PAPER DESK -- PAPER ONLY, simulated maker quotes on the recorded Kalshi tape; not a verdict
How the models work
export 2026-10-22, generated 2026-10-23T19:43:59Z; 2 refusal(s)
Three dates frame this record. Capture history starts 2026-09-24, the first night in the fill-realism table. The first ledger night with paper fills is 2026-10-03. The prospective forward paper series, with schedules written before the games, runs from 2026-10-06 (evidence packet section 5). No calibration verdict yet, for or against the model.
Scoreboard arms (scored on every tick against the outcome)
| arm | what it is |
|---|---|
| A | the venue mid at the tick, untouched (the reference every other arm is scored against) |
| B | walk-forward fit on mid-only features over the state-complete population |
| B_lag | the previous receipt's mid (a staleness control) |
| C | walk-forward fit on the mid plus the sport's game-state features, same population as B |
| D | C restricted to ticks where the pregame model is present; unavailable until a prior model row exists |
mlb model card
pregame rating model (Elo / Poisson / NegBinom / NNLS family) -> leak-free per-sport recalibration (Platt / temperature / isotonic) -> one calibrated p(win) anchor; see docs/PLATFORM.md and domains/<sport>/predictor.py
in-game daily fit no: no daily fit (the fit CLI takes nba, nfl, soccer, tennis); in-game calibration is read from scoreboard arms B and C against the mid arm A
- mid: quote around the captured venue mid
- venue_prior_state_blend: MLB: blend of the frozen venue pregame mid (p0) with the live state
| figure | metric | model | reference | gap | 95% CI | n | verdict | as of | note |
|---|---|---|---|---|---|---|---|---|---|
| moneyline | Brier | 0.2429 | 0.2390 | 0.0039 | [0.0028, 0.0051] | 13992 | TRAILS_CLOSE | 2026-09-24 | |
| in-play win probability | Brier | 0.2377 | 0.2067 | 0.0310 | [0.017, 0.0454] | 227 | BEHIND | 2026-09-14 |
On MLB moneyline the pregame model trails the devigged close by +0.0039 Brier (95% CI [+0.0028, +0.0051], DM p 3.6e-12, n=13,992). (source docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md)
On paired in-play ticks the MLB in-game model trails the contemporaneous Kalshi price (Brier 0.2377 vs 0.2067, delta +0.0310, CI [0.0170, 0.0454], 227 games); the corpus later failed the join-integrity check and the 178-game segmented revision keeps the verdict. (source docs/evidence/ingame/GATE_A0_2026-09-14.md)
nba model card
pregame rating model (Elo / Poisson / NegBinom / NNLS family) -> leak-free per-sport recalibration (Platt / temperature / isotonic) -> one calibrated p(win) anchor; see docs/PLATFORM.md and domains/<sport>/predictor.py
in-game daily fit yes: frozen daily fit from the scoreboard history (live_shadow_fit)
- mid: quote around the captured venue mid
- scoreboard_b: the frozen daily fit of scoreboard arm B (S780)
- scoreboard_c: the frozen daily fit of scoreboard arm C (S780)
- nba_logistic_v1: NBA: the frozen nba_logistic pricer from the state row with p0 = venue pregame mid
| figure | metric | model | reference | gap | 95% CI | n | verdict | as of | note |
|---|---|---|---|---|---|---|---|---|---|
| moneyline | Brier | 0.1735 | 0.1666 | 0.0069 | [-0.0038, 0.0175] | 372 | MATCHES_CLOSE | 2026-09-24 | |
| moneyline | Brier | 0.1735 | 0.1672 | 0.0063 | [-0.0042, 0.0168] | 372 | MATCHES_CLOSE | 2026-09-24 |
On NBA moneyline the gap to the Shin-devigged close, +0.0069 Brier (95% CI [-0.0038, +0.0175], DM p 0.20, n=372), is indistinguishable from zero and underpowered (MDE 0.0152). (source docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md)
Against the proportional-devig close the NBA moneyline gap is +0.0063 Brier (95% CI [-0.0042, +0.0168], DM p 0.237, n=372); same verdict. (source docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md)
nfl model card
pregame rating model (Elo / Poisson / NegBinom / NNLS family) -> leak-free per-sport recalibration (Platt / temperature / isotonic) -> one calibrated p(win) anchor; see docs/PLATFORM.md and domains/<sport>/predictor.py
in-game daily fit yes: frozen daily fit from the scoreboard history (live_shadow_fit)
- mid: quote around the captured venue mid
- scoreboard_b: the frozen daily fit of scoreboard arm B (S780)
- scoreboard_c: the frozen daily fit of scoreboard arm C (S780)
soccer model card
pregame rating model (Elo / Poisson / NegBinom / NNLS family) -> leak-free per-sport recalibration (Platt / temperature / isotonic) -> one calibrated p(win) anchor; see docs/PLATFORM.md and domains/<sport>/predictor.py
in-game daily fit yes: frozen daily fit from the scoreboard history (live_shadow_fit)
- mid: quote around the captured venue mid
- scoreboard_b: the frozen daily fit of scoreboard arm B (S780)
- scoreboard_c: the frozen daily fit of scoreboard arm C (S780)
| figure | metric | model | reference | gap | 95% CI | n | verdict | as of | note |
|---|---|---|---|---|---|---|---|---|---|
| O/U 2.5 | Brier | 0.2465 | 0.2390 | 0.0076 | [0.0059, 0.0092] | 7558 | TRAILS_CLOSE | 2026-09-24 | |
| in-play win probability | Brier | 0.2279 | 0.1427 | 0.0852 | [0.0488, 0.127] | 51 | BEHIND | 2026-09-14 |
On soccer O/U 2.5 the pregame model trails the devigged close by +0.0076 Brier (95% CI [+0.0059, +0.0092], DM p 1.1e-19, n=7,558). (source docs/evidence/pregame/DM_RECOMPUTE_2026-09-24.md)
On paired in-play ticks the soccer in-game model trails the Kalshi price (Brier 0.2279 vs 0.1427, delta +0.0852, CI [0.0488, 0.1270], 51 games); same join-integrity caveat. (source docs/evidence/ingame/GATE_A0_2026-09-14.md)
tennis model card
pregame rating model (Elo / Poisson / NegBinom / NNLS family) -> leak-free per-sport recalibration (Platt / temperature / isotonic) -> one calibrated p(win) anchor; see docs/PLATFORM.md and domains/<sport>/predictor.py
in-game daily fit yes: frozen daily fit from the scoreboard history (live_shadow_fit)
- mid: quote around the captured venue mid
- scoreboard_b: the frozen daily fit of scoreboard arm B (S780)