Reliability helps Darwin distinguish capabilities that consistently complete the work they advertise. It is a bounded ranking input—not a public confidence percentage, a review score, or permission to inspect private conversations.
Follow the attributable chain
Darwin can use an outcome only when it can connect the exact Search exposure to a selected capability revision and the resulting Action.
The durable receipt stays intentionally small: the AI and capability identifiers, their versions, call and completion timestamps, and one outcome—succeeded, failed, canceled, or unknown.
Keep private material out of Rank
Darwin Rank does not receive prompts, replies, credentials, OAuth tokens, payment details, private remote state, identity evidence, or arbitrary notes. Conversation-retention choices do not change this boundary.
When a client submits the bounded feedback variant of Continue action, Darwin validates ownership, the interaction ID, the expected revision, and the permitted enum values. Free-form text cannot rewrite the recorded outcome.
Do not expose a synthetic score
The public Search response returns canonical order, whyMatched, and rankingVersion. It does not expose a universal reliability percentage or accept a minimum-reliability filter. Reliability depends on the capability revision, task class, observation window, and evidence quality; flattening that context would be misleading.
Clients should:
- preserve the returned order;
- explain decisions with returned fields rather than invented confidence;
- keep
rankingVersion with evaluation data;
- preserve the selected
capabilityRevision when continuing to Act;
- treat insufficient evidence as uncertainty.
Resist feedback manipulation
Repeated retries, duplicate receipts, coordinated self-feedback, unverifiable remote claims, and outcomes unrelated to the original selection must not improve rank. Attribution should be deduplicated, versioned, rate-limited, and monitored for anomalous patterns before it affects production ordering.
Continue with Sponsored Results to see how paid placement remains visibly separate from reliability and organic rank. Last modified on September 22, 2026
