What was tested
This is historical integration evidence using ResearchResponse and Profile identities; today’s API uses Run. It covers backend ingestion and replay, not a reporting UI or customer outcomes.
| Tested boundary | Outside this evidence |
|---|---|
| Completed responses → application records | Live answer quality, reporting UI, customer outcomes |
Start with a question a business recognizes
API answers and AI answer-engine observations are different evidence. Two retrieval services may observe the same engine differently.
Which villa rental services do AI answers recommend for a week on the Amalfi Coast? Do we appear, and which sources shape the recommendations?
Librarium collects the facts; the application owns the record
Agentsy assigns the owner and measurement series, then stores an immutable receipt and each successful AI answer-engine observation. Replay safety is an application guarantee, not proof an answer is correct.
| Record | Retained facts |
|---|---|
| Measurement series | Entity and prompt, identified by an opaque key |
| Observation | Effective provider, historical Profile, content, citations, retrieval service, engine, observation time when available |
| Same receipt again | No duplicate observations |
| Same identity, changed facts | Rejected |
Compare like with like
The latest observation stays within one owner, measurement series, and AI answer engine. Retrieval-service disagreement remains inspectable, not independent confirmation or a promise about every account.
| Lookup | Boundary |
|---|---|
| Latest engine observation | Same owner + series + engine |
| Latest per retrieval service | Same boundary, separated by service |
| Newer answer to another question | Does not replace this question’s evidence |
Keep product policy outside the research package
Librarium supplies completed results; Agentsy owns persistence and isolation. Applications can build review screens or agent tools over approved evidence without adopting private runtime state.
| Visibility model | Policy |
|---|---|
| Approved observations | Application-owned records |
| Credentials, raw responses, headers | Excluded |
| Provider metadata, usage payloads, arbitrary errors | Excluded |
What this proves—and what comes next
Faked research with stray HTTP blocked tests ingestion, replay, changed-payload rejection, immutability, and owner/series isolation. Scheduling, retrieval-service selection, reporting, and production operations remain application work.
See the Laravel agent example and queue guide for current Run-based integration.