An open-source AI harness for math research

Research mode is experimental. It can use a lot of tokens: live runs on small problems have used hundreds of thousands, mostly input, and the reason is not yet understood. Set limits before you start.

Manifield’s research mode is an AI harness for mathematical research: it gives models a project library, a fixed goal, tools, assignments and a record of evidence. You choose the models and limits, follow each attempt, and decide what belongs in your theory.

#The idea

When one conversation works on a hard problem, it tends to commit to its first approach and keep pushing. Research mode separates choosing a direction from doing the work. A director decides what to try; separate agents each pursue one approach; independent reviews decide what counts.

#Freezing the goal

Start research… → Hand to the director turns your problem into a theory item and freezes its statement and conventions. From then on, nothing can quietly change what is being proved: a correction to the target keeps the old version and needs its own evidence.

#The director and its assignments

One persistent director conversation searches your library, sends out work, reads the reports and decides what to do next. It uses mathematics for quick checks and understanding, and hands substantive proofs, lemmas and computations to assignments, one to four at a time:

• A direct worker: a focused conversation on one approach, without a mandatory plan.
• A planned branch: a task plan with dependencies, proof review and its own history.

The director can continue an attempt or start a fresh one. A report may honestly contain no result. An unresolved gap in a proof does not unlock the tasks that depend on it.

While it plans, research mode can search arXiv for titles, authors and abstracts. That is a metadata search, meant to reduce invented citations; whether a result is new stays your call.

#Proofs and reviews

Inside research mode, the Prover drafts a proof from your items and attached sources, and must read what it cites. The Reviewer, prompted to be adversarial and able to re-read every citation and run Python spot checks, hunts for flaws; issues go back to the prover, up to a limit you set. The best possible outcome is Review found no errors, never "correct".

A claimed solution must pass independent checks of the contract and of the mathematics before it is accepted. A claim that the problem itself is wrong gets its own mathematical and scope checks.

#Limits

Settings → Limits → Research, and the Limits… control on a new run, set tool calls, tokens for the whole program and for each assignment, elapsed time, plan revisions and proof-repair rounds. Each can be bounded or unlimited, and they are saved with the program. A request already sent may overshoot a ceiling; no further request starts once it is reached. Pause a program to change its limits.

#Following a run

A program's page has four tabs: Overview, Assignments, Evidence and Log. You can pause, continue, or give the director guidance at any point; it keeps the same conversation.

#Why this should be open

Labs are building systems like this one, and most of them are closed. A research harness that mathematicians can read, run on their own problems, criticize and change seems worth having in the open.