Nine stats. Every starter, every week, graded against what actually happened. No averages, no filler — this table is the truth of the project, and it moves only when the model earns it.
How to read it. Each number answers one question: what share of starter-player weeks did we predict this stat close enough to trust? "Close enough" = within ±25% of the actual result (hover any ⓘ for the exact rule and a worked example). Every prediction uses only information available before that week's kickoff. Baseline = predicting each player's own recent average. 2025 was validation-only; the frozen hypotheses shown here are now spent and were never trained or selected on 2025 outcomes. The 2026 season begins in September and is the next prospective scoring window. Verified through the same pipeline shown here.
| Stat | Baseline | Previous | Current | Target i | vs baseline | Rank skill i |
|---|
The RB receiving-yards point model did not suddenly become accurate for every player. A frozen distribution model ranked each week's projections by interval width and selected the lowest-uncertainty 10%. That small cohort cleared the 75% goal again on frozen 2025 validation.
(q90 − q10) / max(|q50|, 15); lower is better, and each week selects the lowest
ceil(10%) with a fixed player-ID tie-break. The rule was chosen on 2024 before 2025 was read.
This is an abstention and role-certainty result, not 82% accuracy across every RB projection.The confidence head mostly found backs with a reliably minimal receiving role. In the selected 2025 cohort, 90/119 outcomes were zero receiving yards (75.63%), and 115/119 point estimates sat within 3.75 yards of zero. Among the 29 selected nonzero outcomes, 10/29 were Tier-A hits (34.48%). Those figures are a post-result diagnostic explaining the mechanism, not a new selection rule or promotion gate.
All seven finished with zero receiving yards. This is a concrete example of the model identifying “little or no receiving role,” not calling seven featured-back yardage totals.
| Player | Team | Point | Central-80 range | Actual | Grade |
|---|---|---|---|---|---|
| Rasheen Ali | BAL | 0.0 | 0.0–1.8 | 0 | Hit |
| Dameon Pierce | HOU | 0.0 | 0.0–2.5 | 0 | Hit |
| Brian Robinson | SF | 0.0 | 0.0–3.2 | 0 | Hit |
| Chris Brooks | GB | 0.0 | 0.0–3.4 | 0 | Hit |
| Chris Rodriguez Jr. | WAS | 0.0 | 0.0–14.5 | 0 | Hit |
| Trevor Etienne | CAR | 0.3 | 0.0–15.4 | 0 | Hit |
| Terrell Jennings | NE | −0.5 | 0.0–15.4 | 0 | Hit |
Display ranges floor negative lower bounds at zero for readability;
signed raw outputs remain in the validation receipt. Frozen evidence:
ws1-phase-b2-2025-week8-public-example.json.
Method receipts: predictions are walk-forward (week W uses strictly pre-W information), scored on the weekly starter pool (32 QB / 64 RB / 96 WR / 32 TE), vendor projections never touch the model, and every number on this page traces to a committed scorecard artifact in the repo. Deep dives: how a prediction is made · the ceiling study. The stricter ±10% grading (Tier B) joins this board per-stat with the next scorecard run.