Observation → hypothesis → falsifiable requirements → baseline → reference implementation → experiment → correctness gate → efficiency measurement → replication → result → maturity update.
Where practical, record model, provider, model version, reasoning effort, tool versions, fixture, prompt, hardware/environment, repetition count, observable tool activity, final result, correctness result, cost/tokens if available, and known confounders.
Never infer or request private chain-of-thought. Use observable prompts, tool calls, files, events, process results, provider-reported usage, and final artifacts.
Incorrect candidates are not “efficient.” They remain in the result set but cannot support an efficiency-superiority claim. Every comparison states the exact correctness and safety gate.
A baseline represents a real available alternative, uses the same fixture/outcome definition, and discloses capability differences. Record exact software/configuration revisions, repetitions, and independent environments. Distinguish replicated mechanism from replicated benefit.
Negative, null, failed, and interrupted experiments are retained. Separate harness failure from system-under-test failure and missing data from zero.
Prefer hypothesis, experimental, observed, measured, preliminary, supported, unverified, and not yet established. Avoid revolutionary, breakthrough, solved, guaranteed, eliminated, or proven without evidence meeting the actual claim.