Team Spectre

Strata: an AI research agent that can’t forget its failures An AI agent that forgets its failures will report its luck as discovery. When an agent tests many trading ideas and reports the best one, the more ideas it tries, the more likely its best result is luck. An agent that forgets earlier sessions undercounts its attempts, sets the bar too low, and calls luck a discovery. Statistical tests also can’t catch an agent that peeks at future data. Strata fixes both at the database layer on MongoDB Atlas. Every attempt goes into a permanent record the agent can add to but never edit or delete. Future market data sits in a sealed Vault the agent’s database role cannot read. The Auditor, plain code with no model, is the only role that can open it. Built today: Six custom Atlas roles that enforce these rules; 21/21 permission checks pass against the cluster. A point-in-time data pipeline: 4,296,065 hourly bars from Binance’s public archive in time-series collections, split into research (2020–2024) and Vault (2025 onward), with every file checksum-verified and 682/682 load checks passing. The pool and file manifest are committed for reproduction. A Modelence app showing the attempt stream and the two bars every result must clear, on fixture data. Planned next: the Auditor’s deflated-Sharpe verdicts counted across every session, live research sessions, a hash-chained record, and a Strategist that learns which search policies produce results that survive.