The Observatory · Prevention Lab
Turn a report into a testable improvement.
Bring a small set of records you are permitted to use. Compare deterministic triage, name an outcome owner, freeze a success rule, and replay it when the evidence arrives.
Start with a bounded evidence packet
Use de-identified product feedback. This form processes at most 24 records and 500 characters per body in memory. It does not call a model, retain a participant row, contact anyone or change the product.
Input format and limits
Supply a question, outcome_owner, evidence_start and evidence_end as ISO timestamps, success.minimum_attempts (at least 50), success.maximum_recurrence_percent (0–100), and records. Each record needs a unique id, body, severity (low, medium, high or critical), and created_at. The past evidence window may cover at most 31 days. Severity and report accuracy are caller-supplied and unverified.
The existing quality gate can hold spam, duplicates or insufficiently specific records. Its exclusions are counted in the result. A recommendation is a proposal for a reversible experiment, never permission to execute it.
Advanced: use a JSON packet with up to 24 records
Agent contract · same planner, same limits
POST the packet with consent: true as JSON to /api/experiments/prevention. For a later replay, POST {"action":"replay","plan":{"source":{},"source_hash":"..."},"measurement":{"control":{"attempts":100,"failures":12},"treatment":{"attempts":100,"failures":4},"window_days":14,"comparable_cohorts":true,"severe_harm":false}}, replacing the source and hash with your saved plan. Responses are private and no-store. Missing denominators abstain; changed source is refused.