Chapter one
A 5.8-million-parameter network, bred rather than trained, that produced spectacular returns — and the audit, commissioned by its own author, that took them apart.
Before the life simulations there was Emergence: a distributed evolutionary system that breeds trained neural networks together to trade equities. Not evolving the architecture, the way the literature does — evolving the learned weights, treating whole neurons as indivisible genetic traits. Four isolated bloodlines bred from a single frozen ancestor, then crossed round-robin, selecting on pool averages rather than outliers so that inbreeding depression could not quietly set in.
It ran. 1,796 offspring from 1,006 distinct parents, a hundred generations of accumulating lineage across 326 cycles. Scores climbed from the mid-200s to 767.
Validation precision 40.58% against a 17.21% base rate — a 2.4× lift, on 5.7 million rows of daily equity data. Modest, measured, and the strongest defensible claim in the project.
EVALUATION_2026-08-01.md §2.1 · verified against the loaded checkpointSupervised training happened once, offline, before every evaluation window. The scaler is fit on training dates only, with children inheriting the parent's. In the auditor's words: "Classic scaler leakage — the thing that kills most quant projects — is genuinely absent. Someone built that carefully."
dataset.py:176-184 · EVALUATION_2026-08-01.md §2.2Row-wise fitness-weighted crossover plus Gaussian point mutation, verified firing in production data at the expected rate. Splits cleanly adjusted across AAPL, TSLA, NVDA, AMZN and GOOGL — no fake collapses. The bones are sound.
brain.py:383-390 · confirmed against the live jello_memory DBIn August, fifteen months after the run, the lab audited itself: four independent adversarial reviews, each reading real code and quoting real line numbers. The verdict was one sentence.
No return figure this system has produced can currently be read as evidence of trading skill.
What fails is the grader and the world — not the evolution.
2,141 of 2,141 symbols end within five trading days of the last date in the file. Zero end earlier. Forty-nine of fifty tested delisted tickers are simply absent — SIVB, FRC, CS, TWTR, ATVI, VMW, CERN, XLNX. The two total losses of the period are invisible to the model. Every fitness number ever computed was measured on a universe pre-filtered for still listed.
data/arbie_universe.parquet · EVALUATION_2026-08-01.md §1.1One receipt, from the run's own results file:
RKLB · signal 2026-04-24 · entry 79.69 · exit 131.16 · +64.59% · +$1,029
The simulation's last day was 2026-05-07, when RKLB closed at 78.58 — a trade of about −1%. The 131.16 is the close on 18 May, eleven days outside the window, after a +34% gap the model could not have seen. That single imaginary trade is +$1,029 of a $20,774 final equity. Three of the five eras end mid-2025 and mark their books 11–13 months in the future, in a rising market.
simulate_portfolio.py:1108-1110, :620-622 · a two-line fixAll the search pressure in the system lived in a fitness function that could be satisfied by inaction. A hundred generations of genuine selection were aimed at a target that does not measure trading skill — and selection, being exactly as intelligent as its test, found that out long before we did.
The 767 peak is the clearest case. Looked at properly, the model had learned to size its buys to capture black swans — a perfect strategy in a world with no transaction costs that settles at future prices, and a ruinous one anywhere else. It was not trading. It was reading the grader.
EVALUATION_2026-08-01.md §0 and §3 · four independent auditsThe project retracted its own headline result rather than defending it — that retraction is in the lab report, dated, with the original left standing above it. Nothing here was hidden and nothing was fabricated. The numbers are real. They just do not mean what they appeared to mean, and the person who paid for finding that out is the person who commissioned the audit.
And this is where the rule came from. Every experiment since obeys a sentence that cost fifteen months to learn:
Evolution is exactly as intelligent as the test you give it.
So stop designing intelligent animals. Design tests that cannot be cheated.
The life simulations exist because of this. Worlds where fitness is the physics, where an animal that doesn't eat dies and there is no score to fool, are a direct answer to a grader that could be satisfied by doing nothing. The market lab closed for the same reason. The whole research programme is one investigation, and it starts here.
← Back to the lab