ACCURACY

Inputs. Quali pace, FP sector deltas, tyre compound history, current standings, circuit-specific wet/dry signals, current weather snapshot — all pulled live from OpenF1. Live mode adds real-time positions, gaps, recent lap pace, stint info, race-control events, and pit stops.

Model. An F1-domain system prompt drives a LLaMA 3.3 70B inference (Groq) constrained to JSON output: winner, podium, wildcards. We never feed it free-form 'guess the result' — every pick must cite numbers from the supplied data.

Validation. Every response is checked: driver acronyms must exist in the actual race grid, at least one podium pick must be real. Invalid responses are rejected (no caching), and stale fallbacks serve the previous valid output if Groq has an outage.

Logging. Every validated pick is upserted into race_prediction_history immediately. After the race finishes, ground truth is auto-inserted by a cron job pulling actual positions from OpenF1. Hit-rate is computed by joining the two — no manual editing, no after-the-fact 'corrections'.

What this is not. Not betting advice. Not a guarantee. Race outcomes are dominated by chaos (weather, retirements, strategy calls). The honest number — live hit-rate beats coin flip — is the claim. Pre-race is closer to coin flip. Both numbers are published above; we don't hide misses.