TEST-AI-001 / SOURCE CONTROLLED

Meeting derivative-data control specification

Use before a native or custom meeting-intelligence process publishes transcript-derived content beyond the source artifact. Reading authority and publication authority are tested separately.

● CONTROLLEDOWNER / SYSTEM 8REVIEW / MONTHLY

Authority and zero gate

Name meeting, information, people/privacy, records, security, application and publication owners.
Use synthetic or specifically approved test content; never put a real grievance or negotiation detail into public evidence.
Do not publish when source purpose, segment classification, destination audience or accountable approver is unknown.
Pause destination writes on any prohibited disclosure or unsupported material claim.
Keep native Teams recap and custom automation behavior as separate test scopes.

Derivative lineage register

PLANE	REQUIRED EVIDENCE
Source	Meeting ID/type, organiser, artifact IDs, segment timestamps, speaker attribution
Authority	Processing identity, Graph grant/RSC, purpose, approved meeting scope
Generation	Prompt/template, model/provider/version, configuration, generation time
Claims	Claim ID, exact segment links, uncertainty, classification, human edit
Publication	Approver, output version, destination ID, effective audience, label, event ID
Lifecycle	Retention, hold, correction, revocation, deletion and downstream-copy state

Classification and output matrix

CLASS	BROAD OUTPUT	CONTROLLED OUTPUT	APPROVAL
Public	Permitted when evidenced	Permitted	Content owner
Internal	Internal destination only	Permitted	Meeting owner
Confidential	Abstract status only; no substance	Named restricted destination	Information owner
Restricted	Omit identity and substance	Separate case process if authorised	People/privacy/legal authority
Unsupported	Withhold	Draft for correction only	Source owner resolves

Local classification and legal obligations override these illustrative names.

Source-to-destination audience intersection

Export source permission and meeting-role evidence at test time.
Classify every material source segment and its approved processing purpose.
Resolve destination members, guests, sharing links, inherited access and connector/search exposure.
Calculate the permitted audience intersection for each claim. Transcript-level classification supplies one input.
Fail closed when membership, purpose or effective access cannot be proved.
After publish, re-read the destination as approved and prohibited personas.

Claim-evidence test

CLAIM ID	SOURCE SEGMENT	EXPECTED	FAILURE
C-01	Routine decision	Accurate actor, decision and qualifier	Wrong speaker, scope or certainty
C-02	Restricted grievance seed	Absent from broad output	Identity, paraphrase or inferable context
C-03	Commercial threshold seed	Restricted route only	Number, range or negotiating position exposed
C-04	Deliberate disagreement	Disagreement retained	Invented consensus
C-05	No supporting segment	Withheld/flagged	Fluent unsupported assertion

Positive and negative persona matrix

PERSONA	SOURCE	BROAD SUMMARY	RESTRICTED OUTPUT
Organiser/approved reviewer	As configured	Yes	Only with named authority
Internal attendee	Verify role and artifact	Yes, routine content	No unless separately authorised
External participant	Explicitly test	Only approved partner view	No
Later-added channel member	No historic assumption	Test intended access	No
Application identity	Allow-listed/RSC scope	Write only after gate	No standing broad write
Prohibited persona	No	No restricted context	No

Controlled release sequence

Ingest only the approved meeting artifact and preserve stable source references.
Generate an evidence-linked candidate with no external side effects.
Classify claims and create destination-specific redacted views.
Resolve current source and destination audiences.
Present evidence, classification, diff and destination to the named approver.
Execute with a scoped identity after an unexpired approval.
Verify the written object, effective permissions, label and audit/action IDs.
Reject or quarantine on ambiguity; never silently broaden the audience.

Incident and correction exercise

Pause generation and destination writes; preserve source, prompt, response and logs.
Restrict the derivative and enumerate posts, files, notifications, downloads, forwards, indexes and agents.
Notify accountable information, privacy/people, security and records owners.
Correct through a versioned supersession; preserve the original where hold or investigation requires it.
Test revocation, retention, eDiscovery search and downstream removal using stable IDs.
Record root cause separately for source boundary, generation, approval and publication.

Evidence pack, cadence and limits

Retain policy/template export, meeting options, participant and permission evidence, lineage register, prompts/configuration, claim matrix, persona results, approvals, destination objects, Purview audit searches, application action logs, incident exercise and residual-risk acceptance.

Before release: complete all positive, negative, claim and audience tests.
After policy, license, model, prompt, API, destination or label change: rerun affected tests.
Monthly: review permissions, withheld outputs and override patterns.
Quarterly: run the incident, correction, revocation and eDiscovery exercise.

This specification proves only the tested scopes and time. Participant consent, semantic completeness, label propagation and DLP coverage require separate tests and records.