EVOLUTION.md sections 3.5, 3.6, and 7; ECONOMICS.md section 8; and the files under simulation/results/.
What was tried, in order
The working hypothesis, not isolated by any single run, is that every extra account carries fixed per-tick costs that consume the edge a small account earns, so splitting capital across more accounts does not pay for itself. Capital-weighted skill, not the number of agents, drives return.
What replaced it
Fixed slots, owner-scoped bounded reallocation under a policy the owner signs, and retirement of starved agents. No genome is copied, so no genome-consent question arises. See Reallocation and retirement.The pre-registered study, and what it found
The replacement was confirmed in simulation, on a model in which genome fields carry an assumed edge. The remaining gate was whether the production strategy’s genomes carry an edge of that size. Issue 102 answered it with a study whose design, hypotheses, thresholds, and decision rule were written down before any market data was fetched (simulation/results/genome-edge-preregistration.md).
It backtested 1,000 genomes of the production strategy on hourly closes built from every on-chain Trade event of the Kuru mainnet MON-USDC market (in-sample 2025-12-23 to 2026-05-31, held out 2026-06-01 to 2026-09-30), net of measured costs. The result is a no-go:
- Edge size. The in-sample top fifth of genomes beat the bottom fifth out of sample by 0.028 market standard deviations per tick (95 percent bounds 0.015 to 0.039), against 0.278 for the modeled edge.
- Rank persistence. The mean rank correlation between adjusted fitness and the next 20 ticks’ net return was 0.066 (bounds 0.027 to 0.103), against 0.416 in the model. At zero cost it is -0.017, so the persistence that exists sorts genomes by how often they trade.
- Calibrated re-run. At the measured magnitude of 0, owner-scoped reallocation was inconclusive against a static population (2 of 12 cells significant, both lower, median gap -0.24 percentage points).
Why this is stated as a rigor feature
A mechanism was built, confirmed in simulation, and then tested against real market data before it was allowed near capital. The test could have come out the other way, and was written so that it could. The negative outcome is therefore evidence that the gate works.Limits of the evidence
- The study measured MON-USDC on Kuru. The owner then decided agents trade WETH/USDC on mainnet, so the result is extended to that market as an assumption, not a finding.
- It covers one strategy signal (a trailing mean), one venue, and costs taken from snapshots, not a spread history. No stress regime occurred in the held-out data, so only the choppy regime could be tested on its own.
- It says nothing about strategies with a different signal, or with information purchases that pay, which could not be replayed on historical data. A future strategy would need its own pre-registered study.
- All earlier simulation results rest on synthetic markets and an assumed genome model. They are statements about the mechanism in that model, not about production returns.