Is the data clean, and are the models valid?

We ripped through the full survivorship-free Sharadar corpus (21,862 US stocks incl. 15,573 delisted, 1998–2026) on two fronts: a data-cleanliness audit, and a model-validity gauntlet that tests every signal against a size + momentum characteristic-matched peer basket — so "beats the market" can't be mistaken for alpha. Then independent frontier LLMs reviewed the whole framework.

1 · Data cleanliness

Streaming every table, hunting the price artifacts that corrupted earlier backtests.

2 · Model validity — the gauntlet

Each signal vs its size+momentum-matched peers (winsorized forward returns, bootstrap significance, 4 subperiods). VALID ALPHA only if it beats matched peers significantly and holds across eras.

3 · Independent LLM review of the framework

Frontier models adversarially judge whether the methodology is sound and whether to trust the verdicts.