FIELD NOTE / TEAMS / DERIVATIVE DATA
Trace every meeting summary claim to an authorised source segment
A fluent summary can conceal a speaker error, missing qualification or restricted source segment. Treat each decision, risk and action as a claim with a source range, confidence, audience and approval state. The publication service then has concrete inputs for deciding whether a derivative can be written to Teams, SharePoint or a line-of-business system.
Acquire the authorised meeting object
Record the meeting ID, organiser, chat/thread ID, transcript ID, recording location, participant roster, start/end UTC and classification. For custom Graph processing, use the documented online meeting transcript endpoint with an application access policy or delegated context that matches the design.
Keep the source transcript in its Microsoft 365 location. The processing job stores identifiers, hashes and permitted excerpts. The design keeps the transcript in its existing controlled repository and stores claim references only.
Normalise speakers and timestamps without changing the record
Convert the transcript into stable segments containing segmentId, speaker object or unresolved label, start/end time, text hash and source URI. Preserve corrections as a new version linked to the original segment.
If diarisation is uncertain, mark the speaker unresolved. If two participants disagree, retain both claims. A summary field cannot manufacture consensus.
Store decisions and actions as typed claims
Each derived claim contains claimId, type, text, source segment IDs, confidence, sensitivity, proposed owner, due date, model/version, prompt version and review state. Accepted, corrected and rejected claims remain traceable.
Use deterministic extraction rules for dates, people and identifiers before an LLM adds interpretation. Validate generated JSON against a schema. A missing source array fails the item before publication.
Calculate the destination audience
The safe audience comes from the approved purpose, the source segment permissions and the destination membership. Resolve guests, channel type, shared-channel tenancy, SharePoint site membership and sensitivity before writing.
For restricted topics, split the derivative or create manual minutes in a smaller location. The broad meeting summary should say that a restricted item exists only when policy and purpose permit that disclosure.
Require human authority for external effects
The system may propose an owner and due date. A named approver confirms the claim, source, recipient and action. The executing identity receives only the API and resource authority required for the approved write.
Use an idempotency key such as meetingId:claimId:destinationId:version. Retrying an approved action returns the existing target record. A failed response after commit enters reconciliation instead of blind replay.
Retain a useful audit chain
Record request ID, model deployment, prompt version, source IDs, claim hash, reviewer, approval UTC, destination object ID and write result. Purview audit and Graph or workload logs support the platform side; the application log joins those events to the claim.
Exclude transcript text, access tokens and unnecessary personal data from operational logs. Store sensitive evidence under the same or tighter classification than the source.
Run claim, audience and replay tests
Prepare supported and unsupported claims, two speakers with similar names, a restricted segment, an external guest, an unrelated user, a rejected action and a timeout after the target write. Verify source links, correction handling, audience denial, one target effect and an explainable queue item.
Measure unsupported-claim rate, missed decisions, speaker corrections, permission denials, duplicate prevention and human edits. A readability score cannot demonstrate safe operation.
Publish the tested boundary
Native Teams recap and custom Graph automation have different storage, permission and audit behaviour. State which path the design covers. Generated minutes remain a derivative and may require separate employment, privacy, records or legal review.
The system can assist with a working record. It cannot confer decision authority or infer consent for later publication.