Thermite for Science
Better results in mathematics and simulation, and literature research you can verify.
Where a program can verify a result, agents can search for a better one: a construction, a faster solver, a more accurate method. Goal campaigns reproduce a paper's results to a tolerance. Research campaigns read a field's literature at scale, with every quotation and every reported value checked against its source.
Mathematical discovery
Agents improve programs that build constructions. The judge runs each program and checks its construction exactly, in integers or rationals, before the bound counts.
For Local AI · Solo builders
- Measures
- The construction's size or score
- Holds
- Checked exactly, in integers or rationals
- Hidden
- A second, independent verifier
Simulation speed and accuracy
Agents change solver code, build settings and solver options. The judge times each version and checks it against reference runs, energy drift and convergence order.
For Solo builders · Startups · Enterprises
- Measures
- Time steps a second or time to solution
- Holds
- The same physics, within tolerances
- Hidden
- Input decks the agents never see
Literature review
Agents read a field's reviews, papers and benchmarks and record claims with exact quotes. Code checks every quote and number against its archived page, and a review panel of models decides each question.
For Solo builders · Enterprises
- Counts
- Verified claims across a field
- Checked
- Every quote and number found in its archived page
- Decided
- A panel of models, by majority
Data from the literature
Agents read open-access papers and record each reported value as a claim, quoted with its conditions. Code checks that every number is in the quotes, and a review panel of models decides each question.
For Solo builders · Startups · Enterprises
- Counts
- Values with their conditions and units
- Checked
- Each number found in its quote from an archived page
- Decided
- A panel of models, by majority
Reproducing published results
Agents make the released code install and run. The judge rebuilds it, reruns each table and figure and checks every number within your tolerance, and the run ends when every goal passes twice.
For Solo builders · Enterprises
- Measures
- Tables and figures reproduced
- Holds
- Each number within your tolerance, with the paper's settings
- Hidden
- Seeds the agents never see
Guides by audience
Twenty minutes to a first verdict.
Install it on one computer, or ask your own AI agent to:
Set up Thermite on this computer for me: follow https://usethermite.com/llm-setup.txt