A causal simulation workbench with bounded execution coverage

The workbench is intended for exploratory wargaming and mechanism discovery. An analyst describes a bounded sociotechnical situation, reviews the compiled configuration, runs autonomous and non-autonomous systems in a shared world, and inspects how local interactions accumulate into system-level outcomes. The long-term target is a general-purpose simulator; the deployed product executes only behavior covered by its reviewed templates and registered mechanisms.

A run supports a conditional claim about an implemented model: under these assumptions, components, information paths, and transition rules, this trajectory occurred. It does not predict how real people or institutions will behave, estimate an intervention effect, or validate a social theory by itself.

The ontology is a set of separations—not a list of agent types

The simulator does not begin by declaring everything an agent. It separates what exists, what can act, what can be attempted, what is permitted, what succeeds, what information is represented, and what an analyst later infers. This prevents narrative convenience from silently becoming world truth.

One typed world—not one narrative per agent

The current runtime stores one strict CausalState. The record collections below are the actual canonical ontology. The graph is a later projection of them; it is not the primary data model.

A car illustrates the distinction. It may be an entity with state, placement, and mechanisms through which a person can operate or move it. It gets its own active-system binding only when autonomous onboard behavior matters to the modeled question.

Configured structure, spatial topology, and causal history use different relation vocabularies

The ontology is not fundamentally a generic property graph. Several typed vocabularies are projected into graph form for inspection.

Agency is an implementation binding layered onto a persistent entity

ActiveSystemSpec binds a scheduled process to an entity. It declares which observation ports the process may perceive, which output ports it may use, its initial private state, optional next activation, interface-withdrawal conditions, and limits on observations, actions, state size, and model spend.

Position expectations, reporting relationships, and authorizations are institutional “oughts.” They are not the person's own values, proof of physical possibility, or commands to choose a result. Relationships, loyalties, distrust, and interpretations may remain natural-language memories rather than privileged numeric trust variables.

Information has a carrier, representation, provenance, route, and recipient-specific observation

The simulator keeps world truth separate from what any actor knows. Copying information creates a new representation with declared parent lineage; it does not teleport one abstract message between minds.

A configured connection is only a directed possibility. A retained effect_routed → state_committed → observation_delivered chain is evidence that a representation reached an actor. It is not evidence that the content was true or persuasive.

An adjudicator may inspect hidden state when its authority permits it. Actor-visible text must still be projected from committed actor-authorized observations—not copied from the adjudicator's explanation. The implementation can guarantee access-control separation more readily than it can prevent an actor from making a plausible inference from permitted evidence.

Capability, competency, affordance, authorization, attempt, and outcome remain separate

The intended general seam does not require a pre-authored verb such as crop_screenshot_and_email_bob. An actor may propose an open semantic intent. A suitable exact, stochastic, LLM, external, scripted, or replay authority interprets it within a declared patch grammar and read/write scope.

The current runtime preserves several of these distinctions through ports, mechanism authority, and validation, but a general capability–affordance resolver is not yet a complete production subsystem.

No actor or narrator directly edits canonical truth

An active system emits an ActionIntent. The current runtime materializes an ActionAttempt, emits and routes an EffectEnvelope, invokes a registered mechanism, validates the proposed change, and commits a StatePatch or a visible rejection.

Every transition authority declares what it may read and write, its assumptions and fidelity, provenance, and invalid questions. An LLM game master is a legitimate coarse-grained authority, not merely a fallback. It is not an omnipotent state editor: it proposes a structured patch that passes through the same generic and subsystem-declared validation boundary.

Configured possibility and realized history are retained separately

New general-world runs use stock Concordia for same-moment actor activation and checkpoint invocation, with a project-owned canonical-world component validating and committing typed transactions. Historical workflow profiles still use ActiveRuntimeSession and CausalSession during migration; they remain compatibility paths, not a hidden scheduler inside the new general runner.

Each event retains logical time, state revision, declared execution parents, action/effect/mechanism/observation identifiers, variance source, invariant results, and any committed patch. Checkpoints preserve canonical state and private active-system state so a run can be resumed or forked. Replay reconstructs from retained state and event evidence; it does not regenerate a model trajectory.

Same-moment participants currently receive inputs derived from one frozen pre-moment checkpoint and may generate proposals concurrently, but accepted proposals are committed in canonical order. General joint conflict resolution and atomic multi-proposal commit are not established. Scenarios whose result depends on simultaneous contention require an explicit allocation mechanism or must remain unsupported until the runtime can perform snapshot → independent proposals → conflict resolution → atomic commit.

Organizations and institutions are representations of organized dependencies—not privileged atoms

At high resolution, an organization can be represented as a sociotechnical network of people, artifacts, software, resources, routines, communication routes, authority claims, sanctions, and technical mechanisms. An AnalyticalBoundary can group these records for inspection without becoming another executor.

At coarse resolution, the same subsystem may be represented by one declared active surrogate when its internal machinery is irrelevant to the research question. Organizationhood itself grants no agency. A coarse surrogate may not independently produce the same causal behavior as a simultaneously active detailed representation unless a coupling contract divides responsibility.

Institutions may then be inferred from stable patterns: which records count as authoritative, who can block resource movement, which commitments are enforced, which routes become obligatory, and which intermediaries become relied upon. The current runtime supports authored flat analytical boundaries; dynamic, overlapping, contested, or inferred institutions remain future work.

Natural language proposes semantic configuration; reviewed machinery remains the executable authority

The analyst describes a world in ordinary language. The authoring model returns a typed, editable V2 bundle whose scenario, run controls, and optional analyses have separate identities and authority. New general-world runs receive only scenario and run contracts; analytical lenses can be attached afterward. Scenario labels such as “election,” “outbreak,” or “port crisis” are analyst descriptions—not fundamental ontology types.

The V2 general-world contract can compose people, world records, active systems, spatial links, information representations, resources, relationships, schedules, and reviewed exact or coarse transition requests without generating Python. Historical template-bound authoring families remain for earlier workflows. Open prose still cannot create arbitrary executable semantics: behavior needs a registered exact authority, a declared coarse LLM authority, or a visible descriptive/unsupported classification.

Coordination and influence are interpretations of trajectories—not hidden causes

Waltzman's From Minds to Coordination motivates attention to how local interactions reshape perceived risks, reliance, dependencies, and collective coordination readiness. Heterogeneous adaptive pressure is especially important in the paper's threat framing, but the substrate can also represent homogeneous broadcasts, segmented campaigns, ordinary disruption, legitimate disagreement, and defensive response.

The irreducible world does not contain magic global variables named “institution,” “fragmented trust,” or “blocked coordination.” A Waltzman analysis should derive directional findings from lower-level evidence: which sources and procedures actors rely on, which risks they name, which prerequisites appear, where dependencies conflict, how positions change, and whether locally reasonable actions remain jointly compatible.

The current V2 Waltzman lens derives evidence-linked trajectories for information exposure, source-reliance proxies, expressed concern domains, shared action dependencies, and coordination-readiness signals. It uses deterministic calculations over retained typed records; it neither calls another model nor changes the simulation. These are transparent operationalizations, not calibrated measures of real trust, risk, or readiness. A single execution can demonstrate that a mechanism is possible inside the implemented world. Detecting an invariant requires comparable trajectories and aligned directional change across heterogeneous interactions; it cannot be established by labeling one cross-sectional snapshot.

A simulator becomes an experimental instrument only through a declared comparison

Before interpreting repeated executions, the analyst must state the question, factors varied, factors held constant, outcome measures, comparison rule, and what would count as an independent replicate. A run, a checkpoint fork, a model call, and an agent decision are different evidentiary units.

Mechanism demonstrations ask whether a process can occur under implemented assumptions. Comparative experiments ask whether outcomes differ across controlled conditions. Robustness studies ask whether a finding persists across plausible specifications, seeds, models, or world assumptions. None of those licenses a claim about people or institutions without separate empirical grounding.

New general simulations run on Concordia; historical workflows remain on the prior runtime during migration

The natural-language general-world path now compiles semantic proposals into a canonical world hosted by stock Concordia. Concordia owns simultaneous actor activation and checkpoint invocation; project components own authorized context, typed transaction validation, atomic world commits, and evidence. Historical scenario workflows remain readable and executable on the earlier runtime until they are deliberately migrated.

What the current system does not establish

This public paper is a current synthesis, not a replacement authority. When it conflicts with current source or an accepted ADR, the source contract and ADR govern.