The −40 update: two public ledgers and a promotion rule we can't fudge

Three weeks ago we audited our own ledger, found a settlement bug that had scored 18 non-runners as losses, and corrected the public record. This is the follow-up we promised — and it comes with two changes to how this site works.

(Housekeeping first: our racing-data subscription lapsed on 16 July and the site ran on stale data until 8 August. Nothing settled and no picks were published in that window — the gap is visible in the ledger, as it should be. The pipeline is back, backfilled, and now alarms loudly instead of failing silently.)

Where the numbers actually stand

Everything below is computed live from the stored settlement ledger — the same rows the performance page reads.

The public card is down. Not "temporary variance" down, not "wait for the good run" down — down 32.5 points after 200 settled picks, and we publish that number on our own homepage. The luck band at v1.0's prices is roughly ±33 points at one sigma, which means this result is consistent with a model that has no edge and got ordinary luck. That is the honest reading and we'd rather write it ourselves than have someone else write it for us.

Meanwhile the shadow model — v1.1, a machine-learned value model that only bets when its probability estimate beats the market price, and has run silently under identical rules since day one — is quietly in profit. Smaller sample, fewer picks by design, +3.9 points.

Change one: both ledgers are now public

Until today the site led with v1.0 and kept v1.1 on a comparison page. From today, /performance shows both ledgers side by side — settled counts, strike rates, P&L and the luck band for each. If the shadow model collapses from here, you'll watch it happen in the same table.

Change two: a promotion rule we're locking in before we know the answer

The obvious temptation with a losing public model and a winning shadow one is to swap them today and pretend the ML model was the headline all along. We're not doing that, because a promotion decided after peeking at the results is exactly the kind of quiet cherry-pick this site exists to reject.

Instead, the rule is pre-registered as of today, 8 August 2026:

v1.1 replaces v1.0 on the public card if and only if both models reach 300+ settled picks and v1.1 leads by 25 or more points at that moment.

Right now v1.1's lead is over 36 points — but v1.0 has 200 settled picks and v1.1 has 152, so nothing promotes. If the lead evaporates before both hit 300, v1.0 stays and the rule will have saved us from crowning a lucky streak. Either way the decision is mechanical, dated, and public before the outcome is known. That's the point.

Change three: the picks are now frozen at the database level

Every published pick is protected by a database trigger: any change to a pick's identity — the horse, the price, the bookmaker, the score — or any deletion is permanently recorded in a public audit table. Settlement stamps (win, lose, void) are the only writes allowed through untracked. The performance page shows the running revision count; it is zero today and the goal is that it stays zero forever.

None of this makes the model better. It makes the record trustworthy while we do the slow work of making the model better — that work is the v2 pipeline, and it gets its own write-up when there's something real to show.

Every number in this post is checkable: the ledger · the CLV audit · the correction post. 18+ · Tips are opinions, not advice · BeGambleAware.org.