mg-setup

/mg-setup

The setup workflow — the four-step sequence for building a MadGraph configuration. The orchestration disciplines it leans on are already in your context, not restated here: dual-spec, regime classification, slice-boundary premises, whole-spec reconciliation, revise-before-caveat, and "a clean run is not evidence" all live in lead-discipline.md. This skill is the setup-specific scaffold over them.

Step 1 — Frame the physics-spec. Capture the user's prompt verbatim (dual-spec discipline). If too vague to act, surface and ask.

Step 2 — Classify the regime. Classify the physics regime (regime-classification discipline) and route to the slices each regime implicates — sub-threshold parent → ma-bw-window-consultant; NLO bracket syntax → ma-nlo-syntax-consultant; EFT operators → ma-eft-consultant; and so on.

Step 3 — Build the candidate. Translate physics-spec + regime classification into a candidate MadGraph configuration. Dispatch the slices the regime named plus any others your reading requires; consultants author within slice, you conduct, iterate as returns reveal what's next.

The simulation-spec is what the configured artifacts do, not what you meant them to do, and this is exactly where the two silently diverge. A model name does not fix its vertices, a restriction card silently prunes couplings, a chain drops a sub-decay the parentheses never bound — each artifact is opaque until its owning slice reads it. So build every load-bearing choice on what the artifact actually carries, read by its owner, in both directions: a contribution the physics needs can be silently absent (a dead process, a removed coupling), and one you never intended silently present (a contaminating topology, an extra coupling). On a don't-run setup both are invisible until something is read.

The build is not done until each load-bearing choice is grounded in what its artifact does — never its name — and the candidate is coherent across stages.

Step 4 — Reconcile and present. Compare the assembled simulation-spec against the physics-spec, then ship.

• Per-requirement. Verify each physics-spec requirement is satisfied by the assembled simulation-spec, given everything else now in it. A DIRECT-cited value carries its own evidence; re-dispatch the owning slice only to Establish a load-bearing value that shipped non-DIRECT.
• Whole-spec. An assembly of individually-correct slice outputs can still be globally wrong — a dead process, a form that computes a different observable, a contaminating topology, a silently dropped coupling. Confirm the emergent cross-slice invariants with a lens that did not build them:
  • Trivial invariants (BR ≤ 1 arithmetic, a count, a sign) you check yourself — no dispatch.
  • Runtime realization — invoke /mg-probe on the assembled commands with a target-derived expectation list (per /mg-probe), drawn from Step 2's regime classification; re-dispatch ma-physics-consultant for the target-state only if that classification is not specific enough. The probe is the independent lens — it acts (reads the generated artifacts), where the build did not; its card carries the cheap-vs-launch discipline.
• Reconcile. A deviation from a target-derived expectation is a finding: revise via the owning slice (revise-before-caveat). A revision does not close the finding — re-verify a substantive fix (per Re-verify a fix in lead-discipline): re-probe once against the revised artifact, the expectation list widened to the fix's blast radius (the deviated expectation plus the ones that passed before the fix); a trivial source-cited fix clears by inspection. Risk gate: if after this a load-bearing part of the spec remains unestablished — a finding that will not settle, a substantive fix the re-probe could not clear, a non-DIRECT load-bearing claim you could not establish, or an invariant the source cannot settle — do not bury it in a caveat: surface it prominently and recommend /mg-deep-verify, leaving the call to the user. Do not loop here, and do not auto-invoke the cascade. The gate is tight — a clean build settles and never trips it.