How the system fits together

Simjecture separates the agent’s freedom to investigate from the authority to accept a scientific claim. The current default is minimal mode, available through the browser workspace and headless launcher.

Human direction and study state

A conversation helps prepare the question, scientific scope, operator requirements, instruments and resource budget. Launching records that brief. The root hypothesis, requirements, completion policy and deadline stay fixed within the study.

Steering adds visible guidance for the next checkpoint. A continuation creates a linked phase with its own agreed brief and budget, preserving the previous phase. See continuation and steering.

Worker, director and reviewers

Role	Responsibility	Authority
Human researcher	Choose the question, assess relevance, guide interpretation	Launch, steer, stop and prepare continuations
Research worker	Plan, implement, run experiments, challenge and repair claims	Submit work and request reviews
Research director	Assess feasibility, time remaining and scientific coverage	Request a recorded replan and stops of named experiments
Methods reviewer	Assess an instrument-backed implementation and commissioning	Approve or return gaps for the proposed method
Claim reviewer	Assess whether evidence supports or falsifies the target	Return an explicit scientific verdict
Host evidence service	Preserve state and enforce recorded rules	Validate identities, receipts, commitments and completion conditions

These are responsibilities and separate contexts, not a requirement to buy several model providers. Worker and reviewer models can be chosen independently. The director does not accept scientific claims; a strategic decision to continue a calculation is not methods approval.

Agent runtime and evidence service

Native agents retain their normal shell, filesystem and configured tools. A compatible-API worker is also available. The service exposes a compact lab client for recorded experiments, prospective repair commitments, review requests, notes and progress. See the minimal-mode API.

The host maintains its own ledger and report. It records scientific state from receipts and review decisions rather than inferring success from an agent’s final message. Review gaps can resume the investigation; transient provider disconnects retry within the same wall budget. Authentication or exhausted-quota failures require attention and are reported explicitly.

Local and remote numerical execution

Experiments can run locally or on selected SSH workers. Their frozen inputs, resource requests, execution outcomes and retained files share the same research record. Detached numerical jobs can survive an agent turn or coordinator restart. An unreachable worker remains operationally uncertain until its job is reconciled.

Slurm submission is not implemented by the current SSH worker adapter. It is a separate integration opportunity. See SSH workers.

Execution and trust

Bubblewrap is the default numerical execution backend where the host supports its namespaces. Explicit cooperative PRoot and native-process backends accommodate restricted trusted hosts; neither is an OS security boundary. Native agent tools run with host-account access outside the numerical sandbox.

The public service uses per-visitor ownership, quotas and its separately configured executor. A private local workspace and a public multi-user service have different trust assumptions. See server deployment and security guidance.

Scientific evidence and completion

The evidence guide explains source/output identity, methods qualification, counterexample search, fresh repair tests and independent review. The saved answer or repair completion policy determines which reviewed outcomes complete the study. Neither elapsed time nor successful process exit establishes a scientific answer.

Other workflows and historical records

Structured and frontier workflows remain explicit launcher choices. DSH and the legacy campaign kernel retain their own contracts and role interfaces. Existing studies keep their recorded mode; they are not migrated silently.

The classic architecture and classic evidence model explain those older records. Their active-time accounting and claim APIs should not be applied to minimal studies, which use an absolute wall deadline and their own evidence API.