Signed evidence for AI-driven underwriting.
A sample of how a global P&C carrier could use The Hive Vault to prove that every model-assisted underwriting decision was accompanied by a signed, regulator-verifiable evidence packet. No real customer data is shown here.
Engagement parameters
Thesis
Carriers cannot defend an AI-assisted underwriting decision in court without a signed chain that ties model version, input features, and human reviewer to the bound policy. The Vault produces that chain at decision-time, not after the fact.
What gets signed
Each underwriting decision generates a receipt with the following bound elements:
- Model version + rubric IDed25519-signed
- Input feature hash (privacy-preserving)blake2b-256
- Reviewer DID + decision-time delegationdid:web:carrier-ops
- Bound policy ID + premium quoteanchored on Base
Sample stats
Illustrative · not a claimed deploymentSample portfolio: Tier-1 P&C carrier, AI-assisted underwriting on 1.2M annual personal-auto + small-commercial decisions. Numbers are illustrative; a real Vault is filled in with the buyer's own deployment.
Sample evidence packet — anonymized fields
Sample data · field shape onlyEach bound policy emits one receipt. This is the field shape; values are illustrative.
| Field | Type | Sample value | Bound to |
|---|---|---|---|
| decision_id | uuid | 01JR4K9... | policy_id |
| model_version | semver | uw-pricing/3.7.2 | rubric_id |
| features_hash | blake2b | 9f3c...a201 | input row |
| reviewer_did | did:web | did:web:carrier-ops | delegation |
| premium_quote_usd | decimal | 1,284.50 | policy_id |
| bind_chain | chain | base-8453 | tx_hash |
| sig_alg | alg | ed25519 | issuer DID |
Sample ROI — defending an AI underwriting decision
Sample ROI · illustrative mathCost-to-defend a single contested AI-priced policy through civil adverse-action litigation (separate from a DOI market conduct exam). Contest rate and settlement reserve are user-adjustable below — your numbers will differ materially by carrier, line of business, and geography. This illustrative model does not represent a commitment to achievable savings.
- Underwriting decision ID · model version · reviewer DID — signed at decision-time
- Adverse-action code bound to receipt; meets state DOI disclosure obligation documentation requirements
- Programmatic receipt retrieval replaces forensic model reconstruction for contested decisions
- Covers personal-auto book at normalized contest rate; reduces per-decision defense cost
Contest rate is user-supplied. Actual rates vary by line of business, state jurisdiction, and carrier size. This model does not represent a commitment to achievable savings.
- Input feature hash (privacy-preserving) · adverse-action code · bound premium quote — full audit trail
- State DOI disclosure flag per decision; NAIC AI Model Bulletin alignment signal
- Covers small-commercial book with elevated adverse-action litigation exposure
- Reviewer delegation chain proves authority level at each underwriting step
Contest rate is user-supplied. Actual rates vary by line of business, state jurisdiction, and carrier size. This model does not represent a commitment to achievable savings.
- NAIC-aligned model audit trail · ML-DSA-ready sig slot · state DOI disclosure epoch — full long-tail coverage
- P&C policies litigated 7-10 years post-bind: receipts re-verifiable under current court-accepted algorithms post-2035
- Covers long-tail liability and specialty lines where re-signing cost under deprecated algorithm is unquantifiable
- Per-record pq_sig and pq_alg reservation enables epoch flip without re-signing underlying decisions
Contest rate is user-supplied. Actual rates vary by line of business, state jurisdiction, and carrier size. This model does not represent a commitment to achievable savings.
Cost of NOT being on this tier
P&C policies routinely surface in litigation 7-10 years after binding. Cost of re-signing decisions under a new algorithm if Ed25519 is deprecated before that litigation surfaces creates an unquantifiable re-issuance gap. E-discovery cost of reprocessing signed receipts if the court system's verification tools do not support Ed25519 post-deprecation applies across the full retention window.
Without Hive Vault
With Hive Vault
Per-decision delta: $191,300. Sample annualized exposure reduction (slider-computed at your contest rate): $25.0M. Vault substrate cost at this volume runs single-digit millions.
Post-quantum readiness
For high-security buyersDesigned-to-support · ML-DSA-ready
PQ-readiness for long-tail liability — receipts must verify in 2030+ even if Ed25519 is later weakened.
P&C policies routinely surface in litigation 7-10 years after binding. The receipt format reserves pq_sig and pq_alg fields on every record so the same evidence packet can be re-anchored under ML-DSA / Dilithium-class signatures during epoch flip — without re-signing the underlying decision.
Sample receipt
This is the structure of the signed receipt your evidence room produces. Every field is verifiable offline against the issuer's public key.
Activation
In a real Vault this section is a live deeplink. The buyer pays USDC on Base; the receipt above is signed at confirmation; the dashboard URL + tenant API key issue inline.
activate · USDC · 0x15184Bf50B3d3F52b60434f8942b7D52F2eB436E · ERC-681 · evidence bundle attached
How a real Vault differs from this sample
A real Vault is invite-only and contains the actual buyer's name, the negotiated economic terms, the live activation deeplink against a specific tx, and the signed receipt chain. Each Vault is gated by a six-word passphrase issued at intake.