Python, CLI and MCP

The engine is pure Python standard library: no keys, no network, no execution. Fetching data is the caller’s job.

Python

from evidence_deflator import effective_window, selection_penalty, months_for_power

w = effective_window(cutoff="2024-10", start="2020-01", end="2025-06", trials=41)
w.open_book_months   # 58  months on or before the cutoff
w.effective_months   # 8   clean months
w.months_required    # 128 clean months a Sharpe-1 result needs, best of 41
w.verdict            # 'underpowered'
w.to_dict()          # everything, with the reasoning as plain text

selection_penalty(41).t_adjusted   # 3.26
months_for_power(2.0)               # 12

Pass cutoff_source= with the provider’s documentation URL; without one the result says the cutoff is unverified. Pass effective_trials= when you have measured how correlated the attempts are.

Command line

git clone https://github.com/Beltran12138/evidence-deflator
cd evidence-deflator
python tools/window.py --cutoff 2024-10 --start 2020-01 --end 2025-06
python tools/record.py --start 2026-03 --end 2026-08 --sharpe 3 --trials 4800

MCP server

For Claude, Cursor and any other MCP client. Tools include knowledge_window, track_record, counterfactual_audit and trace_audit.

pip install "evidence-deflator[mcp]"

{
  "mcpServers": {
    "evidence-deflator": { "command": "evidence-deflator-mcp" }
  }
}

Data

Each leaderboard edition is published as JSON under CC BY 4.0: /leaderboard/2026-10.json. Fields per entry: rank, name, return_pct, days, sharpe, t.

Hosted evidence score

A hosted endpoint that returns an evidence score for a strategy, agent or leaderboard entry is planned. If you run a framework, a competition or an allocation process and would use one, tell us the use case.