Skip to content

rt-execution-bridge

Runs and audits the empirical analysis rather than just advising — maps a DiD/IV/RDD/synthetic-control/DML design to the concrete StatsPAI / Stata MCP tools in the environment and reports the fitted, audited number.

Category: analysis
Field: economics
License: MIT
Updated: 2026-07
Stages: data-analysis · code-generation

Execution Bridge (rt-execution-bridge)

Close the last mile: turn "you should use a heterogeneity-robust DiD / weak-IV-robust CI / multiple-testing correction" into an actual fitted, audited estimate. Full map + orchestration spine + validated worked-examples (DiD / IV / RDD / synthetic-control / DML): shared-resources/empirical-methods/execution-with-mcp.md.

When to trigger

  • You have data and a design and need the actual estimate + its diagnostics.
  • A reviewer objection (real or simulated) needs an executed answer (e.g. "TWFE is biased under staggered adoption" → run Callaway–Sant'Anna + Goodman-Bacon).

The spine (always)

  1. detect_design → preflight / recommend → fit with as_handle=true.
  2. audit_result(result_id) — enumerate the checks the design still owes; run each suggest_function it names.
  3. Design-specific sensitivity from the handle (honest_did_from_result, sensitivity_from_result, evalue_from_result).
  4. bibtex(keys=[…]) for citations — never invent references.

Design → tools (summary; full table in the canonical doc)

  • Staggered DiD: callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result.
  • IV: iv + effective_f_test + anderson_rubin_ci (weak-IV-robust).
  • RDD: rdrobust + rddensity / mccrary_test.
  • Synthetic control: synth / sdid + placebo inference.
  • DML / high-dim: dml + dml_diagnostics (overlap) + oster / sensemakr.
  • Multiple testing / inference: romano_wolf, wild_cluster_bootstrap, twoway_cluster.
  • Exhibits: etable / did_summary_to_latex straight from the handle.

Hard rules

  1. Run, don't claim — every reported estimate/CI/F/bound traces to a tool call.
  2. bibtex is the only citation source.
  3. Method here, placement there — where the result goes (body vs. appendix, page limit, house table style) is the target pack's skills + official-source-map.md.
  4. Degrade honestly — if StatsPAI/Stata are not connected, adapt the code/ skeleton and flag any unverified number.

Output format

Text Only
【Design】DiD / IV / RDD / SCM / DML / …
【Estimate】point [CI] (estimator)
【Key diagnostic】(first-stage/effective F, pre-trends p, overlap, placebo p, …)
【Audit gaps run】…
【Magnitude】interpretable units
【Next】rt-submission-readiness / the pack's tables-figures skill