The Record

Flat-$2 = one $2 win bet on every flag. Losses are counted identically to wins.

Why the counts on this page differ. 184 scored selections = 47 wins + 125 losses + 12 voided by scratch. The 172 settled bets behind the ROI exclude the voids, because a scratched runner is money returned, not a bet lost. The closing-line table below is smaller again. It only covers selections we have a closing price for. Calibration is larger than all of them: it scores every runner we priced, not just the ones we flagged.

Read the range, not the headline. On 172 settled bets the ROI figure could comfortably be anywhere from -32% to +21%, and on this record there is roughly a 70% chance the true long-run figure is negative. A few months is nowhere near enough racing to tell a real edge from a hot streak, and anyone quoting the headline number at you , including us, is quoting the least reliable part of this page.

The record is concentrated: 2026-07-26 contributed +39.98 units across both tracks. Every other race day combined — 27 of them — comes to -62.60 units. We would rather you learned that here than worked it out yourself.

The card is also published on TwinSpires as Edge Card Plays. Edge Card is an independent publication, not a TwinSpires product, and the record here is the whole record either way.

Our flags are straight win bets. In a pari-mutuel pool your payout is set by the final odds no matter when you bet, so there is no price to hold you to . This table is a model diagnostic, not a bet rule. Overlay at close counts the flags whose closing odds still beat our fair number: the value we saw was real at post. Value at the close is the stricter test. the close also reached the reference price we recorded before the race. When these run low, it means our fair line is too long, usually because it is anchored to a generous morning line rather than the real market. That is the known v0 weakness the version notes track. ML→close is the opposite signal: how far the crowd moved toward our flags after we published.

ML→close = to-1 points the price moved from the morning line to post. Positive means it shortened. This is descriptive only: because your payout is the final price regardless of when you bet, our “play at” figure is a reference price and not a rule. There is no earlier price to catch and nothing to hold you to. To be unambiguous, because it matters: a flag that closes below that figure is still graded on the public record, win or lose. It does not become a pass after the fact. A run of “short” closes is a sign our number was set against a loose morning line, not the real market.

Our call is an early hold/shorten read set from each flag’s committed board price. Long-priced flags the crowd tends to leave alone hold, short ones get bet down. It is a heuristic on a small sample, shown next to what actually happened so you can watch it prove out or fail in public. Not a bet-timing tip: you are paid the closing price whenever you bet, so a flag we call “likely bet down” is simply a likely pass. The read is versioned: cards locked before August 24, 2026 carry the call they were published with, from one pair of thresholds used at both meets, and cards from that date use thresholds set separately for each track. This column therefore mixes two versions on purpose. A call we have already published is not restated.

Across every runner we priced, not just the flags, binned by our model win probability against how often they actually won. The probabilities were hash-chained before post, so this is an honest reliability check, not a backtest.

Brier score: lower is better (0 = perfect). Ours and the morning line’s sit within a thousandth of each other, and that is the honest headline: a standalone Brier comparison cannot tell a real model from a plausible one. Both numbers are dominated by the fact that most horses lose, and a model that did nothing but copy the board would score the same as the board.

So we ask the question that separates them: does our number add anything the board does not already contain? Fit the outcome on the market’s estimate alone, then again with ours added, and test whether the second fit is genuinely better.

The answer is suggestive, at the edge of significance, on a sample too small to lean on. We are publishing all of it rather than the flattering part.

The test was chosen and written down before we ran it, by someone other than the person who built the model. That matters more than the result. Anyone can fit models until one of them clears a threshold and then publish that one; the only defence is to fix the method in advance and be bound by whatever it returns. The estimator below, a conditional logit on race choice sets, was specified, written and frozen by a separate reviewer before any of these numbers existed, and we are reporting exactly what it produced, including the two checks that go against us.

Frozen analysis, as of 2026-07-31. Every figure in this section is pinned to the ledger as it stood that day and is deliberately not re-fit as new races are graded. Re-fitting a pre-registered test to newer data is the thing pre-registration exists to prevent. The race and runner counts below are therefore smaller than the calibration counts above, and that is expected, not an error. When it is re-run, it will be re-run in full and dated again.

The pre-registered test is a conditional logit over race choice sets, 138 races, 1,170 runners. Our estimate carries a coefficient of +0.6643 (standard error 0.3013) once the board is already in the model, likelihood-ratio χ²(1) = 4.833, p = 0.0279.

Two robustness checks disagree with that, and they matter more than the headline. Allowing for the fact that horses in the same race are not independent gives a cluster-robust p of 0.0548, the wrong side of the conventional 0.05 line. And when whole races are resampled and the test refit, 41% of those refits fail to reach significance at all. A result that disappears two times in five under resampling is not a finding to build on.

An earlier version of this page reported p = 0.026 from the same data. That number was overstated and we are correcting it. It treated all 1,181 runners as independent observations. They are not: exactly one horse wins each race, and the probabilities are worked out relative to the others in that race, so the real number of independent units is the 138 races, not the runners. Ignoring that makes any result look stronger than it is.

Superseded figure below. The current record is 47-125, -22.62 units, -6.6% ROI. The paragraph that follows is the frozen snapshot from the pre-registered analysis and is not our current performance. It is kept because deleting a figure we once published is not something we do.

Alongside it, the money, as measured on 2026-07-31, when this analysis was run. It is a smaller sample than the running record above and does not match it, because races have been graded since. The running record is the current figure; this one is frozen so the pre-registered analysis cannot be re-fit to newer data. Over 58 settled bets across 16 race days, flat-$2 profit is +7.24 units, an ROI of +6.2% with a 95% range of -46% to +68%, and roughly a 43% chance the true long-run figure is negative. Whole race days were resampled rather than individual bets, because one day carries most of the profit.

We are deliberately not printing a yes or no here. The method and every number are published together so you can weigh them yourself, and the full output is in the machine-readable record. Judge it again at the end of the meet, when there is more of it.

The public ledger is one continuous SHA-256 chain: every card’s lock entry references the exact bytes of the entry before it, so a back-edit anywhere breaks the walk. Chain status: verified ✓ · 949 entries · 54 cards.

Corrections and verification

Two integrity defects were found on 2026-07-26 and are recorded here in full. No card, metadata file, ledger entry, recorded hash, or timestamp proof was rewritten to make either go away.

How to verify a card

Download card.json from any card page and hash it (sha256sum card.json). That value should equal the downloadable artifact hash shown in that card’s “Verifying this card” box. The card’s OpenTimestamps proof commits to those same bytes.

There are four independent evidence layers, answering different questions:

1. The append-only ledger and Git history: when the card was committed, and that nothing was edited afterwards.
2. The card lock and publication timestamps: that the card existed before first post.
3. The OpenTimestamps calendar commitment: independent third-party evidence that those exact bytes existed when we say they did.
4. A Bitcoin attestation, once the proof is upgraded: the same evidence anchored in a block.

Layers 1 and 2 establish that a pick was published before post. A Bitcoin confirmation arriving later strengthens the external evidence; it is not what makes the pre-post claim true, and a pending proof does not weaken it.

Card hashes: recorded lock hash vs downloadable file

Cards published before 2026-07-26 were generated on Windows. The lock hash was calculated from CRLF-formatted bytes before Git stored the same JSON with LF line endings, so the recorded hash does not match the file you can download. The selections and JSON values are unchanged. The difference is exclusively line endings, proven byte-for-byte for every affected card. This broke reproducible verification, which is why it is disclosed here rather than described as a formatting detail.

Provenance display defects

Cases where the evidence was intact but the site did not show it. Recorded to the same standard as anything else here: what a reader saw, what actually caused it, and how long it lasted.

Machine-readable: provenance_defects.jsonl

Machine-readable: card_hash_corrections.jsonl (hash-chained, append-only) and proof_status.json.

Metadata proofs

A card’s meta.json holds presentation detail: publish time, weather, per-pick notes. For 2026-07-17 (both tracks) that file was edited after its proof was created, so the original proof commits to the metadata as first published, not to the current file. Both are kept: meta.json.ots is the original-publication proof, and meta.current.json.ots commits to the current bytes, stamped 2026-07-26 and labelled as such. The card hash embedded in the metadata never changed, and the card.json proofs are unaffected. Metadata is presentation, not locked evidence. The card is the locked evidence.

Result entries: provisional grades superseded by chart grades

Provisional results were later superseded by enriched chart results. Both entries remain in the append-only ledger, while the canonical record counts each selection once. Settlement values (win/loss and P/L) agree in every case; the later entry only adds finishing position, closing odds and drift.

Machine-readable: record.json. It distinguishes raw ledger entries from effective results, so the de-duplication rule is visible rather than something you have to reconstruct.