Simjecture

Quickstart · Documentation · Recorded demo · Contributing · Releases

Hypothesize. Simulate. Falsify.

Simjecture is a research workspace for computational science. Work with an agent to turn a question into a bounded study, run experiments, search for counterexamples, and test repairs to failed hypotheses. The harness records what ran and requires independent review before a scientific claim can close.

Use a compatible API model or an installed native agent CLI. Start in the browser workspace; terminal and headless workflows are available when you need automation.

A recorded Gray–Scott study found a finite-domain counterexample. You can inspect its code, evidence and review without an API key or new simulations. Simulation results remain conditional on the model and numerical scope; see scientific limitations.

Install and open the workspace

On Linux or inside WSL, install Simjecture 0.5.3:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.3/install.sh | bash

The installer verifies the release bundle's checksum, sets up uv, Python 3.12 and workspace dependencies, then starts the local GUI. Git and a CLI agent are not required. For Python-package installation, SSH forwarding or source setup, see installation options.

1. In Connections, add a compatible API endpoint and key, or select an installed, authenticated CLI in the message composer
2. Start a conversation with your question, paper or simulation outputs
3. Open Autonomous research, prepare a study with the agent, and review its question, evidence requirements and budget before starting
4. Follow the recorded experiments and independent review in the study view

Reopen the workspace with ~/simjecture/start-workspace. Research files stay under ~/simjecture/artifacts/projects, separate from versioned application folders. Keep the server terminal open or use a session manager. Ubuntu/Debian may request sudo to install Bubblewrap; restricted hosts may need additional execution setup.

Follow the workspace walkthrough for agent selection and your first investigation, or try the no-key recorded demo first.

What you can do

• Investigate a question: work interactively, then launch a bounded autonomous study with recorded evidence and independent methods/claim review
• Challenge and repair hypotheses: search for counterexamples, preserve claim ancestry, and test committed predictions with fresh evidence
• Use scientific instruments: commission optional WarpX, operator-supplied FLASH, EOS/opacity tools and ITER diagnostics, or add your own capability
• Run local or SSH experiments: select numerical workers and track resource reservations, execution receipts and recovery
• Continue an investigation: prepare a linked phase with a new budget and selected prior files, or send advisory guidance at a checkpoint
• Compare coding agents: run task-scoped benchmarks over recorded diagnostics and inspect completion, time, cost and missing measurements

Minimal mode is the default: the agent chooses its research strategy while the host manages evidence, review and deadlines. Structured and frontier modes are also available. See research modes and evidence and architecture.

Numerical experiments use Bubblewrap by default. Native agents are trusted same-account processes with host tools; the experiment sandbox does not enclose them. Explicit cooperative PRoot execution is not an OS security boundary. Read security guidance before running sensitive workloads and restricted-host setup when needed.

Find your next step

• Use the workspace: first investigation, continue or steer a study
• Automate from a terminal: headless studies, command reference, optional terminal dashboard
• Prepare an instrument: guided commissioning, runtime deployment, ITER ecosystem pack
• Add compute: SSH workers
• Evaluate models: benchmark and community submissions
• Inspect the science: evidence and claims, evaluation status, research and release archive

0.5.3 brings SSH experiment workers, the measured coding-agent leaderboard, ledger and recovery repairs, and shared conversation/study/evidence navigation. Release changes live in CHANGELOG.md; checks and limits are in the 0.5.3 acceptance record and archive. The documentation index covers the full guide set.

Contributing

Researchers, developers and first-time contributors are welcome. Help improve an example or guide, reproduce a bug, add a tested diagnostic or scientific capability, review numerical methods, or share benchmark results.

Pick a small task in CONTRIBUTING.md, which includes setup and checks for each kind of change. Open an issue to discuss an idea or reproducible problem; discuss large changes before implementing them. For security-sensitive reports, follow SECURITY.md.

Citation and license

Use CITATION.cff and cite the exact version and Git commit used for a result. The concept DOI identifies the software release series.

Copyright 2026 Bowen Zhu and contributors. Licensed under the Apache License 2.0. See third-party notices for upstream attribution and redistribution requirements.