simulation/README.md, ECONOMICS.md sections 3 and 8, and simulation/results/.
Evidence chain for the issue 102 study: a test pre-registered before any data was fetched, a backtest on real Kuru MON-USDC data, a negative result (the top fifth of genomes beat the bottom fifth by 0.028 sd per tick against 0.278 modeled), and a no-go recorded in a core document that keeps reallocation off in code. The study covers one strategy plugin only. Here a filled dot means a stage that was run and reported. Source: ECONOMICS.md section 8 and EVOLUTION.md section 7.
The harness
simulation/ is a standalone package. It does not talk to a wallet, a chain, or the API. It calls the same fitness, reallocation, and risk rules the production code uses, rather than carrying second copies, and takes its constants from src/protocol.ts, which imports the constants the worker uses. A run with a given seed is deterministic, and two random streams share the seed, one for the market and one for agents, so a market path does not change when population size changes.
Markets are synthetic and not historical Monad prints. Four regimes are tested: a trending market, a mean-reverting market, a choppy sideways market, and a single stress shock halfway through the run. Each is run at populations of 10, 100, and 1,000, which gives a 12-cell comparison.
Assumptions, labeled
Each agent in the early model has a hidden skill drawn once. That device has no production counterpart, because skill is not carried by the genome, so no selection mechanism could act on it. From issue 84 on, an experiment-only genome model gives genome fields a regime-dependent edge ofmagnitude * (score - 1/3), where the score measures how close timeHorizon, entryThreshold, and the traded pair are to an assumed per-regime optimum. Magnitudes of 0.006 and 0.009 passed a static validation. 0.003 did not survive costs even in a static population. Every part of this model, the optima, the weights, the form, and the magnitude, is an assumption chosen to give selection something to act on, not a measured property of any strategy.
The decision rule is fixed first
The rule for a go or no-go was fixed before the runs were read as a verdict. An evolutionary arm counts as outperforming only if, in a majority of the 12 cells, its mean ROI is higher, the two-sided Welch p-value is below 0.05, and the median gap across cells is positive. The stress regime is reported on its own and not averaged away. If fewer than half the cells reach p below 0.05 in either direction, the call is inconclusive, not a finding that the mechanism hurts. Runs use 50 seeds per cell and report distributions, not point estimates.Pre-registration
Two studies were pre-registered: the out-of-sample confirmation of owner-scoped reallocation (issue 92, seeds 51 to 100), and the genome-edge study on real market data (issue 102). In each, the hypotheses, arms, settings, thresholds, and decision rule were committed before any run or data fetch, and any deviation is stated before the results. The genome-edge study’s design file issimulation/results/genome-edge-preregistration.md, and its deviations are listed in genome-edge.md ahead of its results.
Safety checks in the harness
Tests check that no reallocation or replacement creates or destroys capital and that no capital crosses an owner boundary. The adversarial run (simulation/results/adversarial.md) forces eligibility at every evaluation and a winning-tick book, and no capital invariant broke. That is a safety result, not evidence that evolution improves returns.
Limits of this method
Synthetic markets, an assumed genome model, a 300-tick horizon with no regime change inside a run, and a cost model whose Kuru fee and slippage line was an assumption. Section 6 ofECONOMICS.md lists what a production-grade simulation must do that this one does not. A simulated result is a statement about a mechanism in a model.