What the Swarm Knows and What It Does Not:
Declared Epistemic Limits of a Gated Alpha-Research Process

In one line: I tested 212 ideas, held back the two that worked, and published five things I still cannot know.

Abstract. A systematic alpha-research process is examined through Fejes Tóth’s (1964) result on the honeycomb, in which bees build a provably optimal partition of the visible plane yet fall a fraction of one percent short where the comb closes in the dark. The process reported here has evaluated 212 hypotheses under fixed gates (Newey–West statistics, margin-of-safety reruns, and multiplicity deflators), with every statistic produced by deterministic code rather than a language model; two crowding effects, at t of 4.1 and 3.9, were held significant but not promoted. Against these established results the paper sets the process’s structural limits, published as standing policy: survivorship in the universe, regimes absent from the sample, capacity assumed rather than proven, crowding unmeasured beyond long public filings, and near-fitting strategies at t = 2.3 that fail the promotion bar. The claim is that performance is inexpensive and certainty costly, and that the durable edge favors a process that accounts honestly for what it cannot know.

1. Introduction

In 1964 the Hungarian mathematician László Fejes Tóth published a paper titled What the Bees Know and What They Do Not Know. The bees are optimal where the structure is visible. The hexagonal comb is provably the cheapest way to divide a plane into equal cells, though establishing that against every curved and conspiring alternative took until 1999. At the base of each cell, where two layers of comb meet out of sight, Fejes Tóth showed the bees’ three-panel cap can be beaten by a fraction of one percent of wax. The architecture is optimal in the light and a whisker off in the dark, and no bee registers either fact. The research process behind this museum is named the swarm, so the question transfers directly: what does the process know, and what does it not? Most research programs publish only the first half. This note sets out the second.

2. What the process has established

The process reports what it has tested. To date, 212 hypotheses have passed through identical gates: Newey–West statistics, margin-of-safety reruns, and multiplicity deflators. Every verdict is recorded on the public wall, including the founder’s own retired namesake model. The process distinguishes significant from promoted. This month it identified two real crowding effects, at t of 4.1 and 3.9, and promoted neither, because the threshold does not move. Its arithmetic is not an opinion: the statistics are produced by deterministic code, and no language model in the shop computes so much as a Sharpe ratio. When referees spent four years on the sphere-packing proof and reached only 99 percent certainty, Thomas Hales spent a further decade having a machine verify every line. Performance is inexpensive; certainty is costly.

3. Declared limits

The following limits are stated as plainly as the process can see them.

1. The universe flatters the results. The panel comprises today’s index constituents, the companies that survived. Every Sharpe ratio on the wall is an upper bound for this reason, and is labeled as such.
2. Unmet regimes. The sample contains a few crises; the future is not obligated to draw from the same urn. Correlations measured in calm are fiction in a storm. The process stresses for this, but a stress test is only a rehearsal.
3. Capacity is assumed, not proven. Every backtest presumes that the process’s own trading does not move the price. At current size that is nearly true, which reflects the size more than the method.
4. Unmeasured crowding. The in-house 13F study measures crowding among long positions in public filings, excluding shorts and swaps. It captures the floor of the crowd and misses the ceiling.
5. Tempting near-fits. Newton held that twelve spheres can touch a central one and a thirteenth never quite fits, though the remaining slack looks inviting; the proof took 250 years. Each week the pipeline meets a strategy at t = 2.3 that appears to almost fit. The slack is real, and it still does not fit.

4. Discussion: the bee and the mathematician

The bee needs only to be right. The mathematician must be certain against every competitor, including the infinitely many nobody has drawn. A backtest that works is the bee: impressive and locally optimal, yet blind to its own caps. A claim that survives pre-registration, stress, and deflation is a step toward the mathematician’s standard. It never arrives there, because markets are not geometry. The wager of this program is that in a market full of confident bees, the durable edge belongs to whoever keeps honest accounts of the dark.