Assurance Evidence Schema Governance requires that organisations standardise the evidence objects, fields, digital signatures, and retention rules used to prove conformance across all governance controls. Every control dimension in the Agent Governance Standard produces evidence — blocked action logs, configuration snapshots, test results, audit trails. Without a standardised schema governing these evidence artefacts, each control produces evidence in ad hoc formats, making cross-control analysis impossible, automated validation unreliable, and auditor review prohibitively expensive. The evidence schema MUST define the structure, mandatory fields, signature requirements, and retention periods for every evidence type, and MUST be versioned and change-controlled alongside the taxonomy and dependency register.
Scenario A — Heterogeneous Evidence Formats Block Automated Audit: An organisation implements 47 governance controls across 8 AI agents. Each control was implemented by a different team, and each team chose its own evidence format: some produce JSON logs, others CSV exports, others PDF reports, and two produce proprietary binary formats from commercial governance tools. When the annual conformance assessment begins, the assessor spends 340 hours (£68,000 at external audit rates) simply normalising evidence into a comparable format before substantive assessment can begin. Three controls produce evidence that cannot be correlated with evidence from related controls because timestamps use different formats (Unix epoch vs. ISO 8601 vs. local time without timezone), agent identifiers use different naming conventions, and action references are not cross-indexed.
What went wrong: No evidence schema existed. Each implementation team made independent format decisions. Evidence objects lack common mandatory fields (standardised timestamps, agent identifiers, control references). Consequence: £68,000 in normalisation costs per assessment, 3 controls with uncorrelatable evidence, overall assessment confidence reduced because cross-control analysis was impossible.
Scenario B — Unsigned Evidence Undermines Regulatory Submission: An organisation submits conformance evidence to a financial regulator as part of its AI agent authorisation process. The evidence includes 18 months of blocked action logs from AG-001 enforcement. During the regulator's review, the regulator asks: "How do you demonstrate that these logs have not been modified since creation?" The organisation cannot answer — the logs are plain text files on a shared network drive with no digital signatures, no hash chains, and no tamper-evident storage. The regulator declines to accept the evidence, requiring the organisation to re-implement its evidence pipeline with integrity controls before resubmitting. The re-implementation takes 6 months. During this period, the agent authorisation is suspended.
What went wrong: The evidence schema did not specify integrity requirements (digital signatures, hash chains, tamper-evident storage). Evidence was produced and stored in a format that could be modified without detection. Consequence: 6-month authorisation suspension, £420,000 in re-implementation costs, reputational damage with the regulator.
Scenario C — Inconsistent Retention Rules Create Legal Exposure: An organisation faces litigation related to an AI agent's actions 4 years ago. The organisation's legal team requests all governance evidence for the relevant period. Three controls retained evidence for 7 years (meeting financial services requirements), four controls retained evidence for 3 years (meeting the minimum default), and two controls retained evidence for only 1 year (set by the implementation team based on storage cost concerns). The evidence for the two short-retention controls has been purged. These controls are material to the litigation because they would have demonstrated that the agent was operating within governed parameters. The organisation cannot prove its governance posture at the time of the incident.
What went wrong: No standardised retention rules existed. Each implementation team set retention independently, without reference to regulatory or legal requirements. Consequence: Inability to demonstrate governance posture for a material litigation, adverse inference drawn by the court, settlement at £1.8 million — 4x the estimated judgment with evidence.
Scope: This dimension applies to all evidence artefacts produced by any governance control within the Agent Governance Standard. "Evidence" means any data object created to demonstrate that a governance control is implemented, operating, or effective. This includes but is not limited to: action logs, configuration snapshots, test results, assessment reports, approval records, audit trails, and performance metrics. The scope covers the schema definition (structure and fields), integrity requirements (signatures and tamper evidence), retention rules, and access controls for all evidence types. Organisations using third-party governance platforms must verify that the platform's evidence output conforms to the organisation's evidence schema or must implement a schema-conforming extraction layer.
4.1. A conforming system MUST define a versioned evidence schema specifying the structure, mandatory fields, data types, and validation rules for every evidence type produced by implemented governance controls.
4.2. A conforming system MUST include the following mandatory fields in every evidence object: unique evidence identifier, creation timestamp in ISO 8601 format with timezone, source control dimension identifier (AG number), source agent identifier, evidence type classification, and schema version reference.
4.3. A conforming system MUST apply digital signatures or cryptographic hash chains to evidence objects at creation time, enabling subsequent verification that the evidence has not been modified since creation.
4.4. A conforming system MUST define and enforce retention periods for each evidence type, aligned with the most stringent applicable regulatory, legal, and contractual requirement, with a minimum default of 3 years for unregulated contexts.
4.5. A conforming system MUST validate evidence objects against the schema at creation time, rejecting or flagging evidence that does not conform to the defined structure and mandatory field requirements.
4.6. A conforming system MUST version the evidence schema using immutable version identifiers, with full change history, and must maintain backward compatibility or provide explicit migration paths when the schema evolves.
4.7. A conforming system SHOULD implement a centralised evidence repository or federated evidence index enabling cross-control evidence correlation using the standardised fields.
4.8. A conforming system SHOULD generate automated alerts when evidence production for any control ceases unexpectedly, drops below expected volume by more than 50% for 24 hours, or fails schema validation at a rate exceeding 5% of evidence objects.
4.9. A conforming system MAY implement evidence provenance chains — metadata recording the complete lifecycle of each evidence object from creation through any transformations to archival or purge.
Evidence is the currency of assurance. A governance control that cannot produce credible, structured, verifiable evidence of its operation provides no assurance value — it is an assertion without proof. The Agent Governance Standard's Section 7 (Evidence Requirements) and Section 8 (Test Specification) of every control dimension specify what evidence is required and what tests must pass. AG-221 governs the evidence infrastructure itself — ensuring that evidence produced across all controls is structured consistently, integrity-protected, and retained appropriately.
Three forces make evidence schema governance essential. First, scale: an organisation with 20 agents and 50 implemented controls may produce millions of evidence objects per year. Without a standardised schema, these objects cannot be indexed, queried, or correlated at scale. Second, integrity: evidence submitted to regulators, auditors, or courts must be demonstrably unmodified. Ad hoc evidence formats without integrity protection are vulnerable to challenge. Third, interoperability: evidence produced by different controls must be correlatable — a blocked action log from AG-001 must be cross-referenceable with the configuration version from AG-007 that defined the limit. This requires common identifiers, common timestamp formats, and common reference fields.
Evidence schema governance is also a precondition for automated conformance assessment. As governance frameworks mature, manual assessment becomes unsustainable. Automated assessment tools require structured, schema-conforming evidence as input. Without a governed schema, every automated tool must implement custom parsers for every evidence format — a combinatorial problem that does not scale.
The evidence schema should be implemented as a formal schema definition — JSON Schema, Protocol Buffers, Avro, or equivalent — that can be used for automated validation. The schema should define: (1) a common envelope containing the mandatory fields required by 4.2 for all evidence types, and (2) type-specific payloads for each evidence category (action logs, configuration snapshots, test results, etc.).
Recommended patterns:
Anti-patterns to avoid:
Basic Implementation — The organisation has defined an evidence schema specifying mandatory fields and data types for all evidence types. Evidence objects are validated against the schema at creation time. Retention periods are defined for each evidence type and enforced through manual processes with periodic verification. Digital signatures are applied to evidence objects at creation time. The schema is versioned with change history.
Intermediate Implementation — The evidence schema is machine-readable (JSON Schema or equivalent) and integrated with evidence production pipelines for automated validation. A centralised evidence repository or federated index enables cross-control evidence correlation. Retention lifecycle management is automated — evidence is archived and purged on schedule with audit trails. Schema evolution follows defined compatibility contracts. Evidence production monitoring generates alerts when production volume drops unexpectedly or validation failure rates exceed thresholds.
Advanced Implementation — All intermediate capabilities plus: evidence provenance chains record the complete lifecycle of each evidence object. The evidence schema is independently audited annually for completeness and regulatory alignment. Automated evidence analysis tools consume schema-conforming evidence for continuous conformance monitoring. Evidence integrity is verified periodically (e.g., hash chain validation) to detect storage-layer tampering. Cross-organisation evidence exchange uses the standardised schema for supply-chain assurance.
Required artefacts:
Retention requirements:
Access requirements:
Test 8.1: Mandatory Field Presence Validation
Test 8.2: Schema Validation at Creation
Test 8.3: Evidence Integrity Verification
Test 8.4: Retention Period Enforcement
Test 8.5: Cross-Control Evidence Correlation
Test 8.6: Schema Version Immutability
| Regulation | Provision | Relationship Type |
|---|---|---|
| EU AI Act | Article 12 (Record-Keeping) | Direct requirement |
| EU AI Act | Article 17 (Quality Management System) | Supports compliance |
| SOX | Section 802 (Criminal Penalties for Document Alteration) | Supports compliance |
| SOX | Section 404 (Internal Controls) | Supports compliance |
| FCA SYSC | 9.1 (Record-Keeping) | Direct requirement |
| NIST AI RMF | GOVERN 1.4 (Documentation) | Supports compliance |
| ISO 42001 | Clause 7.5 (Documented Information) | Direct requirement |
| DORA | Article 10 (ICT-Related Incident Reporting) | Supports compliance |
| GDPR | Article 5(2) (Accountability Principle) | Supports compliance |
Article 12 requires providers of high-risk AI systems to ensure automatic recording of events ("logs") with sufficient detail to trace the system's operation. AG-221 implements this requirement at the governance level — standardising the format, integrity, and retention of all governance evidence ensures that regulatory record-keeping obligations are met consistently across all controls, not just for operational logs but for the full evidence portfolio that demonstrates governance effectiveness.
Section 802 establishes criminal penalties for knowingly altering, destroying, or concealing documents or records with intent to impede an investigation. For AI governance evidence, this means evidence integrity is not merely a governance best practice but a legal obligation. AG-221's requirements for digital signatures, hash chains, and enforced retention directly support compliance by making evidence modification detectable and premature destruction technically prevented.
Clause 7.5 requires the organisation to manage documented information needed by the AI management system, including ensuring its availability, suitability, and adequate protection. AG-221 provides the schema-level governance that ensures documented information (evidence) is structured consistently, integrity-protected, and retained appropriately.
SYSC 9.1 requires firms to arrange for orderly records to be kept of their business and internal organisation. For AI agent governance, this extends to the evidence demonstrating that governance controls are operating effectively. The FCA expects records to be producible in a format that enables regulatory analysis — ad hoc evidence formats do not meet this expectation.
| Field | Value |
|---|---|
| Severity Rating | High |
| Blast Radius | Organisation-wide — affects the credibility of all governance evidence across all controls |
Consequence chain: Without evidence schema governance, evidence produced by individual controls is heterogeneous, unverifiable, and uncorrelatable. The immediate failure mode is assessment inefficiency — auditors and assessors spend disproportionate time normalising evidence rather than evaluating it. The more severe failure mode is evidence integrity failure — evidence submitted to regulators or courts that cannot be demonstrated to be unmodified is vulnerable to challenge and may be rejected. The downstream consequence is loss of governance credibility: an organisation that cannot produce structured, integrity-protected evidence of its governance operations cannot demonstrate conformance. In litigation, missing or unreliable evidence triggers adverse inferences. In regulatory proceedings, it triggers findings of inadequate record-keeping.
Cross-references: AG-219 (Control Taxonomy Governance) defines the control set for which evidence schemas must exist. AG-007 (Governance Configuration Control) governs the versioning of the evidence schema itself. AG-056 (Independent Validation) consumes schema-conforming evidence as input to validation activities. AG-157 (External Conformance Assessment) requires structured evidence for external assessor review. AG-153 (Control Efficacy Measurement) produces evidence objects that must conform to the evidence schema.