Conjectory is a tool that uses AI to do mathematical research, and a record of what the research finds. This site is the public part of that record.

Records

The most recent results — from the workbench and from the community. Each row links to the full paper.

All records →

What it is

Conjectory is two things at once: a research tool and a research record.

The tool runs on the researcher's own computer. You give it a goal in plain language. An AI agent then works toward the goal in small steps: it forms conjectures, runs experiments, checks claims, and writes up what it finds. The researcher stays in charge the whole time.

The record is a plain git repository. Every step, whether a person or the agent took it, is saved as a commit, so anyone can see who did what, when, and what it cost. A claim climbs a ladder as evidence builds: idea, conjectured, preregistered, measured, replicated, derived, machine-verified. A claim can also end in one of three ways: refuted, superseded, or abandoned. Failed tries stay in the record.

A closer look at the workbench →

How it works

1. Set the goal. A project starts with one file, GOAL.md. It says what to look for, what counts as success, and when to stop. Changing direction means editing this file, and every edit is saved.
2. Let it work. The agent works in small steps and commits each one. It cannot spend more money or time than the limits set before the start.
3. Watch and steer. The graph of the work and its journal update live. You can interrupt, ask questions, revise the goal, or go back to an earlier commit and try a different way.
4. Keep it honest. Every commit records whether a person or the agent made it. Decision rules are locked in before the data they judge. Refutations are published the same way as proofs.