L20-NAIL-MSP v0.3

Source-Aware Preseason Minute-Share Projection is the production successor to v0.2. Its implementation and validation history are tracked in GitHub issue #1.

The current baseline relies primarily on a player's total minutes in the immediately preceding season. That is a useful availability-and-role signal, but it conflates three distinct questions: how much a player played when available, how often the player was available, and whether the player was trusted to start. It also incorrectly turns a full injury absence into a zero role signal.

Player States

Every preseason roster candidate will use exactly one evidence pathway:

A missed season is missing role evidence, not evidence that a player should receive zero minutes. For a gap of (k) seasons, the last-observed role input (x) is attenuated before it enters the allocation model:

The separate gap-state and gap-length features allow the ridge-softmax model to learn how much additional role adjustment that attenuation requires. The decay and cold-start pathway use only historical preseason information, never realized target-season minutes.

Allocation Contract

The model will convert the source-aware player role scores into a constrained team allocation. The frozen evaluator assigns probability across every transaction-derived opening-roster candidate, matching its first-20-game target. The website then applies an explicit product constraint: it retains the 15 highest-scoring players for each team, renormalizes their shares to 240 regulation minutes per game, and starts all remaining rostered players at zero. Preseason NAIL remains a modest within-roster signal rather than a claim that coaches mechanically allocate minutes by player value.

Manual minutes overrides remain the final product layer for known current information, including an injury restriction, delayed availability, or an explicit coaching-role judgment.

Replacement-Token Experiment

The opening roster is necessarily incomplete over a 20-game window: trades, two-way conversions, signings, and emergency call-ups can receive minutes after opening day. The baseline treats those players as zero-probability outcomes. That is correct for the known-player allocation but makes the model incapable of representing the missing roster mass.

The current experiment augments each team distribution with one anonymous replacement token:

For a frozen target, the token target is the actual minute share of all players absent from the transaction-derived opening roster. The token has a single unpenalized global intercept; it does not know the future identity of an arrival. This first slice tests whether making that uncertainty explicit helps the full team allocation before adding any roster-churn predictors.

Player-level diagnostics remain calculated over real players, so the token cannot conceal individual allocation mistakes. Team total variation, Brier score, and cross-entropy are calculated on the augmented player-plus-token distribution. The experiment is isolated from the v0.3 production artifact and will be reported here after the frozen replay completes.

Result

The replacement-token replay completed all ten frozen holdouts, 2016-17 through 2025-26. It modestly improved the augmented team allocation, but it does not yet identify which individual teams will need replacement minutes well enough to change the product model.

Total variation improved in eight of ten holdouts, but Brier improved in only five. The token's mean prediction was 5.52% against an actual mean outside-opening-roster share of 5.08%; its median prediction was 5.40% while the actual median was only 2.71%. The actual and predicted token shares had a 0.47 team-level correlation, so the token has some roster-specific variation but remains close to a pooled league reserve.

Accordingly, the experiment is not promoted. It validates the missing-mass contract and produces a small team-allocation gain, but a useful product model should also forecast which teams are exposed to arrivals and departures. That requires a separately validated roster-churn feature set rather than treating the global token intercept as a finished availability model.

Artifact: artifacts/rotation/l20_source_aware_preseason_minute_share_replacement_token/l20-source-aware-replacement-token-20260909T052927Z-f726c5c.

Frozen Evaluation

The target remains each player's cumulative minute share over a team's first 20 regular-season games. Every evaluation fold must use the opening roster and information available before that season begins. v0.3 will be compared with v0.2 overall and separately for continuous players, gap returners, and cold starts. It is eligible to replace the current baseline only if it is at least competitive overall without creating a material failure in any player-state group.

Promotion Result

The first full rolling replay evaluated ten target seasons, 2016-17 through 2025-26. Each holdout used only earlier target seasons to select the role-rate shrinkage, gap decay, and ridge penalty. It then compared v0.3 with the similarly rolling v0.2 control.

v0.3 improved allocation total variation in all ten frozen holdouts. The initial artifact is artifacts/rotation/l20_source_aware_preseason_minute_share/l20-source-aware-20260909T005439Z-0cf9b7b.

Player-State Diagnostics

The candidate is trained conditionally on the transaction-derived opening roster. Players who later appeared but were absent from that roster remain an explicit evaluation-only outside_opening_roster category; no opening-game reconciliation is used during training.

The gap-returner prediction is intentionally conservative but is still mildly high on average. That is visible in the artifact rather than concealed by a zero-minute fallback.

Expanded Shrinkage Grid

The initial selector landed at the lower bounds for both the ridge penalty and role-rate shrinkage. A second full replay therefore expanded the candidate grid to (q \in {1, 2, 3, 4, 5, 15, 30}) games and \(\alpha \in \{0.01, 0.02, 0.05, 0.10, 1, 10\}\), while retaining the same gap-decay choices.

The expanded grid selected (q=1) in every frozen holdout. It selected \(\alpha=0.10\) in nine of ten holdouts, establishing that the ridge penalty is not at its lower boundary. Gap retention remained conservative early, typically (0.50), and reached (0.75) for the final two holdouts.

q=1 is intentionally accepted as the minimum practical role-evidence threshold. It leaves a 10-game MPG observation at 91% of its raw value and a 60-game observation at 98%, while avoiding the claim that a one-game role sample should be trusted with no attenuation. The selected result is artifacts/rotation/l20_source_aware_preseason_minute_share/l20-source-aware-20260909T013520Z-cd15741.

Production Artifact

For the 2026-27 forecast, the final model selected one configuration by running time-ordered internal validation over all eleven completed source seasons, then fitting the allocation model on all of them. It does not read 2026-27 game outcomes.

The frozen production artifact is artifacts/rotation/l20_source_aware_preseason_minute_share/l20-source-aware-production-20260909T042413Z-550aa6a. It is the allocation baseline used by the local Win Projections page; explicit user minute overrides remain a product-layer adjustment above this model.