The fine print, in plain sight

Verification & disclosures

This service’s claim is not that the picks always win — the record shows plainly that they don’t. The claim is that the record itself is honest. This page sets out exactly what our published evidence proves, what it doesn’t, and every known gap.

What the SHA-256 anchors prove

Every day’s selections are published to the public proof archive with a SHA-256 hash. The hash proves byte-for-byte file integrity: if a selection file were altered after publication — a losing pick removed, a percentage massaged — its hash would no longer match, and anyone can recompute it from the file alone. What a hash does not prove, by itself, is when the file was published.

Timing evidence comes from GitHub: a file publicly visible before a race is run cannot have been written after it. For the historical record this relies on GitHub’s public availability evidence where available. Durable independent third-party timestamping is being added for future picks, so pre-race timing will not depend on any single platform.

Known gaps in the record

  • 13–22 June 2026: the earliest selections were archived after racing, not before. They carry no pre-race timing evidence and are excluded from the verified record for that reason.
  • 69 win selections (13 June – 5 July 2026) predate the public archive window and are not publicly reconstructible from the current archive. They exist in the operator’s internal log but you cannot independently verify them — so we don’t count them, for better or worse, in any figure on this site.
  • Two records exist: the public verified record (from 6 July 2026 — everything on the record and returns pages), and a longer internal operator record that includes the unverifiable early weeks. Where the two differ, the public verified record is the only one we publish or advertise.

Frozen model manifest

The selections are produced by a fixed model, frozen on 13 June 2026 and unchanged since. These are the SHA-256 fingerprints of the deployed artifacts — if the model were quietly retrained or swapped, these values would change:

ArtifactSHA-256
Win model (LightGBM)d79bd56c162f9259d24ba12a04298243aea60d6df0f212148634842a920f2f92
Place model (LightGBM)f5d62b4adb98473ad971d62ff85ccd983cb8b4e9c1108bbde277516f8ce4d93a
Feature list (135 features)2a1f781e8535e6a763ef0d816376d4a9761e504c588f41fc50b855b1c089b3b6

Manifest recorded 30 August 2026 from the deployed artifacts (files dated 13 June 2026). The scoring wrapper code around the frozen model is maintained separately; its changes are recorded in the archive’s decision notes. A signed manifest file in the public archive is planned so this table can be verified without trusting this page.

Questions about any of this — or found something that doesn’t verify? We want to know first.