[Use Case] Agent-team delivery of regulated medical-device compliance work — authority, verification and evidence semantics in production
fg-tida-use-cases-use-case-agent-team-delivery-of-regulated-medical-device-compliance-work-authority-verification-and-evidence-semantics-in-production-80
| Type | FG-TIDA document |
|---|---|
| Group | Trust and Identity for Humans and Agentic AI |
| State | Submitted |
| Authors | Not entered |
| Last updated | 2026-10-23 15:11 UTC |
Formats and resources
Abstract
Complete every field. Replace the angle-bracket placeholders. Tick boxes with [x].
1. Identification
Title: Agent-team delivery of regulated medical-device compliance work — authority, verification and evidence semantics in production
Submitting organization: SynomosAI (LGD theory initiative); production reference: MedXpert
Contact name: Steven Zhao (independent expert, ORCID 0009-0001-0512-1237)
Contact email: zhaoxinghua06@126.com
Sector (tick all that apply; a cross-industry case may tick more than one):
- [x] Health
- [ ] Education
- [ ] Displaced persons / humanitarian
- [ ] Financial services
- [ ] Telecom
- [ ] Media and publishing
- [ ] Logistics
- [ ] Energy
- [ ] Critical infrastructure
- [ ] Smart home / IoT
- [x] Other (specify): Professional regulatory-compliance services (cross-industry applicability)
2. The situation
Plain language description (3 sentences max): An AI agent team prepares medical-device regulatory compliance deliverables (technical files, clinical evaluations, submission dossiers) for manufacturers under NMPA, FDA, EU MDR and JP PMDA regimes. Clients and, downstream, regulators must decide whether machine-produced compliance evidence can be trusted without a traditional human-authorship paper trail. No external verification standard exists today for agent-produced professional work products.
Actor: Medical-device manufacturer (client); AI agent team (multi-agent delivery); receiving parties (client QA/regulatory staff; regulators)
Action: Agents research, draft, verify and assemble regulated compliance deliverables and hand them to the client for use in official submissions
Decision required: May this agent team's compliance deliverable be accepted as regulatory-grade evidence, and under what verification conditions?
Problem encountered: No agreed evidence semantics or verifier-side obligations exist for agent-produced professional work; provenance, failure states and accountability cannot be externally checked
Current mitigation: Internal operational governance laws (LGD) with a cryptographic provenance chain (Sigstore Rekor transparency log, OpenTimestamps, DOI/SWHID archival) applied to every deliverable package
Residual gap: External parties have no standard way to verify provenance, integrity and failure semantics of agent work products; accountability after irreversible actions (e.g. a filed submission) is not standardised
3. Actors and context
Taxonomy roles involved
- [x] Principal — possesses the authority being exercised; source of the mandate
- [x] Relying party — must decide whether to accept the agent's authority
- [x] Builder
- [x] Deployer / Owner
- [x] Agent instance
- [ ] User — interacts with the agent; may or may not be the principal
- [x] Infra provider
- [x] Attestor
- [ ] Other (specify):
Mandates in this case
| Grantor | Grantee | What is conferred | Governing regime |
|---|---|---|---|
| Medical-device manufacturer (client) | AI agent team (delivery firm) | Mandate to prepare regulatory deliverables under SOW/NDA | Medical device regulations (NMPA / FDA 21 CFR / EU MDR 2017/745 / JP PMDA) + contract law |
| Delivery firm owner | Individual agent instances | Operational mandate to research, draft, verify, publish internally | Internal operational governance laws (LGD) + professional liability regime |
Cross-border?
- [x] Yes — jurisdictions involved: China (NMPA), USA (FDA), EU (MDR), Japan (PMDA)
- [ ] No
Embodied (physical actuation or sensing)?
- [ ] Yes — physical context:
- [x] No
Agent action type (required; tick all that apply. Where several apply, the most consequential governs the case)
- [ ] Read-only
- [x] Consequential (reversible)
- [x] Irreversible
Crosses organisational boundary? (required)
- [x] Yes (specify): agent team → client organisation → regulator (three trust boundaries)
- [ ] No
Risk level (required)
- [ ] Low
- [ ] Medium
- [x] High
4. Theme relevance
| Theme | Yes | Primary | Justification (required if checked) |
|---|---|---|---|
| Dynamic Identity | [ ] | [ ] | |
| Continuous Trust and Attestation | [x] | [x] | Core question is whether a relying party can continuously verify agent-produced evidence (provenance chain, Rekor log, timestamps) — i.e. attestation consumption on professional work products |
| Delegation | [x] | [ ] | Client-to-agent-team mandate spans three organisational boundaries and four regulatory regimes; standing and scope of the mandate is exactly the delegation problem |
| Discovery and Cross-Border Trust | [ ] | [ ] | |
| Runtime Enforcement (Control Plane) | [ ] | [ ] | |
| Embodied AI Identity and Trust | [ ] | [ ] |
Related theme proposals: #27 — Operational governance laws for autonomous AI agents in regulated industries; #7 (verifier-side requirements and failure semantics); #22 (RA-AAI consumption side)
5. Requirements
| # | Type (choose one) | Requirement description | Criticality (choose one) |
|---|---|---|---|
| 1 | Technical | Evidence semantics for agent-produced work products: what a relying party may check, and what "PASS" must mean | Must |
| 2 | Technical | Verifier-side obligations and failure semantics (distinguishing "verified" from "not checked") for agent deliverables | Must |
| 3 | Technical | Interoperable provenance chain for work products: transparency-log entry + cryptographic timestamp + archival identifier, independently reproducible | Must |
| 4 | Technical | Accountability attribution after irreversible agent actions (e.g. a filed regulatory submission): who/what is answerable, and how it is recorded | Must |
| 5 | Business | Mandate representation that is regime-aware across multiple jurisdictions (conferral, standing change, continuing effect) | Should |
| 6 | Business | Deployability under regulatory constraint: governance toolchain must run offline on commodity hardware with low cost and short setup (days, not months) | Should |
| 7 | Technical | Revocation and correction semantics when an agent deliverable is later found defective | Should |
| 8 | Legal | Boundary conditions for accepting agent-produced evidence in regulatory contexts per regime | Nice-to-have |
6. Assessment criteria
Success criteria: An independent third party can verify the provenance and integrity of an agent-produced compliance deliverable without trusting the vendor, and every irreversible action is attributable and auditable end-to-end.
Measurable metric: Independent reproduction of the verification chain (Rekor logIndex lookup, timestamp check, archival resolution) succeeds 100% of the time; audit shows 0 unattributable irreversible actions across production deliveries.
7. Duplication check
Existing standards or SDOs: IETF RATS (attestation message formats); W3C DID/VC (credential syntax); EU AI Act transparency obligations; NIST AI RMF; ISO/IEC 42001 (AI management systems); Sigstore (artifact signing).
Overlap notes: RATS and W3C cover attestation formats and credentials, not evidence semantics for agent-produced professional work products; EU AI Act and NIST AI RMF state obligations but no verification mechanics; ISO/IEC 42001 is a management-system standard without per-artifact provenance. None addresses the relying-party decision for agent-delivered regulated work.
8. Maturity
Pilot status
- [ ] Hypothetical
- [ ] Pilot in progress
- [x] Deployed in production
Reference implementation: https://github.com/zhaoxinghua09-cell/uibc-core (DOI 10.5281/zenodo.22821835; Sigstore Rekor transparency-log entry logIndex 2883389783; OpenTimestamps; SWHID swh:1:ori:819530c8…) — in daily production use for medical-device compliance delivery
9. IP and confidentiality
Confidentiality level
- [x] Public
- [ ] FG-internal only
- [ ] Redact before publication
IP notes: LGD theory stack publicly anchored (Zenodo DOI); client and employer identifying details withheld; only publicly verifiable artifacts referenced.
10. Assets
Link to assets:
- [ ] Dataset
- [x] Code — uibc-core; silent-failure-catalog; assayance
- [x] Protocol — LGD operational governance laws; evidence-chain specification
- [ ] Other:
Revisions
- No file revisions have been uploaded.