Research Experiment Plan

Quick start

1. Freeze the claim, decision rule, and anti-claims before listing runs.
2. Classify the intended evidence as exploratory, confirmatory, independently verified, or operational/high-stakes; use ../research-pipeline-planner/references/epistemic-assurance-contract.md when promoting beyond exploratory status.
3. Decide whether this is a standalone plan or the experiment stage inside an orchestrated suite.
4. Initialize the full experiment pack with scripts/init_experiment_pack.py whenever the plan needs tracked execution.
5. Build the minimum decisive experiment blocks, not a benchmark wishlist or a paper-defence script.
6. Run a cheap non-vacuity preflight before expensive or confirmatory execution.
7. Separate must-run from nice-to-have runs, attach explicit decision gates, and emit bridge-ready outputs for later execution.
8. For orchestrated execution, declare inputs, evaluator artifacts, required outputs, allowed lineage relations, and gate-conditioned downstream activation using ../research-pipeline-planner/references/experiment-execution-binding.md.

Constants

• MAX_PRIMARY_CLAIMS = 2 unless multiple inseparable claims are truly required.
• MAX_CORE_BLOCKS = 5 unless the evaluation setting forces more.
• MAX_BASELINE_FAMILIES = 3 unless the field norm demands additional families.
• DEFAULT_SEEDS = 3 when stochastic variance matters and budget allows.