VerifAIdentity

Methods

Every number has a method you can check.

Each method is versioned. A change to weights, rules or thresholds is a new version, recorded with the evidence it produced.

Posture score PS-1.0

40% control health + 35% exposure + 25% evidence coverage

  • Control health: share of compliance clocks not failing; fixed late counts half; clocks that cannot be verified are left out here and lower coverage instead.
  • Exposure: 100 minus open findings, weighted critical 25, high 12, medium 4, low 1.
  • Evidence coverage: share of per-system checks backed by fresh evidence.
  • Left out: informational and simulated findings, retired clocks, systems not connected.

Always shown with its coverage. In our lab: 82, from control health 96, exposure 77 and coverage 68%, read as "few problems in what we can see".

Access layers AL-1.0

Each account in five layers: account state, sign-in possible, sessions, use after the trigger, what it holds. The conclusion is the strongest the evidence supports:

  • Used: a successful sign-in after the trigger was observed (proven access)
  • Can sign in: active, and nothing we can read blocks sign-in (possible access)
  • Blocked: active, but a lock or a block-all policy we can read stops sign-in
  • Off, record only (governance, HR) or cannot tell

Layers a vendor does not expose read-only are reported as not observable, never assumed.

Evidence chain EC-1.0

entry = SHA-256(previous | seq | time | kind | record | content)

  • Each entry is signed with HMAC-SHA256; the key never enters the database.
  • Verification checks links, signatures, and that every current record matches its latest entry.
  • Against a head kept elsewhere, a shortened or rewritten chain fails.
  • Ten automated tests cover edits, rewrites, deletions, truncation and key handling.

Identity types IG-1.0

  • Human: member, guest, emergency (break-glass), governance identity
  • Workload: service principal or managed identity (from Entra, high confidence); service account by name with no person match (low confidence)
  • AI agent: Entra agent tag (medium confidence) or name only (low confidence)
  • Owner: person, owned, multiple, unowned, vendor, or unknown when owners were not collected

Reconciliation always reports matched and unmatched accounts with their denominators.

Answer evaluation

ExpectedCasesResult
Should answer3434 correct, 0 wrong, 0 wrongly declined
Should decline66 declined, 0 answered anyway

Every evidence answer must cite its records (34 of 34 did). Limits: we wrote the set, on lab data; in a pilot the design partner's expert writes and judges it.

Detection test cases

Each detection is tested with simulated events in our lab tenants only (never on real data), at the default thresholds shown on the Platform page.

  • Covered by automated tests: spray, fail-then-success, one person failing in two systems, leaver use, failure spike
  • Not tested yet: brute force against one person, admin change by an unknown identity
  • Not measured yet: false-positive rates on real traffic, and detection rates on attacks we did not write ourselves
  • Severity follows volume: for example, fewer than five sign-in attempts for an unknown username is low

See it on your own identity data.

A read-only pilot with Okta, Microsoft Entra ID and your HR export. 30 to 60 days, measured against targets we agree up front.