rd-author

rd-author — compile the brief into a model document

The brief's slots table is your input; a validate-clean .rdj.json per arm is your output. A model is data — one JSON document, no Julia, closed node vocabulary. That is not a stylistic preference: the document is what makes the model reviewable, diffable, replayable, and safe to accept from an agent.

Read agent/reference/model_schema.md now. It is generated from the engine, so it is the authority on every key, column, operator and verb. This skill is the judgement layer on top of it; when the two disagree, the generated reference wins.

The shape

{
  "rd_format": "reactive-dynamics-model", "version": 1.0,
  "meta":        { "tspan": 48.0, "dt": 1.0 },
  "params":      [ { "name": "hit_rate", "value": 0.55 } ],
  "places":      [ { "name": "scientist", "init": 4.0 } ],
  "transitions": [ { "id": "trial", "rate": 1.5, "cycletime": 4.0 } ],
  "arcs":        [ { "transition": "trial", "place": "scientist", "side": "lhs" } ],
  "population":  [ ]
}

meta.tspan is required. transitions[].id is the FK every arc points at. Arcs are a flat list, not nested under transitions — which is what lets you diff one arc without touching anything else.

Work in this order

1. Places — one per slot in the brief's resource and outcome rows. init, and cost/reward where value flows.
2. Transitions — one per process. rate, cycletime, prob_of_success, priority.
3. Arcs — the wiring, with modality. This is where the modeling lives and where the errors are.
4. Population — starting tokens, if the study has structured items.
5. Validate, repair, repeat until zero diagnostics.
6. Arms — copy the baseline, change the lever, nothing else.

Modality: the axis that decides whether your model is right

Every lhs arc carries three independent axes. Omitting modality means consumed.

{ "transition": "trial", "place": "scientist", "side": "lhs",
  "modality": { "allocation": "upfront", "return": "conserved", "blocking": "block" } }