Research · the reads index
Every model question this program has asked was pre-registered before any fit, judged on fixed gates, and recorded with its verdict — including the failures. Nothing ships without passing its gate; a FAIL stays visible here with its diagnosis. This page is parsed from the research log at bake, so it cannot drift from the record.
The RAPM campaign, closed
Eight pre-registered reads on identical gates. The two best designs correlate 0.97 — estimation is exhausted — and the gap to the bar is information, not math: the benchmark surface carries per-shot defender context that public feeds don't. Ratings stay internal-descriptive; the gate re-opens only with a new data tier.
Data gates
The reads · 21 sections, newest last · expand for the verdict
§15bgravity as a diet shift: the held-out confirmationPASS
§15's volume form failed on construct, not noise (G1 0.431 certified a stable effect); the usage-neutral diet form showed gravity's predicted direction in exploration (+0.196 with usage) and earns a clean held-out test. ### §0f update (202
Verdict: PASS (2025 full season and 2027 live; §1 sub-walk gates remain FAIL by design of the lineage).
§15gravity: teammate 3PA rate, a WOWY expression on the stint spineFAIL
Track C's first gravity expression measured teammates' 3PA volume per team possession; possessions are zero-sum, so it ranked usage displacement (Spearman −0.532 with usage) — the diet form carries the test to held-out data (§15b).
Verdict: FAIL — no surface.
§8the τ-swept tracking-SPM priorFAIL
§7's own diagnosis: the SPM prior is strong (oof r 0.778 vs its target) but τ = 250 — fixed for a box prior — swamps it at the CV λ; §7 named the next read: pre-register τ, scored on held-out game margin, not on the benchmark. This is that
Verdict: FAIL — τ-grid at (τ*=2000, λ*=3000): stability 0.494 (bar 0.60), benchmark 0.658 (bar 0.70) — the campaign closes at the response-side ceiling.
§7Tracking SPM prior in the xPTS ridgeFAIL
§6: the RAPM family's ceiling at the box tier is ~0.66-0.68 vs the 0.70 bar, and the gap was attributed to information (tracking), not estimation. The BDL trial landed that tier (hoops_mart `bdl_mart.py`, `bdl_season.parquet`: 47 season-ave
Verdict: FAIL — the tracking-SPM prior target loses to §5's Game Score prior.
§6Prior-the-prior: xPTS ridge shrinking toward the multi-season ratingFAIL
§5's Game Score prior delivered +0.018 (0.658 → 0.676); the remaining variance is estimator noise that a *rating-quality* prior absorbs better than a box prior. The two best designs correlate 0.97 — the §2 multi-season rating is a high-qual
Verdict: FAIL — benchmark 0.658 at λ=30k (§5's Game Score prior: 0.676), within noise of each other.
§5Box-score priors in the xPTS ridgeFAIL
§4: benchmark 0.658 with stability passing; the remaining 0.04 is partly LUMA's priors. The standard xRAPM practice: shrink each player toward a box-score prior rating instead of toward zero — the ridge recovers player impact faster for low
Verdict: FAIL — benchmark 0.676 (bar 0.70) at the peak; stability 0.572 there (bar 0.60), passing at λ=100k where the benchmark is 0.661.
§4Richer xPTS surface: shot-type contextFAIL
§3 confirmed the xPTS lever (single-season 0.464 → 0.621) with a distance-only surface; the remaining 0.08 to the bar is surface richness first. The Events text carries the shot type on every FGA line ("3PT Pullup Jump Shot", "Cutting Dunk
Verdict: FAIL — benchmark 0.658 at the stability-passing λ (bar 0.70).
§3xPTS responseFAIL
§2's verdict: multi-season raw-points RAPM peaks at 0.618 vs the 0.70 bar; the remaining distance is the response — LUMA's xpts_rapm uses expected points from shot quality, ours used raw points. §3 replaces the response: same design, same b
Verdict: FAIL — but the xPTS lever is confirmed: 0.621 on identical gates.
§2Multi-season RAPM 2022-23 → 2024-25FAIL
§1 v3's verdict: single-season raw-points RAPM peaks at 0.464 vs the 0.70 benchmark bar; the literature's main lever is multi-season data. The backbone (pbpstats intervals + possessions) is per-season pullable and both its gates passed for
Verdict: FAIL — multi-season RAPM peaks at 0.618; the YoY-stability gate confirms the estimator.
§1RAPM from ESPN PBP stintsFAIL
The spec's on/off, lineup and team-impact surfaces all need a baseline-plus-minus spine, and LUMA's `xpts_rapm` gives an external benchmark. First model of the program; the possession/stint derivation it needs ships only under this registra
Verdict: FAIL — single-season raw-points RAPM peaks at 0.464 vs the 0.70 bar.
§14env-normalized targets + drift-aware conformalFAIL
§13's two FAILs named exactly this read: per-36 targets carry era drift the linear `era` term only partially absorbs (H3), and pooled cross-era residuals under-cover when scaled to a later era (M4's naive conformal .72-.84 at nominal .90).
Verdict: FAIL, 1/5.
§0gmarket-layer data gate—
No explicit verdict line recorded in the log.
§13the senior-MLE rebuild: validation, v3 features, conformal intervalsFAIL
The owner-directed senior-ML review (`docs/MODELING-AUDIT-2026-11.md`) found the pipeline needed what it now has, in this order: **validate first** (the audit's H4), then features (H1/H2/M1/M2), then the model — with intervals (M4) rather t
Verdict: FAIL, 0/5.
§12cthe availability modelFAIL
Verdict: FAIL — availability is not forecastable from public history at this tier.
§12btracking-feature lagsFAIL
Verdict: FAIL, 0/5.
§12agradient boosting on the §11 feature matrixFAIL
Verdict: FAIL, 0/5.
§11ridge with age + lagsPASS
Verdict: PASS — 5/5 folds under the 0.98 bar vs all three §10 baselines
§10projection foundation: baselines read—
No explicit verdict line recorded in the log.
§0flineups/on-off stint gate—
No explicit verdict line recorded in the log.
§0ehistory-tier consistency gate—
No explicit verdict line recorded in the log.
§0dteam-tier analytics gate—
No explicit verdict line recorded in the log.
parsed from pipeline/results/RESULTS-HOOPS.md at bake — the internal research log, verdicts included