Seconds instead of overnight, and gates a strategy clears before it sees real money.
What it is
Two halves, built together.
The lab replays a trading idea against history: 89 instruments, 250 sessions, 24
variants, in 11.1 seconds. Specialist agents own the separate jobs, reading sentiment,
producing a prediction, sizing risk, remembering what happened last time, placing the
order, and every decision lands in a pgvector telemetry store, so why did it do that
has an answer weeks later. Features in PostgreSQL, models in XGBoost and scikit-learn.
The gates are the half worth arguing about. A strategy does not advance toward real capital because it looked good in a backtest. It advances because it cleared a threshold written down before the run.
Three venues are wired and confirmed in the source: Alpaca for equities, Kalshi and Polymarket for event contracts. 59,393 lines, 668 tests.
The speed number, and why it is the important one
The full replay runs in 11.1 seconds against a 60-second design target. Certified July 16, 2026.
That sounds like a performance stat. It is a research stat.
At sixty seconds you run the sweep, go do something else, and come back, so you test the ideas you were already confident about. At eleven seconds the sweep is inside your thinking loop, so you test the idea you are unsure about, because being wrong costs eleven seconds. Hypotheses examined per hour is the actual output of a research system, and it is bounded by how long you wait.
Fast enough to stay in the loop is a categorical difference, not an incremental one.
The graduation gates
A strategy that looks good on historical data is not a strategy. It is a hypothesis nobody has attacked yet.
So promotion is staged, and each stage has an exit criterion decided in advance. Same idea as the requirements database in JARVIS and the approval interlock in VERONICA, in a domain where skipping it is measured in dollars:
Deciding the acceptance criterion after seeing the result is not analysis. It is narration.
Nothing has cleared them, so every position is paper. It does not predict prices and is not trying to. It measures whether a strategy survives contact with historical data, and enforces what a strategy has to prove before it is allowed to act at all. Nothing here is investment advice.
Why event contracts sit next to equities
The design choice I would defend hardest.
An event contract settles to a known truth at a known time. Did the thing happen, yes or no. That makes it a clean environment for asking whether a model is calibrated: when it says seventy percent, does the thing happen seventy percent of the time?
An equity return gives you nothing that clean. You have to risk-adjust it, pick a benchmark, choose a window, and argue about all three. Running both checks the messy domain against one where being wrong is unambiguous.
Where it stands
Certified, and not the thing I have been working on lately. FRIDAY and VERONICA took priority when agent authority turned out to be the more interesting problem.
The lab is fast, the gates are defined, and the honest read is that the infrastructure did exactly what it was built to do: it is telling me nothing has earned promotion yet. A research system whose first answer is not this one is a research system worth trusting with the next one.