A small green week, and cutting off the long shots

This week's picks went 8W 0P 13L from 21 races, 38% winners, for +£3.17 at flat stakes — a +13% ROI week. That's a good week on its own, but I want to be honest about scale: it's £3.17 on 1-point stakes, and the all-time picture is still 156 picks settled at 31% win rate and -8% ROI. One decent week doesn't undo four months of data. It just tells me the model can string together a positive stretch, which it needs to keep doing a lot more of before -8% turns into something respectable.

The biggest win was Acting Lady at Newmarket on 10 July, 2/1, +£3.33 — more than covering the week's biggest loss, Claymore at Newmarket on 4 July, 14/1, -£2.00. Two Newmarket runners, one week apart, opposite outcomes. Nothing deeper to read into that, it's just what happened.

Auto-bet stops touching anything over 4/1, and drops each-way entirely

The real change this week wasn't in the model, it was in what real money actually backs. I went back through every auto-bet placed so far and split it by price band: at 4/1-and-under, it's +£98 across 76 bets. At 9/2-and-over, it's -£155 across 42 bets. Same inversion shows up across every published tip, not just the staked ones, so it's not a small-sample fluke I'm chasing. Each-way had its own separate leak — -£167 across 41 bets, against win-only's +£125 on 97.

So from now on, auto-bet has a 4/1 ceiling and never goes each-way, full stop. The model itself is untouched — it still publishes whatever price and bet type it thinks is right, and you can see it on the card same as always. This only changes what gets staked with real money behind it. It's a staking rule, not a model change, and I want to be upfront that I don't fully know why the longer prices have underperformed this badly. Could be variance still working itself out over more bets, could be something structural in how the model prices outsiders. I'm not going to pretend I've solved it, I'm just going to stop bleeding money on it while I figure it out.

Also fixed something uglier this week: a data collision where the automated card fetch had silently overwritten Royal Ascot's and York Ebor's real fixture data with an unrelated same-day ordinary meeting at the same course. Both festivals' actual data is restored, and there's now a check that stops a festival's slug being claimed by an unrelated fixture just because it shares a course name. I don't love finding that kind of thing after the fact, but I'd rather find it than not.

Next week I'm watching whether cutting off the 9/2+ bets actually improves the real staked P&L or just narrows the sample enough that a couple of bad weeks looks like a good decision. Six months from now I want to be able to say which one it was.