Every evening, before the next day's racing, two separate things happen. The engine reads each race and produces a file. Then those files are turned into the words you read on the card.
They are different jobs, done by different things, and it is worth knowing which is which.
What the map looks at
The read is built from the runners' own past races. For every horse in a field, the engine takes every qualifying previous run and asks where that horse actually was while the race was being run — not where it finished, but where it sat.
Alongside that, it considers:
- how each of those past races was itself run — whether the pace was strong, contested or slow
- the draw, and what it is worth at this course over this trip
- the trainer and the jockey
- today's going, trip and class, and whether a horse's form was built in conditions that resemble them
- the market, as a rival opinion to be aware of rather than an input to follow
What the engine does with those elements — how they are weighted, where the thresholds sit, which combinations it treats as meaningful — is the method. That part is not published, for the same reason no analyst publishes their working: it is the thing being offered.
What comes out is a file
The output of all of that is not prose. It is a structured file for each race: the running style assigned to every runner, the shape the race is likely to take, the horses the engine expects to matter under that shape, and the scenarios it ranks as most likely, possible and outside.
By the time that file exists, the read is finished. Every judgement the map makes has already been made, and it has been made from the data, in code, before a single word has been written.
Who writes the words
The written scenarios on each card — the titles, and the couple of sentences under each one — are produced by a language model working from that file, under a fixed set of instructions that does not change from race to race. Automated checks then reject any write-up that contradicts the file it came from.
The model does not make the predictions. It describes them.
It is not given the form, it is not asked for an opinion, and it has no view about which horse should win. It is handed a read that is already complete and asked to put it into English that a person can follow. Where it is told to be careful, it is told so for a reason: a title may say how the race is likely to be run, and it may not say who wins, because those are claims of very different strength and the card should not blur them.
Why say so at all
Because a reader deciding whether to trust an analysis is entitled to know which part of it is analysis.
The judgement in a PaceMap card is the engine's, and it is the same judgement whether the words around it are good or bad. The words are a translation layer. Saying so costs nothing, and it means the writing can be judged as writing and the read can be judged as a read.
How often the read is right is a separate question, and it is answered with numbers rather than prose. Every map published is scored against what actually happened, and the record — including the months that were worse — is on the accuracy page.