Strix · Security Review Proof Pack

AI Prior-Authorization Governance

Execution control + independently verifiable evidence for AI agents in prior-authorization workflows.

strixgov.com

One-pager · v1

Strix evaluates each state-changing action an AI prior-authorization agent takes at the moment it executes, blocks the ones that veer, and records an Ed25519-signed evidence record that a third party can verify against a public key — without trusting Strix or the vendor. This pack is the short version for a security or compliance reviewer.

1 · The artifact you can verify

Every governed action produces a signed evidence record over a canonical payload. The public half of the signing key is published as a JWKS (RFC 7517); verification uses standard Ed25519 primitives.

Signature: Ed25519 (EdDSA)
Sample record: 5686
Public keys: /.well-known/strix-jwks.json

2 · Verify it yourself (no Strix dependency)

Run the open-source verifier against the public JWKS. The verdict comes from the cryptography, not from a Strix dashboard.

npx @strixgov/verifier@latest 5686

Or fetch the keys directly at https://www.strixgov.com/.well-known/strix-jwks.json and check the signature with any Ed25519 library. A live, in-browser check is at strixgov.com/solutions/healthcare-prior-authorization.

3 · What Strix governs on a prior-auth agent

Failure mode	Ungoverned	Strix decision
Duplicate 278 (retry/replay)	Two identical authorizations reach the payer	Blocked — REPLAY (single-use authority)
Out-of-scope write-off	$1,250 adjustment no one authorized is posted	Blocked — UNAUTHORIZED (no token for that action)
Injected CPT code (out of scope)	A different procedure than approved is authorized	Blocked — SCOPE_MISMATCH (outside approved scope)
Legitimate downcode in approved family	Submits, but with no record it was in scope	Admitted — within the approved code family, and recorded
High-risk submission	High-cost study auto-fires, no human in the loop	Held — INTERCEPTED (routed for human approval)

4 · Compliance mapping

Each flag is derived from the verification outcomeof the record — never asserted by the application. Strix asserts only the controls that follow from cryptographic evidence; administrative and physical safeguards remain the covered entity's responsibility.

EU AI Act Art. 12 — record-keeping / traceability	Hash + chain + signature valid → tamper-resistant, traceable record.
EU AI Act Art. 14 — human oversight	Actor bound in the signed payload; high-risk actions require human approval before execution.
EU AI Act Art. 28 — provider obligations	Strix signs as provider with keys published on the JWKS; auditors verify against it.
HIPAA §164.312(b) — Audit Controls	Signed record present with actor + action bound (audit log integrity).
HIPAA §164.312(c)(1) — Integrity	Ed25519 signature detects any post-hoc alteration of the record.
HIPAA §164.312(d) — Person/Entity Authentication	Actor identity bound into the signed payload, verified with the signature.
NIST AI RMF — Govern / Map / Measure / Manage	Risk-tiered capability registry, content-addressable policy version, signed per-action evidence, human approval on high-risk.

5 · Scope & honest limitations

Strix governs the action — capability, actor, intent, context — and records the decision. It does not make the clinical or medical-necessity determination.
Member and claim data stays in your environment, isolated at the database with row-level security. Strix records the decision, not the PHI.
The interactive demo runs on synthetic patients and a mock payer — no PHI, no live payer, clearinghouse, or EHR.
Verification is independent: the verifier is open-source and checks signatures against the public JWKS with no dependency on Strix.