A Lifecycle Trust Framework for Agentic AI

Registry, Evidence, and Gates — the LGD Model

Source: Steven Zhao, Independent Researcher (SynomosAI / LGD framework)
Contact: zhaoxinghua06@126.com
Meeting: 1st FG-TIDA Meeting, Paris, 30 November – 1 December 2026
Document type: Input document

1. Summary

Trust in agentic AI is not a single event — it is a property that must be established, demonstrated and re-evaluated across the entire lifecycle of an agent: from birth (identity issuance) through operation (consequential acts) to evolution (capability changes) and retirement.

This contribution proposes the LGD lifecycle trust framework, built on three executable laws:

• LGD-I · Registry: every agent is registered at birth — identity, creator, governance owner, records;
• LGD-II · Evidence: every consequential act leaves replayable evidence — decision logs, risk registers, human-oversight proof, incident ledgers;
• LGD-III · Gates: every capability evolution passes a gate — trigger, assessment, release, retrospective review; high-stakes authority stays with humans.

The framework is a civil-society reference model, deliberately aligned on top of existing standards (ISO/IEC 42001, EU AI Act, ISO 13485/14971, NIST AI RMF, UNESCO Recommendation) rather than competing with them. It directly supports FG-TIDA's work item on trust frameworks and lifecycle (assurance) models.

2. Problem Statement

Mainstream AI-governance efforts address trust mostly at single points:

1. Identity — "who is it" (passports, registries, algorithm filing);
2. Evidence — "prove it" (audit trails, explainability, transparency);
3. Gates — "who lets it grow" (change control, conformity assessment).

Almost no framework chains these points into one unbroken lifecycle thread. As a result, an agent can be well-identified at birth but unverifiable in operation, or heavily audited in one phase but ungoverned in another. Trust degrades silently across lifecycle gaps.

3. Proposal — the LGD lifecycle trust model

3.1 Three laws as a lifecycle thread

3.2 Isomorphic reference model

The regulated world already runs this chain in practice: medical-device lifecycle regulation (classification → design control → risk management → registration → clinical evidence → quality systems → post-market surveillance → retirement). The framework transfers this proven model to agentic AI, giving FG-TIDA a concrete, real-world-tested assurance spine rather than an abstract policy layer.

3.3 Lifecycle stage = executable instrument

Each stage maps to a checkable tool or evidence generator, so the framework is testable and benchmarkable:

• identity issuance and verification (see companion contribution C1);
• risk classification at birth;
• change-gate assessment for capability upgrades;
• incident ledgers for post-deployment monitoring;
• retirement checks that close the chain.

4. Mapping to FG-TIDA Work Items

5. Relationship to Companion Contribution C1

• C1 (verified-agent-identity) answers "who is the agent, and whom does it act for?" — identity infrastructure.
• C3 (LGD) answers "under what conditions should that entity be trusted to act, across its lifecycle?" — trust management and assurance.
• Together they cover the two halves of the FG-TIDA mandate: identity infrastructure and trust management.

6. Proposed Action for FG-TIDA

1. Consider LGD as an input to the trust frameworks and lifecycle (assurance) models work item.
2. Use the three laws (Registry / Evidence / Gates) as a candidate checklist for the lifecycle assurance model under discussion.
3. Reference the medical-device paradigm as a proven real-world template for lifecycle governance.
4. Align with the companion contribution (C1) so identity and lifecycle trust are specified coherently.

7. References

• Zhao, X. (2026). The Lifecycle Governance Doctrine (LGD). DOI: 10.5281/zenodo.22456647. CC BY 4.0.
• Companion contribution: Decentralized Agent Identity Infrastructure (ERC-8004, iden3, Attestation Registries).
• ISO/IEC 42001, EU AI Act 2024/1689, ISO 13485/14971, NIST AI RMF — alignment references.
• ITU-T FG-TIDA Terms of Reference (TD295/P, SG17).