/mstack:preregister

Stage: design · Voice: preregistration-clerk

Before fielding a survey, before analyzing a sample you have not touched, or whenever confirmatory must be credibly separable from exploratory. Expects /mstack:hypothesis-design, /mstack:identification-review, and /mstack:power-analysis to have run, plus /mstack:survey-build if survey-based. The bar throughout: could two researchers, given only this document and the data, produce the same primary table without talking?

Procedure

1. Load .mstack/hypotheses.md (estimands, SESOIs, decision rules), .mstack/identification-review-*.md (the tests to pre-commit), .mstack/theory.md (rivals), .mstack/power-analysis.md, .mstack/research-question.md, .mstack/lit-map.md; .mstack/survey-design.md if a survey is involved; paper/sections/theory.tex and methods.tex if drafted.
2. Read ${CLAUDE_PLUGIN_ROOT}/references/preregistration-guide.md for the specificity standard and the vague-to-specific examples, and ${CLAUDE_PLUGIN_ROOT}/references/robustness-protocols.md for what to pre-commit.
3. Registry and data status. OSF by default (the Secondary Data template when the data already exist); AsPredicted for a simple experiment; EGAP for a field experiment in political science; the AEA registry if an economics audience expects it. Record in the header what has already been seen of the data (nothing; descriptives only; an exploratory split), because a plan that hides prior looks is worse than none.
4. Draft prereg/osf-prereg.md from ${CLAUDE_PLUGIN_ROOT}/skills/preregister/assets/prereg-template.md. All nine sections are required:
   1. Hypotheses — each with direction, estimand, SESOI, and the decision rule (support / evidence against a meaningful effect via the equivalence region / inconclusive); one primary, the rest explicitly secondary; the rival-discriminating hypotheses from hypotheses.md included.
   2. Sample — population, frame, provider or dataset with version, target N justified by /mstack:power-analysis, stopping rule with no interim looks, expected attrition and its handling.
   3. Exclusions — every rule with a threshold decided before data; from the probe manifest if survey-based; how excluded cases are reported.
   4. Measures — exact items and stimulus text attached; scoring, reverse-coding, index construction fixed now; the covariate list with each justified as pre-treatment; the exhaustive moderator list.
   5. Primary analysis — the equation with variable names as they will appear in code, the estimator for the design (a heterogeneity-robust DiD estimator under staggered adoption, rdrobust for RD, and so on), sample, standard-error type and cluster level, test and alpha, software with versions.
   6. Secondary analyses — each with its estimand; the multiplicity policy for the family.
   7. Robustness — the checks committed to in advance, including the sensitivity analysis named in the identification review.
   8. Deviations policy and SOP — deviations reported alongside the original plan; standard operating procedures for unanticipated problems adopted and linked (Lin & Green 2016).
   9. Data and code availability — repository, timing, restrictions, and the blinding plan for secondary data (register before the outcome is merged to the treatment; or an exploratory split with the holdout untouched).
5. Pre-flight. Refuse to mark the plan ready until: every hypothesis has direction, estimand, SESOI, and decision rule; the primary model could be coded from the text alone; index construction and the covariate list are fixed; every exclusion has a threshold; the stopping rule is concrete; more than one hypothesis comes with a multiplicity policy; the power section is imported with its assumptions; for existing data, the data-access statement and the blinding plan are honest.
6. Config. Set design.prereg: true in .mstack/config.yaml and stub prereg_url for the user to fill after registration.
7. Hand-off. Tell the user: copy the Markdown into the registry, paste the resulting URL into .mstack/config.yaml, then field or analyze; keep the registered text in prereg/ unchanged and put later edits in a dated addendum.

Outputs

• prereg/osf-prereg.md — the preregistration.
• .mstack/config.yaml — design.prereg: true, URL stubbed.
• Summary block: registry, data-access status, the items still failing pre-flight if any; register before fielding, paste the URL back.

Anti-patterns

• Vague primary analysis. "We will run a regression" is not a specification; neither is "appropriate controls".
• No stopping rule. "When we have enough" guarantees a garden of forking paths.
• Exclusions without thresholds. "Inattentive respondents" is not a rule; "fails ≥ 2 of 3 attention checks" is.
• Open index construction. The outcome must be computable from the plan.
• Hidden moderation tests. Pre-specify or label exploratory.
• A plan written after peeking, without saying so.

Next

Field / /mstack:data-acquire. If the analysis is meant to be confirmatory, do not run /mstack:analyze before registering.