What we measured, and what we killed

50.8% against the spread at our primary threshold, 95% CI [48.7, 52.9]. Breakeven at −110 is 52.4%. The interval contains a coin flip and the point estimate is below breakeven.

What we did: No NFL game-line model ships. THE LINE launched with no model at all, by decision.

54.1% ATS at the pre-registered threshold, 99% CI [48.3, 59.7] on 501 bets — an interval that still contains a coin flip. The shape across thresholds was incoherent (50.7 → 54.1 → 51.3 → 46.4).

What we did: Refused. 54.1% is exactly the number the pre-registration was written to defend against, and the holdout was not opened. (Attempt 3 was later spent on the 2026 thesis — see below.)

546 ideas searched on 2006–2022; three frozen and tested once on sealed 2023–2025 seasons against a bar set in advance. None cleared it. All three were positive there (blowout reversal +6.1% on 48 bets, Wong teasers +3.7% on 99 tickets at −120 but −1.5% at −135, wind unders +4.3% on 108 bets) with every 95% interval spanning a loss. Each shrank by half or more from where it was found.

What we did: No NFL betting formula is published as a winner. The sealed seasons have had their last look; only new seasons, graded at prices we recorded, can promote a survivor.

62–59–1, −9.4u (−8.4%, 95% CI −14 to −2) on 112 plays at an average −117, weeks 1–5 of 2026. Every game carried both sides at the better book, so a pair only profits on a middle: 1 of 60 landed, against about 8% needed.

What we did: The plays stay what they were built as — execution notes on where the better number is — and are not presented as a betting record.

Brier 0.356 against the market’s 0.246 over 372 captured prices; its confident calls (≥70% at −150 or better) went 14–21, and its closing-line value was negative.

What we did: It stays admin-only and behind no public number.

Receiving yards: real but miscalibrated (skill +1.28% Brier skill score, reliability deciles off by more than 10 points at n≥100). Receptions: −0.61% — no better than the player’s own season history.

What we did: Phase B never ran. No prop probability, projection or recommendation ships. We still capture and grade prop PRICES, because storing a market price is not forecasting an outcome.

AUC 0.548 against the book price’s 0.558. The market ranks hitters better than we do, and the constraint is discrimination — no amount of recalibration fixes it.

What we did: No hits probability is ever a recommendation, and the bet picker still shows the market’s de-vigged number alone. The model figure appears in exactly one place — the Edge Board’s analysis lane — where it renders beside the price’s own number, is sorted by the measured band rather than by raw edge, and carries this result and its reliability curve on the surface itself.

Hits priced heavier than −160 lost: −9.1% over 271 legs once hit prices were read correctly. We once reported that the same market lighter than −160 made +39% (n=231). That number was a pricing bug: 218 of those 231 legs were priced off the singles market before 2026-06-03 (see the hits-pricing receipt). With correct prices, only 18 hits ever came in lighter than −160 — too few to say anything.

What we did: A date-aware −160 cap on single hit plays. It still stands, because heavy hits lose either way, but the "+39% light band" is withdrawn as evidence for anything.

They did not. Before 2026-06-03 our odds feed read the price of a "singles" bet (pays only on a single) as the price of an anytime hit, then graded the bet as an anytime hit — Bobby Witt Jr. at −118 when the real price was −270. Hit plays published under the bug: +90.7 units, +33% (n=272). After the fix: −22.2 units, −7.7% (n=289). Two slates, 5/16 and 5/17, account for +70.3 of the season board’s +71.4 units.

What we did: Every hits result before 2026-06-03 is void as evidence. The classifier was fixed on 2026-06-03 (commit 118f84b), and from June onward hit plays returned almost exactly what their prices implied.

Six independent searches covered about 1,000 slices of June–August. The nine best formulas were frozen and tested once on September–October data none of them had seen, against a pass bar written before the test. All nine failed. The Daily Card, our published product from late June, finished −26.4 units. As a strikeout forecast, the market line beat SALCI: average miss 1.78 Ks against our 1.86.

What we did: No MLB pick is sold or presented as an edge. For 2027 the projection must beat the market line on held-out starts before it may disagree with one, and every test is registered before opening day.

Did not survive its closing-line check.

What we did: The lane was removed from the bet registry rather than left running quietly.

Failed its own baseline.

What we did: Kept as a research ledger, never promoted to a play surface.

Nine markets captured pregame and settled silently for months — n=306,491 settled bets. Every one of them lost to its own closing price: home runs −41.1% ROI, RBIs −22.3%, total bases −18.8%, runs −18.3%, batter strikeouts −18.0%, anytime hits −13.1%, and the three pitcher lanes −8.1% to −12.3%. Average CLV across all of it: 0.00pp — indistinguishable from noise.

What we did: None of these markets was ever promoted to a play surface, and four of them stopped being archived on 2026-09-03. The ledger is now a 14-day rolling window rather than a growing table, because it was re-confirming a settled verdict at ~5,650 rows a day. The full-history numbers above are frozen in the artifact so the result survives the rows.

From late August FanDuel posted a 60-rung ladder of ALTERNATE totals (23.5 to 82.5) alongside each game's real total, and our archive read every rung as if it were the total. Five week-1 plays were published at FanDuel lines that were never its real total (the worst: under 70.5 on a game whose total was 40.5), and five more were published at DraftKings because its real line looked better than one of those fake rungs. As graded, the ten went 6–4.

What we did: All ten are VOIDED — both halves, because neither claim was true, and removing only the losers would have flattered the record. They stay on the ledger, marked void, with their original prices. The archive now keeps only a book's real total (chosen by price, closest to even), and every surface that reads totals reduces the history to it first. Every other week-1 total play was checked against the book's real total at the moment it was published; all 22 matched. An audit on 2026-09-26 found two more errors in how the ledger was scored, and both were corrected in place. First, the voided plays were still counted in the headline figures: line value read +2.6 points a play instead of +1.0, and beat-the-close read 30 of 60 instead of 25 of 50. Second, four ARI@LAC spread plays had their closing line read from the wrong side of the market (+9.5 instead of −9.5), which gave them closing-line values of about ±20 points instead of +2, −1, +1 and 0. No play's result changed.

Measured at 105 inferred routes / 4.6 games — refuted in both units, so it is not an artifact of converting between them.

What we did: We publish our own number and say where the other one came from. YPRR also turned out to be mostly a usage metric: it is TPRR x yards-per-target, and its stability sits beside the usage factor (0.867) not the efficiency one (0.425).

Reversed. Passing xTD per attempt reads r=0.205 and stabilizes at >20.4 games; the raw passing touchdown rate reads 0.244 at >15.4 games.

What we did: The index uses the raw passing rate for quarterbacks — the one place a raw touchdown rate survives. A passer’s touchdown rate carries persistent skill that field position cannot see, which is exactly what xTD removes by design.

Measured over 135 season-weeks, paired: an index of pure opportunity LOSES (WR/TE -0.016 [-0.022, -0.009]). Adding production makes it +0.005 [0.001, 0.010] — but that version contains points per game as a component, so it beats a baseline it is built from.

What we did: The grade ships as a description of where a player stands, never as a ranking of next week. Quarterbacks clear nothing in any configuration tried.

Withdrawn on re-measurement. The two providers agree exactly where both record a player (r=1.0000, 100% exact match). The apparent 14% population gap is entirely weeks 16-18 of one file — coverage is 100.0% in weeks 1-15 and 44/20/6% after.

What we did: We had banned league-wide averages and percentile ranks across 2023 on the strength of an undiagnosed 14%. They are allowed; what is banned is mixing the two 2023 chartings, which double-count.

They are not. `target_share` has 60 non-null rows in 2003, 168 in 2004, ZERO in 2005, 71 in 2006 and 452 in 2007, against ~16,500 a season either side; `air_yards_share` is absent 2003–2005. Our aggregator fell back to 0, so six seasons of "commanded 0% of his team’s targets" entered a training set as measured data.

What we did: Caught because the air-yards projection came out 14.6% WORSE than doing nothing. Nulling the missing values moved target share from +2.0% to +7.9% and its R² from 0.304 to 0.519. Missing is now null and the row is dropped — never a zero that arithmetic accepts.

Both wrong, and wrong the same way. Snap counts start in **2013** (23,800 rows; the original probe tested only 1999, 2006 and 2016). Expected touchdowns cover **2009–2025**, ~5,000 player-games a season — the "three seasons" figure came from a probe that only asked about 2023–2025.

What we did: Each stale bound had kept a real metric out of the projection model — snap share, the fastest-stabilizing usage metric in the sport, and expected touchdowns. Both now project (+7.3/+6.0% and +16.0/+16.5%). The bound is read from `lib/nfl/coverage.ts` in one place instead of being restated in migration headers and script guards, and the 2013–2015 snap seasons were backfilled so the claim is true in the database too.