When Is an AI Agent Allowed to Pay Again?

Imagine an AI agent paying for a resource, service or payout. It sends the request. The network times out. The agent does not know whether money moved. That is the familiar duplicate-payment problem. But when we tried to turn that problem into an executable verification standard, every public contract we examined removed one hidden assumption from the model.

The question changed from “What happened?” to “What exactly authorizes the next monetary action?”

The first model was too simple.

payment attempt→timeout→reconcile→success / failure→retry / stop

This hides several independent questions inside reconcile and failure. A locally reasonable retry can still violate the economic intent.

Next-action invariant

The prior outcome needs sufficient evidence.

The prior financial action needs a known enough finality state.

A new monetary action needs its own authority.

If value moved, fulfillment or compensation may also need reconciliation.

Correction one: evidence authority is separate.

In the Crossmint public-contract study, the reviewed public surfaces exposed useful transaction evidence but did not establish a canonical precedence rule across all potentially conflicting evidence surfaces.

multiple plausible observations
        ↓
no documented canonical precedence
        ↓
EVIDENCE AUTHORITY = UNRESOLVED
        ↓
fail closed

Evidence existence ≠ evidence authority. Several observations do not automatically tell an agent which observation is allowed to decide the financial future.

Correction two: final failure is not retry authority.

The PayRam public-contract pass exposed another assumption. A terminal state can tell us something about the prior execution without documenting whether a new monetary action is safe or authorized.

Finality is evidence about the past. Retry authority is permission for the future.

prior attempt FAILED≠retry authorized

A retry may require new approval, a new request identity, business reconciliation, delay or provider-specific recovery semantics. When those semantics are unresolved, failure does not silently become permission.

Correction three: payment finality is not fulfillment finality.

x402 pushed the model further. Payment settlement can be established while the final application response or paid-for resource is still not observed because transport fails afterward.

PAYMENT FINALITY
      ≠
FULFILLMENT FINALITY

COMMITTED(payment)
∧ UNKNOWN(fulfillment)
→ RECONCILE / HOLD
→ NO REPURCHASE YET

If an agent interprets “resource not observed” as “payment did not work,” a second purchase can be valid in isolation and still violate the original economic invariant.

The four-question model.

Agent payment decision coordinates

Evidence authority
Which observation is allowed to establish the state?

Payment finality
Did the prior monetary action commit, fail, remain pending or remain unknown?

Retry authority
Is another monetary execution actually permitted?

Fulfillment finality
If money moved, was the expected result delivered, refunded, compensated or unresolved?

There is no universal rule saying failed → retry, timeout → retry, or even committed → done. The next action depends on the causal contract of the workflow.

From research notes to a decision gate.

AUTHORITY
   ↓
EVIDENCE AUTHORITY
   ↓
PAYMENT FINALITY
   ↓
RETRY AUTHORITY
   ↓
FULFILLMENT FINALITY
   ↓
ALLOW / HOLD / STOP / RECONCILE / COMPENSATE

ALLOW is intentionally hard to earn. Missing or contradictory evidence should request more evidence, reconciliation or compensation rather than fabricate certainty.

Why a risk score is not enough.

“87% safe” does not answer the operational question at the money boundary. A consequential agent needs to know what action is allowed now and why.

observed state→invariant→missing / satisfied proof→decision

The implementation therefore produces a deterministic Evidence Pack: canonical input, recomputed machine decision, customer-readable summary and integrity manifest. The summary can be checked against the underlying state instead of trusted as narrative.

What the three public signals actually prove.

They do not prove that Crossmint, PayRam or x402 is insecure. They are not production audits, endorsements or vulnerability claims. Each study is bounded to the public contract reviewed at the time.

Crossmint

Do not invent evidence precedence when the public contract leaves it unresolved.

PayRam

Do not turn a terminal failure into retry permission without documented retry semantics.

x402

Do not equate committed payment with completed fulfillment.

A useful market signal does not merely validate the product. Sometimes it invalidates an assumption inside the product — and makes the model better.

The model is broader than one payment API.

The same causal shape appears in escrow release, payout, refund, programmable-wallet spend, delegated approval, stablecoin settlement, machine-to-machine purchase and compensation flows.

actor
→ authority
→ action
→ state transition
→ observed evidence
→ finality
→ downstream obligation
→ next-action authority

The next real workflow

Where could your agent make the wrong next financial move?

Bring one safely generalized workflow involving a retry, payout, release, refund, repurchase, settlement or programmable-wallet action. We are looking for the exact proof that should gate the next monetary transition.

Actor: Who or what initiates the financial action?
Authority: What proves it is allowed now?
Ambiguity: Which timeout, conflict or partial failure creates uncertainty?
Truth source: Which evidence is authoritative when observations disagree?
Acceptance: What must be proven before the next action is safe?

See the Agent Payment Verification Pilot → Describe one workflow →

Do not send credentials, private keys, personal data or identifying production details. Provider-specific claims require provider-specific evidence.

The product boundary.

RESONANCE is not claiming a universal payments algorithm. ContractGraph-QA is not a wallet, payment processor or authorization service. The current product is narrower: independently pressure-test a bounded financial state machine, preserve uncertainty, and return evidence for whether the next monetary action is allowed.

Before an AI agent pays again, the system should explain not only what happened — but why the next financial move is authorized.

RESONANCE evidence trail

Provider studies are bounded analyses of reviewed public contracts. They are market-driven model corrections, not claims of provider vulnerability, endorsement, affiliation or exhaustive production behavior.