These figures retain the original AI selection and pricing rules. The current Power picks have a separate live record.
Every AI tip · profit to 1pt level stakes
Every tip we've ever published, every tier, win or lose — nothing held back.
420/1293
Winners, all priced tipsrate depends on price
+28.01
Points, all priced tips1pt level stakes
Return at 1pt level stakes
+4.8%
The dot is the measured return. The line is where the true figure plausibly sits.
95% range across 1047 settled picks, resampled from the actual results rather than assumed to be bell-shaped. It includes zero, so this record is not yet distinguishable from breakeven.
Over 1047 picks at these prices, a swing of about ±103 points either way is ordinary luck, not a signal. Judge the record against that band, not against zero.
1293 settled picks · every published tip with a scoreable price, funded or not
Another 90 published picks went out at SP or without a price — 17 of them won. With no price there is no return to measure, so they appear in the pick history below and in none of the figures above. Counting them in the tiles while striking the return without them put two populations under one label, which is how this page briefly read 1,046 in one tile and 959 in the next.
The price basis: struck at the best of around 29 bookmakers and settled at — the better of that price and the SP. Both flatter the figure. Without BOG the same bets return about 5.8 points a bet less, measured across 842 of them, which is more than our selections have ever been measured to add. How to make sure you have it.
32%
Win-only strike rate · 382/1178 picks
tips: 102 of 205 returned something — a win or a place. Not a win rate, and not counted as one anywhere in the record above.
Closing Line Value · since 1 Jul 2026
Do our prices beat the Betfair closing price? The average can be dragged up by a few big drifters, so the honest number is how OFTEN we beat the close. Under 50% counts against us, whatever the average says.
Measured on the bets we placed, not on every published tip — they are the only ones carrying a closing price.
-0.8%
Avg. CLV
vs Betfair close
43.8%
Beat the close
57/130 bets
CLV across 130 settled bets, shown win or lose, measured against the Betfair close and reported whichever way it goes.
Recent Form
Last 30 days
-20.29pts
97/297 · 33% · -7% ROI
Record
1383 tips settled
Cumulative P&L
+28.01pts
Worst run — the deepest fall from a high point on this curve was −100.28pts, from 15 Jul 2026 to 26 Aug 2026, and the longest run of consecutive losers was 14. Measured on all published tips at 1pt level stakes.
Where we are · 13 August 2026
The line above peaked on 15 Jul 2026 and has fallen since. Over the last 30 days every published tip, at 1pt level stakes, comes to -20.29pts across 297 settled picks, a return of -7%. Across the whole record, the same population is +28.01pts from 1293 priced settled picks — so the drawdown has taken back a large part of a gain, not opened a hole underneath it.
The narrower board record — the picks that cleared the staking gates, which is the Headline picks tab above and not the whole bet history — is +29.32pts from 63 settled, +46.5% all time. That gap between the two populations is the board filter doing the job it exists for, and it is the most useful thing in these numbers. It is also an all-time figure against a 30-day one — the board has flattened over the same recent stretch. Month by month, it is on the Board picks tab.
What we make of it: a losing stretch of this size was always inside the range. The resampled interval printed at the top of this page includes zero, which is another way of saying a run like this is what a record scoring like the market looks like from the inside. We do not think it shows the model got worse.
We have not changed the selection method in response to it, and we want to be exact about that, because we did change the prompt on 12 August. That was a bug: a rule told the model to compare its own estimate against the chance the offered price implies, a number inflated by the bookmaker’s margin, when the same prompt says elsewhere that the only meaningful comparison is against the de-vigged fair price. It was firing on races it should never have touched. We fixed it because it was wrong, not because the curve was falling, and we wrote down beforehand what we expect it to do — fewer passes, a slightly larger board, and no improvement in return. If that reasoning turns out to be wrong, this note will say so rather than disappear.
Monthly Breakdown
Every published tip, 1pt level stakes.
| Month | Tips | W-P-L | Win% | P&L (pts) | |
|---|---|---|---|---|---|
| Sep 26 | 119 | 39-0-80 | 33% | -7.85 | -7% |
| Aug 26 | 504 | 163-0-341 | 32% | -34.96 | -7% |
| Jul 26 | 366 | 108-7-251 | 30% | -20.07 | -5% |
| Jun 26 | 394 | 127-40-227 | 32% | +90.89 | +22% |
ROI by month
Confidence Calibration
When the AI estimate on a pick card says X%, how often does that pick win?
Measured on picks published with an AI estimate between 2 Jul 2026 and 16 Aug 2026 — 607 of the 1383 settled picks on this page, because the estimate was only added to the cards partway through the record. The bands below hold 592 of those; the rest fall in bands too thin to show.
said 65.4% · 44 picks, 24.5 effective
said 47.4% · 124 picks, 67 effective
said 34.3% · 154 picks, 87 effective
said 24.4% · 180 picks, 99.7 effective
said 14% · 90 picks, 49.3 effective
Vertical line = what we actually predicted in that band. Bar = win rate since, weighted toward recent results, so the effective sample is smaller than the raw count. Bands with fewer than 25 settled picks are left out rather than shown thin — a win rate off a handful of races is not evidence of anything, and the highest bands are always the smallest.
Board record by price band
Never pooled into one figure. A band of short favourites wins more often because it is short — only the gap against what the market expected compares across rows.
Samples per band are small — read the gap, not the strike rate, and read neither as settled.
Several figures appear across this site and they measure different things. Here they are together, each one put through the same calculation at 1pt level stakes.
Roughly, each row accepts more than the one above it: one pick a day is our single strongest call, a handful more make the board, more again get funded one way or another, and every race gets a published read. The narrower the row, the shorter its record — and the less its return has been tested.
Headline picks
One a day: the board horse the model rated highest. Scored 1pt win-only at the taken price, so a place counts as a loss here — the Board row settles each-way bets as they were actually struck. Much the shortest record on this page.
Board picks
The picks we actually put money on. Fewer than qualified — a daily cap until 3 July, and since then prices that moved or fell outside the band we bet.
Bets we funded
Every bet actually struck, including ones outside the board — midday additions and days the selection rules would not have backed. This is the population the execution comparison lower down is drawn from.
Every published tip
Every pick we published, including the ones that never qualified. Most of these were never bettable.
Attributed to a course
The total across the course list — board-priced picks only, counted under the course page each appears on. Picks at meetings with no course page (international cards) appear on none, so this can sit slightly below the published-pick count.
Closing line value sits outside this table because it is not a return: -0.8% mean against the closing exchange price across 130 priced bets, beating the close on 43.8% of them. Measured against Betfair SP rather than bookmaker SP, which carries the overround and flatters this figure. How we measure.
Recomputed on every visit · every AI tip · profit to 1pt level stakes
Put this record on your site
A live scoreboard — wins and losses, updated automatically. Free to embed anywhere.
A record isn't an edge. See how we measure ourselves — every test we've run on our own tips, including the ones that failed. Or read the race-by-race archive — what we said before each race, next to what won.
Same 212 bets, same horses, same days — settled two ways. The only thing that changes is where the bet was placed.
Measured on 12 August 2026, over bets settled between 16 June and 8 August — the 212 of our funded bets that carry a Betfair starting price to compare against. That is the Bets we funded row above, less the bets with no exchange price to settle at. Figures are frozen at that date rather than recomputed on each visit, so a published finding cannot drift quietly.
Just under 20 points of stake, on bets we had already decided to make. At £10 a point that is about £200 we left at the till; the horses were identical. It is the largest number we have ever measured about our own betting, and it has nothing to do with picking.
Not a promise, and not an edge: it is the gap between two ways of settling the same bet, and it is population-specific — yours will differ. The full working, including where Best Odds Guaranteed beats the exchange and where it doesn’t, is on how to bet better.
We’re thinking about building this for your bets. Log them in the bet tracker and we’d show the same two lines for your own record — what you took, what you’d have had at BOG or Betfair SP, and the running total. If that sounds useful, tell us — we’d rather build it because someone asked than because we assumed.
Up to five ranked Power picks with AI commentary from 8am UK time once published, plus results with pending BSPs clearly labelled. Unsubscribe any time.
One step per settled pick, not equally spaced by date. The dashed grey line across the chart is break-even; the red dot is the record’s worst moment, -3.09pt on 16 Jun; the strip underneath is each pick’s own result.
Odds-on picks and anything above 6/1 are excluded from 26 July 2026, when the board’s price floor and cap began applying — they are still published, on their own strand, but were never board bets. Picks before that date are counted exactly as they went out.
+4.8% ROI (95% range -5.3% to +14.4% — not yet distinguishable from breakeven) across 1047 settled picks at 1pt level stakes — every published pick carrying a scoreable price, funded or not. It answers whether the selections are any good, which is a different question from how our own bets have done: from 17 Aug we place exactly one of these a day, and that strand has its own tab on the record. Full record, both populations and the ranges.
The price basis: struck at the best of around 29 bookmakers and settled at — the better of that price and the SP. Both flatter the figure. Without BOG the same bets return about 5.8 points a bet less, measured across 842 of them, which is more than our selections have ever been measured to add. How to make sure you have it.
