Try online · Documentation · Research loop · Recorded studies · Contributing
Hypothesize. Simulate. Falsify.
Simjecture runs autonomous numerical investigations from a human-defined scientific question. An agent designs experiments, searches for counterexamples and tests revised hypotheses. The evidence system preserves what ran and a separate reviewer examines whether the results justify the conclusion.
You set the question, physical scope and budget. The agent chooses how to work. You can inspect intermediate results, steer the investigation and continue it in a new phase. The result is a research record you can follow from hypothesis to code, measurements, counterexamples and review.
A counterexample must survive the relevant numerical controls. Under the repair completion policy, the agent then proposes the smallest justified change to the hypothesis, commits its predictions before testing them, and collects fresh validation evidence. Missing evidence sends the work back through the loop.
The answer policy can also finish with a reviewed falsification of the original claim. New browser studies default to answer; the CLI defaults to repair. Both use the minimal workflow: flexible agent strategy, recorded experiments and explicit scientific acceptance. A deadline can leave the question unresolved.
The optional research director can stop unaffordable experiments, request a replan and tune remote instrument probe deadlines and retries. Operators can also delegate storage increases for large builds and checkpoints. The monitor shows failures, before/after limits and the reasons for changes; scientific acceptance and the campaign deadline remain separate. See director control and recovery.
Conjecture 1 is an
optional scientific-attention coprocessor
(packages/simjecture-monitor)
that observes public study state and proposes when an investigation may need help or
director review, while keeping scientific review and repair authority with the host.
The package also provides offline SFT and measured-return RL training and evaluation
tools; see scientific attention.
Read how the loop works or how evidence is checked.
The 48 hour FRC campaign on eight GPUs ended with its merger hypothesis unresolved. Its report shows the retained plasma fields, numerical controls, model usage, API cost estimates and the controller failures exposed by two days of research. Public tables reproduce the accounting.
Magnetic-island coalescence with FLASH shows a fresh v0.7.0 investigation that falsified a reconnection-rate slope band. Follow the evolving plasma fields, numerical qualification and accepted counterexample. The proposed replacement law remains unvalidated.
FLASH → WarpX kinetic follow-up connects a fresh MHD current sheet to local kinetic calculations. Two new particle seeds and numerical controls support the finite pressure prediction, with independent claim and report acceptance. The earlier unresolved records remain available alongside both fresh examples.
From hardware design to 3D plasma (Qin-1) connects CAD → mesh → circuits and electromagnetics → 3D plasma → visualization. An agent revised the coil and feed geometry, solved current sharing with Elmer, and transferred the checked device field into native 3D WarpX. The selected paired-feed design gives 33–34% more upstream transverse field while retaining 97.8% of the array current across matched refinements, while five plasma controls expose 11–71% bulk-flux variation that leaves quantitative shielding unqualified. Explore the interactive results or verify the retained record (an operator-assisted local toolchain study, not a formal harness-adjudicated claim).
These plasma and multiphysics investigations include preserved records, downloadable figures and explicit distinctions between a completed simulation, an approved method and an accepted scientific conclusion.
The orbital-accuracy walkthrough uses the current minimal workflow to investigate a concrete question: does small energy error imply an accurate trajectory? It connects the operator's brief, actual numerical experiments, hypothesis revisions and independent review. It found five counterexamples in the original case matrix, then supported a four-case repair using fresh simulations and independent review. The retained record can be verified without a model call.
| Question you should be able to ask | Recorded material |
|---|---|
| What was being tested? | Original hypothesis, operator requirements and committed repairs |
| What actually ran? | Frozen source and inputs, command, runtime identity and execution receipt |
| Where did this number come from? | Output files, hashes and the experiment cited by the claim |
| What challenged the result? | Counterexamples, controls, failed attempts and review gaps |
| Why did the investigation finish? | Independent review, completion policy and budget or stop reason |
A repaired claim cannot reuse old exploratory results as fresh validation. Changed committed source, missing committed cases and altered recorded outputs are checked by the service. Numerical adequacy and scientific interpretation remain review judgments. See the evidence guide.
Open simjecture.com for the hosted workspace with supplied inference and compute within usage limits. Signed-in visitors can upload files and retain projects. Guest conversations and artifacts are temporary.
For your own Linux or WSL machine, install the latest stable release:
curl -fsSL https://github.com/tomzhu0225/simjecture/releases/latest/download/install.sh | bashThe installer verifies the release bundle and prepares Python and the browser
workspace. Reopen it with ~/simjecture/start-workspace. Research files live under
~/simjecture/artifacts/projects. See installation options.
In the local workspace, open Knowledge → Connect folder to make research files searchable across new conversations. The library also indexes completed workspace discussions and study reports, with versioned source citations. Indexing needs no embedding service. See the knowledge library guide.
- Choose a compatible API model or an authenticated native agent in the workspace.
- Discuss your question and prepare a study brief with Autonomous study.
- Review its hypothesis, evidence requirements, completion policy and budget.
- Start the investigation and follow its experiments and Evidence & review.
The first-study tutorial walks through the full counterexample-and-repair workflow. Headless studies cover terminal launches and automation.
For longer-running work, optional projects and workspaces group conversations and shared context, or keep separate local lab environments. The default remains a single personal workspace. CLI and HTTP interfaces expose the same project operations for external agents and scripts.
Ordinary Python studies work without an external solver. Optional capabilities include WarpX, operator-supplied FLASH, CAD, meshing and FEM tools (FreeCAD, Gmsh, Elmer), Quantum ESPRESSO DFT, EOS/opacity and FAC atomic-physics tools, ITER diagnostics and GPU cylinder flow. The cylinder solver and its validation records were contributed by Zifei Meng.
An experimental Windows application interface shares job execution, input snapshots and artifact collection across product adapters. It supports CST magnetostatic jobs, native CST Python scripts, and an initial SOLIDWORKS part rebuild/export adapter. These use trusted Windows processes and record exploratory jobs; SOLIDWORKS still needs native model commissioning.
- Commission a scientific instrument or add your own capability.
- Configure optional scientific attention with Conjecture 1.
- Run experiments on local or SSH workers.
- Steer or continue an investigation.
- Compare agents on recorded diagnostic tasks.
- Operate a hosted service.
Native agents retain their existing tools and run with their host-account access. Numerical experiments use the selected execution backend, Bubblewrap by default. Read security guidance and restricted-host setup when deploying workers.
Help us test the research process on meaningful questions. Contributions can add scientific instruments, validated diagnostics, reproducible studies, numerical review, documentation or infrastructure. Start with CONTRIBUTING.md or open an issue to discuss a study or a substantial change.
Release notes · Release checks · Research status · Historical records
Use CITATION.cff and identify the exact version and Git commit used for a result. The concept DOI identifies the release series.
Copyright 2026 Bowen Zhu and contributors. Licensed under Apache 2.0. See third-party notices for upstream attribution and redistribution requirements.