, roster-2026-09-21: 1286, regular-2025-26: 776, former-regular: 2), withBorn: 2304, withPosition: 2304, bornConflicts: 0, bornConflictSamples: nonei

sourceCounts

rosterSnapshot: 1286, rosterFiles: 1286, leagueRegulars: 776, formerRegulars: 2, prospectPools: 1489, rosterHistory: 1286, prospectHistory: 1456, wikipediaTitles: 1090, sportradar: 1278i

crosswalks

wikipedia: (linked: 1090, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 27, 0.92-0.99: 0, 0.99-1: 1063)), sportradar: (linked: 1278, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 0, 0.99-1: 1278)), ELH: (linked: 21, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 11, 0.99-1: 10)), KHL: (linked: 104, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 0, 0.99-1: 104)), NL: (linked: 13, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 13, 0.99-1: 0)), SHL: (linked: 81, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 0, 0.99-1: 81)), Liiga: (linked: 51, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 0, 0.99-1: 51)), DEL: (linked: 6, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 5, 0.99-1: 1)), WHL: (linked: 154, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 1, 0.99-1: 153)), OHL: (linked: 195, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 0, 0.99-1: 195)), USHL: (linked: 136, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 1, 0.99-1: 135)), QMJHL: (linked: 79, histogram: (0-0.5: 0, 0.5-0.75: 0, 0.75-0.92: 0, 0.92-0.99: 1, 0.99-1: 78)) … +2 fieldsi

wikipedia

cachedTitles: 1090, cachedNulls: 196, keysOutsideSpine: 0, ruleViolations: 0, expected: 0, violators: none, rule: title must contain the surname AND start with the first two letters of the given name (build-white-noise.mjs playerTitleOk)i

sportradar

players: 1427, withReference: 1425, referenceInSpine: 1278, nameAgree: 1278, referenceOnly: 0, referenceOutsideSpine: 147, samples: (referenceOnly: none)i

otherLeagues

files: (DEL: (5 fields); (5 fields), ELH: (5 fields); (5 fields), ELJ: (5 fields); (5 fields), KHL: (5 fields); (5 fields), Liiga: (5 fields); (5 fields), NL: (5 fields); (5 fields), OHL: (5 fields); (5 fields); (5 fields), QMJHL: (5 fields); (5 fields); (5 fields), SHL: (5 fields); (5 fields), U20SM: (5 fields); (5 fields), USHL: (5 fields); (5 fields); (5 fields), WHL: (5 fields); (5 fields); (5 fields)), byLeague: (DEL: (entities: 504, withBirthdate: 0, withAgeOnly: 331, noBirthInfo: 173, candidatesScored: 1228, linked: 6, queued: 0, conflicts: 0), ELH: (entities: 475, withBirthdate: 294, withAgeOnly: 0, noBirthInfo: 181, candidatesScored: 1188, linked: 21, queued: 1, conflicts: 1), ELJ: (entities: 538, withBirthdate: 538, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 286, linked: 3, queued: 1, conflicts: 0), KHL: (entities: 1001, withBirthdate: 1001, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 541, linked: 104, queued: 1, conflicts: 0), Liiga: (entities: 662, withBirthdate: 662, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 170, linked: 51, queued: 1, conflicts: 0), NL: (entities: 512, withBirthdate: 0, withAgeOnly: 0, noBirthInfo: 512, candidatesScored: 2032, linked: 17, queued: 0, conflicts: 0), OHL: (entities: 998, withBirthdate: 998, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 790, linked: 195, queued: 0, conflicts: 0), QMJHL: (entities: 967, withBirthdate: 967, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 593, linked: 79, queued: 0, conflicts: 0), SHL: (entities: 530, withBirthdate: 530, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 220, linked: 81, queued: 0, conflicts: 0), U20SM: (entities: 792, withBirthdate: 792, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 143, linked: 28, queued: 0, conflicts: 0), USHL: (entities: 1123, withBirthdate: 1116, withAgeOnly: 0, noBirthInfo: 7, candidatesScored: 836, linked: 136, queued: 2, conflicts: 0), WHL: (entities: 1110, withBirthdate: 1110, withAgeOnly: 0, noBirthInfo: 0, candidatesScored: 696, linked: 154, queued: 0, conflicts: 0)), entities: 9212, pairsAtOrAboveQueue: 881, linkedEntities: 875i

queue

size: 6, byLeague: (ELH: 1, KHL: 1, Liiga: 1, USHL: 2, ELJ: 1), first20: (league: ELH, entityId: 10097211, entityName: Tomáš Galvas, entityBorn: —, entityAge: —, entityTeams: 1 item, candidate: Tomas Galvas, candidateNhlId: 8470500, candidateBorn: 1978-05-30, candidateClub: PIT, score: 0.923, evidence: 6 fields … +1 fields); (league: KHL, entityId: 44358, entityName: Fyodorov Viktor S., entityBorn: 2008-02-21, entityAge: 18, entityTeams: 1 item, candidate: Viktor Fedorov, candidateNhlId: 8486078, candidateBorn: 2008-02-21, candidateClub: SEA, score: 0.91, evidence: 6 fields … +1 fields); (league: Liiga, entityId: 40432390, entityName: Eelis Uronen, entityBorn: 2008-06-24, entityAge: 18, entityTeams: 1 item, candidate: Anttoni Uronen, candidateNhlId: 8486310, candidateBorn: 2008-06-24, candidateClub: CBJ, score: 0.899, evidence: 6 fields … +1 fields); (league: USHL, entityId: 9982, entityName: Xavier Veilleux, entityBorn: 2006-05-23, entityAge: 20, entityTeams: 1 item, candidate: Xavier Veilleux, candidateNhlId: 8485078, candidateBorn: 2006-03-23, candidateClub: NYI, score: 0.814, evidence: 6 fields … +1 fields); (league: USHL, entityId: 9902, entityName: Austin Baker, entityBorn: 2006-02-11, entityAge: 20, entityTeams: 1 item, candidate: Austin Baker, candidateNhlId: 8485098, candidateBorn: 2006-02-12, candidateClub: DET, score: 0.814, evidence: 6 fields … +1 fields); (league: ELJ, entityId: 10100131, entityName: Adam Trachta, entityBorn: 2007-04-10, entityAge: 19, entityTeams: 1 item, candidate: Adam Benak, candidateNhlId: 8485367, candidateBorn: 2007-04-10, candidateClub: MIN, score: 0.807, evidence: 6 fields … +1 fields)i

teams

count: 32, withNhlId: 32, withSportradar: 32, utah: (ids: (4 fields); (4 fields), predecessors: (5 fields), notes: Schedule files on disk: id 59 in 2024–25; id 68 in 2025–26, 2026–27., The build brief said id 68 is valid from 2026–27; the data on disk shows 68 already in the 2025–26 schedule (preseason games included) and 59 only in 2024–25. Recorded what the data shows; both ids resolve to UTA with their season ranges.)i

params

prior: -2.944, name: (strong: 3.379, medium: 0.251, weak: -4.554, strongAt: 0.97, mediumAt: 0.92), surnameRule: a pair whose folded surnames share no token can reach the queue but never a link (cap 0.91); twins and same-birthday club-mates are the reason, linkCapWithoutSurname: 0.91, birthdate: (exact: 5.779, sameYearDifferentDay: -3.504, yearFromAge: 0.642, unknown: 0), position: (agree: 0.663, disagree: -2.813, unknown: 0), shoots: (agree: 0.642, disagree: -2.303, unknown: 0), history: (leagueSameSeason: 2.14, clubBonus: 1.099, leagueOtherSeason: 1.386, leagueAbsent: -0.944, noHistory: 0), linkAt: 0.92, queueAt: 0.75i

at

—

summary

ranAt: 2026-09-29T15:08:07.772Z, dry: no, params: (limit: 40, gapMs: 2100, userAgent: CapOrCup source verifier (contact [email protected]), allowedDomains: nhl.com, tsn.ca, sportsnet.ca, espn.com, nbcsports.com, cbssports.com, apnews.com, thehockeynews.com, thescore.com, yahoo.com, dailyfaceoff.com, dailyfaceoff: news articles only, banned: /an excluded site|an excluded site|an excluded site|an excluded site|an excluded site/i, robotsTtlDays: 7), pool: (urls: 1395, allowed: 1232, banned: 33, other: 130, invalid: 0, facts: 2170, byHost: (www.nhl.com: 968, thehockeynews.com: 66, www.dailyfaceoff.com: 75, www.tsn.ca: 26, www.espn.com: 48, www.thescore.com: 15, sports.yahoo.com: 17, www.nbcsports.com: 3, ca.sports.yahoo.com: 2, www.cbssports.com: 6, africa.espn.com: 1, www.sportsnet.ca: 5), queueAllowed: 1232, neverVerifiedBefore: 992), run: (fetched: 40, verified: 40, found: 32, notFound: 9, totalOnly: 0, noAmountToCheck: 0, changed: 0, dead: 0, blocked: 0, errors: 0, redirectOffDomain: 0, robotsDisallowed: 0 … +3 fields), cumulative: (urlsInStore: 280, verified200: 280, found: 205, notFound: 95, dead: 0, blocked: 0, errors: 0, robotsDisallowed: 0, changedEver: 0), ledgerFacts: (total: 751, judged: 200, standing: 157, pending: 529, atRisk: 43, unverifiableHost: 22), factsAtRisk: (nhlId: 8477496, player: Elias Lindholm, team: BOS, expected: 7750000, tier: certified, status: amount-not-found, sources: 1 item); (nhlId: 8483489, player: Fraser Minten, team: BOS, expected: 879166, tier: certified, status: amount-not-found, sources: 1 item); (nhlId: 8477493, player: Aleksander Barkov, team: FLA, expected: 10000000, tier: reported, status: amount-not-found, sources: 1 item); (nhlId: 8481556, player: John Beecher, team: FLA, expected: 850000, tier: reported, status: amount-not-found, sources: 1 item); (nhlId: 8483424, player: Owen Beck, team: MTL, expected: 853333, tier: certified, status: amount-not-found, sources: 1 item); (nhlId: 8483728, player: Jared Davidson, team: MTL, expected: 850000, tier: reported, status: amount-not-found, sources: 1 item); (nhlId: 8484984, player: Ivan Demidov, team: MTL, expected: 940833, tier: reported, status: amount-not-found, sources: 1 item); (nhlId: 8483686, player: Adam Engstrom, team: MTL, expected: 896667, tier: certified, status: amount-not-found, sources: 1 item); (nhlId: 8484170, player: Jacob Fowler, team: MTL, expected: 923333, tier: reported, status: amount-not-found, sources: 1 item); (nhlId: 8478424, player: Jansen Harkins, team: TBL, expected: 850000, tier: reported, status: amount-not-found, sources: 2 items); (nhlId: 8483752, player: Dominic James, team: TBL, expected: 910000, tier: reported, status: amount-not-found, sources: 1 item); (nhlId: 8477149, player: Scott Sabourin, team: TBL, expected: 850000, tier: certified, status: amount-not-found, sources: 2 items) … +31 more, factsOnUnverifiableHosts: (nhlId: 8483464, player: Marco Kasper, team: DET, expected: 886666, tier: reported, sources: 2 items); (nhlId: 8478043, player: Sam Lafferty, team: FLA, expected: 850000, tier: reported, sources: 1 item); (nhlId: 8480995, player: Pontus Holmberg, team: TBL, expected: 1550000, tier: reported, sources: 1 item); (nhlId: 8479520, player: Brandon Duhaime, team: TOR, expected: 2600000, tier: reported, sources: 1 item); (nhlId: 8480002, player: Nico Hischier, team: NJD, expected: 7250000, tier: reported, sources: 1 item); (nhlId: 8475231, player: Casey Cizikas, team: NYI, expected: 2500000, tier: reported, sources: 1 item); (nhlId: 8482157, player: Will Cuylle, team: NYR, expected: 3900000, tier: reported, sources: 1 item); (nhlId: 8481668, player: Arturs Silovs, team: PIT, expected: 2800000, tier: reported, sources: 1 item); (nhlId: 8477479, player: Tyler Bertuzzi, team: CHI, expected: 5500000, tier: reported, sources: 1 item); (nhlId: 8480448, player: Parker Kelly, team: COL, expected: 1700000, tier: reported, sources: 1 item); (nhlId: 8480800, player: Quinn Hughes, team: MIN, expected: 7850000, tier: reported, sources: 1 item); (nhlId: 8479619, player: Michael Carcone, team: UTA, expected: 1750000, tier: reported, sources: 1 item) … +10 more, nextUp: (url: https://www.nhl.com/news/topic/free-agency/jaden-schwartz-signs-3-year-contract-with-colorado-avalanche, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/news/vancouver-canucks-trade-dakota-joshua-to-toronto-maple-leafs, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/oilers/news/release-oilers-extend-shakir-mukhamadullin-spencer-stastney, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/oilers/news/release-oilers-re-sign-kasperi-kapanen-to-one-year-contract, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/oilers/news/release-oilers-re-sign-max-jones-to-one-year-contract, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/oilers/news/release-oilers-sign-mathieu-joseph-to-one-year-contract, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/oilers/news/release-oilers-sign-ryan-shea-to-five-year-contract, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/panthers/news/florida-panthers-agree-to-terms-with-forward-eetu-luostarinen-on-three-345422684, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/panthers/news/florida-panthers-agree-to-terms-with-forward-sam-reinhart-on-an-eight-year-contract-extension, lastFetchAt: —, amountFacts: 2); (url: https://www.nhl.com/penguins/news/penguins-acquire-defenseman-erik-karlsson-from-the-san-jose-sharks-in--345527572, lastFetchAt: —, amountFacts: 2), definitions: (verified: HTTP 200 on an allowed host after redirects, found: the expected cap hit appears in one of the written forms ($8.5 million, $8,500,000, 8.5M, AAV of $8.5); 1.1 never matches 1, totalOnly: only the contract total (cap hit x term) appears, changed: content SHA-256 differs from the previous fetch of the same URL, dead: HTTP 404/410/451 or the host no longer resolves, atRisk: a ledger cap hit whose allowed sources have all been checked and none shows the amount (source-dead when every one is gone))i

deadOrChanged

none

at

—

fields

capHit: (compared: 307, agree: 276, disagree: 31, agreementPct: 89.9, excludingAdjudications: (compared: 304, agree: 276, disagree: 28, agreementPct: 90.8), pairwise: (gap-round1 | wiki-finder: 3 fields, nhl-extract | researcher: 3 fields, current-contracts | researcher: 3 fields, gap-round1 | researcher: 3 fields, current-contracts | gap-round1: 3 fields, correction | nhl-extract: 3 fields, nhl-extract | verify-adjudication: 3 fields, correction | verify-adjudication: 3 fields, current-contracts | wiki-finder: 3 fields, gap-round1 | nhl-extract: 3 fields, gap-round1 | gap-round2: 3 fields, correction | researcher: 3 fields … +7 fields), disagreements: (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields) … +19 more, tolerance: 25000), lastSeason: (compared: 447, agree: 390, disagree: 57, agreementPct: 87.2, excludingAdjudications: (compared: 445, agree: 390, disagree: 55, agreementPct: 87.6), pairwise: (gap-round1 | nhl-extract: 3 fields, nhl-extract | wiki-finder: 3 fields, gap-round1 | wiki-finder: 3 fields, nhl-extract | researcher: 3 fields, current-contracts | nhl-extract: 3 fields, current-contracts | researcher: 3 fields, gap-round1 | researcher: 3 fields, gap-round2 | nhl-extract: 3 fields, gap-round1 | gap-round2: 3 fields, current-contracts | gap-round1: 3 fields, current-contracts | gap-round2: 3 fields, current-contracts | wiki-finder: 3 fields … +9 fields), disagreements: (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields) … +45 more), termYears: (compared: 319, agree: 291, disagree: 28, agreementPct: 91.2, excludingAdjudications: (compared: 317, agree: 290, disagree: 27, agreementPct: 91.5), pairwise: (gap-round1 | nhl-extract: 3 fields, nhl-extract | wiki-finder: 3 fields, gap-round1 | wiki-finder: 3 fields, nhl-extract | researcher: 3 fields, current-contracts | nhl-extract: 3 fields, current-contracts | researcher: 3 fields, gap-round1 | researcher: 3 fields, gap-round2 | nhl-extract: 3 fields, gap-round1 | gap-round2: 3 fields, current-contracts | gap-round1: 3 fields, current-contracts | gap-round2: 3 fields, current-contracts | wiki-finder: 3 fields … +8 fields), disagreements: (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields) … +16 more), signedOn: (compared: 257, agree: 165, disagree: 92, agreementPct: 64.2, excludingAdjudications: (compared: 256, agree: 165, disagree: 91, agreementPct: 64.5), pairwise: (nhl-extract | wiki-finder: 3 fields, nhl-extract | researcher: 3 fields, current-contracts | nhl-extract: 3 fields, gap-round1 | nhl-extract: 3 fields, gap-round2 | nhl-extract: 3 fields, gap-round1 | gap-round2: 3 fields, current-contracts | gap-round1: 3 fields, current-contracts | gap-round2: 3 fields, current-contracts | researcher: 3 fields, gap-round1 | researcher: 3 fields, gap-round1 | wiki-finder: 3 fields, current-contracts | wiki-finder: 3 fields … +4 fields), disagreements: (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields) … +80 more, within3Days: (agree: 220, compared: 257, agreementPct: 85.6, note: strict agreement plus disagreements whose dates lie within 3 days of each other), gapHistogram: (1-3 days: 55, 4-30 days: 13, over 400 days (a different deal): 14, 31-400 days: 10)), expiryStatus: (compared: 84, agree: 81, disagree: 3, agreementPct: 96.4, excludingAdjudications: (compared: 84, agree: 81, disagree: 3, agreementPct: 96.4), pairwise: (gap-round1 | researcher: 3 fields, gap-round1 | gap-round2: 3 fields, gap-round2 | researcher: 3 fields), disagreements: (7 fields); (7 fields); (7 fields)), clauseStatus: (compared: 45, agree: 20, disagree: 25, agreementPct: 44.4, excludingAdjudications: (compared: 45, agree: 20, disagree: 25, agreementPct: 44.4), pairwise: (clauses-r3 | current-contracts: 3 fields, clauses-r1 | gap-round2: 3 fields, clauses-r1 | researcher: 3 fields, clauses-r1 | clauses-r3: 3 fields, clauses-r1 | gap-round1: 3 fields), disagreements: (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields); (7 fields) … +13 more)i

overall

—

coverage

totals: (contracted: 1068, priced: 743, certified: 562, reported: 181, clauseKnown: 495, regulars: 776, regularsPriced: 705, pricedPct: 69.6, clauseKnownPct: 46.3, regularsPricedPct: 90.9), perClub: (abbr: BOS, slug: boston-bruins, rows: 43, contracted: 42, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: BUF, slug: buffalo-sabres, rows: 44, contracted: 25, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: DET, slug: detroit-red-wings, rows: 32, contracted: 32, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: FLA, slug: florida-panthers, rows: 35, contracted: 35, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: MTL, slug: montreal-canadiens, rows: 42, contracted: 32, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: OTT, slug: ottawa-senators, rows: 27, contracted: 27, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: TBL, slug: tampa-bay-lightning, rows: 49, contracted: 37, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: TOR, slug: toronto-maple-leafs, rows: 47, contracted: 20, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: CAR, slug: carolina-hurricanes, rows: 42, contracted: 36, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: CBJ, slug: columbus-blue-jackets, rows: 26, contracted: 26, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: NJD, slug: new-jersey-devils, rows: 37, contracted: 37, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields); (abbr: NYI, slug: new-york-islanders, rows: 44, contracted: 41, capHit: 6 fields, fields: 4 fields, clause: 6 fields, regulars: 3 fields) … +20 more, thinnest: (byPricedShare: (2 fields); (2 fields); (2 fields); (2 fields); (2 fields), byCertifiedShare: (2 fields); (2 fields); (2 fields); (2 fields); (2 fields), byRegularsPriced: (2 fields); (2 fields); (2 fields); (2 fields); (2 fields), byClauseKnown: (2 fields); (2 fields); (2 fields); (2 fields); (2 fields), byExpiryKnown: (2 fields); (2 fields); (2 fields); (2 fields); (2 fields), bySignedOnKnown: (2 fields); (2 fields); (2 fields); (2 fields); (2 fields), namedBiasClubs: (4 fields); (4 fields); (4 fields))i

thinnest

none

at

—

seasons

2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025i

arenas

none

note

—

Season 2025-26i, built 2026-09-29i.

Structural findings, as measured:

• drive: finding a_driveVsClubStrength · question Do team-relative measures compress players on deep clubs? Pearson r between the player's measure and his 2025-26 club's season xG share (xg-2025-26 teams.xgShar · n 621 · r 0.003 · ci95 -0.073, 0.077 · slopePerSharePoint 0i
• driveOff: finding a_driveVsClubStrength · question Do team-relative measures compress players on deep clubs? Pearson r between the player's measure and his 2025-26 club's season xG share (xg-2025-26 teams.xgShar · n 621 · r 0.053 · ci95 -0.025, 0.126 · slopePerSharePoint 0.015i
• driveDef: finding a_driveVsClubStrength · question Do team-relative measures compress players on deep clubs? Pearson r between the player's measure and his 2025-26 club's season xG share (xg-2025-26 teams.xgShar · n 621 · r 0.057 · ci95 -0.015, 0.123 · slopePerSharePoint 0.009i
• shutdown: finding a_driveVsClubStrength · question Do team-relative measures compress players on deep clubs? Pearson r between the player's measure and his 2025-26 club's season xG share (xg-2025-26 teams.xgShar · n 589 · r -0.041 · ci95 -0.116, 0.035 · slopePerSharePoint -0.005i
• driveOffAdj: note no structural lean was tested for this measure; see the top-level findings
• twoWayAdj: note no structural lean was tested for this measure; see the top-level findings
• startPM60: note no structural lean was tested for this measure; see the top-level findings
• setupXg60: finding b_setupVsTeammateFinishing · question Does Setup ride on the shooters? Teammates' finishing = the club's shooters' (goals - xG) per 100 unblocked shots with the player himself removed. Players with · setupXgPerPrimaryAssist n: 491, r: 0.02, ci95: -0.076, 0.109 · setupXgPer60 n: 491, r: 0.049, ci95: -0.046, 0.144 · reading An assist only exists when the shot went in, so a passer whose linemates convert low-xG chances collects assists on low-xG shots (per-assist falls) while his toi
• finish: finding d_finishGsaxVsHomeArena · question Do shot-location recording habits at a rink leak into Finish and GSAx? One-way spread of the controlled residual across the 2025-26 club's home arena, with a la · n 540 · arenas 32 · eta2 0.065 · eta2UnderNull 0.057i
• gsax: finding d_finishGsaxVsHomeArena · question Do shot-location recording habits at a rink leak into Finish and GSAx? One-way spread of the controlled residual across the 2025-26 club's home arena, with a la · n 71 · arenas 32 · eta2 0.388 · eta2UnderNull 0.444i
• aPM60: note no structural lean was tested for this measure; see the top-level findings
• realPoints60: note no structural lean was tested for this measure; see the top-level findings
• turnover60: finding e_puckBattlesVsHomeArena · question Hand-recorded counts differ by rink (the deployment method says so). One-way spread of per-60 counts, after position and log minutes, across the home arena; ska · limit Season totals mix 41 home and 41 road games, so a rink's scorers are visible here only as the half-diluted club residue, and a club's playing style sits in the · byStat hitsPer60: (n: 458, arenas: 32, eta2: 0.054, eta2UnderNull: 0.068, permutationP: 0.797, controlR2: 0.131, lowestArenas: (3 fields); (3 fields); (3 fields), highestArenas: (3 fields); (3 fields); (3 fields)), blocksPer60: (n: 458, arenas: 32, eta2: 0.08, eta2UnderNull: 0.068, permutationP: 0.228, controlR2: 0.502, lowestArenas: (3 fields); (3 fields); (3 fields), highestArenas: (3 fields); (3 fields); (3 fields)), takeawaysPer60: (n: 458, arenas: 32, eta2: 0.08, eta2UnderNull: 0.068, permutationP: 0.234, controlR2: 0.076, lowestArenas: (3 fields); (3 fields); (3 fields), highestArenas: (3 fields); (3 fields); (3 fields)), giveawaysPer60: (n: 458, arenas: 32, eta2: 0.072, eta2UnderNull: 0.069, permutationP: 0.379, controlR2: 0.009, lowestArenas: (3 fields); (3 fields); (3 fields), highestArenas: (3 fields); (3 fields); (3 fields)), dzGiveawaysPer60: (n: 458, arenas: 32, eta2: 0.059, eta2UnderNull: 0.068, permutationP: 0.681, controlR2: 0.752, lowestArenas: (3 fields); (3 fields); (3 fields), highestArenas: (3 fields); (3 fields); (3 fields))i
• whiteNoise: finding c_whiteNoiseVsMarketAndWikipedia · question Is attention a function of where a player plays and whether he has an English Wikipedia article (a pageview component only exists when a title resolved)? · marketSize n: 912, rLogPopVsScore: 0.025, ci95: -0.039, 0.086, rLogPopVsResidual: 0.026, ci95Residual: -0.038, 0.089, byTier: (small: (n: 317, meanScore: 53.968, meanResid: -2.58), mid: (n: 279, meanScore: 58.265, meanResid: 1.248), large: (n: 316, meanScore: 58.032, meanResid: 1.486)), clubField: current (2026-27) club, since attention is measured now · wikipediaArticle nWithArticle: 905, nWithout: 7, meanScoreWith: 56.886, meanScoreWithout: 31.429, diff: 25.458, ci95: 11.882, 39.35, diffAfterControls: 23.286, ci95AfterControls: 5.932, 42.052, pointBiserialR: 0.08 · wikipediaArticleAllListed nWithArticle: 1090, nWithout: 196, meanScoreWith: 54.474, meanScoreWithout: 25.684, diff: 28.791, ci95: 25.617, 32.116, pointBiserialR: 0.359, note: raw percentile scores, no minutes control: most players without an article have no 2025-26 NHL ice time · unknownWikiStatus 0i

• 2025-26 club = the club a player logged the most regular-season shift seconds for in data/shifts/shifts-2025.json (the same charts the measures are built from); the 2026-27 camp roster is NOT used for club, arena or market tier except for White Noise, where attention is measured now. Cross-check: of 778 players with a league-production teams string, the shift club matched the first-listed club 756 times and the last-listed 718 times, so that string leads with the majority club more often than it ends with it; the disagreements are traded players.i
• Traded players carry their majority club; team-relative measures (Drive, driveOff, driveDef, adjusted drive) already compare each game against that night's club, so a traded player's number mixes two clubs' baselines.
• Home arena is the modal regular-season home venue in schedule-20252026.json; it is a one-to-one relabeling of club in 2025-26 (FLA, NSH, PIT and TBL each played one home game elsewhere), so arena slices repeat club slices and share their q rather than enlarging the FDR family.i
• Control fit: OLS of the measure on isD (isG where goalies are present), log(volume) and, where the measure is not already deployment-adjusted, the offensive-zone start share (mean-imputed with an indicator when the site itself would not show one). Residuals are what the slices average.
• Flagged = the 2,000-resample percentile bootstrap interval of the group's mean residual excludes 0. q = Benjamini-Hochberg over the one-sample t p-values of every slice of the measure (n >= 5), with the arena slices excluded as duplicates; survivesFdr = q <= 0.10. Groups under 5 players get no interval and a reason.i
• Position appears as a slice for completeness but is a control dummy, so its residual mean is zero by construction; the raw means are the informative part of that row.
• Finish uses xg-2025-26 shooters (goals minus xG per 100 unblocked shots, >= 80 shots). advanced.json goalsAboveExpected sums (1 - xG) over goals scored only, so it is not net of misses and was not used as Finish.i
• Shutdown/driveOffAdj/twoWayAdj: universe = every skater with the field in deployment.json (forwards and defence; fitted separately by the builder), which is the rankings-desk lens; the stats-desk headline Shutdown list is defence only, and the position slice gives the D-only raw mean.
• White Noise is a percentile among all 1,286 listed players; this audit keeps the players with 2025-26 NHL ice time so minutes can be controlled, and its club, arena and market tier are the current club's.i
• Representation: top list by measure value (top 100; top 30 for goalies); ratio = share in the top over share in the universe, interval from resampling the top list and the universe independently.i
• Simpson checks compare raw group means pooled and within F and D (both sides >= 10 players); a reversal is a sign flip of the raw difference. Residual differences are given alongside.i
• Club posterior: normal-normal shrinkage of club mean residuals with tau2 by method of moments; tau2 = 0 means the between-club spread is no wider than sampling noise and every club posterior is 0.i

Findings

• measured In the league's play-by-play a goal is stamped about one second later than a saved or missed shot. The shift is in 2023-24, 2024-25 and 2025-26 and not in 2018-19, 2019-20, 2020-21, 2021-22 and 2022-23.i
• measured A timing input with no hockey meaning lowered held-out log loss by 0.000661 per shot on the raw clock (95% range −0.000881 to −0.000466). On the corrected clock the change is +0.000008 (−0.000002 to +0.000016) and it no longer lowers it.i
• measured The champion's rebound and rush inputs are short windows on that clock. With goal times moved back 1 s the champion scores 0.22883 against 0.22861 on the same 105,288 shots: +0.000225 per shot (95% range +0.000107 to +0.000351). Part of its published accuracy was the clock.i
• measured 70% of the shots flagged as rush in 2025-26 follow the club's own blocked attempt and are not rushes.i
• measured The flurry idea on the corrected clock: −0.000003 per shot (95% range −0.000033 to +0.000029). The desk's reading: adds nothing: the gain challenger C registered on the raw clock was the clock.i
• measured Seven challengers were registered before being fitted, trained on 2023-24 and 2024-25 and judged on 2025-26 by a bootstrap over games (1,000 resamples). Passed the registered rule: strength-state models (four fits) (−0.000491, 95% range −0.000713 to −0.000248), flurry features added (−0.001587, 95% range −0.001970 to −0.001228) and rink-adjusted + strength-state fits + flurry (−0.001965, 95% range −0.002436 to −0.001503).i
• measured The gain of the strength-state models (four fits) does not come from the clock inputs: with rebound and rush removed from both sides its difference is −0.000489 (95% range −0.000707 to −0.000240). It is withheld because it keeps the champion's raw-clock flags, and the timing input still lowers its log loss by 0.000651 (−0.000865 to −0.000456).i
• measured By strength state, judged on all games, the split has a lower log loss than the champion on the power play and with the net empty; at 5-on-5 and shorthanded the range spans zero.i
• measured Replica check: the desk's copy of the champion scores 0.22830 per unblocked shot on 112,091 shots from 1,312 games (published 0.2281) and matches the site's stored values on 112,089 of 112,089 shots.i
• decision No challenger is put forward. The desk's reason: no candidate passes the registered rule, is attributable and shows no placebo gain, so the desk puts none forward. B passes the rule but its flags are built on the raw clock. The clock finding goes to whoever owns tools/pull-pbp.mjs.
• caveat The first stage of the study put forward strength-state models (four fits). The clock follow-up withdrew it. In the desk's words: WITHHELD: passes the registered rule, but a clock placebo still lowers log loss on the clock its flags are built on; its counterpart on the corrected clock is G.
• decision Desk note: Two things for the hub to weigh. (1) The rule measures every challenger against a champion that gains about 0.000244 per shot from the clock, so a challenger built on the corrected clock starts that far behind. On a level clock the strength split gains 0.000486 (G against F, 95% range [-0.000742, -0.000247]); against the published champion G is -0.000261 with range [-0.000543, 0.000026], which misses the rule at the top of its range (95.7% of resamples at or below zero). (2) Correcting the clock in the champion is an integrity decision, not an accuracy one: F removes information the model should never have had, so it cannot pass a rule that asks for a lower log loss. The desk suggests deciding that first. If the corrected clock becomes the baseline, the strength split should be registered again against it and judged on a season not yet read; its empty-net fit is unstable (trained on one season it lost to a single fit).i
• caveat Seven challengers have now been judged on the same proof season; read a single pass with that in mind.
• measured Arena scorers, 2025-26: 8 of 32 buildings record shot distance differently from the same clubs' road shots elsewhere (25% at p below 0.01, against 3.5% when the building is nothing special).i
• measured The largest is STL at Enterprise Center: shots recorded 1.89 ft farther (about +0.87 to +2.92 ft). Across its five seasons on file the mean is −0.93 ft, the opposite direction to this season, with four of the five seasons on the side of that mean.i
• measured From one season to the next the arena effects correlate at r 0.519 since 2021-22 (the registered threshold was 0.5). The spread between buildings has shrunk from 1.71 ft in 2021-22 to 1.00 ft in 2025-26, against a noise floor of about 0.56 ft.i
• measured The distance correction was not shown to help on games it had not seen: 22 arena-seasons repaired against 20 broken (exact p 0.88).i

Challengers

Trained on 2023-24 and 2024-25, judged on 2025-26. Rule: adopt only if the bootstrap 95% range of (challenger - champion) log loss lies below zero; otherwise reject; adoption itself is a hub decision card, never automatic. Champion held-out log loss 0.22830 on all 1,312 games; 0.22861 on the 1,232 games with an events file, where the clock follow-up is judged.i

A negative difference is a lower (better) log loss than the champion. A pass needs the whole range below zero.

Known limitation: the event clock

Every shot in the league's play-by-play carries a clock time. Goals sit about one second later than saved or missed shots, measured against whatever happened just before; the likely reason is that a goal takes the time the clock stopped and other shots the time the scorer logged them. This is a property of the feed from 2023-24 on; in 2018-19, 2019-20, 2020-21, 2021-22, 2022-23 goal times and shot times agree. In 2023-24 and 2024-25, shots logged in the same second as the previous event were goals 1.0% of the time and shots logged one second after it 3.0%, against 10.8% at three seconds and about 7% overall. That gap is not hockey: it appears after hits, faceoffs, giveaways and takeaways alike. Moving goal times back 1 second makes the timing of goals and of other shots line up (after a hit, the distance between the two patterns falls from 0.07 to 0.021; moving them back 2 seconds overshoots). Two inputs of our expected-goals model are short windows on that clock: 'rebound' (within 3 seconds of the same club's previous attempt) and 'rush' (within 4 seconds). Because a goal stamped a second late can fall outside the window, both were short of goals and the model learned to mark them down (rush coefficient -0.4216 on the raw clock, -0.1317 on the corrected one). A timing input with no hockey meaning improved the model's score on unseen games by 0.000661 per shot on the raw clock; on the corrected clock it changes it by +0.000008 and no longer improves it. The corrected champion (F) scores 0.22883 against the published champion's 0.22861 on the same 2025-26 shots: +0.000225 per shot, 95% range [0.000107, 0.000351]. It is WORSE, and the range excludes zero. Part of the champion's published accuracy was the clock, not hockey: against the same fit on the raw clock the cost of correcting it is +0.000244 [0.000127, 0.000372]. Separately, 70% of the shots flagged 'rush' in 2025-26 are not rushes at all: they are follow-ups to the club's own blocked shot, flagged because the feed credits a blocked shot to the shooter's club but gives its zone from the blocker's side.i

Cost of correcting the clock in the champion: +0.000225 log loss per shot against the published model (95% range +0.000107 to +0.000351), +0.000244 against the same fit on the raw clock (95% range +0.000127 to +0.000372), on 105,288 shots in 1,232 games.i

Shots flagged as rush that follow the club's own blocked attempt: 2018-19 48%, 2019-20 55%, 2020-21 57%, 2021-22 55%, 2022-23 61%, 2023-24 55%, 2024-25 69%, 2025-26 70%. Why: the feed credits a blocked shot to the shooter's club and gives its zone from the blocker's side, so "previous event by the same club outside the offensive zone" is true for every shot that follows an opponent's block.i

Arena scorer effects

Positive means shots in that building are recorded farther from the net than the same clubs' shots elsewhere. Ranges are the mean difference plus and minus 1.96 standard errors, computed here from the desk's figures. The standard error treats shots as independent, so the ranges are somewhat too narrow.i

Scorer effect: declared by the registered rule (mean consecutive-season r from 2021-22 = 0.519 against a threshold of 0.5; 0.25 of arenas at KS p < 0.01 in 2025-26 against a placebo rate of 0.035). The margin on r is thin and the effect has shrunk: the spread of arena mean differences was 1.713 ft in 2021-22 and is 0.996 ft in 2025-26, against a noise floor of about 0.564 ft. Correction: by the letter of the registered rule the held-out share of arenas at p < 0.01 fell (0.175 to 0.1625), but the change is 22 arena-seasons repaired against 20 broken (exact p = 0.8776), so it is NOT a demonstrable improvement. Carried from one season to the next the unshrunk map moved the share from 0.3481 to 0.4367.i

Study stamp 2026-09-29 13:24 UTCi. Read from data/models/xg-challengers.json, data/quality/decisions-request-xg.json, data/quality/xg-clock-finding.json, data/quality/arena-profiles.json.

Findings

• measured 20.1% of these players graduated by 23 (383 of 1,902).i
• measured Rank correlation with NHL games through 23, skaters: the formula at 18 0.448, at 19 0.530, at 20 0.606; draft position 0.605; raw points per game in the draft year 0.152.i
• measured Formula minus draft position on the same skaters: −0.159 at 18 (−0.203 to −0.113), −0.074 at 19 (−0.114 to −0.029), +0.001 at 20 (−0.040 to +0.041).i
• measured Taken apart at 18: percentile alone 0.356, times the league ladder 0.431, times the age factor 0.431 and plus the NHL bonus (the full formula) 0.448.i
• measured Graduation by score at 18: 3.3% in the bottom tenth (6 of 184) and 48.4% in the top tenth (89 of 184); the share rises with the score but not at every step.i
• measured Relative age: births by calendar quarter are not spread evenly over the year (chi-square 161.9, p below 0.001). In the score: no detectable carry in this sample: the difference between the oldest and youngest birth quarter in league-adjusted score at 18 has an interval that includes zero.i
• measured Measured growth in points per game, same league, outside the NHL: 1.110 times a year from 18 to 24 on average, against 1.081 implied by the hand-set age line.i
• caveat Whether a steeper age line would rank better: the two horizons disagree (three seasons ahead: steeper ranks worse; through age 23: no detectable difference), so the choice of slope cannot be settled by ranking on this sample; the differences are a few hundredths of rho either way.i
• measured Ladder: nine leagues re-measured from 694 moves to the NHL; the site's value sits inside the range for seven.i
• measured Every pick in the class lists, goalies included: 1,902 of 1,903 have a known outcome, and draft position alone ranks NHL games through 23 at 0.588 (95% range 0.552 to 0.621).i
• measured Graduation by round: 72.7% of round-1 picks graduated by 23 (197 of 271), against 1.5% in round 7 (4 of 271).i

Validity against draft position

Whole classes: graduation by round

1,902 of 1,903 picks, goalies included. Draft position alone ranks NHL games through 23 at 0.588 (95% range 0.552 to 0.621).i

Outcome: NHL regular-season games and points summed over every season up to and including the age-23 season (age = whole years on 15 September of the season's start year). Graduated = 82 NHL games within any two consecutive seasons, both no later than the age-23 season. gpAfter / ptsAfter at an age count only the seasons after that age's season.

Graduation by score at 18

The league ladder, re-measured

Study stamp 2026-09-29 13:57 UTCi. Read from data/assay/prospect-backtest.json.

Findings

• measured What the market pays for, other things held. Each extra 0.1 points per game in the season before: +12% (95% range +10% to +14%). Each extra tenth of the schedule played in the season before: +10% (95% range +8% to +12%). Each extra year of term: +17% (95% range +15% to +19%). A defenceman against a forward with the same points: +22% (95% range +13% to +31%).i
• measured Not distinguishable from zero once those are held: a goalie's goals saved above average (p 0.98), age (p 0.31), age squared (p 0.39), the free-agent flag (p 0.84) and the second-contract flag (p 0.12).i
• measured For goalies the market prices games played and not save performance: in four goalie-only fits on 67 deals the performance term was never distinguishable from zero (p 0.32 to 0.86).i
• measured Backtest: fitted on 204 deals starting 2019-20 to 2024-25, tested on 462 deals starting later. On 3- to 8-year deals (201) the mean error was $1.24M, against $1.49M for production-quantile matching and $2.69M for the position median.i
• caveat The training deals held no 1- and 2-year contracts, and on those the fit prices too high: 40% on 1-year deals (77) and 33% on 2-year deals (184).i
• caveat The backtest prices a deal of known length. Without the term the mean error over all 462 test deals is $1.49M, against $1.06M with it.i
• measured Bands: the 80% band held 58.0% of test deals (268 of 462): 80.6% on terms seen in training and 40.6% on terms not seen.i
• measured Where the residuals lean: 55 groups tested, five survive false-discovery control: CAR (club) −16.6% against what the fit prices (22 deals), round-1 picks +7.0% against what the fit prices (247 deals), cap hits of the reported tier −8.6% against what the fit prices (154 deals), CHI (club) +44.6% against what the fit prices (15 deals) and undrafted players −11.5% against what the fit prices (89 deals).i
• measured Nothing survives by nationality, market size or position.
• measured The site's constant, "+20% a point, 2024→2026": Reproduced (+19.6%). A rise in dollars is supported: 7 of 10 readings exclude zero (the site's own reading does not). A premium is not: no reading separates the rise from the cap's own +18.2%, and once terms are matched the open market rose as much. '+20% a point' describes what the salary cap did, not a second-contract premium.i
• measured In numbers: the site's arithmetic gives +19.6% (95% range −3.3% to +39.8%) while the salary cap's upper limit grew +18.2%; measured against the cap the change is +1.2% (−18.2% to +18.3%).i
• measured The ledger behind it: 1,205 rows, 449 without a cap hit, 76 labelled entry-level, 10 set aside for a signing date that cannot belong to their seasons, 666 fitted.i
• caveat 14 rows labelled entry-level carry a term an entry-level contract cannot have, so the site's market panel skips second contracts it means to count.i

Backtest against simple baselines

Fitted on 204 deals starting 2019-20 to 2024-25 (terms 3, 4, 5, 6, 7, 8 years only). Median error is the middle absolute error as a share of the real cap hit.i

Band coverage

What the market pays for

Where the residuals lean

Study stamp 2026-09-29 12:51 UTCi. Read from data/assay/pricing-backtest.json.

Findings

• measured 160 questions were written against the corpus of 1,269 passages: 90 answerable, 30 unanswerable, 20 adversarial and 20 paraphrases in 10 pairs.i
• measured Recall at 1, 3, 5 and 10: 57.8%, 76.7%, 81.1% and 86.7%. Mean reciprocal rank 0.683 (95% range 0.599 to 0.760). A passage that states the answer reaches the model for 92.2% of questions.i
• measured Weakest topics at five: long-term injury 50% (3 of 6), minimum salary 50% (3 of 6), performance bonuses 60% (3 of 5) and waivers 67% (4 of 6). Each rests on a handful of questions: they say where to look.i
• measured Paraphrase: on the same 10 facts the top five holds the right passage for 80% of questions in the agreement's words and 30% in a fan's words; the two wordings share a first passage on 10% of pairs.i
• measured Refusing by retrieval score alone: the best threshold refuses 26 of 30 unanswerable questions (86.7%, range 70.3% to 94.7%) and wrongly refuses 16 of 90 answerable ones (17.8%, range 11.3% to 26.9%). 25 of the unanswerable questions score above the lowest-scoring answerable one, so no threshold separates them.i
• measured Mirror check: on 15 questions sent to the live route the top five passages matched on 100%, all passages handed over on 100% and the fallback text on 100%; 21 of 21 mirrored constants match the app's source.i
• caveat The fallback does not refuse: it answered 19 of 30 unanswerable questions with quoted passages. The quotes are real; they are not answers.i
• measured One defect in the app's own code was found while measuring and is listed below; the study changed nothing in the app.
• measured Against the earlier run: recall at five was 81.1% and is 81.1%.i
• measured Held-out questions (60 answerable, written after the ranking was frozen): the right passage is in the top five for 53.3% and first for 25.0%, against 81.1% on the questions the ranking was developed on.i
• decision Six changes to the ranking were tried against the held-out questions; none was adopted, so the site keeps its ranking.
• decision Model and judge: not run: no credit and the hub does not allow model spend.

Retrieval

Mean reciprocal rank 0.683 (95% range 0.599 to 0.760). A passage that states the answer reaches the model for 92.2% of questions. The range on recall at 5 is the desk's. The others are Wilson 95% ranges computed here from the desk's counts; the same arithmetic reproduces the desk's range.i

Questions written after the ranking was frozen

Refusing by retrieval score

Defects found in the app

• The contract-term rule, /\bextension\b|\bterm\b|how long|length|how many years|.../, appends "SPC term years extension greater than six seven years". It fired on 24 of the 110 answerable and paraphrase questions, triggered by "how long" (12), "term" (7), "length" (1), "how many years" (2), "extension" (2). "\bterm\b" matches inside "long-term", so 6 long-term-injury questions are searched as if they were about contract length, and "how long" does the same to "How long is the waiver claim period?". With the rule the gold chunk ranks lower on 13 of those 24 and higher on 2 (sign test p = 0.0074). Removing this rule alone brings 3 into the top 5 (ans-006, pp-09-a, pp-09-b) and drops 0.i

Study stamp 2026-09-29 14:08 UTCi. Read from data/assay/assistant-eval.json, data/assay/assistant-eval-mirror.json, data/assay/assistant-eval-items.json, data/assay/assistant-eval-heldout.json, data/assay/assistant-eval-heldout-results.json.

Findings

• measured Goals that get a start: 5,127 at 15 s, 6,369 at 25 s and 7,082 at 35 s, over 695 players in 2025-26.i
• measured Top-thirty overlap between windows (1 means identical lists): 15 s against 25 s, 0.571, 15 s against 35 s, 0.596 and 25 s against 35 s, 0.800.i
• caveat One entry in the top-ten lists is a player id with no name on file.

Study stamp 2026-09-29 12:11 UTCi. Read from data/assay/sensitivity-start.json.

Findings

• measured 17 event types; 10 have the full 40 windows in the sample, the rest have fewer windows than that on file.i
• measured The sample touches 32 of the 32 games read. It is the same every time it is drawn on the same day (seed 20260929).i
• caveat About 138 reviewed windows per event type bound precision within ±5 points at 80% confidence when precision is near 0.7; two reviewers on the same windows give Cohen's kappa for the task ceiling.i

How to review: Open /video/<gameId> and jump to the window's start second. Record every entry (carried/dumped/denied), exit, takeaway, giveaway, board battle, shot and save you see, with jersey numbers, BEFORE revealing the model's events. Save the flags; the morning build applies them (wrong → drop, edited → retype, correct → confidence 1, missed → add) and reports precision and recall per type on /integrity.

Study stamp 2026-09-29 12:20 UTCi. Read from data/condensed/review-sample-2026-09-29.json.