Flow 03 · Actium + DeepTap + Veritize

Grounded, verified, routed.

Actium drives the agent loop. DeepTap grounds each call in the private corpus. Veritize verifies against ground truth. RelayGate enforces policy and signs the composed receipt.

Request path

Actium agent loop RelayGate deeptap-ground (pre) veritize-check (post) policy + metering composed receipt Model Claude / GPT DeepTap private corpus Veritize drift check

End-to-end call

// Actium task step
const result = await relay.chat({
  model: "claude-sonnet-4",
  messages: [{ role: "user", content: userQuestion }]
});
// RelayGate pre-request: DeepTap ground
//   - looks up top-k passages from tenant corpus
//   - attaches them as system grounding
// RelayGate post-response: Veritize check
//   - scores response against corpus
//   - annotates receipt with drift score
//   - blocks or flags if drift > threshold

Policy that makes it work

# config.yaml
workers:
  - id: deeptap-ground
    stage: pre-request
    config:
      corpus: tenant.corpus_id
      top_k: 5
  - id: veritize-check
    stage: post-response
    config:
      threshold: 0.35
      on_drift: annotate    # or 'block'
  - id: metering
    stage: post-response

Observability output

  • Grounding trace: top-k passage ids attached to the receipt.
  • Drift score: Veritize returns a numeric drift score per response, recorded in the receipt.
  • Composed receipt: a single Ed25519-signed record covers grounding, response, verification, and metering.

← back to flows · DeepTap · Veritize · R1