One question, measured nine ways: how often is a starter's projection close enough to trust?
Success bar: Tier A = within 25% of what actually happened, for 75% of starters · Tier B = within 10%, for 40%. Starter pool: top 32 QB / 64 RB / 96 WR / 32 TE weekly, walk-forward. Touchdowns are a 10th governing metric, scored probabilistically (Brier skill + calibration), not by closeness.
Honesty box. The governing nine are the founder's 2026-07-25 metric law: the
eight stat metrics plus rb_rec_yds, with anytime_td removed to a probabilistic
10th metric. Both tiers are open. On 2023 (tune) the champion scores 53.4% Tier A /
34.7% Tier B against a 75% / 40% bar; the naive baseline is 44.36% / 23.36% on 2024 tune.
The receptions rows clear their bar; the six yardage rows (27–35 against a 75 target) are the open problem.
2025 is validation-only (confirm), 2026 is sealed.
Tier A per metric, 2023 tune window — three models on identical rows. The receptions rows carry the model; the yardage rows are where the 75% target lives. Hover any bar for detail.
Source: snapcore/eval/results/composite_v1b_2023_gate.json (Nash condition 1,
free 2023 replication, walk-forward, 3-seed). The dashed line is the 75% Tier-A success bar. Aggregate
53.4 / 34.7 is the ratified pool-level figure, not a simple mean of these rows.
Composite v1b Tier A on the three reception rows, 2023.
RB 89.8 · WR 85.7 · TE 90.4 — all at or near the receptions success bar. Volume (who gets the ball) is forecastable; this is the model working.
Composite v1b Tier A on the six yardage rows, 2023, vs the 75% bar.
qb_pass 54.0 is the one yardage row with real signal; the other five (qb_rush 28.4, rb_rush 35.1, rb_rec 30.2, wr_rec 27.7, te_rec 30.3) sit 27–35 against a 75 target. Weekly efficiency is ~42% of a receiver's variance and statistically a coin flip — that is the ceiling, and it belongs to the sport.
Week-to-week persistence (correlation r). This is a football fact, not a model result — it sets the ceiling for any pregame model, free or paid.
Weekly yards = volume × efficiency. Volume (who gets the ball) is forecastable; efficiency (what each touch yields) is ~42% of a receiver's weekly variance and statistically a coin flip. No pregame information touches that 42%. That is why the reception rows clear and the yardage rows don't.
| Metric | Naive A | Prior A | Composite A | Δ vs prior | ~Rows |
|---|
Tier A per metric, 2023 tune window. Source: composite_v1b_2023_gate.json. Aggregate pool-level
Tier A / B: composite v1b 53.4 / 34.7, prior champion 53.4-pool-comparable, naive 44.36 / 23.36
(2024 tune). Touchdowns (probabilistic, 10th metric): BSS 0.1276, ECE 0.0137 — Tier B cleared
on 2024 tune, holding on 2023 and pooled 2021–24. Paired net +387 (2023) / +395 (2024 tune); Nash condition 1 cleared.
rb_rec_yds, retired by the 2026-07-25 metric law and the 2026-08-04 scale correction. Those
scoreboard figures, and the lookup-table / ceiling-band / per-metric-Tier-B breakdowns built on them, have
been removed: they were a higher, easier scale (one fewer hard row, one inflated row) presented as a
scoreboard position. The per-metric Tier-B ladder is being rebuilt on the governing nine; until then this
page reports Tier A per metric plus the ratified aggregate. No number here is carry-over — every cell is
regraded from current result files.
Method notes: walk-forward only (every projection for week W uses strictly pre-W information); TUNE = 2024, CONFIRM = 2025 (validation-only, never used for selection), 2026 sealed. The champion is a tune/2023-free-replication result — promising, not yet 2025-confirmed at scale. Public benchmarks (RotoWire / ESPN / FantasyPros) are graded on identical rows and remain benchmark-only — never model inputs. See Progress for the program view.