56.7%
Under rate in the 2023-2025 backtest · n=598
Case study · from the graveyard
Our best angle hit 56.7% over 598 games. We killed it anyway.
This is the full life of the getaway-day UNDER: found on June 21, validated on June 22, stress-tested on June 23, and killed on July 1 when 2026 out-of-sample data reversed the signal. Eleven days, five experiments, every number below is a real row in our research database. A tout would have sold you this angle. We are showing you its autopsy.
61.1%
Road-team subtype, 2024-2025 · n=234
41.4%
2026 out-of-sample · n=70 · the reversal
Killed
July 1, 2026 · model deprecated, hypothesis to the graveyard
June 21, 2026Discovery · hypothesis registered
A dead angle points at a live one
We were killing a different idea, the plain outdoor day-game UNDER, when the segmentation showed the whole signal was hiding in one subtype. Standard outdoor day games were sub-breakeven in the 2025 holdout at 49.7% (n=378). But games where either team had to fly to another city within 24 hours, the classic getaway day, went under 56.7% of the time across 598 games from 2023 to 2025.
The mechanism read clean: travel-day logistics split attention, managers compress pitcher usage to protect arms before a flight, and the road team plays against a departure clock. Published circadian-travel research pointed the same direction. So it entered the pipeline as hypothesis 54: a claim with a threshold, not a play.
Getaway games 56.7% under · n=598
All other outdoor day games 49.7% · n=378
June 22, 2026Validation · Phase 1 backtest, four gates
It passed everything we threw at it
The formal backtest replicated the seed at 56.8% on 518 decided games. The purest version of the mechanism, road team traveling, hit 61.1% in 2024-2025 (n=234). 2023 was flat at 50.9%, consistent with the pitch-clock adjustment year suppressing the fatigue mechanism, while 2024-2025 combined ran 59.5% with a bootstrap 90% confidence floor of 56.2%, far above the 52.4% a bettor needs to beat the vig. The line barely moved open to close, so the books were not pricing it. Four gates, four passes.
Replication 56.8% · n=518
Road-only 61.1% · n=234
CI floor 56.2% vs 52.4% breakeven
If we sold picks, this is the day the sales page would have gone up.
June 23, 2026Stress test · the devil's advocate blocks it
Our own adversarial review said no
Every validated angle here gets a devil's-advocate pass whose only job is to kill it. It found what the averages hid: the edge was concentrated in a handful of parks. Kauffman Stadium read 88.9% under (16 of 18), an outlier so extreme it got its own data-integrity check, while Yankee Stadium went 0 for 10, Dodger Stadium 40%, Wrigley 42%. A universal fatigue mechanism should not care which stadium you are in. Re-run with venues balanced, the confidence floor slipped to 51.87%, half a point below the gate. Deployment: blocked. The public verdict stayed "validated, but not bettable as specified."
Kauffman 88.9% under · 16/18
Yankee Stadium 0% · 0/10
Venue-balanced CI floor 51.87% · gate 52.38%
June 29, 2026The tiebreaker attempt · not enough data
We refused to call it on 16 games
One honest way to break the tie: extend the sample with live 2026 games. But only 16 venue-balanced 2026 games had closing-line coverage, and 57 of 73 qualifying games were missing odds. Sixteen games is a coin flip's shadow, so instead of a verdict we ordered a data backfill and waited. The hypothesis sat unchanged: no upgrade, no kill, no play.
July 1, 2026The reversal · out-of-sample verdict, kill
2026 said no, loudly
With the backfill in, all of 2026 became a true out-of-sample test: games the angle had never seen, scored against the rule as registered. Result: 41.4% under on 70 venue-balanced games (44.8% across all 87), a full 14.7 points below the 2025 holdout. The combined confidence floor fell to 49.2% against the 52.4% gate.
Before killing it we checked ourselves: re-running the historical-only data reproduced the June 23 result to three decimal places, so the pipeline was intact and the reversal was real, not a data artifact. Same day: model deprecated, hypothesis 54 moved to the public graveyard. An edge that was probably real in 2024-2025 did not survive contact with 2026, and no one can know that in advance. That is what perishable means.
2026 OOS 41.4% under · n=70
Drop vs 2025 -14.7 pts
Pipeline integrity check passed
Season by season · the same rule, scored honestly (gold line = 52.4% breakeven)
Two seasons comfortably above breakeven, then the first season the angle was scored on data it had never seen, it fell below a coin flip. This is why a backtest is never a record and why we grade forward, out-of-sample, before anything is treated as an edge.
Why we publish our kills
A 56.7% angle over 598 games is exactly the kind of number the pick-selling industry retires on. It would have made a great sales page right up until it lost you money at any price in 2026. The backtest was not a lie. It just was not the future.
What survives this story is not the angle, it is the process: register the claim before looking, let an adversary attack it, refuse verdicts on thin samples, verify the pipeline before trusting a reversal, and kill in public. Every angle that reaches our board has passed through the same machinery that killed this one. We kill our own angles. That is the product.
Provenance: hypothesis 54 (proposed Jun 21, 2026 · rejected Jul 1, 2026), experiments 39, 41, 42, 56 and 61, model 29 (deprecated Jul 1, 2026), all rows in the Edge Labs research database. Numbers verified against the live rows on Jul 2, 2026. Nothing on this page is a pick, 21+, research and context only.