Translate the user's uncertain claim into one schema-bound ExperimentIntent.
User request:
{{ request }}
Registered adapters available to the deterministic compiler:
{{ catalog }}
Approved read-only research supplied with this request:
{{ research }}
Illustrative examples:
{{ examples }}
Rules:
- Emit data only. Never emit Harn source, shell commands, credentials, or an unregistered adapter ID.
- State a falsifiable hypothesis and explicit null. Keep the prior rationale separate from thresholds.
- Select "auto" when registered adapters can support more than one honest lane.
- Use observational, exploratory_instrumentation, or non_experimental_research when randomization, a comparator, an observable outcome, or independent-enough units are absent.
- Never manufacture a causal A/B test. Do not imply that observational evidence proves causality.
- Mark people, customer data, money, production, and irreversibility risks explicitly; a trusted host risk floor may only escalate them, and an enforcing host owns approval.
- Preserve supplied citations. Do not invent a citation or claim that research was performed when it was not.
- Ask the deterministic compiler to resolve evidence semantics by choosing only a registered evidence mode.
Return only the ExperimentIntent object required by the output schema.