Exposure-Gated Cold-Start Prior

This frozen preseason prior applies only to first-NBA-season players. It keeps the completed 2024-25 lagged RAPM state for returning players and zero for other no-prior players. For every first-year player, it combines the draft RAPM rate with the cold-start exposure gate and the historical pooled replacement-token estimate:

\[ \widehat R_i^{cold} = p_i^{low}\,R^{replacement} +(1-p_i^{low})\,\widehat R_i^{draft}. \]

Here \(p_i^{low}\) is the probability of finishing under 5% of team possession opportunities, \(R^{replacement}\) is the pooled low-exposure token RAPM, and \(\widehat R_i^{draft}\) is the draft-profile rate estimate. This is a continuous mixture, not a hard player classification.

Frozen 2025-26 Inputs

The revised prior is artifacts/models/exposure_gated_cold_start/2025-26/exposure-gated-cold-start-2025-26-20260806T014656Z-cdafdec8/. All inputs end with 2024-25:

Component	Frozen source	Value / role
Draft rate	1996-97 to 2024-25 first-year players	Player-specific \(\widehat R_i^{draft}\)
Exposure gate	Same historical cutoff	Player-specific \(p_i^{low}\)
Replacement token	1996-97 to 2024-25 regular seasons	\(R^{replacement}=-4.740\) RAPM

The frozen evaluation is artifacts/models/frozen_prior_evaluation/2025-26/frozen-exposure-gated-cold-start-prior-2025-26-20260806T015348Z-089d2a36/. It evaluates realized 2025-26 lineup exposure, but does not use target-season scores, possessions, player coefficients, or playoff outcomes to form priors.

Results

Cohort	Possession RMSE	Game-margin RMSE	Team NetRtg RMSE	Pythagorean win RMSE
Regular season	1.198952	14.7413	4.6989	10.2466
Playoffs	1.192980	17.6725	-	-

Against the prior draft-cold-start ablation, the blend improves every listed regular-season metric. It is slightly worse in the 85-game playoff sample, so the Frozen Preseason Leaderboard continues to show both cohorts separately.

Revised 2025-26 Rookie Rankings

The sortable table ranks the blended preseason RAPM prior, not observed 2025-26 performance. It exposes both rate and exposure components so the change from the original draft-only ranking is inspectable. The full 100-player table is revised_rookie_rankings.parquet in the immutable prior run.

Top 25 Exposure-Gated Rankings

Rank	Player	Pos.	Pick	Draft rate	P(low exposure)	Blended RAPM prior
1	Collin Murray-Boyles	F	9	-0.60	5.6%	-0.83
2	Cedric Coward	G	11	-0.56	6.9%	-0.85
3	Dylan Harper	G	2	-0.85	2.2%	-0.94
4	Derik Queen	C	13	-0.55	9.3%	-0.94
5	Kon Knueppel	G-F	4	-0.83	3.0%	-0.95
6	VJ Edgecombe	G	3	-0.90	2.5%	-1.00
7	Walter Clayton Jr.	G	18	-0.50	12.0%	-1.01
8	Tre Johnson	G	6	-0.87	3.8%	-1.02
9	Egor Demin	G	8	-0.81	5.6%	-1.03
10	Cooper Flagg	F	1	-1.00	2.4%	-1.08
11	Carter Bryant	F	14	-0.74	8.8%	-1.09
12	Khaman Maluach	C	10	-0.77	8.1%	-1.10
13	Yang Hansen	C	16	-0.47	14.8%	-1.10
14	Ace Bailey	F	5	-0.96	3.9%	-1.11
15	Jeremiah Fears	G	7	-1.02	3.5%	-1.14
16	Noa Essengue	F	12	-0.90	8.3%	-1.22
17	Nolan Traore	G	19	-0.79	12.5%	-1.28
18	Nique Clifford	G	24	-0.47	19.2%	-1.29
19	Kasparas Jakucionis	G	20	-0.73	14.8%	-1.32
20	Joan Beringer	F	17	-0.74	15.2%	-1.34
21	Drake Powell	G-F	22	-0.72	17.5%	-1.42
22	Jase Richardson	G	25	-0.75	18.1%	-1.48
23	Will Riley	F	21	-0.72	20.9%	-1.56
24	Asa Newell	F	23	-0.61	23.6%	-1.58
25	Danny Wolf	F	27	-0.39	28.4%	-1.63

Rocco Zikarsky falls toward the bottom of the revised table because the gate assigns him a 71.7% low-exposure probability. That is the intended correction: draft-rate estimates no longer treat an exposure-risk player as a favorable one-number cold start.

Artifacts

File	Contents
`revised_rookie_rankings.parquet`	Draft rate, gate probabilities, replacement value, and blended prior
`metadata.json` / `manifest.json`	Three source runs, temporal boundary, and integrity hashes
Frozen evaluation artifacts	Possession, game, team NetRtg, and win metrics with the exact prior vector