DAFE Trading Systems

AI writes trading indicators for millions of people, and the code compiles, passes the validator, and is quietly wrong. A 20-bar breakout whose window includes the current bar can never fire. A moving average with a zero length computes nothing. A misspelled builtin silently returns zero. The chart looks fine. Existing checkers only catch bug classes somebody already thought of. OakGuard is a verifiable agent layer for AI-written trading code. It runs each script against runtime oracles on four synthetic markets and proves when it is broken: a marker that fired zero times on 2,400 bars, a series with no finite value on 600 bars. It then asks Muse Wonder for a rule that would have caught that bug before it ran. The rule is data, never code: schema-constrained JSON in a three-family rule language that cannot touch the chart, the engine, or an order. A referee the model cannot see then decides. It grades the rule against broken variants of the same bug, a third of them held out from the model, plus 46 known-good scripts, including contrast scripts that sit right next to each bug class. A rejected rule goes back to the model with the reason and a counterexample. Only rules that pass every gate are promoted and enforced in the editor for every script the agent writes from then on. On the first live run, three rules caught 174 of 174 broken variants, 58 of 58 held out, with zero false alarms on the known-good corpus, for $0.10 to $0.12 of model usage. The referee rejected all nine deliberately flawed rules in its self-test. Pointed at our own engine, the same oracles found two interpreter bugs that had been silently killing shipped indicators. Built on DAFE Charts, my trading terminal, in ten days (Oct 26 to Nov 4, 2026). Every claim reproduces headlessly without an API key, and the judge build runs in a browser with no backend. It certifies the code, not the returns.