/research-design

Role: Principal Investigator (study design) Pipeline position: mid-chain

Operationalizes the hypotheses into testable predictions and designs the study or experiment — variables, control conditions, experimental unit, allocation and masking, sample, instruments, power analysis, threats-to-validity — then produces a preregistration that freezes the analysis plan before data collection.

Canonical invocationCopy Anchor Link

/research-design

With arguments:

/research-design [--suite-name NAME]

InputsCopy Anchor Link

_inputs/proposal.md (the plan-of-record from /research-proposal, the predecessor stage), _inputs/synthesis.md (the SOTA map + gap statement), and _spec/research-spec.md (the hypotheses).

OutputsCopy Anchor Link

_inputs/study-design.md (operationalized predictions, variable specification, control conditions, experimental unit, allocation and masking plan, sample, instruments, power analysis, threats-to-validity) + _inputs/preregistration.md (the frozen analysis plan, including the control conditions with their pass criteria, the allocation and masking plan, the missing-outcome handling, and any interim looks).

DownstreamCopy Anchor Link

/research-experiment (downstream consumer)

Workflow phasesCopy Anchor Link

The command follows the standard /research pipeline workflow:

1. Load context — Read the synthesis and the spec's hypotheses.
2. Verify the Sequence Gate — Confirm _inputs/proposal.md (and the upstream _inputs/synthesis.md and _spec/research-spec.md) exists; otherwise halt with Blocked: run /research-proposal first.
3. Design the study — Operationalize predictions; specify variables, control conditions, and instruments; fix the experimental unit, the seeded allocation, the masking plan, and the run order; size the sample for the planned analysis; enumerate threats-to-validity.
4. Preregister — Freeze the analysis plan, control conditions, allocation, and masking before any data collection.
5. Update the handoff manifest — Record the study-design and preregistration paths.
6. Emit the design — _inputs/study-design.md + _inputs/preregistration.md.

Structure and sizingCopy Anchor Link

• Structure is fixed in the design, not the analysis. The design names the experimental unit and the unit of analysis, the allocation procedure with its seeded schedule and, when units enroll one after another, its concealment, who is masked (or why masking is infeasible), the nuisance factors and how each is blocked, randomized, or held, and the run order. It plans the comparator and the positive and negative controls the experiment will execute. A control added after the freeze is a disclosed deviation.
• One-factor sweeps cannot show an interaction. Where components or parameters are suspected to interact, the design adds a joint layout over the suspect set and preregisters the interaction it claims.
• The smallest effect of interest comes from the spec. It traces to the spec's success metrics or a practical-significance threshold the synthesis grounds. A prior study's reported effect bounds the plausible range and never stands in as the smallest effect of interest.
• Size for the analysis that will run. Each primary claim's goal (superiority, non-inferiority, or equivalence) and sidedness are fixed in the design, and each is powered at its own smallest effect of interest at the per-test threshold the declared correction leaves. The required N is reported as a range across the plausible effect; when feasibility caps N, the smallest detectable effect at that N is reported and the smallest effect of interest is stated as not powered.
• Allowances and simulation. N counts independent units and is inflated for within-cluster correlation; the collection target comes from the analyzed N through a recorded loss rate. Where no closed form matches the planned analysis, power is estimated by simulation with the tool, version, seed, and Monte Carlo error recorded, and a sequential rule fixes its number of looks and boundary in the preregistration.

Rigor floorCopy Anchor Link

For a comparison-bearing study the design fixes, and preregisters, the empirical floor: one unified budget across all compared methods (equal wall-clock and, where meaningful, equal evaluation count), at least thirty repetitions per instance for stochastic methods, an ablation factor that disables one component at a time, sensitivity sweeps over each parameter, and automated configuration on a training split evaluated on a disjoint test split. The floor depth scales to the target-venue ambition.

Failure modesCopy Anchor Link

ExamplesCopy Anchor Link

# Design the study and preregister the analysis plan
/research-design --suite-name <name>

Cross-referencesCopy Anchor Link

• Pipeline overview
• Commands reference