Our Prime predictions use an AI layer to read each runner's form and rate it — then the app turns those ratings into a win chance by blending them with the live market and an independent form rating. Crucially, the system keeps testing and re-tuning itself on real results, so it gets sharper over time. Here's the whole loop, in plain English.

These are the Vector Prime predictions
On the boards and race cards this tip is tagged Prime — it's what your Vector Prime plan includes. It runs alongside Vector Core (tagged Core), the rules-based model that prices every field from the market, our own rating and form, and Vector Boost (tagged Boost), an experimental machine-learning model — see How the Core predictions work and How the Boost predictions work.

1. The AI scores each runner — it doesn't pick the winner

For every runner, the AI gives a 1–10 score on a set of racing factors. It's judging each factor in isolation — not choosing the winner. That's deliberate: keeping the AI to "rate the evidence" and letting the app do the maths makes the whole thing consistent and checkable.

The factors span the form itself (recent runs, class, distance and the going), the people (trainer and jockey), and the shape of the race (where a runner is likely to sit and what its sectionals say). The exact set is not fixed — it changes as the model learns, which is what the rest of this page is about.

2. The app turns those scores into a win chance

a
Weight the factors. Each factor score is multiplied by a learned weight — how much that factor has actually mattered for winners in the past — and added up into a single model score.
b
Blend with the market & a second opinion. The model score is mixed with the live market price (what punters are backing) and an independent AI form rating. That three-way blend is what produces each runner's win %, place % and fair odds on the card.

Because it blends in the market, a runner nobody is backing won't be over-rated on the AI's opinion alone — and where the AI and the market disagree, that's exactly where "value" flags appear.

3. It grades itself on every result — and re-tunes

This is the part most tipping services don't do. After every race resolves, the system re-checks whether its factor weights are still the best ones:

1
Back-test on races it didn't train on. It re-runs its picks over past races using candidate weightings — but always judged on a held-out set of newer races it hasn't learned from (a walk-forward test), so it can't just memorise the past.
2
Only change if it's genuinely better. A new weighting is adopted only if it would have found more winners and places than the current one on that held-out set. If nothing beats it, nothing changes.
3
Watch live, and undo if it slips. After a change, a guardrail watches the next stretch of live races. If the real win rate drops materially, it automatically reverts to the previous settings. So a tweak can help, but it's never allowed to quietly hurt.

4. It goes looking for new factors

The model isn't stuck with a fixed list of factors. In the background it scans the data it already collects and measures how strongly each signal separates the runners that placed from those that didn't. Anything that reliably points at placegetters becomes a candidate worth adding.

5. New factors are tested so they can only ever help

A new factor can never make your predictions worse
When a promising factor is adopted it starts at zero influence — so predictions don't change at all. It then has to earn its weight the same way everything else does: by beating the current model on races it hasn't seen. And the same live guardrail is watching, ready to switch it back off if it ever hurts. Upside only, downside blocked.

Watch the brains working

All of this runs quietly on every race result — grading, re-tuning, hunting for new edges, and stress-testing them before they ever touch your tips. The upshot: the model you're looking at today has been shaped by every race that came before it, and it'll be a little sharper tomorrow.

For the other two flows scored alongside it — the rules-based model that prices every field, and the experimental machine-learning model — see How the Core predictions work and How the Boost predictions work.