government-decision-record-audit-stream-reference

AGPL-3.0 reference implementation of government-decision-record-audit-stream. Runs the canonical PRFSA × VendorG GovDecide v3.x trajectory end-to-end against a mock Federal AI Use Case Inventory + classification-clearance gate. Proves THE FIRST Suite audit-stream with three orthogonal invariants (human-agency-officer + Federal AI Use Case Inventory entry + classification-clearance per E.O. 13526) works end-to-end in code.

Part of the Kinetic Gain Protocol Suite.

Sibling to fhir-resource-access-audit-reference (HealthTech), matter-decision-record-audit-stream-reference (LegalTech), grid-decision-record-audit-stream-reference (EnergyTech), and defense-decision-record-audit-stream-reference (DefenseTech).

What this proves

The GovTech spec ships with three independent invariants that must hold on every event. This is the most ambitious invariant design in the Suite — most verticals enforce 1-2 invariants. GovTech requires all three:

1. human-agency-officer-required — every event must carry agent.agency_officer_id_tokenized. No anonymous AI decisions in government settings. Maps to OMB M-24-10 §5(d) human-review minimum practice for rights-impacting AI.
2. Federal AI Use Case Inventory entry — every event must reference a registered agent.ai_use_case_inventory_entry_id that exists in the federal inventory AND whose agency matches the resource. Maps to OMB M-24-10 §3(a) — agencies must register all rights/safety-impacting AI use cases before deployment.
3. classification-clearance — every event's agent.clearance_level must be ≥ resource.classification along the ordered ladder (UNCLASSIFIED < CUI < CONFIDENTIAL < SECRET < TOP-SECRET). Maps to E.O. 13526 — Classified National Security Information.

The reference impl proves all three interlock: the vault enforces them at request-time; the verifier independently validates them on the produced event stream.

Architecture

orchestrator.mjs
   │
   ├─ requests access via federal-vault.mjs (enforces all 3 invariants)
   │
   ├─ builds hash-chained event via event-builder.mjs (canonical-JSON SHA-256)
   │
   └─ emits to examples/prfsa-govdecide-reference-stream.ndjson

verifier.mjs (independent, post-hoc)
   │
   ├─ chain integrity (each prev_hash = prior hash)
   ├─ invariant #1: human-agency-officer
   ├─ invariant #2: Federal AI Use Case Inventory entry
   └─ invariant #3: classification-clearance per E.O. 13526

Run it

git clone https://github.com/mizcausevic-dev/government-decision-record-audit-stream-reference
cd government-decision-record-audit-stream-reference
npm install
npm start    # orchestrates + writes the stream + runs the verifier
npm test     # 10 unit tests including vault-denial + verifier-trip cases

Expected output:

Built 3 events → examples/prfsa-govdecide-reference-stream.ndjson
OK · 3 events · chain ✓ · 3 invariants ✓ (human-agency-officer + Federal AI Use Case Inventory + classification-clearance)

Canonical trajectory — PRFSA (fictional Pacific Region Federal Services Agency)

1. Benefit eligibility pre-screened — UNCLASSIFIED, rights-impacting under OMB M-24-10 §5(d). Inventory entry PRFSA-AI-2026-014, agency officer ID required. Agent at UNCLASSIFIED clearance is sufficient.
2. Permit application classified — UNCLASSIFIED, rights-impacting. Inventory entry PRFSA-AI-2026-022. Agent at CUI clearance (higher than needed; agents are allowed to operate on lower-classified material).
3. FOIA response triaged — CUI tier, neither rights- nor safety-impacting (but STILL requires inventory entry per §3(a)). Inventory entry PRFSA-AI-2026-031. Agent at SECRET clearance comfortably ≥ CUI.

Vault denial scenarios (covered by tests)

The mock federal vault rejects requests with crisp reasons matching the four real-world failure modes:

Failure mode	Trigger	Reason in test
Unknown inventory entry	Agent references AI use case not registered with OMB	"Federal AI Use Case Inventory has no entry..."
Cross-agency mismatch	PRFSA officer trying to use ANOTHER_AGENCY's inventory entry	"Inventory entry agency does not match resource agency"
Clearance < classification	UNCLASSIFIED agent attempting SECRET resource	"Agent clearance X < resource classification Y per E.O. 13526"
Missing officer identity	Event lacks agency_officer_id_tokenized	"Missing agent_agency_officer_id_tokenized"

Why a separate AGPL-3.0 reference impl

• The spec repo (government-decision-record-audit-stream) is MIT and contains schema + example data + static verifier. It does NOT run end-to-end.
• This repo wires the federal vault + audit-stream + verifier into a runnable trajectory. AGPL-3.0 because reference implementations carry a stronger copyleft posture than the specs themselves — standing Suite rule.
• Sibling reference impls follow the same pattern. Specs MIT, reference implementations AGPL-3.0.

Composes with

• government-decision-record-audit-stream — the spec this implements
• omb-m24-10-readiness-evidence-bundle — evidence bundle that ingests events produced here
• government-ai-incident-card-profile — any verifier failure becomes a published Incident Card
• state-government-ai-disclosure-tracker — regulatory-lifecycle context (OMB memos, agency policies, state government AI laws)
• citizen-data-vault-contract-profile — vault contract for citizen data accessed by the AI
• Kinetic Gain Protocol Suite — umbrella

Compliance posture

Reference implementation readiness scaffolding for OMB M-24-10 + OMB M-24-18 + AI Bill of Rights + Section 508 + Privacy Act + FOIA + NIST AI RMF + FedRAMP. Does NOT constitute OMB compliance attestation, agency Authority-to-Operate (ATO), FedRAMP authorization, or classification-handling approval. The mock federal vault + inventory are in-memory — production deployments must use the real Federal AI Use Case Inventory + agency identity provider (PIV/CAC) + classified environment infrastructure. Per the standing Suite public-language guardrail: readiness · evidence · posture · controls · scaffolding — never "compliant" / "certified" without external attestation.

License

AGPL-3.0-only. Spec repos this depends on remain MIT.