Evaluation status¶
Simjecture provides an autonomous research loop and a recorded evidence system. Its documented studies demonstrate specific workflows and numerical outcomes; they do not establish unrestricted scientific reliability. Simjecture was first built to attack simulation-gated plasma conjectures. The claim ledger, sandbox, commissioning protocol, and capability framework are now domain-neutral enough to extend to other fields that have a sharp question, a commissioned instrument, and a cheap verification path. Plasma remains the origin domain.
Demonstrated¶
Natural-language hypothesis input and autonomous agent-authored computation.
A current minimal-mode Kepler counterexample-and-repair study with accepted reviews, preserved arrays and independent numerical reproduction.
Domain-neutral claims, prospective evidence contracts, and guarded closure.
Sandboxed Python plus separately installed CPU and CUDA capabilities.
Instrument commissioning across representation, physics, boundaries, diagnostics, and numerical regime.
Recorded program, input, command, artifact and claim provenance, with model usage reporting and explicit incomplete-usage flags.
Durable long-running execution, cancellation, replay, and recovery.
A planted matched-moment kinetic-stability counterexample verified by analytic, independent PIC, and WarpX paths.
Real magnetic-mirror, nonlinear-Landau, Gray–Scott, and two-dimensional collisionless GEM autonomous campaigns, including a preserved 12-run CUDA GEM demonstration.
The fresh FLASH island-coalescence audit also preserves an unresolved outcome and a post-run correction: a half-cell error in the agent’s diagnostic caused false eligibility rejections and was missed in methods review. Recorded decisions are inspectable judgments, not automatic scientific correctness.
The two-solver FLASH-to-WarpX study also completed real state transfer and numerical controls while leaving its claim unresolved. The demo refresh audit records the observed agent errors, harness fixes and remaining limitations separately.
Not demonstrated¶
Reliable execution of arbitrary acceptable natural-language hypotheses.
Empirical truth beyond the installed simulation model family.
Consistently correct autonomous statistical interpretation.
A model advantage over equal-budget conventional scientific strategies.
Next¶
The current target remains methods and infrastructure. The same claim ledger, sandbox, commissioning protocol, and capability framework are ready to extend to other simulation-gated fields. The next scientific step is to use that tooling to hunt independently confirmed new results — a compact law, a matched counterexample, or a scoped impossibility — in the origin domain and beyond.