Turn a business objective into an evidence-backed experiment design, an implementation contract, and, when requested and supported, an AE experiment draft.

Read only the references needed for the current request:

references/platform-operations.md

references/experiment-creation.md

references/platform-operations.md

references/metric-readiness.md

references/sdk-integration.md

references/sdk-index.md

references/experiment-sdk-contract.md

references/client-experiment-sdk.md

references/server-experiment-sdk.md

references/hybrid-experiment-sdk.md

references/exposure-contract.md

references/sdk-troubleshooting.md

References are a curated fast path, not the whole documentation set.

Extract:

Convert these into a falsifiable hypothesis. Clarify only missing facts that materially change the design. For a conversion goal, establish the population, denominator or exposure behavior, numerator behavior, attribution window, and analysis unit.

Do not silently invent a target population, conversion definition, or technical platform.

Pass the project gate in . With , establish candidate exposure and outcome events, assignment identity and join path, timestamps, exact saved-metric definitions, and—when available—baseline and eligible traffic.

If the project is unavailable, accept user-provided schemas or definitions and label all platform-dependent conclusions as unverified.

Apply . Classify candidates as , , , or ; recommend one primary event metric and confirm its calculation code before planning sample size or duration.

Define the Feature key, type, typed default, ownership, stable assignment unit, one control group, treatment groups, group values, traffic, allocations, layer, targeting, and exclusions.

Require allocations totaling , experiment traffic in , type-correct values, a stable exposure-to-outcome identity join, and at least one primary metric. Resolve real Features and layers with before reuse or creation.

Follow this order: confirm the primary metric and calculation code → obtain its baseline, MDE, and any required variance → calculate the sample target with

scripts/calculate_experiment_plan.py

→ derive duration from the sample target and effective eligible daily units. Apply the preregistered planning policy in . Use fixed and two-sided testing, policy-default , and Bonferroni planning for multiple treatments. Require the MDE type and direction; never default MDE to 5% or assume variance for a continuous metric.

When experiment traffic is already confirmed, calculate its duration. When it is not confirmed, obtain verified layer capacity and let the script recommend the smallest absolute traffic candidate that reaches the target within the maximum runtime. Default to at least seven days and full-week alignment. Return the actual infeasible duration instead of truncating it. Explain the baseline, MDE, power source, allocations, multiplicity rule, traffic evidence, sample targets, duration adjustment, and any native-report mismatch. Do not return a definitive plan when required evidence is unavailable.

For an explicit draft-creation request, follow and . Create only authorized draft assets, verify the saved result by reading it back, run supported readiness and conflict checks, and return the compact receipt and experiment link defined there.

Submitting, starting, changing live traffic, pausing, ending, or deleting requires separate explicit confirmation. Never turn draft creation into launch.

Treat project resolution, metadata discovery, candidate-event searches, metric comparison, Feature and layer inventory, capability discovery, schema inspection, and command execution as internal working context.