simulation

PAIDF (Physical-AI Data Factory) Simulation. One skill, two tracks selected by a Stage-0 router. All per-track stage detail is in references/<track>/.

user prompt
    |
    v
Stage 0 — Router
    |
    +── single-flow ──► references/single-flow/stages.md
    |     flows: good / good_fixed / defect / missing / lighting / paired
    |
    +── roi ─────────► references/roi/stages.md
          variants: Day-0 (synth only) / Day-1 (synth + real + MI)

Instructions

Apply the Stage-0 router to the user prompt first, then hand off to the matching reference doc under references/<track>/. The router emits a small JSON intent record; downstream stages parse the user yaml(s), surface a summary, gate on approval, and emit docker run commands. The skill never executes docker on its own — it only emits the commands and waits for the user.

Stage 0 — Router

Apply these rules to the user's prompt in order, first match wins. Rule 1 is the safety-critical gate that prevents silent over-delegation to the ROI track (the failure mode that reverted commit 18cd8d5).

1. ROI track gate — explicit keywords only. Trigger if the prompt contains any of:
   • cad2roi, usd2roi, usd-to-real, usd to real
   • ROI (case-insensitive; standalone token), per-ROI, per-component crop
   • MI registration, mutual information, align to real photo, register to <photo>
   • bridge crops, bridge between pads / solders
   • anchor on <component class>, crop <class> ROIs
   • scan_grid plus per-cell crop, pure-synthetic ROI
   • mesh-level semantic rules → track = roi. Then sub-route by photo:
     • prompt names a real photo (--photo-path, real.png, "AOI screenshot", "PHOTO_DIR=...") → Day-1
     • otherwise → Day-0
     • ambiguous → ASK "Do you have a captured real PCB photo to align against, or only the CAD-derived USD?" → Hand off to references/roi/stages.md.
   
   NOT triggers for the ROI track (these stay single-flow):
   • A bare photo path mentioned without any crop / register / align verb (e.g. "generate 9 good images for /tmp/photo.png").
   • A component class mentioned without a crop verb (e.g. "render images of capacitors"). The user must say what to DO with the photo or class. This is the 74eef40 rule, preserved verbatim.
2. Paired / ChangeNet — prompt contains paired, golden+defect, golden / defect, ChangeNet, siamese, pair dataset → track = single-flow, sub-mode = paired. Run the good flow + the defect flow with the same random_seed; post-process via build_pair_dataset.py. Hand off to references/single-flow/stages.md (Stage 1 Rule 1).
3. Defect intent — prompt contains defect, pose defect, shift, tombstone, sideflip, polarity, reverse polarity, missing, hidden component, two-pass → track = single-flow, flow = defect (or missing for the missing/hidden keywords).
4. Zoom / close-up / fixed-camera — prompt contains zoom, close-up, single <component>, fixed camera, solder fillet randomization, tin pad, vantablack → track = single-flow, flow = good_fixed.
5. Lighting demo / variant (explicit demo intent) — prompt contains lighting demo, lighting variant plus a named variant (ring light, dome light, scene lights, preserve color) → track = single-flow, flow = lighting.
6. Default catch-all — good, clean, defect-free, no defects, or anything else (including bare generate / render / make images) → track = single-flow, flow = good.
7. Ambiguous — if no rule matches confidently OR Rule 1 / 5 needs disambiguation (variant unclear, lighting variant unclear), ASK ONE question before any further work. Do NOT emit docker commands or write YAML in the disambiguation turn.

Lighting parameter is orthogonal to the flow in the single-flow track — if the user names a lighting mode anywhere in the prompt, apply it as an override on top of the chosen flow's base config. Default when nothing is named: scene-authored lights. Full table in references/single-flow/stages.md §Stage 1.

Normalized router output

The router emits this record before handing off:

{
  "track":   "single-flow | roi",
  "flow":    "good | good_fixed | defect | missing | lighting | paired",   // single-flow only
  "variant": "day0 | day1",                                                // roi only
  "raw_intent": { "...": "rest of the canonical intent schema for the chosen track" }
}

Then load references/<track>/stages.md and execute the rest of the pipeline from Stage 1 onward.

Track summary

Single-flow crop.enabled defaults: true for good / defect / missing (crops are usually the training target); false for good_fixed (camera is already on one component) and lighting (full-board visual QA). Override via prompt: "no crop" / "with crop". Full knob list in references/single-flow/overrides.md §crop block.

Both tracks share the same Docker image (paidf-simulation:<tag>), the same pcba_target.yaml shape, the same <usd-assets>/ mount convention, and the same approval-gate discipline.

Examples

References