Errors get corrected in public
Published articles are permanent. Every revision ships with a dated public note, and recalled numbers stay on display. That is what makes the wins believable.
Revision notes
Relabeled October 8, 2026: the series is now called Deep Dives (it was Articles), because The Gravity Report now publishes two formats, Deep Dives (long, with the full method) and Articles (short). Same number, same address. No numbers, findings, or forecasts changed.
Relabeled October 8, 2026: the series is now called Deep Dives (it was Articles), because The Gravity Report now publishes two formats, Deep Dives (long, with the full method) and Articles (short). Same number, same address. No numbers, findings, or forecasts changed.
Relabeled October 8, 2026: the series is now called Deep Dives (it was Articles), because The Gravity Report now publishes two formats, Deep Dives (long, with the full method) and Articles (short). Same number, same address. No numbers, findings, or forecasts changed.
Oct 6 2026 — relabeled: an article typeset in LaTeX, not a research paper; GRAVITY-RS now defined as GRAVITY Retrospective; no numbers changed. Earlier: Jul 31 2026 — working-paper status labels added to every page (PDF footer + status notice, also on the web edition); no content changes — still v3.
Oct 6 2026 — relabeled: an article typeset in LaTeX, not a research paper; no content changes. Earlier: v2, Jul 11 2026, after external review — model output separated from editorial scenario weights (each weight now anchored to an observable rate); on/off language corrected to cross-season persistence, and the shrunk-input counterfactual re-run properly: #53–55, not ≈#48 (the harshest reading grazes the claim); comparables disclosed as 33 seasons / 22 players (32% per-player rate) + cutoff-sensitivity table; grading tightened: #51–70 = near miss, not push. The registered 75% is unchanged — that is the bet. Then Jul 31 2026 — working-paper status labels added to every page (PDF footer + status notice, also on the web edition); no findings or numbers changed.
Oct 6 2026 — relabeled: an article typeset in LaTeX, not a research paper; no content changes. Earlier: v2, Jul 11 2026, after external review — absence ledger corrected (the shoulder cost 48 games, not 73; 79 of 246 = 32.1%), inference language tightened, forward scenarios → availability × impact matrix, sensitivity appendix added (#112–#147 across all variations). Full note on the last page. The errors were ours. Then Jul 31 2026 — working-paper status labels added to every page (PDF footer + status notice, also on the web edition); no findings or numbers changed.
Method changes
The 30 now ranks teams with TORQUE, a new team model, and lives on its own page. The July method added up player projections weighted by last season's minutes; it ignored rookies and returnees and squeezed every player's minutes equally on deep rosters. TORQUE prices every player and hands out minutes like a real rotation. Tested on 2011–2026 seasons it was never tuned on, it ordered the standings at 0.736 against 0.700 for the July method: a real but modest gain. The July 2026 edition stays published unchanged, and forecast E1 is graded on it as registered.
Recalled: Jimmy Butler at #13 (v1)
Jul 10 2026, 10:20 AM · The 644 (Rankings Desk)
The sender recalled this ranking on Jul 10, 2026. Reason: the model rated 38 games of elite per-minute play as "availability risk." It didn't know the season ended with a torn ACL on Jan 19. With the same injury discount every rehab case gets, Butler moved #13 → #58 , and the correction shipped the same day, in public. In plain terms: the trash folder is part of the methodology. Errors get corrected in public and kept on display — that's what makes the wins believable.
Forecast amendments
Grading tightened against ourselves, following external review: №2's registered #51–70 band, originally labeled a "push," will be graded a near miss — a failed prediction owing a diagnosis, not partial credit. The claims, probabilities, and thresholds above stand exactly as registered.