Simjecture

Hypothesize. Simulate. Falsify.

Simjecture is a research harness for computational science. An agent writes and runs experiments, searches for counterexamples, and proposes small, testable repairs to failed hypotheses. The harness records what actually ran and requires independent review before a scientific claim can close.

v0.5.2 is the current stable release. Minimal mode is the default: agents keep their native tools and choose their research strategy, while Simjecture manages recorded evidence, review and deadlines. Structured and frontier modes remain available. The system does not establish physical truth from a successful simulation or promise that a long agent run will produce a useful result.

The workspace includes visible Continue investigation and Send guidance actions in the study view and experiment monitor. Continue with a new budget and selected prior files, or send advisory guidance at the next checkpoint. See continuation and steering.

The 0.5.3rc4 preview includes local/SSH experiment workers and refreshed research interfaces. In Machines, connect with an SSH address and password; hardware and worker setup are detected automatically, with optional advanced settings and agent preparation. Availability updates every 30 seconds. Minimal studies can dispatch parallel experiments with GPU reservations and recover existing jobs after SSH interruptions or coordinator restarts. Agents and independent review stay on the coordinator; Simote remains optional.

Research tools now use compact cards with build lists, diagnostics and logs in a separate Details panel. The preview also includes rc1's campaign audit repairs, request-level usage records, native DeepSeek tool-message continuity and experimental Simjecture Bench (task pack 0.3.0). See SSH setup, benchmark usage, and rc4 scope and validation. The stable installation commands below remain on 0.5.2.

Rc4 integrates the independent rc3 audit: concurrent evidence-ledger writes now preserve the hash chain, reserved multicore experiments retain their wall-time allowance, and worker polling, transfer recovery and usage-log reading are more efficient. It also fixes conversation navigation/submission races and adds small interface and accessibility improvements. Minimal remains the default; the research strategy and completion rules are unchanged. These repairs have direct reproductions and regression checks; a new long research campaign is still needed to measure their effect on scientific delivery.

The 0.5.3rc4 preview includes a task-scoped leaderboard: compare verified completion, cost and elapsed time, explore Pareto tradeoffs, import timed trial grades and export a summary. Run your model starts fresh trials for any native CLI or configured API model, with visible progress and downloadable grades. Actual results from 42 configurations are bundled with the task pack: 252 scheduled attempts, with quota/expiry observations kept separate from inference results. Earlier pilots and qualifications remain in the audit records; they are absent from the public leaderboard. Missing measurements stay explicitly labelled. See the leaderboard protocol. The leaderboard first shipped in 0.5.3rc3 and is retained in rc4; it is absent from the older 0.5.3rc2 preview.

The leaderboard opens on Official Simjecture results, with two task tabs, time/cost ranks, published API tariffs and a cost-versus-time plot. Each task has ranked time bars, ranked cost bars and a Pareto plot with the upper-left corner preferred. Unfinished runs appear in red and retain recorded cost bars; configurations without inference are omitted. Model colours are consistent, and only same-model effort settings are connected. Methodology is a link beside the plots. Current community and local trials have a separate view; older qualifications remain in the audit records. A standalone read-only publication needs no GUI, API key or imported data:

simjecture llm-benchmark export-page --output ./leaderboard.html

To try rc4, preserving existing research data:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.3rc4/install.sh | bash
# Or install only the Python package:
uv tool install 'simjecture[workspace,tui]==0.5.3rc4'

Install and open the workspace

On Linux or inside WSL, run:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.2/install.sh | bash

The installer downloads the versioned release, verifies its checksum, installs uv, Python 3.12 and the workspace dependencies, and starts the local GUI. Git and a CLI agent are not required. Enter a compatible API endpoint and key in Connections, or select an existing local CLI from the message composer. Choose Interactive research for a task, or prepare an Autonomous research study together with the agent.

Software lives under ~/simjecture/app/<version>; conversations, inputs, simulations and studies stay in ~/simjecture/artifacts/projects. Reopen the GUI with ~/simjecture/start-workspace. On SSH hosts the installer prints forwarding instructions for your local browser. Keep the server terminal open, or run the launcher in your usual session manager. Ubuntu/Debian may request sudo to install Bubblewrap; restricted containers can open the GUI but may need a separately configured execution backend.

See installation options and the workspace walkthrough.

Start a study from an external CLI

Use Linux with Python 3.11+, uv, and an installed, authenticated native agent CLI. Bubblewrap is the default numerical execution backend.

uv tool install 'simjecture[tui]==0.5.2'
simjecture install core
simjecture doctor --execution-backend bubblewrap

simjecture study --campaign ./runs/my-study \
  --hypothesis-file hypothesis.txt --instructions-file instructions.md \
  --backend codex-glm --model glm-5.3 --wall-seconds 3600

Write a falsifiable hypothesis in hypothesis.txt. In instructions.md, specify physical scope, available instruments, resource limits and the evidence needed to accept or reject it. Configure provider credentials in your agent's own credential store, outside the study directory. The model name must match that configuration.

A terminal launch shows activity, backend/model, remaining time and experiment/review counts. Inspect the same study in either interface:

simjecture web ./runs/my-study
simjecture tui ./runs/my-study

Resume using the same campaign path and launch options. Its recorded mode and original wall deadline remain fixed. A normal agent exit or inconclusive result does not mark research complete. Transient provider disconnects retry with backoff until the deadline; authentication, permission and exhausted-quota errors pause for attention. Waiting for a provider counts against wall time.

See minimal-mode usage, terminal progress and the browser interface.

Choose the agent, workflow and execution backend separately

Use --mode structured or --mode frontier for another workflow. Minimal lets the agent organize its investigation around a small evidence API; structured uses explicit claim-scoped roles and contracts; frontier gives a persistent researcher control of the classic claim tree. Existing campaigns retain their recorded semantics.

Native agents retain their own shell, search, file and configured provider tools. They are trusted, cooperative processes on the study host; they are not enclosed by the numerical experiment sandbox. Simjecture's supplied tools record experiments, provenance and reviews. Native scratch work is not automatically scientific evidence.

If namespaces are unavailable, explicitly select --execution-backend proot-cooperative on a dedicated non-root account. It provides execution bookkeeping and limits, not an OS security boundary. Simjecture never silently falls back to it. See restricted containers.

DSH integration remains available with native research tools and a governed MCP boundary. Simote integration is optional; Simjecture can run independently on a local or remote host. Deployment to a remote host does not itself create a multi-machine experiment scheduler.

What minimal mode enforces

The agent chooses hypotheses, code, diagnostics and the next discriminating experiment. The host maintains the scientific record:

• Immutable question and scope. Operator requirements cannot be waived by a worker or reviewer. Repairs preserve ancestry and explain each changed assumption or bound.
• Recorded execution. Commands, declared inputs, runtime identity and output hashes accompany each experiment. Interrupted attempts remain visible.
• Commissioning before evidence. Instrument-backed studies receive source-bound methods review. Exploration and supplied examples cannot be relabelled as new evidence.
• Counterexample search and fresh repairs. Support includes an actual challenge attempt. Repaired predictions are committed before their fresh tests.
• Independent acceptance. A methods approval permits evidence collection; it does not accept a scientific claim. Unconverged, censored or missing results are not physical counterexamples.
• Persistent context. The host maintains an experiment journal and bounded research brief. Evidence-linked notes and comparisons help retain failed approaches and findings.
• Bounded work. The original deadline survives restart. Cancellation uses durable job records and reports unresolved cleanup rather than assuming that agent exit stopped every computation.

Under the default CLI repair policy, a study completes when its original claim is independently supported, or when a falsified original leads to an independently supported repair with the required ancestry. The workspace defaults new studies to answer, which also permits an independently accepted root falsification to complete the investigation. Select the policy in the study brief or with --completion-policy; resumes retain it. A deadline may leave the campaign unresolved; that is recorded, not promoted to success.

See methods and progress oversight, research memory and evidence and claims.

Begin with a working scientific instrument

For expensive solver setup, provide a guided commissioning package: exact working source and command, output reader, observable, validation checks, runtime and known limitations. Reproduce that anchor in a separate bounded readiness run before spending hours on hypothesis testing.

simjecture study --campaign ./runs/guided-study \
  --hypothesis-file hypothesis.txt --instructions-file instructions.md \
  --guided-commission /path/to/guided_commission.json \
  --capabilities /path/to/installed-capabilities \
  --backend codex-glm --model glm-5.3 --wall-seconds 3600

Guided examples are starting instruments, not fresh evidence for the new hypothesis. An executable smoke test does not validate a new geometry or diagnostic. The guided commissioning guide explains readiness budgets, immutable anchors and instrument versus production review.

Optional instruments include WarpX CPU/CUDA, operator-supplied FLASH, and EOS/opacity capabilities. For example:

simjecture install warpx-cpu
simjecture doctor --profile warpx-cpu

The release also includes an ITER ecosystem pack: CHERAB/Raysect synthetic diagnostics, CHERAB-ITER/IMAS adapters and IMAS data tools. Use Research tools → ITER pack → Install, then Run demo for a monitored example with plots, or run simjecture install iter-pack. JOREK, SOLPS-ITER and DINA-PS have separate Install with agent entries and source-based guidance; they are not silently included in the diagnostics installation.

FLASH source is not redistributed. Solver source inspection can be configured separately from a writable problem/build area. See runtime deployment and the scientific skills.

Research core introduced in v0.5.1

This release builds on v0.5.0's minimal default and shared interfaces:

• Methods/progress oversight, stalled-session recovery and deadline-bounded provider retry.
• Automatic research journaling, bounded context and evidence-linked comparisons.
• Guided anchors in native studies and explicit cooperative execution for restricted hosts.
• Updated FLASH/Python skills with guided-study, HDF5 and decision-rule lessons.
• Preserved instrument ownership across classic assignment rollover.
• Explicit unmet review prerequisites; conditional approval cannot open the evidence gate.
• Numerical checks for contradictory prospective acceptance bounds.
• Output eligibility annotations exposed for review without forcing metadata-only solver reruns.
• Durable cancellation cleanup independent of CLI-response parsing.

The changelog and repair report include validation and remaining limitations. General reuse of raw solver data under independently versioned analysis still needs a fuller provenance API.

Measured evidence and limitations

Minimal is the default by design preference, not a demonstrated universal performance advantage. The latest MiMo v2.6 Pro / DeepSeek V4.1 Flash comparison found faster useful progress with DeepSeek in a guided minimal FLASH study, but no independently accepted scientific conclusion from either model configuration. Unequal duration, different provider tool protocols and one trial per configuration limit that comparison. Token totals distinguish cached input, uncached input and output; they are not billed-cost estimates.

The historical FLASH narrative recorded an exponent-band falsification with an interval that overlaps the allowed band. That overlap alone does not establish such falsification; the earlier summary is not independently endorsed here. Provenance enables inspection, but cannot make a faulty scientific decision correct.

Replay a completed example without an API key or new simulations:

git clone https://github.com/tomzhu0225/simjecture.git
cd simjecture
uv sync --frozen
uv run python demos/gray_scott_counterexample/verify_record.py
uv run simjecture web demos/gray_scott_counterexample/record --read-only

Documentation and development

Start at the documentation index. Linux release assets include a launch archive with matching Python/DSH packages, setup scripts and SHA-256 checksums. Native-agent setup does not require installing DSH.

uv sync --all-groups --extra tui --extra dsh
uv run ruff check .
uv run pytest
uv run simjecture schemas --output schemas --check
uv run --group docs sphinx-build -W -b html docs docs/_build/html

conjecture-solver and acs remain compatibility aliases. The Python package remains conjecture_solver. See CONTRIBUTING.md, SECURITY.md and THIRD_PARTY_NOTICES.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.