An adversarial stance toward our own ideas.
Every paper claims an edge somewhere. Most evaporate under honest cost accounting, a single regime change, or a different random seed. Our working assumption is that every candidate signal is noise until proven otherwise — and we spend more effort trying to break an idea than to defend it. We know this works because we ran the audit on ourselves: our own early promotions turned out to be selection on noise, and the current process was rebuilt on the other side of that finding.
Promotion now runs through a pre-registered three-gate funnel — Grid A (signal structure under combinatorial purged cross-validation), Grid B (execution-cost sweep on locked predictions), Grid C (a fifteen-seed causal walk-forward that could actually have been traded). Every threshold is written down before the first candidate runs; the deploy object is the seed median, never the best seed; and when the backtest can't separate two variants, paper trading adjudicates — not a third sweep. Only a human moves anything from paper to live capital. The full protocol is public: the Grid A/B/C paper.