Alvance Company Simulator

Alvance Company Simulator is a deterministic, resettable enterprise digital world for model and agent interaction. It simulates state, permissions, policies, business side effects, and audit history; it does not select a model, plan tasks, or contain an autonomous agent loop.

Any client can interact through HTTP, MCP, the Python API, or the CLI.

What it simulates

The built-in incident scenario composes seven optional components:

Component	Application surfaces	Current fidelity
identity	organization, employees, RBAC	stateful roles and capabilities
collaboration	documents, chat	classified reads and auditable writes
work_management	projects, issues	assignments, labels, workflow state
source_control	repository	versioned files and commits
service_management	ITSM, approvals, policy	causal change and incident gates
customer	CRM	accounts, restricted contacts, notes, tags
analytics	warehouse	column-restricted read-only SQL telemetry

Scenarios load only the components they need. Component dependencies are resolved automatically and tables for unloaded components are not created.

Architecture

Any model, agent, or program
      │
      ├── HTTP JSON
      ├── MCP stdio
      ├── CLI
      └── Python API
              │
Composable world kernel
      ├── scenario manifest + component registry
      ├── SQLite state + deterministic logical clock
      ├── RBAC, policy gates, and side effects
      └── optional read/write audit trace
              │
Optional evaluation layer
      ├── state / trace / summary artifact bundle
      ├── one-way handoff to a trusted verifier
      ├── reward.json + CTRF diagnostics
      └── Harbor with Docker or E2B

Quick start

python3 scripts/company_sim.py --db runtime/company.db reset \
  --scenario scenarios/incident-response-v1

python3 scripts/company_sim.py --db runtime/company.db components

python3 scripts/company_sim.py --db runtime/company.db search \
  --actor EMP-ALICE timeout

Load the small collaboration-only world:

python3 scripts/company_sim.py --db runtime/minimal.db reset \
  --scenario scenarios/minimal-collaboration-v1

The scenario manifest chooses components, durability, read auditing, request budgets, and component-specific limits. See docs/composable-worlds.md.

Model-neutral interfaces

• HTTP: python3 -m company_simulation.server --db runtime/company.db
• MCP: python3 scripts/mcp_server.py
• MCP example config: examples/mcp/stdio.json
• CLI: python3 scripts/company_sim.py --help
• Python: CompanyWorld.reset(...), then search, read, warehouse_query, and action

MCP is an interface protocol, not a dependency on a particular model vendor.

Artifact and verification contract

python3 scripts/company_sim.py --db runtime/company.db export-artifact \
  runtime/submission --replace

python3 scripts/company_sim.py verify-artifact runtime/submission

The bundle contains:

submission/
├── manifest.json
├── checksums.json
└── payload/
    ├── world.db
    ├── events.jsonl
    └── summary.json

The optional Harbor runner kills remaining agent processes, captures this bundle as root, and uploads hidden tests only afterward. Verifiers emit reward.json, reward.txt, ctrf.json, and a domain report.

Evaluation example

python3 scripts/harbor_secure.py run \
  -p harbor/tasks/incident-response-cross-system -a oracle -e docker

For E2B:

export E2B_API_KEY="..."
python3 scripts/harbor_secure.py run \
  -p harbor/tasks/incident-response-cross-system -a oracle -e e2b --force-build

Harbor and E2B are optional evaluation infrastructure, not part of the world kernel. The secure remote path has been verified with Oracle 1.000 and nop 0.000.

Design and limits

• docs/simulation-audit.md: what is and is not simulated
• docs/composable-worlds.md: component and scenario contract
• docs/performance.md: storage, latency, and scale limits
• docs/evaluation-contract.md: artifact and reward protocol
• docs/open-source-landscape.md: projects reviewed and adoption decisions
• docs/threat-model.md: trust boundaries and residual risks

License

MIT