sim_generate
Write a Petri-net model from a plain-language description with the configured LLM, with the levers that moved the pass rate in the benchmark behind it, each optional. The model is checked as it arrives (decode, validator, then the builder's behavioural checks: disconnected parts, transitions that never fire, places nothing fills, and any place_ids you named); what is wrong is shown back to it and it answers again, up to retries times. A net that decodes and passes the validator is stored as a NEW model you own, exactly as sim_code_to_flow stores one, even when behavioural problems remain; clean says whether any do and problems lists them. notes lists what is right for some nets and wrong for others, and is never fed back: a state where nothing can fire, a place that grows without limit. A net that cannot be stored is returned as model with the reasons. Levers: retries (0-4, default 2) re-asks with the findings and is the strongest; best_of (1-5, default 1) starts again from scratch when a sample ends with problems and stops at the first clean one; selfcheck asks the model to play its own net forward and replays that against the net (it catches inconsistency, not a wrong idea); example is "real" (default, a worked order workflow) or "placeholder" (the older schema with placeholder names, which small models copy literally); model picks one of the deployment's allowed models; thinking caps the model's hidden reasoning tokens (0 turns it off); temperature is 0-1. Cost: every model request is one call and one request makes at most 6 (best_of x (1 + retries), plus one for selfcheck); a request over that is refused before it runs. The calls it really makes are counted against your daily LLM allowance, the same one the builder uses, and the response reports calls, tokens (thinking tokens included) and the allowance left. Needs sign-in; refused when this deployment has no LLM provider configured.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| name | No | name for the model; filled in when the LLM leaves it out | |
| model | No | one of the models this deployment allows; omit for its default (the error for an unknown name lists them) | |
| best_of | No | independent samples, tried until one is clean, 1-5 (default 1) | |
| example | No | "real" (default) or "placeholder" | |
| retries | No | feedback rounds after an answer with problems, 0-4 (default 2) | |
| thinking | No | cap on hidden reasoning tokens, 0-24576; 0 turns thinking off; omit for the model default | |
| place_ids | No | place ids the net must contain; a missing one is fed back like any other problem | |
| selfcheck | No | one extra call: the model plays its net forward and the server replays it (default false) | |
| temperature | No | 0-1; omit for 0.2 | |
| requirements | Yes | what the model should represent, in plain language: the process, its resources and their sizes |