Sample record

From signed record to evidence pack.

This sample shows how selected signed records are assembled by hand, with context, for one review. It isn’t output from an export feature Glacis offers today.

What a record does and doesn’t establish is set out beside the one below.

Anatomy of a record.

A signed record from one supervised AI action. A valid signature confirms the record is intact and shows which key signed it, and the record reads back what the control reported. It doesn’t tell you whether the note was clinically correct, whether every encounter was captured, or whether the system is safe or compliant, and it can’t confirm that events happened the way, or at the time, the record says.

Service-witnessed
Glacis operational record

OVERT 1.1.0 · signed with Ed25519 · countersigned by a Glacis service-operated witness

Record
glc_receipt_0197…f135
Standard
OVERT 1.1.0
Signed
Ed25519, operator + witness
Chain
genesis, position 1
Recorded
2026-06-11
{
  "receipt_version": "2.0",
  "overt_version": "1.1.0",
  "id": "glc_receipt_019765f0-3a1c-7d42-9e08-c2b4d6a8f135",
  "timestamp": "2026-06-11T09:14:02.317Z",
  "methodology_version": "1.0.0",
  "demonstration_note": "Demonstration workflow data. The cryptography is real: every signature and hash in this receipt verifies in your browser.",
  "subject": {
    "organization_id": "org_example_health",
    "deployment_id": "dep_clinical_scribe_prod",
    "workflow": "ambient-scribe-draft-note",
    "model": "claude-sonnet-4-6",
    "provider": "anthropic"
  },
  "evidence": {
    "input_hash": "31ee69bf5bd68b21a89872cbd319f4d71860c23d31da5aa75b39cf77ae248981",
    "output_hash": "157cf56f011961904304e359a0f9779152de75d97eaf540f4cd4c8e6beb3a161"
  },
  "signals": {
    "binary_hash": "8445ef214d71984f41f6fad9199b2dc41d036804df5fe75bbc1e94e0fb32913f",
    "network_state_hash": "02d0770849f806acc045ab5ce99c81938c3ca276c779bed0f2977416f4a5be42",
    "epoch_token": "1781082842317",
    "context_window_hash": "ed1f4f976b6013e321979922351ee454c6dfb077019981b6bb022acbfdbe9d7e",
    "composite_hash": "64a74d744d8af6fe0805a438556209eeaa3b62c31810f52a538f4ccc98ee2626"
  },
  "controls": {
    "guardrail_action": "allow",
    "nonconformity_score": 0,
    "rules_evaluated": 14,
    "rules_triggered": 0,
    "policy_mode": "enforce",
    "phi_egress_check": "pass"
  },
  "attestation": {
    "operator_signature": "10caced1f9ca157495eacdc5fd018ab0369c20e543c10c0403e9a80532531931b6e58c450555327eecef7f8dc683300c0d6dafd115a103efbe0932fef4f3150b",
    "operator_public_key": "9c8eb1b83a59662ff76b84075dcb7f8c6b391e2e018666675211d93fc94723d8",
    "witness_signature": "999464e4c04c22744519cd01e3f7b14cb876954d8765bcc2c579c20aad0acee4d8a29d97316c689ac026419c3379bd9021f5b40317dfbb770302486998354901",
    "witness_public_key": "7c038b40b8c6183c749c92079709a32e19b0d4572befcc23e0a348e7e4df39a5",
    "previous_receipt_hash": "genesis",
    "chain_position": 1
  }
}

Verify this record yourself. The workflow data is a demonstration; the cryptography is real, and the checks run in your browser.

Verify a record

What each part means.

Each part of the record answers one question a reviewer will ask.

Policy decision

“What did the control decide?” The policy mode, the guardrail outcome and the rules the record says were evaluated. In this sample: enforce, allow, 14 rules and none triggered.

Safety controls

“Which checks ran?” Rules evaluated and triggered, the PHI egress check result and a nonconformity score, for the one event this record covers.

Model identity

“Which AI handled this?” The model, provider and configuration identifiers as reported at event time.

Cryptographic integrity

“Can anyone tamper with this?” The record is signed with Ed25519 and its contents are hashed, so a later change is detectable. This portal example also carries a countersignature and inclusion proof from a witness Glacis operates.

What leaves your environment

“What left your environment?” Hashed commitments, outcomes, signatures and metadata. No prompts, no responses and no PHI.

Timestamp and epoch

“What time was reported?” The timestamp and epoch token bound into the signed record, at the precision the record states.

From records to an evidence pack.

Selected records are assembled by hand into a scoped review packet. Which fields a reviewer needs, and whether they accept the packet, is theirs to decide.

glc_receipt_0197…f135
Allow
Individual records
Local Glacis runtime
Evidence pack assembly inside your environment
Evidence pack
assembled by hand
Scoped review packet

Coverage summary

Which workflows and controls were in scope for the period, how many in-scope decisions were recorded, and what was excluded, stated so a reviewer can reproduce the count.

Control execution report

Which controls ran on the in-scope events, with pass and fail counts. Evidence of the events the pack records, rather than a policy document saying the controls exist.

Framework mapping

How records map to ISO 42001, NIST AI RMF and EU AI Act controls. One evidence base for several frameworks. The mapping is for the reviewer’s reference and isn’t a certification claim.

Selected record index

The records in the packet, organized by date, system, reported outcome or control identifier.

OSCAL mapping reference

A mapping reference for a scoped review. It isn’t a live OSCAL export or an automatic feed into a GRC platform.

An explicit record data boundary.

Your team keeps the protected context. Reviewers receive a scoped record with what they need to check the control event.

What your team sees

  • Request and response content, kept by your team
  • Policy evaluations for the workflow
  • The attestation chain and dashboards
  • Signatures, hashes and the Glacis service witness countersignature
  • Payload data, kept in the operator environment

What this record profile carries

  • Hashed commitments (HMAC)
  • PHI and PII excluded from this profile
  • Request content excluded from this profile
  • Response content excluded from this profile
  • Raw model outputs and prompts excluded from this profile

These records are structured and queryable. Where they’re stored, and which fields are kept, depends on the deployment configuration.

Bring one real workflow.

Tell us where your AI already acts, and we’ll walk through how a packet like this one would be assembled by hand for that workflow.

Talk to us