Measured against the same games
Which model holds up?
Compare original gameplay forecasts with their frozen recent-rate baseline. This measures forecast quality—not betting profit or an AI analyst’s persuasive explanation.
Comparing scheduled captures from 6 hours before tip only. Browsing a game does not enroll it. Expected, confirmed and user-edited scenarios are kept separate. No model is declared a winner until the registered validation and publication gates pass.
Complete captured window: 2026-09-11 through 2026-10-10, by scheduled tip date in UTC. No sampling or truncation.
0 selected team captures · 0 excluded · 0 provisional outcomes · 0 final outcomes. These are different units; pending captures are not a count of players.
Registered before the checkpoint
4 prospectively enrolled capture revisions across the period’s checkpoints. 0 older, unregistered or other-protocol revisions are excluded from these model scores, out of 4 observed revisions. Revisions are not independent players or games.
Registration identity and exclusions
nrp_c6396045aa0d5391474ffb38
Only this protocol is compared. Registration freezes the descriptive research rules. It does not qualify a model for official recommendations.
Did the scheduled runs happen?
8 protocol-enrolled team checkpoints across all source-lineup cohorts. 8 upcoming
32 observed checkpoints were not enrolled and are excluded from this denominator. These execution counts are separate from scored players. Overall schedule completeness remains unknown: unresolved tips and source dates cannot be counted as zero missing games.
Coverage accounting rules
- Only schedule identities actually retrieved by Research are registered.
- A failed date fetch or invalid/TBD game cannot be counted from team opportunities; this report does not claim complete official-schedule coverage.
- Retrospective or late-first-seen slates are unobserved, not missed pre-tip opportunities.
- This is opportunity coverage, not a forecast score or recommendation ledger.
No settled paired forecasts yet. Upcoming captures remain visible in the denominator; historical outcomes are not retroactively turned into pregame predictions.
Population and comparison rules
- Public model comparisons score only prospectively protocol-enrolled captures; unregistered historical revisions are separately counted.
- League comparisons cover only the explicit UTC tip-date window; older_window navigates the full archive without sampling or truncation.
- Only captured opportunities are measured; schedule watchdog coverage is separate.
- Expected-lineup and user scenarios never enter the confirmed-lineup cohort.
- All scores use original frozen distributions and latest available official correction receipts.
- No model ranking is released without a registered sufficiently powered promotion protocol.