Run your first investigation

This tutorial takes a scientific question through counterexample search and a fresh test of a revised hypothesis. Use it to see how the autonomous worker and evidence system cooperate. The recorded orbital study provides an example you can inspect before spending model usage.

Prepare the workspace

Install Simjecture locally and select a compatible API model or an authenticated native CLI in the composer. The full repair-loop walkthrough below uses the local workspace. The hosted trial supplies its configured model and compute and currently uses the answer completion policy; it can demonstrate recorded experiments and review of the original claim.

The following study uses ordinary Python, NumPy and SciPy. No external simulation package is required. For the complete source-checkout recipe and the exact recorded operator input, see the demo page.

Start with a testable question

Paste this into a conversation:

Help me prepare a study of orbital accuracy. For a planar Kepler orbit with GM=1 and semimajor axis 1, does keeping relative energy error below 0.001 guarantee position error below 0.01 over twenty periods? Use fixed-step kick-drift-kick leapfrog, eccentricities 0, 0.3 and 0.6, and 64, 128, 256 or 512 steps per period. Compare with an independently checked reference orbit. Search for a counterexample. If the claim fails, propose a justified minimal repair and commit its predictions before collecting fresh validation evidence. Prepare a twenty-minute study with the repair completion policy.

This is a finite, educational numerical question. The agent still has to implement and check its computations, choose its experiments, explain any repair and submit its evidence for independent review.

Review the brief and launch

Open Autonomous research and choose Draft from this conversation, or Grill me if you want the agent to ask focused preparation questions.

Before pressing Start research, check:

  • The original claim, finite case family, units and error definitions.

  • The reference calculation and numerical controls the investigation needs.

  • Minimal mode and Seek a supported claim or tested repair completion.

  • A wall budget permitted by your local resources and model allowance.

New browser briefs normally use Answer the question, which can finish after a reviewed root falsification. Choose the repair policy explicitly for this tutorial. With a shorter budget, review or repair may remain unfinished; you can still inspect the partial record.

Follow the investigation

The study view shows active experiments and remaining time. Open Evidence & review to inspect the original claim, recorded computations, methods and review requests. The worker can use its existing tools and change its plan; source and outputs used as evidence retain their recorded identities.

When a candidate counterexample appears, inspect its energy and position errors at the same physical times. Then inspect the controls used to justify the reference. A plot alone does not establish that the antecedent was satisfied.

If the original claim is accepted as falsified, look for a committed replacement and new experiment IDs validating it. The rejected claim stays in the record. A request for more evidence should lead to additional work while the budget allows. See the research loop for the roles and feedback paths.

Read the outcome

Open the final narrative alongside the host-generated study ledger. Identify:

  1. The exact claim accepted or left open.

  2. The experiments and numerical controls cited by the reviewer.

  3. Whether a replacement was committed before its validation runs.

  4. The domain, remaining limitations and reason the study stopped.

A supported repair is conditional on its declared numerical scope. A study that runs out of time has partial findings, even when some individual experiments succeeded. You can steer or continue the investigation with a new question or budget while retaining the previous phase.

For other working styles, see the workspace guide and headless studies.