Investing

Research running · daily tracking in build

Runs on Brain →

A research desk of separated AI roles that argues every stock pick, checks it against hard risk rules, and tracks it to the end. Only I place trades.

  • 31specialist AI roles
  • 8free data sources
  • 255offline tests
  • ~8,600lines of Python

HOW IT WORKS

Architecture

Two streams find candidates, isolated analysts argue them, code enforces the limits, and every pick is tracked until its time runs out. The last step is always mine.

Interactive diagram. Select a card to highlight what it feeds and what feeds it; press Escape to clear.

Who does the work:

UNDER THE HOOD

What it actually does

  • 31 roles in 6 tiers

    Analysis, intelligence, memory, engineering, orchestration and interface. Every role has one job.

  • Each role is blind on purpose

    Every role has a written list of what it may not see, and a test fails if the code and that list drift apart.

  • Bull and bear never meet

    They run as separate processes. Their first run produced two genuinely different arguments, not one argument with the sign flipped.

  • Sandboxed at the file level

    Each role runs in a folder holding only its own input, so isolation is enforced, not requested.

  • Fake work is rejected

    Every output is hashed against all earlier runs, and impossibly fast model calls are refused.

  • Every fetch leaves a receipt

    Each outside call is time-stamped and hashed. A missing source, a missing key, a forbidden request and a provider error are four named states, never a silent blank.

  • Routed by the cost of being wrong

    5 roles use a frontier model, 16 run on local models on my Mac, and 10 are plain code.

  • Hard limits, not warnings

    A position whose worst-case loss cannot be computed is rejected. There is a sector cap, a never-list and a full stop at a set drawdown.

  • Two gates before any automation

    An automated trader would need 50 forward paper positions across two market regimes, beating the benchmark on return and drawdown. Even then it authorises nothing by itself.

  • Picks are checked at open

    Each pick needs reachable kill criteria and a key macro input re-checked against a live source, or it is void.

  • Tracked whether or not I trade

    A simulated sleeve and a real sleeve run side by side. Placing the trade is one field and changes nothing about scoring.

  • Built to doubt itself

    It starts from the evidence that AI stock picking has not beaten buy-and-hold, and is built to measure an edge rather than assume one.

HOW I DESIGNED IT TO WORK

Built to measure, not to assume

Three lines from the rules I wrote for the desk.

“Different agents see different things.”

On isolating the roles

“Built to measure whether I have an edge, not to assume one.”

On why it exists

“Placement is not a condition of tracking.”

On scoring every pick

WHAT'S NEXT

Next, in order

  1. Build the daily job that marks open picks to market, checks stop conditions and scores each pick at its time limit.

  2. Widen the screener to run several candidates at once, inside the sector limits.

  3. Restore the broker price feed with automatic failover, then re-run the live data tests.