Researchers built an AI system that generates live F1 race commentary and architecturally cannot publish a claim it cannot verify.
Pitwall produces real-time Formula 1 strategy briefings in English, Spanish, and Portuguese. Its core mechanism: every generated sentence is broken down into typed factual claims — positions, gaps, tyre state, pace, overtakes — and each claim is checked against a probabilistic race model before the sentence can go out. That model is a vectorized Monte Carlo engine running 2,000 simulated race continuations per lap, trained on 126 races from 2018 to 2024 and validated on fully held-out 2025 and 2026 seasons. On 155 backtests, it placed the eventual winner in its top-3 predictions 90.3% of the time, with a held-out Brier score of 0.0745. The same verifier shapes fine-tuning data: of 3,045 model-generated training targets, the 81.9% whose every claim is state-supported were kept; the remaining 18.1% were discarded and replaced with provably faithful templates, ensuring the model is never trained on ungrounded text. The system ran live at two consecutive Grands Prix — Austria and Britain (Silverstone) in 2026. At Silverstone, a timestamped probability trace committed to disk before the race ended had already locked onto the eventual winner ten laps before the flag.
The wider significance is less about Formula 1 and more about the hallucination problem that dogs every AI system operating on live data. Most grounding approaches treat faithfulness as a post-hoc filter; Pitwall bakes it into both generation and training. The paper also documents a concrete failure mode: fine-tuning on richer, more vivid targets improved fluency but caused hallucinations when the grounding state was sparse — a finding traced to base-model instruction adherence, not model scale.
Sports commentary is probably not where this architecture matters most — but any domain where a fabricated number causes real harm is a more compelling test.