Proofscope

Organizations are introducing payment rails for software agents while spending limits are commonly configured by people. The supplied research confirms several live card or payment products using human approval or predefined controls and reports no reviewed product deriving spend scope from measured task performance. That gap supports an assurance layer, but not the source idea's stronger claim that demonstrated capability should itself authorize spending.

Proofscope would preserve organization, agent identity assertion, model, prompt or policy version, tool set, task class, task-battery version, task case, hidden fixture, sandbox environment, allowed effect, attempt, observation, expected outcome, evaluator rule, evaluator version, score, uncertainty, failure mode, suspected gaming, reviewer finding, capability assertion, signed record, signer, validity window, revocation condition, proposed spending scope, policy owner, approval, payment-rail configuration, transaction request, merchant or counterparty, amount, authorization result, provider acknowledgment, settlement, dispute, incident, re-evaluation and revocation as distinct records.

A sandbox battery samples behavior under selected conditions. Passing does not prove competence in novel contexts, trustworthy intent, security, policy compliance or safe handling of money. Scores can be gamed, fixtures can leak and behavior can drift when models, prompts, tools or vendors change. A signature supports signer and record integrity only. Proofscope must never grant spending authority automatically, infer a legally accountable principal, let a score override merchant, purpose or amount policies, or call a provider authorization successful settlement.

The pilot should use synthetic procurement tasks, fake merchants and an isolated payment simulator. The likely buyer is an AI-platform, security, treasury, procurement or risk leader operating a fleet of purchasing agents, but the input lacks a complete role, organization-size, budget and current-alternative quartet. Task validity, evaluator independence, policy ownership, acceptable false-pass rates, rail integration, re-certification cadence, liability and willingness to rely on an external harness remain unverified.

An AI-platform, security, treasury, procurement or risk leader responsible for controlling a fleet of agents that can request purchases.