Nine experiments, two LLM agent designs, two universes, a decade of daily bars, and 372 tests produced no configuration that beat buy-and-hold. That alone would be unremarkable — most such projects end this way, and most of them end without knowing why.

The contribution here is the why, and it is measurable rather than anecdotal. The limiting factor is not signal quality: an IC of 0.014–0.020 is an ordinary quantitative IC, and professionals build viable products from less. The limiting factor is breadth. Thirty-one signals drawn from six distinct-sounding families of the anomaly literature deliver 5.6 independent bets, and at a monthly rebalance that supports an information ratio of 0.11 — three hundred years to statistical significance.

Successful quantitative funds do not solve this with better mathematics on closing prices. They solve it with breadth: thousands of instruments across asset classes, data that is not public, and institutional financing applied to a small edge. A retail account can replicate their mathematics exactly. It cannot replicate their bet count, and the mathematics is worthless without it.

The apparatus outlived the hypothesis. It caught a 40× data error, a test suite silently running two-thirds of its tests, a safety check that checked nothing, a simulator that permitted negative equity, and — twice — a claim of ours that was stronger than the evidence. That is what it was built for, and it is the part worth keeping.