simulate

name	simulate
description	Run a Monte Carlo simulation study to evaluate finite-sample properties (bias, RMSE, coverage, size, power) of a statistical estimator. Use this skill whenever the user asks to "run a simulation", "Monte Carlo", "simulate a DGP", "check coverage", "finite-sample properties", "calibration study", or invokes $simulate explicitly. Handles both Workflow 11 (new estimator + simulation evidence) and Workflow 12 (simulation-only on an existing estimator). Enforces three-way pipeline isolation — code, test, and simulation agents never see each other's specs.
metadata	{"short-description":"Monte Carlo simulation study — workflow 11/12"}

Skill: Simulate — Monte Carlo simulation study

You are the StatsClaw leader. The user has triggered $simulate (or described a simulation-shaped task in natural language).

The authoritative procedure lives in skills/simulation-study/SKILL.md — load that file and follow it exactly. This skill is the entry point that routes to the right workflow.

Workflow selection

Determine the workflow from the user's intent:

• Workflow 11 (Simulation + Code) — the user is introducing a new estimator AND wants simulation evidence.

  • Dispatch order: planner → [builder ∥ simulator] → tester → scriber → [distiller]? → reviewer → shipper?
  • Builder and simulator dispatch in parallel (same message from leader).
  • Builder gets only spec.md; simulator gets only sim-spec.md.
  • After both complete and their worktrees merge back, dispatch tester with only test-spec.md. Tester validates unit tests AND runs the full simulation, comparing results against acceptance criteria.

• Workflow 12 (Simulation Only) — the estimator already exists, the user wants a Monte Carlo study on it.

  • Dispatch order: planner → simulator → tester → scriber → [distiller]? → reviewer → shipper?
  • No builder.

Three-way pipeline isolation (absolute)

• Builder never sees test-spec.md or sim-spec.md.
• Tester never sees spec.md or sim-spec.md.
• Simulator never sees spec.md or test-spec.md.
• Planner bridges all three — only planner holds the full comprehension.

See skills/isolation/SKILL.md for the worktree mechanics.

Tolerance integrity is absolute

Tester MUST NEVER relax tolerances or acceptance criteria to make a failing simulation pass. The only valid response to a genuine failure is BLOCK. Reviewer cross-audits every tolerance in audit.md against test-spec.md and sim-spec.md.

If the user's request is ambiguous

If it's unclear whether the user wants Workflow 11 or 12 (e.g., "simulate this estimator" when the estimator may or may not already exist), ask a numbered-options question before dispatching planner.