Supports Good Proof Stamps
AI GOLD for Duty of Care
Evidence that holds up when outcomes are appealed.
AI GOLD is a programme-scoped duty-of-care evidence workflow referenced by a Good Proof Stamp. It helps you prove what was considered, within scope, within an evidence window — without dumping sensitive data.
Fail-closedMinimal disclosureAppend-only history
Not a certification. Scope-limited verification. Acceptance depends on counterparty or programme requirements.
Why buyers like this
Appeal-ready evidence
Structured artifacts that hold up when outcomes are challenged.
Minimal disclosure by default
No raw prompts, logs, or PII unless programme requires.
Verifiable outside your perimeter
Counterparties check the Status Link—no portal, no login.
Definitions
AI GOLD
A programme-scoped duty-of-care evidence workflow that produces structured evidence referenced by a Good Proof Stamp. Not a model; not a certification.
Evidence Window
The time-bounded period during which the evidence bundle applies (start/end timestamps).
Evidence Bundle
The structured set of evidence items produced by AI GOLD for a given action, referenced by the Stamp.
Minimal Disclosure
Default mode where raw prompts, logs, and PII are excluded. Disclosure is scope-limited and buyer-configurable.
IDA Evidence Pack
A time-stamped snapshot of evidence artifacts that can be forwarded to counterparties without portal access.
Status Link
The machine-checkable verification endpoint returning current status (VALID, NEEDS_REFRESH, WITHDRAWN, EXPIRED, NOT_VERIFIED).
What AI GOLD is
A structured evidence workflow that produces a duty-of-care evidence bundle referenced by the Stamp.
What AI GOLD is not
Not a model. Not a certification. Not a guarantee of correctness. Not a replacement for clinical/legal judgment.
What's in the AI GOLD Evidence Bundle
Minimum set (no sensitive content exposed by default).
action_class (what type of action was stamped)
stamp_id reference (links to the verifiable Stamp)
evidence_window (start/end timestamps)
policy_version / programme_id
model/tool inventory identifiers (not raw prompts)
applicable rules/checks invoked (names/IDs)
human_involvement indicator (yes/no + role if applicable)
decision_rationale summary fields (structured, non-PII)
exceptions/overrides recorded (if any)
verification metadata: signer, verified_at, expires_at, scope_hash
Optional / programme-defined
- •Sector-specific attestations (e.g., vulnerability flags for utilities, safeguarding checks for education)
- •Lane-specific rule references (e.g., hardship thresholds, eligibility criteria IDs)
- •Guardian involvement reference (panel ID, finality timestamp) when human review occurred
- •Extended audit fields per programme compliance requirements
Security & Chain-of-Custody
Integrity establishment
- •Evidence bundle is cryptographically referenced via
scope_hashbinding - •Verification metadata includes
signer,verified_at, andexpires_at - •Each evidence artifact is time-stamped and correlated to
stamp_id
Counterparty verification (no special access required)
- •Query the Status Link to confirm status, scope, and expiry
- •Receive the IDA Evidence Pack (time-stamped snapshot) for audit records
- •Confirm
scope_hashmatches expected evidence binding
Minimal Disclosure & Data Protection
What is NOT included by default:
- Raw prompts or model inputs
- Full system logs or debug output
- PII, PHI, or case-specific identifiers (unless programme requires)
- Internal system identifiers or infrastructure details
Disclosure is scope-limited and buyer-configurable. Programmes define what must be captured for their duty-of-care lane.
This supports DPA-friendly data minimization; buyers define what must be captured for their duty-of-care lane.
When you need AI GOLD
High-impact action classes that benefit from duty-of-care evidence.
Best fit (duty-of-care heavy)
Hardship determinations
Eligibility decisions affecting access to essentials
Irreversible actions with human impact (closures, bans, denials, disconnections)
Exceptions/overrides where a human would normally be expected
Appeal-required outcomes
Sometimes (only when the lane requires duty-of-care evidence)
Claims adjudication (life-impact / hardship-linked claims)
High-value transfers when a duty-of-care obligation exists (e.g., vulnerability protections, safeguarding rules)
What this replaces
From manual proof gathering to structured, verifiable evidence.
Screenshot packs and narrative memosOne evidence workflow that stamps reference
"Trust us" emails to auditors/partnersStatus Link + IDA Evidence Pack
Repeated rework per appeal bodyReusable snapshot with defined scope
Silent drift after policy/model updatesNEEDS REFRESH / WITHDRAWN semantics
Example: appeal / oversight check
What a reviewer actually checks, cites, and how fail-closed applies.
- 1
Reviewer queries the Status Link
Checks: status == VALID, scope matches action, expires_at > now. No login required.
- 2
Confirms scope + evidence window + signer authority
Validates the decision was within defined boundaries at time of action.
- 3
Cites the IDA Evidence Pack in the appeal record
Time-stamped snapshot filed as audit evidence. stamp_id, scope_hash, verified_at all referenced.
If NOT_VERIFIED, NEEDS_REFRESH, WITHDRAWN, or EXPIRED
Decision cannot be relied on. Reviewer escalates or blocks reliance. Fail-closed.
Responsibilities (RACI-lite)
Supplier responsibilities
- •Produce workflow outputs (evidence bundle) for defined action classes
- •Expose verification references (stamp_id, Status Link, scope_hash)
- •Maintain Status Link availability and accuracy
- •Transition status on material changes (NEEDS_REFRESH) or withdrawal
- •Provide audit evidence on request per [NOTICE PERIOD]
Buyer responsibilities
- •Define action classes requiring AI GOLD-backed Stamps
- •Configure evidence requirements per programme/lane
- •Operate the Gate (enforce fail-closed semantics)
- •Retain audit logs per [RETENTION PERIOD]
- •Update allowlists on verifier/signer changes
Procurement & Contract Hooks
AI GOLD is referenced by Good Proof Stamps and can be required for selected action classes in your vendor agreements. When you require AI GOLD-backed Stamps, you're requiring structured duty-of-care evidence with verifiable chain-of-custody.
See Clause Pack for template language including evidence windows, scope binding, and minimal disclosure terms.
Sample RFP language (paste-ready):
"High-impact action classes SHALL require AI GOLD-backed Stamps with verifiable Status Link and IDA Evidence Pack. Actions taken without VALID status SHALL be treated as unverified."
Control Mapping (Non-Claiming)
Reference alignment to common control frameworks. Mapping only; not a certification claim.
SOC 2 (Type II)
| Control | Family | AI GOLD Mapping |
|---|---|---|
| CC7.1–7.2 | Change Management | Evidence bundle versioning, NEEDS_REFRESH triggers |
| CC7.3–7.4 | Logging & Monitoring | Evidence window, verification metadata |
| CC6.1–6.3 | Logical Access | Scope binding, minimal disclosure |
| CC8.1 | System Operations | Status transitions, withdrawal events |
ISO 27001:2022
| Control | Family | AI GOLD Mapping |
|---|---|---|
| A.8.15–8.16 | Logging & Monitoring | Evidence bundle, audit metadata |
| A.5.19–5.23 | Supplier Relationships | Verification references, scope binding |
| A.8.24 | Communications Security | Status Link integrity, HTTPS-only |
NIST SP 800-53
| Control | Family | AI GOLD Mapping |
|---|---|---|
| AU-* | Audit & Accountability | Evidence bundle, verification logs |
| CM-* | Configuration Management | Policy/model versioning, refresh triggers |
| AC-* | Access Control | Scope-limited disclosure, programme-gated access |
| SC-* | System & Comms Protection | Status Link integrity, chain-of-custody |
| SI-* | System & Info Integrity | Fail-closed semantics, withdrawal propagation |
Not a certification. This mapping shows alignment with control families. Good Proof does not claim SOC 2, ISO 27001, or FedRAMP certification. Buyers should assess fit-for-purpose based on their own compliance requirements.
What counterparties receive
Verifiable reference (Stamp ID plus Status Link)
Scope boundaries plus evidence windows
Signer authority reference (system or Guardian panel when required)
Minimal disclosure by default
Counterparty Acceptance Checklist
Status Link resolves and verifies from Official Verifier
Evidence window + scope match the action class
Evidence bundle integrity checks pass (hash/signature reference)
If NEEDS_REFRESH / WITHDRAWN / NOT_VERIFIED: stop-rely and re-verify
Frequently Asked Questions
Make duty-of-care evidence shippable.
Define your high-impact action classes. Require the Stamp. Let the Status Link do the rest.
Not a certification. Scope-limited verification. Acceptance depends on counterparty or programme requirements.