Autonomous research with inspectable evidence

Simjecture · Autonomous numerical research

Bring a hypothesis and the conditions that matter. Let an agent design experiments, search for counterexamples and test a better explanation—with recorded evidence and independent review.

Run your first investigation → Explore recorded studies Try online ↗

How research moves forward

  1. Frame the questionYou and the agent agree on the hypothesis, physical scope and budget.
  2. Run experimentsThe agent chooses methods and runs recorded numerical computations.
  3. Look for a failureChallenge the prediction and check candidate counterexamples.
  4. Commit a repairPreserve the failed claim and state a justified replacement before testing it.
  5. Collect fresh evidenceRun the committed tests and challenge the revised prediction.
  6. Review the conclusionA separate context examines the claim, source, results and numerical controls.

Review gaps return the agent to investigation. Human steering can change the next experiment. The selected completion policy determines whether a falsification finishes the study or starts a repair; the deadline can leave it incomplete.

Understand the loop · See what counts as evidence

中文：课题组安装与开发指南

See the process in a real record

The 48 hour FRC campaign on eight GPUs records two days of solver development and numerical controls. It ended with the merger hypothesis unresolved; its public accounting includes model usage, GPU allocations and API pricing scenarios.

The fresh v0.7.0 FLASH island-coalescence study falsifies a specified reconnection-rate slope band after numerical qualification and independent review. Its proposed repair remains unvalidated.

Fresh FLASH fields and the accepted scaling contrast

The fresh FLASH-to-WarpX investigation connects an MHD current sheet to local kinetic calculations. Two prospective particle seeds and numerical controls support the finite pressure prediction. Its independent claim and report reviews are accepted; historical records retain their original outcomes.

The Qin-1 hardware-to-plasma study follows CAD and feed redesign through coupled circuit/EM calculations into native 3D plasma dynamics. Its interactive results show both the stronger cold-conductor field and the numerical limits on predicted leakage. This is a local toolchain study, with its execution mode and qualification limits stated explicitly.

For a short, fully inspectable example, the orbital-accuracy investigation found five counterexamples and supported a narrower claim with fresh numerical tests and independent review. Ordinary CPU Python is enough to reproduce its experiments.

Start investigatingFrom a conversation to a tested claim

Prepare a brief, choose a completion policy and follow a study through experiments and review.

 Understand the recordWhy should you trust this result?

Trace a conclusion to its numerical inputs, outputs, counterexamples and review decisions.

 Bring a scientific toolStart from a working instrument

Give the agent a verified example, source access, diagnostics and clear physical limits.

 Keep exploringGuide the next experiment

Steer at a checkpoint or open a linked continuation with a revised brief and new budget.

Choose your working environment

Use the hosted workspace with supplied inference and compute within usage limits, or install locally and choose your own API or native CLI agent. You can run ordinary Python studies, add scientific capabilities, and use SSH workers for numerical experiments. Server mode supports an operated service with per-visitor allowances.

The sidebar identifies the documentation build version. Current walkthroughs and historical records are labelled separately. See research status, release checks, or contribute a tool or study.