metaculus-bot-review

Review how a Metaculus forecasting bot performed on the questions it forecast.

bot-review builds a table of how your bot's forecasts turned out and attaches the reports it posted, so you can see what it was thinking on the questions it got wrong. Everything is read-only and requires no LLM spend. All it needs is METACULUS_TOKEN.

It reads the reports your bot published as Metaculus comments, so it works whether or not the bot saved anything locally. It parses only the format forecasting-tools itself writes, so it works on any bot built on the framework.

Install

pip install metaculus-bot-review

Or with poetry:

poetry add metaculus-bot-review

Set METACULUS_TOKEN in your environment or in a .env file in the directory you run from. That token is what decides which bot gets reviewed.

Scoring your questions

bot-review review --tournament <slug-or-id> --output review.json --summary review.md
bot-review review --resolved-since 30              # anything that resolved recently
bot-review review --post 44328 44326               # specific questions
bot-review review --from-json review.json --top 3  # re-render without refetching

The markdown summary gives your leaderboard standing, how many questions were forecast and scored, and the best and worst questions ranked on whichever score the leaderboard uses — spot peer for the AI benchmark tournaments, time-averaged peer for the Metaculus Cup.

The JSON adds per-question detail: the official scores, the bot's forecast, and one RunTrace per run taken from the comment that run posted — each forecaster's prediction, the run time, and, when the bot publishes its cost metadata, what the run cost and how long it took. QuestionOutcome.trace picks the run that was standing when the question was spot scored, which is the one that earned the score.

Reading the reasoning

Reasoning text is not stored in the table. Pull it a piece at a time:

bot-review show 44328 --section research
bot-review show 44328 --forecaster R1:F3 --comment 921582  # a specific run

Sections are summary, research and forecasts. A research section runs to tens of thousands of characters, so on a question whose resolution you already know it is usually cheaper to search it than to read it:

bot-review show 44328 --section research | grep -niE "death toll" | head

From Python

from metaculus_bot_review.outcomes import get_tournament_outcomes
from metaculus_bot_review.summary import build_summary
from metaculus_bot_review.traces import attach_traces, get_trace

table = get_tournament_outcomes("minibench-2026-06-29")
attach_traces(table)
print(build_summary(table))
print(get_trace(44328, section="research"))

Agent skill

skill/SKILL.md drives the whole loop with an agent which builds the table, decides on the questions worth investigating, reads the reports and writes the review.

Copy it into your bot's repo to use it with Claude Code:

mkdir -p .claude/skills/review-bot
curl -o .claude/skills/review-bot/SKILL.md \
    https://raw.githubusercontent.com/LouisP96/metaculus-bot-review/main/skill/SKILL.md

Then ask it to review the bot.

Rate limits

Metaculus rate-limits at roughly 5 requests per second. The CLI spaces requests by 0.7s; change it with --seconds-between-requests.

License

MIT