canli-mcp / The desk, built into canli-mcp

Red-team a strategy, counting every attempt

Run one strategy tool and judge the result against everything this session has tried: deflated Sharpe at the session's true count of strategies (every attempt and every variant in it), overfitting probability for a grid, against buying and holding, and track-record length. Returns what failed and what to change, and the attempt log. Locally, every attempt is kept in the Null Zoo (zoo_search), and a later session on the same data is judged with the tries before it. Rewriting until it passes is a search, so every attempt is added to the count it is judged against; the final check is holdout_test on years the loop never saw. Clearing the checks is not admission to anything and not a forecast.

The tool is red_team in canli-mcp 0.22.0, free and MIT-licensed, with no account and no API key. It is part of canli-mcp itself, so it is there as soon as canli-mcp is added.

Arguments

As the server declares them
ArgumentRequiredType
nameyesstring
argumentsnoobject
changenostring
barno{deflated_sharpe?, overfitting?}

How a model reaches it

Ask in plain words. The model finds red_team with canli-mcp's find_tool, reads its arguments with describe_tool if it needs to, and runs it with run_tool. The call has this shape, with your values in place of the angle brackets:

{
  "name": "red_team",
  "arguments": {
    "name": "<string>"
  }
}

Words that find it

Besides its name, title and description, find_tool matches a question to red_team on these words: red, team, auto, correct, loop, rewrite, strategy, until, valid, self, correcting, agent, backtest, audit, overfitting, deflated, sharpe, pbo, cscv, data, mining, bias, feedback, fix, next, attempt.

More

Nothing here is investment advice.