A research factory that cannot lie to itself
Every idea must survive a process built to prove it wrong — pre-registered, statistically judged, and honestly scored.
The historical ledger contains 53 tests. The current six sleeves begin again at zero approved edges: a target return is not evidence, and shadow measurement must earn every promotion.
Follow a sleeve from raw data to approval
Five layers stand between a hypothesis and broker execution. Every one is designed to expose weak evidence early.
Collect point-in-time market data
Completed intraday and daily bars are collected incrementally from Alpaca. The live partial bar is excluded, pagination is explicit, and each sleeve keeps an isolated replay database.
Freeze the strategy before judging it
Every sleeve has an exact thesis, disjoint stock pool, deterministic entry and exit rules, risk limits, semantic version, and parameter hash before shadow measurement begins.
Shadow every decision without touching the broker
The six runtimes generate and persist signals, rejection reasons, risk decisions, and hypothetical outcomes while a broker write-lock prevents submit, cancel, or close calls.
Grade against independent truth
Account equity, positions, orders, fills, and closed trades are mirrored from the paper broker. Portfolio performance comes from that reconciliation spine, not from a strategy grading its own output.
Promote only robust, cost-adjusted evidence
A sleeve must clear sample depth, positive expectancy after costs, walk-forward consistency, drawdown, and operational safety gates before a human may enable paper execution.
The first conversation is about process, not promises.
Every sleeve is reviewed on reconciled evidence; no model can promote itself. Ask to see the scoreboard — including everything that failed.
Page data as of 2026-07-20 — The operating board reflects the frozen VWAP v1.0.0 registry and broker-read-only shadow rollout.