No guaranteed winners. Picks are statistical output. Favourites lose most races; so do we. The only claim is that the record is real and checkable.
No inflated test. The model was fitted on 1,312 French races and tested walk-forward on 712 races out of sample — each block scored with weights fitted only on earlier races. On its own the model does not beat the pool, and combined with the pool it did not pass either (-0.3 points on top-3 coverage). We publish the ranks anyway, labelled as unproven, so the open record can decide.
No third-party notarisation (yet). The written-at times on our pages are printed by our own generator. What does not depend on our word: past pages sit unchanged in the archive, and results are public everywhere — so any claimed pick can be checked against any day you like. Judge the record over hundreds of races, not one day.