Frontier AIs bring methods from mathematics, physics and information theory. A public protocol tries to kill every one. What survives is committed before it trades, and every test is counted.
Our research is generated and run by AI agents, and we say so openly. Several frontier models from different vendors play separate roles, so no model ever grades its own homework.
Frontier models carry methods across fields, from physics to markets, and turn them into testable hypotheses.
cross-fieldA skeptic from a different vendor than the proposer hunts for leakage, look-ahead bias and wishful thinking.
vendor-separatedA tool-using agent fetches every cited source and checks the math symbolically. If a quote or a derivation doesn't match, the idea dies before it touches data.
sources + symbolic mathEvery idea is logged in the trial ledger, and the kills are logged too, so the multiple-testing math stays honest.
nothing hiddenConsensus is never evidence. Models share training data, so agreement between them is correlated opinion. Only data the models have never seen can support an idea.
The council makes ideas cheap. Rigor is what makes them worth anything. Every hypothesis runs the same gauntlet, in the open.
The hypothesis, data, metrics and kill criteria are frozen and hashed before any test runs.
Holdout data is run once with logged access and burned if touched early. History from before a model's training cutoff can kill an idea but never support it.
Every version of the pipeline must find nothing in pure noise and recover signals we planted, before it may judge a real idea. Our validation code is locked: the AI can read it, never change it.
We report the Probability of Backtest Overfitting (PBO via CSCV) and the Deflated Sharpe Ratio (DSR) against the full effective trial count.
Every variant ever tried is counted, failures included. A survivor only means something once you know how many died.
Before every order is sent, a salted SHA-256 of it is published and timestamped. We can't rewrite history, and no one can read our orders from the hash.
Aggregated daily results are published on a delay. Full order files and salts go to verifiers under NDA, and become public when a strategy retires.
No numbers until they're measured. Here is when each one arrives.
Protocol and validation tooling published
First live trade, committed before it was sent
First council ledger (proposed, killed, live)
Scoreboard opens with 6 months of verified record
// counters go live with the first published ledger. No numbers until they are measured.
Our goal: beat hedge funds where they can't fit.Status: day 1. The scoreboard is public.
Funds win on data, capital and execution. We are betting on something else: many small, uncorrelated strategies, a very low cost per experiment, and research anyone can reproduce. We will measure it, not claim it.
From December 2026, qualified researchers and allocators can request verification access: commitment logs, salts and read-only exchange records, under NDA. We don't manage outside money and don't sell signals.
Who's behind it. Testrian Lab, an AI-run trading research lab.
contact@testrian.com
One email a week covering what the council proposed, what the protocol killed and what we learned. Free, with no signals and no hype.
We'll email you a link to confirm. Unsubscribe anytime. Privacy
Or read the lab notes first.
The Pro pack is our validation harness as software, with no performance claims. The Backtest Overfitting Check is a free web tool: upload strategy returns, get PBO, Deflated Sharpe and MinTRL. Join the waitlist and we'll write when each one opens.
You'll also get the weekly lab note. Unsubscribe anytime. Privacy