<!-- Generated from .claude/skills/replicate/SKILL.md by tools/sync_agents.py. Edit the source, then run the script. -->

/replicate — Replicate

Category: Validation & QA · Primary agent: replication-auditor

Replicates results from the replication package on a clean environment and traces published numbers to code and data.

Arguments: whatever the user wrote after the command name.

Purpose

Replicate published or reported results from the replication package in a clean environment, compare outputs within tolerance, and trace every table and chart number to the script and data version that produced it. The output is a replication report with a discrepancy log and root causes.

Use it when results are about to be published or used in advice, after model changes, or when reported numbers are questioned.

Do not use it when the replication package does not exist yet; build it first with the reproducibility engineer.

Before you start

1. Check required inputs. For each input below record FOUND or ABSENT (never 'assumed'):
   • Replication package
   • Published outputs If a required input is ABSENT, that part is NOT ASSESSED — NO DATA: do not estimate or invent it; name what is missing and which skill produces it. If all required inputs are absent, stop and say so.
2. Check confidentiality. Identify the data classification level of the inputs (.claude/docs/confidentiality-and-ai-use.md). Do not proceed with restricted or confidential data in a way the policy does not allow.
3. Collaboration. Follow .claude/docs/collaboration-protocol.md: ask questions first, present options, get approval before writing files, and update production/session-state/active.md after each approved section.

Agents

• replication-auditor — primary agent; coordinates the work and integrates results
• reproducibility-engineer — Guarantees reproducibility

Each agent is defined in .claude/agents/<name>.md. If your tool runs subagents, run these agents as subagents, in parallel where the steps are independent; otherwise read each agent's file and apply its instructions in turn. Subagents only write to the paths named in your instructions.

Key design decisions (present options, let the user choose)

Decision	Typical options	What it changes
Environment	container, fresh virtual environment, clean machine	Independence
Tolerance	exact, numerical tolerance by output	Pass criteria
Scope	all outputs, key tables	Effort

Workflow

1. Build the clean environment from lock files or containers
2. Run the one-command replication and compare outputs with tolerance
   • Record runtime and any manual steps needed (each is a finding)
3. Trace each table/chart number to its script and data version
4. Log discrepancies and root causes

Checkpoints with the user

• After scope: confirm outputs and tolerances
• After the run: review discrepancies before root-cause work

Common pitfalls

• Running on the author's machine
• Accepting "rounding" without checking
• Ignoring numbers quoted only in the text

Output

validation/<model>/replication-report.md

The report documents the environment, run results, the number-tracing map and the discrepancy log with root causes.

Gate / verdict

• Outputs reproduced within tolerance, or discrepancies logged with cause
• Manual steps needed to replicate are reported as findings

Suggested next steps

• /validate-model
• /model-release