System architecture

The system deliberately separates scientific autonomy from execution authority.

Workflow scope

New studies default to minimal mode. Native CLI agents retain their host tools; the built-in API agent also supports minimal mode. Minimal studies record their own research ledger, experiment receipts and independent reviews, with an absolute wall deadline that survives restart. Their saved completion policy distinguishes answer (a reviewed negative answer can complete the root) from repair (a falsified root needs a supported repair). See minimal research.

The kernel/DSH architecture below describes the classic campaign route. Its contracts and active-time accounting must not be assumed to describe the minimal study launcher. Structured and frontier modes remain available explicitly.

Human problem contract

The operator supplies the root hypothesis, operational instruction, resource limits, installed capabilities, and any guided commissioning package. These inputs are recorded before the agent begins.

Reasoning harness

The model decides how to operationalize the hypothesis, which diagnostics to write, which sub-hypotheses to register, and which admissible experiment to run next. The built-in runner and the DSH profile are interchangeable reasoning front ends; neither is the authority for scientific state or evidence.

Native tool boundary

Under DeepSeek Harness, scientific role filters restrict authoritative Simjecture MCP operations. Native shell, filesystem, web, skills, workflows and configured plugins remain available for research in a separate working directory. DSH owns the provider, conversation, retry policy, compaction and model-facing session. The profile uses DSH’s workspace-write policy and non-interactive never approval policy; it does not OS-isolate arbitrary plugin code. CampaignKernel remains the scientific mutation and evidence gate. A composite adjudication tool freezes the case, runs a fresh tool-free DSH child with a strict verdict schema, and commits the verdict through raw endpoints hidden from researcher roles. See DSH setup and trust boundaries.

Campaign kernel

The model-independent Python kernel validates typed actions and owns the root hypothesis, hypothesis graph, claim ledger, evidence contracts, commissioning, skills, capability registry, provenance, and workspace policy. The legacy MVP runner delegates to this same kernel, so changing the reasoning harness does not create a second scientific implementation.

The autonomous loop exposes three operational roles. The Falsifier commissions tools and searches for counterexamples. After a valid falsification, the Scientist proposes a minimal repairs successor. When a bounded search finds no counterexample, an independent Judge sees only the auditable evidence package. The Judge reports two separate facts: whether the record is complete enough for a terminal decision (sufficient or insufficient), and, when it is sufficient, the scientific disposition such as supported, instrument_limited, or unresolved. Record completeness never implies support. Roles are explicit workflow contexts, not claims of different underlying model providers.

An accepted verdict does not itself end the process. Guarded finalization checks the entire scientific frontier and writes the report only when no open claim or unrepaired counterexample remains. A falsified scientific frontier therefore requires a minimal repair child; an honestly documented instrument limit may terminate as instrument_limited, and an irreducible bounded ambiguity may terminate as unresolved.

Sandbox and capabilities

Recorded experiment code defaults to a network-isolated Bubblewrap workspace. Installed capabilities are harness-owned executable environments mounted read-only into the same sandbox. Skills explain interfaces and numerical practice but grant no authority by themselves. Native research agents are trusted host processes, not enclosed by that numerical sandbox. The explicitly selected proot-cooperative backend supplies execution bookkeeping without OS-enforced filesystem or network isolation; see restricted hosts.

Scientific state

The claim ledger stores root, scientific, instrument, diagnostic, and control claims with explicit lineage and disposition. Evidence contracts freeze the observable, decision rule, uncertainty criterion, inconclusive conditions, machine validations, source identity, and allowed commands before decisive execution. A claim_decision contract is allowed to support or falsify its claim. A terminal_record contract documents why a bounded campaign cannot decide the scientific proposition and may only end as instrument_limited or unresolved. This distinction prevents a contract written to recognize a failed antecedent or blocked instrument from being treated as positive evidence for the hypothesis.

Durable operational state

The reasoning harness retains its model session. Under DSH, a stable session ID and event-sourced log live inside the campaign; pause, process restart, and resume reopen that session instead of constructing a blank conversation. Simjecture journals simulator intents before dispatch and stores content hashes, operation identifiers, cancellation records, process identities, and authenticated worker receipts. Long jobs survive an MCP restart. A missing or unverifiable receipt remains an unknown operational outcome and cannot become scientific evidence. The bounded snapshot exposes durable jobs, operation bindings, and remaining budgets to a fresh session. One root runner/MCP process holds the campaign ownership lease; detached workers hold explicit active-job writer leases. For this classic campaign route, budget accounting is cumulative active execution time, not elapsed calendar time between sessions. Minimal studies instead retain an absolute wall deadline, including pauses and provider waiting.

Human interfaces

The browser dashboard, Textual dashboard, status, and watch consume one UI-neutral monitor projection. That projection derives claim status, current typed action, loop stage and role, token usage, heartbeats, and terminal state from durable files; the clients do not maintain a competing scientific database.

The local web server provides a localhost API around this projection and the existing reviewed launch/pause/resume/stop functions. Its hypothesis graph uses the same scientific-versus-validation claim classifier as the TUI. This boundary also allows a future reasoning engine to change without rewriting the scientific interface or campaign record.

Trust boundary

The system can establish that a declared computation ran through the expected instrument and satisfied registered gates. It cannot make an imperfect diagnostic scientifically correct, turn a simulation result into empirical truth, or eliminate the need for independent interpretation.

Recorded evidence snapshots and completed adjudications are audit artifacts; operational state and append-only journals continue to evolve during a study. If later review finds a semantic or scientific error, Simjecture preserves the original bytes and attaches a corrective audit record instead of rewriting history.