insurance-decision-record-audit-stream-reference

AGPL-3.0 reference implementation of insurance-decision-record-audit-stream. Runs the Coastguard Insurance × VendorI ClaimsTriage v3.x trajectory end-to-end. 2 invariants: human-adjudicator-required SCOPED (adverse-action-capable kind AND adverse recommendation — deliberate contrast to PropTech's universal rule) + NAIC Model Bulletin on AI 90-day backward bias-monitoring window.

Part of the Kinetic Gain Protocol Suite.

Completes coverage of all 10 verticals with reference implementations:

Vertical	Reference impl
HealthTech	fhir-resource-access-audit-reference
LegalTech	matter-decision-record-audit-stream-reference
EnergyTech	grid-decision-record-audit-stream-reference
DefenseTech	defense-decision-record-audit-stream-reference
GovTech	government-decision-record-audit-stream-reference
FinTech	financial-decision-record-audit-stream-reference
HR Tech	employment-decision-record-audit-stream-reference
EdTech	student-data-access-audit-stream-reference
PropTech	mortgage-decision-record-audit-stream-reference
InsurTech	this repo

What this proves

InsurTech sits between FinTech (scoped human rule) and PropTech (universal human rule) on the human-required dial:

1. Invariant #1 — human_adjudicator_required SCOPED to adverse-action-capable kinds AND adverse recommendations. A "fast-track-approve" triage outcome doesn't need an adjudicator. A "partial-pay" or "investigate-for-fraud" does. The deliberate contrast with PropTech shows the Suite's range — same shape, parameterized differently per vertical's regulatory posture.

2. Invariant #2 — NAIC Model Bulletin AI 90-day bias-monitoring window (backward-looking, must-precede-event). Adopted Dec 2023; ~25 US states have adopted/in-progress. Encodes "AI bias-monitoring must have been completed within the prior 90 days" as a hard precondition for any adverse-action-capable event.

Architecture

orchestrator.mjs (4 steps: filed → triage → fraud-flag → partial-pay)
   │
   ├─ requests access via insurance-vault.mjs (scoped adjudicator + 90-day bias window)
   │
   ├─ builds hash-chained event via event-builder.mjs (canonical-JSON SHA-256)
   │
   └─ emits to examples/coastguard-claimstriage-reference-stream.ndjson

verifier.mjs (independent, post-hoc)
   │
   ├─ chain integrity
   ├─ invariant #1: SCOPED human-adjudicator (only on adverse rec on adverse-action-capable kind)
   └─ invariant #2: NAIC 90-day bias-monitoring window

Run it

npm install && npm start
# → Built 4 events → examples/coastguard-claimstriage-reference-stream.ndjson
# → OK · 4 events · chain ✓ · 2 invariants ✓
npm test    # 11 unit tests

Canonical trajectory — Coastguard Insurance × VendorI ClaimsTriage v3.x

Step	Event	Adverse-action-capable?	Recommendation	Adjudicator required?
1	Claim filed	No	—	No
2	AI triage — fast-track approve	Yes	fast-track-approve (non-adverse)	No (scoped rule shields it)
3	AI fraud-flag recommendation	Yes	investigate-for-fraud	Yes — both invariants fire
4	Partial-pay determination	Yes	partial-pay	Yes — both invariants fire

All adverse-action-capable events also carry a bias_monitoring_window_completed_at stamp pointing at 2026-09-15 — well within the 90-day window for events through 2026-11-15.

Adverse-action-capable kinds (invariant scope set)

Kind	Trigger
triage-recommended	First-pass claim disposition
application-decision-recommended	New-business underwriting
rating-decision-recommended	Pricing/rate-up calls
fraud-flag-recommended	SIU referral

Adverse recommendations (scoped invariant #1 trigger)

decline · rate-up · approve-with-conditions · deny · partial-pay · investigate-for-fraud

Vault denial scenarios (covered by tests)

Failure mode	Reason
Adverse rec on adverse-action-capable kind WITHOUT adjudicator	"requires human_adjudicator_id_tokenized (NAIC + unfair-discrimination posture)"
Adverse-action-capable kind WITHOUT bias-monitoring window	"requires bias_monitoring_window_completed_at (NAIC Model Bulletin on AI bias monitoring)"
Bias-monitoring window completed >90 days before event	"NAIC bias-monitoring window stale"
Bias-monitoring window timestamped AFTER event	"monitoring must precede the AI decision"

Composes with

• insurance-decision-record-audit-stream — the spec this implements
• naic-ai-bulletin-readiness-evidence-bundle — evidence bundle
• unfair-discrimination-incident-card-profile — invariant failures
• state-insurance-ai-disclosure-tracker — state-by-state NAIC adoption lifecycle
• policyholder-data-vault-contract-profile — Decision Card vault contract
• Kinetic Gain Protocol Suite — umbrella

Compliance posture

Reference implementation readiness scaffolding for NAIC Model Bulletin on AI + state unfair-discrimination + unfair-claims-settlement + anti-fraud statutes. Does NOT constitute NAIC audit qualification, state DOI examination clearance, or substitute for an insurer's own ML model risk management framework. The mock vault is in-memory — production deployments must integrate with the policy admin system + claims platform + SIU workflow + actuarial bias monitoring pipeline. 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.