Generate and Optimize

This workflow composes two research packages through public OptPilot contracts:

1. DEVS-Gen turns a natural-language specification into a runnable discrete-event simulator with declared metrics and, when requested, an editable decision-policy hook.
2. LLM-Guided Heuristic Design consumes that hook, repeated-simulation metrics, and exact-replay traces to improve the executable policy.

Neither package imports the other. The generated Environment describes what a Candidate may edit and how it can be replayed; the Method declares which pieces of that contract it requires. Studio enables the pairing only when they match.

Which route to use

Use LLM-Guided Heuristic Design when the decision is most naturally an executable rule that can improve through repeated simulation and trace review. Use COOPA when the decision is most naturally a mathematical program with explicit variables, constraints, and an objective. A project can use both: COOPA can propose a plan, while a DEVS-Gen Environment evaluates that plan under operational uncertainty.

For the solver route, open the solve-or-problem Run setup and fill its required problem input. A usable statement names what to decide, what to optimize, and every constraint, for example:

Minimise the total distance travelled by two AGVs. Decide which AGV serves
each transport request and in what order. Each request must be served once;
an AGV may carry only one load at a time; pickup must precede delivery; and
the schedule must satisfy all battery and shift-length constraints.

1. Generate the simulator

Open DEVS-Gen in Catalog and launch DEVS Simulation Generator Interface, or run its headless generate Resource action. Describe both the system and the decision that should be optimizable—for example:

Jobs arrive at one machine. Whenever the machine becomes idle, a dispatch policy chooses one waiting job. Expose that decision as an editable policy.

The generated devs.simulation.v2 manifest declares metrics and the policy entrypoint. A system without an optimizable decision remains a valid simulation Environment but will not match policy-design Methods.

2. Register the Environment

From the generated Workspace, choose Set up for Catalog. For a declared policy hook, Studio creates:

• a file-Candidate Environment;
• baseline candidate files and policy instructions;
• policyValidation rules;
• seeded evaluator settings and metrics;
• an exact_seed_replay capability that produces a bounded SQLite trace.

DEVS-Gen includes two reviewable reference outputs:

Environment	Policy-hook style	Location
Dispatch station	policy.py factory function	catalog/devs_gallery/environments/dispatch_station/
Triage clinic	Editable DEVS component class	catalog/devs_gallery/environments/triage_clinic/

3. Pair the Method

In Studio's Run setup flow:

1. Select the generated Environment—or one of the reference Environments.
2. Select Trace-guided policy design (language model) from the LLM-Guided Heuristic Design package.
3. Review the compatibility checks, objective, budget, and seed.
4. Save the Run setup and launch it.

The Method evaluates the baseline, replays the worst seed, lets a manager query the event trace, asks parallel editors for complete policy revisions, validates them, and keeps only improvements. The source Environment and prior results are never modified.

4. Inspect and reuse the result

The Run retains every Candidate, trial, observation, trace, and generated policy file. Open the best Candidate to replay it, compare it, save it to the Shortlist, or create an editable Workspace.

The same boundary supports other combinations. A COOPA-produced formulation or solver policy can be retained as a Candidate and evaluated by a compatible DEVS-Gen Environment; translating a mathematical solution into a simulator decision hook remains an explicit modeling step rather than hidden glue.

See LLM-Guided Heuristic Design for the Method contract, DEVS-Gen for generation, and COOPA for provenance-aware OR formulation and solver routing.