Propose Comparison

Skill: propose-comparison

Create a comparison: a small, branch-scoped analysis unit realized as a named DataLad branch plus a project-log entry. Comparisons span a rigor spectrum: an exploratory quick query (default) that lives only as a branch/run, or a confirmatory one whose spec is frozen and registered before execution via govern/preregister. The record here is already the spec-centric research object of STAMPED §3.12.1.

When to use

• The user wants to run any analysis/plot/comparison against the dataset.
• Do NOT use to initialize a project (project/new-project) or to execute an already-proposed

comparison (analyze/run-comparison).

Steps

1. Capture the comparison — get from the user (short, one-liner each):

• what — the comparison in a sentence ("group difference in outcome Y")
• why — the question it answers (optional but encouraged)
• expected inputs and outputs (best guess; refined at run time)
• rigor: exploratory (default). If the user wants confirmatory, create the branch (steps

2–4) and then route to govern/preregister to freeze + register the spec before the analysis is run — do not execute it as confirmatory until the spec is frozen.

1. Choose a clear branch name — cmp/ (e.g. cmp/group-diff-y). The user owns

naming conventions; suggest one and confirm. Clear, stable names are how comparisons stay navigable (the harness does not track branches for you beyond the log).

1. Create the branch — delegate to the datalad doer:

> "Create and switch to git branch cmp/ in this dataset (leave the working tree clean)."

1. Log it — append one entry to project.yaml:

{ ts, op: propose-comparison, stage: analyze, note: "cmp: ()", branch: cmp/ }.

1. Report — the branch name and that the next step is writing the analysis script, then

analyze/run-comparison to execute it with provenance.

Constraints

• One comparison = one branch. Do not stack unrelated comparisons on the same branch.
• Keep project.yaml append-only.
• Do NOT choose the analysis, model, or figures for the user — the harness scaffolds the edges of

the work (branch, provenance, log), not the scientific question itself. Writing the analysis script is the user's job; run-comparison wraps whatever they wrote.

• Delegate the branch operation to the datalad doer; never call git/datalad directly.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: neuroanalytics
• Source: neuroanalytics/data-science-harness
• License: MIT

Install and usage instructions live in the source repository linked above.