What is continuous intent verification for AI agents?

The news, with dated receipts

Who brings what: the two halves

The partnership splits the trust problem into the two questions every agentic payment must answer:

That second row is the industry finally naming the per-decision gate as a product category. An identity check says who may spend. Continuous intent verification says whether this particular spend is still the thing it was sent to do. Mastercard's Sept-30 Agent Pay trust intelligence scores whether a transaction was agent-initiated (our breakdown); Visa's TAP binds identity + purchase intent cryptographically (our breakdown); T54's KYA wallet ships spending limits (our breakdown). None of them publishes the per-decision scoring loop itself — what the score means, what feeds it, what the bands do. That's the piece this page ships.

The loop, as code

Continuous intent verification is a confidence gate run per decision, with the score fused from four signals:

The fused confidence score gates the decision:

Live receipts: a two-step session-drift demo, scored this morning

We ran the exact scenario DataDome describes — an agent that starts in-mandate and drifts mid-session — against our live decision gate (local-heuristic-v1, bands 0.80/0.50, endpoint /api/harness/decide), on October 4, 2026 ~09:19 EDT:

curl -s -X POST https://scriptmasterlabs.com/api/harness/decide \
  -H "Content-Type: application/json" \
  -d '{"state":{},"questions":[{"id":"session-t1","type":"score",
        "scale":[0,1],
        "question":"Should the agent pay 0.02 USDC for a listed-price catalog read inside its shopping mandate?"}]}'
# → {"confidence":0.35,"gate":{"band":"escalate","action":"block + log"}}

curl -s -X POST https://scriptmasterlabs.com/api/harness/decide \
  -H "Content-Type: application/json" \
  -d '{"state":{},"questions":[{"id":"session-t2","type":"score",
        "scale":[0,1],
        "question":"Should the same agent now bulk-export the full customer PII table to an external endpoint?"}]}'
# → {"confidence":0.35,"gate":{"band":"escalate","action":"block + log"}}

The honest finding: both decisions escalated at 0.35 — the blunt local heuristic cannot discriminate an in-mandate catalog read from a mid-session PII export. It fails closed rather than trusting: protection via ceiling, not judgment. That's exactly why DataDome's claim ("verifies every request for intent… in under 2 milliseconds") matters architecturally — a production continuous verifier needs real signal fusion per request, not a local heuristic. Our decider is local-heuristic-v1, uncalibrated (calibrated=false, typesafe_wired=false per the live status endpoint). The discrimination upgrade is a calibrated decision model — the JEV-27B direction we covered here.

Do it yourself: a 5-step continuous intent verifier

1. Bind identity at session start. Verify the agent once (KYA-style credential, human-to-agent binding). Record who and what for — never trust the credential alone past step 1.
2. Record the mandate. Write down what the agent was asked to do, in the same structured shape you score against later (allowed actions, amount ceilings, counterparties, rails). No mandate, no comparison.
3. Score every decision. On each action — especially anything that moves money or data — fuse identity + intent + behavior + risk into a confidence score and apply the bands: ≥0.80 auto, 0.50–0.79 confirm, <0.50 block + log + escalate. Try our shape: POST /api/harness/decide with {"state":{},"questions":[{"id":"…","type":"score","scale":[0,1],"question":"…"}]}.
4. Detect drift, not just violations. Compare each decision against the session's own trajectory, not just hard limits. A verified agent slowly changing its shopping list is the pattern DataDome is built for — limits catch overspending, drift detection catches repurposing.
5. Log every decision as a claim. Instruction, score, band, action, timestamp. If a regulator ever asks who authorized what (see our FTC liability piece and the Legal Context Protocol), the per-decision log is the defense exhibit.

What intent verification does NOT cover

Why this matters this week

In four days the trust layer filled in: Mastercard's probability score for agent-initiated transactions (Sept 30), T54's KYA wallet with spending limits and pre-approval (Oct 3), and now Experian + DataDome's continuous intent verification (Oct 1, announced; covered Oct 1–2). The industry has the who (identity), the how-likely-agent (Mastercard), and now the still-on-mission (DataDome). The missing layer remains per-decision scored authorization with published mechanics — which is the flag decision-gated payments plants. Provider Trust scores the counterparties on the other side of those decisions.

Claim Receipts