The Bet Train started, like most things in this story, as a message.
At 17:19 on 13 July, Andrew wrote: "Now, one last feature. We're going to do a bet train tab. So I will pick selections and we will track them over a number of bets, build each leg up over time to a certain goal. We will have targets of £100, £500, £1,000, £10,000, £50,000, £100,000. We will show if we have banked cash and track each journey."
The idea is a staking chain. Start small. If the bet wins, the whole return rides on the next one. Keep going until you reach a target - or until one loss ends the journey. It's simple, it's brutal, and it makes every single result matter.
Claude built it that evening. The first bug turned up the same evening: a void leg - a postponed match - was busting the entire journey. By the end of the night there was a one-click button to put the top consensus pick on the train, and the Bet Train was part of the paid tier.
About half an hour
On 21 July the train went automatic. At 13:04 Andrew asked for two trains every month: a small one aiming for £100 to £1,000, and a big one aiming for £10,000 to £1,000,000, never picking the same match twice in a month, each posting to its own channel.
At 13:36 the big one was gone. "Let's remove the big bet train and just keep the small one... we will just try to reach each milestone at a time." One train, climbing a ladder of targets up to £10,000, banking at each one and carrying a quarter of it on to the next.
At 13:44: "Let's wipe all the bet trains we used for testing and let's begin for real this week."
Changing your mind in half an hour is cheap when the thing you're changing took twenty minutes to build. That's one of the quiet advantages of building with an AI partner: an idea can be tried, looked at and thrown away before lunch, and nobody's afternoon is wasted.
No invented prices
The same day brought a rule that runs through everything since. Andrew noticed that the odds shown beside the Model vs Model picks weren't real bookmaker odds. Where no price existed, the system had been filling the gap with one worked out from the model's own probability. "Can we remove our odds and only have live, up-to-date odds on this and each prediction?"
It was removed everywhere. The reason was written into the code: a made-up number dressed up as odds is actively misleading, because it doesn't match what a real bet would pay. If there's no price, the record says so.
Scoreboards everywhere
July was the month the system learned to watch itself. Accuracy by confidence band - how often a HIGH-confidence pick is right compared with a MEDIUM or LOW one. Accuracy by agreement - "If all 7 agree, I want to know how many times this comes up." Which exact combination of models is right most often. Andrew's reaction: "Excellent - this will allow us to build a picture over time and combine models together to maximise results."
On 23 July he built one of those combinations himself: Avenger, a blend of the models. The public Discord channel stopped carrying picks and started carrying a morning digest instead - "Good morning! Here are the results for..." - with the day's accuracy broken down by confidence.
And on 21 July, almost in passing, a milestone. Andrew had started posting the picks to X by hand, and had a Discord server running. His question was the one every new account asks: how many followers before this earns anything? The answer involved affiliate deals, a limited company that didn't exist yet, and a decision: "Let's wait and see where we are in six months."
The lab book
- Tested: two automatic trains with very different targets. The big one lasted about half an hour.
- Didn't work: odds invented from the models' own probabilities. Removed on 21 July - real prices or nothing.
- Tested: predicting the Champions League and the European cups, added on 21 July on the theory that the power rankings could compare teams from different leagues. (Reversed three weeks later.)
- Tested: seven more leagues - Portugal, Greece, Poland, Denmark, Turkey, Austria and Switzerland. (One of the quietest bugs of the build was hiding in this. It wasn't found until August.)
- Still testing: which combination of models is best. The tracking started here.
Next stop: one very good week.