/research-analysis

Role: Principal Investigator (analysis) Pipeline position: mid-chain

Analyzes the raw data per the preregistered plan — runs the planned statistical tests on the independent unit of analysis, reports effect sizes with confidence intervals (credible intervals under a Bayesian plan; not p-values alone), applies multiple-comparison correction where applicable, states how missing data was handled, runs robustness / sensitivity checks, and discloses every deviation from preregistration.

Canonical invocationCopy Anchor Link

/research-analysis

With arguments:

/research-analysis [--suite-name NAME] [--override] [--correction METHOD] [--alpha LEVEL]

InputsCopy Anchor Link

The raw data at the host-natural location the experiment stage recorded, plus _inputs/preregistration.md (the frozen analysis plan).

OutputsCopy Anchor Link

_outputs/analysis.md (the preregistered tests with effect sizes, CIs, unit of analysis, and a per-domain bias judgment for each headline result; the multiple-comparison correction; the missing-data handling; the robustness checks; and disclosed deviations) + figures / tables at a host-natural location.

DownstreamCopy Anchor Link

/research-paper (downstream consumer)

Workflow phasesCopy Anchor Link

The command follows the standard /research pipeline workflow:

1. Load context — Read the raw data and the preregistration.
2. Verify the Sequence Gate — Confirm the experiment outputs exist; otherwise halt with Blocked: run /research-experiment first.
3. Analyze — Run the preregistered tests on the independent unit of analysis; diagnose assumptions without letting the diagnostics choose the test; report effect sizes + CIs; correct for multiple comparisons; state the missing-data handling; run robustness / sensitivity checks.
4. Disclose deviations — Record every departure from the preregistered plan. A move to a fallback the preregistration names is logged as a planned adaptation; any other substitution is a deviation.
5. Update the handoff manifest — Record the analysis path and the figures / tables.
6. Emit the analysis — _outputs/analysis.md + figures / tables.

Statistical disciplineCopy Anchor Link

• The test follows the design. The preregistered primary test stands unless the preregistration's own fallback trigger fires; a screening test that fails to reject does not show an assumption holds.
• n counts independent units. Repeated measures, sub-samples, clusters, and seeds are aggregated to the unit or modeled as dependence, never counted as independent replicates.
• A null is read three ways. The effect-size interval, the equivalence result against the smallest effect of interest, and a sensitivity statement (the smallest effect detectable at the achieved analyzed n) are reported together. Power computed from the observed effect is never reported.
• Bayesian plans on their own terms. A preregistered Bayesian plan reports the posterior summary with its credible interval, runs prior-sensitivity, computation-diagnostic, and posterior-predictive checks before any posterior is reported, and uses a pre-fixed practical-equivalence region for nulls.
• Result-level bias and figures. Each headline result carries a low / some-concerns / high judgment per domain, seeded from the design's threats-to-validity ledger and never summed into one score, and every figure meets the encoding, uncertainty-disclosure, and accessibility floor.

Rigor floorCopy Anchor Link

A stochastic method is run for at least thirty independent repetitions per instance or condition, more when the spread across repetitions is wide relative to the differences under test, until the summary statistics are stable. Its results are reported as best, mean, median, standard deviation, and worst across repetitions, with convergence curves and distribution plots. Comparisons use nonparametric tests — a paired signed-rank test pairwise, an omnibus rank test with post-hoc correction across methods — reported with a stochastic-superiority effect size and a critical-difference diagram. A result advances to the paper only when it clears the advancement gate: it beats its comparators convincingly and significantly, and the clearance is logged in an auditable runbook.

Failure modesCopy Anchor Link

ExamplesCopy Anchor Link

# Analyze the raw data per the preregistered plan
/research-analysis --suite-name my-study

Cross-referencesCopy Anchor Link

• Pipeline overview
• Commands reference