Architecture

worldseq evaluates CEO planning and replanning over simulated SaaS companies. The engine executes typed strategic commitments; prose is an explanation, never an instruction to manufacture a business outcome. The core is independent of the applications and agent providers that use it.

Ownership

• src/catalog: attributed source claims, declarative companies, tasks, admission.
• src/evidence: typed public sources, claims, reference cases and parameter ensembles.
• src/validation: bound assertions, private campaign registration and paired statistics.
• src/engine: virtual time, business transitions, public projections, replay and verification.
• src/agents: per-agent policies/prompts, strict protocol schemas, adapter runtime.
• src/runtime: episode lifecycle and canonical persistence.
• src/http: player and control API handlers.
• src/evaluation: registration, attempts, budget accounting and results.
• src/integrity: canonical hashes, seals and safe artifact handling.
• dashboard: small human evaluation interface using shadcn components.

PostgreSQL owns service episodes. Each mutation locks its episode row and commits state plus revision atomically. Pure in-process episodes support deterministic tests and local scripted evaluations. Business days and real runtime budgets are separate clocks. A closing episode rejects new player writes before it is reconciled and sealed.

The player listener exposes only episode-scoped public operations. The separate control listener creates episodes and owns private records, seals and evaluation reports. Tokens are scoped to one boundary. Agent containers receive neither host mounts nor database, control or signing credentials.

Release boundaries

1. Workspace and strict agent contracts.
2. Deterministic engine, connected business records, accounting and attack tests.
3. Sourced family/company/task catalog, explicit behavioral sensitivity and admission.
4. Durable service, remote/container adapters, evaluation and submission CLI, dashboard.
5. Full corpus validation, reproducibility report, contribution and operator documentation.

The public corpus contains distinct business families and companies, long-horizon tasks, short crisis/liquidity tasks and source-bound mechanism cases. Public source code cannot constitute a sealed holdout. Official registrations must provide separately held worlds and seed schedules. Task count is coverage, not evidence of realism. Dataset and mechanism provenance, executable semantics, negative controls and admission results accompany the corpus.

Evidence and assurance

Documented business rules, measured traces, aggregate calibration and authored behavior hypotheses are different evidence classes. Hypotheses require sensitivity runs; aggregate fit alone does not validate an intervention. Runtime validation and cryptographic seals prevent specific manipulation classes; they do not establish universal reward-hijacking immunity or real-world causal validity.

Local results are self-reported until deterministic replay verifies them. Replay does not prove that a named model produced the decisions. Only maintainer-controlled runs signed by a separately pinned trusted key can establish official provenance. Remote agents report hidden token use; they belong to a distinct assurance track.

Operating kernel

src/engine/business separates declarative specs, causal conditions, interventions, financial transitions, product fulfillment, observation publication and assessments. The core executor shares company and department capacity. Financial, product and revenue verifiers independently reconstruct primitive events; they do not call the transition implementation or trust its daily cached totals. Separate verifiers also check staff allocation, customer readiness, contractual renewal, irreversible losses, customer-batch exposure and fixed-cohort retention. Published financial calendars expose contractual dates and settled status while withholding future collection-delay draws.

Legacy episodes without operating/scenario specifications retain engine1 behavior. Operating episodes declare engine2 semantics. Both are parsed explicitly; source, company, task and execution identities bind registrations. The existing legacy source hash remains separate from the broader execution identity, which also covers scripts, data, dependency locks, migrations, runtime versions and immutable images.

Campaigns register once, preserve every attempted or unlaunched case, propagate cancellation to workers and owned containers, and retain failures. Dashboard heartbeats are operational metadata. Imported submissions pass independent replay before their complete public projection can enter history.

Environment verification and operator observation

Admission executes canonical transitions and invokes two separate hooks. The operator observer receives detached, allowlisted public frames and bounded pages; its PostgreSQL read model backs the existing run history and task rollout routes. Writes use backpressure, a cumulative telemetry limit and coalesced heartbeats. This state is observational and cannot award admission. Historical imports lock the canonical batch and verify retained receipts before publishing its projection.

The trusted witness-completion hook receives the full factual journal and seal only after a feasible path completes. It stores private compressed evidence and checks the independent inventory/case/source/seed binding, replay and all required constraints before issuing a certificate. Failed certification preserves the behavioral result as inconclusive. Neither hook exposes the private witness bundle to the player or the operator rollout API. The underlying development reference policy remains public source; do not describe these development instances as confidential holdouts.

A feasibility certificate establishes existence for its exact registered instance. Protected issuance requires that instance’s current certificate and trusted bindings; development episodes remain a separate mode. The shared finalizer enforces declared assessment and counterfactual requirements. These bounded guarantees do not establish feasibility for unseen instances or a universal mathematical proof. See private instances and release qualification.

New admission inventories freeze witnessRetentionVersion: 1. Their completed receipts must retain the matching certificate inventory; omitting it or rehashing a result without it is rejected. Older inventories remain readable as historical evidence and are not upgraded to protected-witness certification.