Summary

  • Revision 01 of an individual IETF agent-audit draft adds a record for cumulative spending limits: amounts already spent, amounts reserved for unsettled permits, the new request and the version of the consumption state read by the decision point.
  • A lost response does not free a reservation. A retry of the same request must use its held place; a different request based on an older state can expose a double-spend window across records.

The Missing Reply Is Not a Refund

An agent asks a paid tool to do something. The tool's response vanishes. Should the agent's spending allowance open up again? Sankalp Gilda's revised audit-record draft gives a terse answer: keep the money held until a record of the same request shows what happened. A missing reply says nothing about whether the action executed.

The text was announced on 27 October as an individual Internet-Draft. Its Datatracker entry describes a proposal for an audit format, not a deployed payment system. The important change from revision 00 is new Section 5.9, “Consumption under a cumulative limit.” It moves the audit question from “Was this one call permitted?” to “What did the decision point believe was already spent or promised when it permitted the next one?”

That distinction matters because a cumulative ceiling can be crossed without any single payment looking large. The draft says the decision depends on money already committed, amounts reserved for earlier permits that have not settled, and the current request. A timestamp and a freshness label alone cannot show the sum. So its proposed consumption-state input carries the limit document's digest, a currency unit, the limit, consumed, reserved, requested, a list of held request digests and amounts, and stateRevision—the version of the balance the decision point read.

One Retry, Two Different Bills

The draft supplies its own arithmetic. Under a limit of 50,000 pence, 30,000 has been consumed. An earlier permitted call for 12,000 pence has not settled, so 12,000 remains held. If that same call is retried with the same request digest, the retry is within the ceiling at 42,000. Counting the held request again would punish the caller for a lost response. A different 12,000-pence request would bring the total to 54,000 and should be denied. That example is a worked design case, not a reported transaction.

The held amount has only two exits. A record for the same request with an observed effect of occurred moves it into consumed. A record showing none releases it. Without such a settlement record, the reserve stays put. “No response” does not silently become “no effect.” This may inconvenience a caller whose tool really did nothing; it also stops an uncertain call from lending its unused-looking budget to the next decision.

Now suppose a second call reads an older stateRevision that predates the first hold. Its own calculation may look neat while omitting the earlier promise. The draft asks a party holding records under the same limit document to compare them in revision order. A later permit must show an unsettled earlier permit in held or consumed. An old-state permit that does not is the detectable window in which one ceiling can be spent twice. A release from held with no matching record of none is another detectable gap.

The format deliberately allows a well-formed record to say a decision point permitted a call above its limit. Rejecting the record as malformed would make the failure disappear from the audit trail. The record's job is to expose a breach of delegated authority, not to certify that every decision was good. This is also why the person reviewing it needs several records, not a screenshot of one green approval.

Who Owns the Balance?

stateRevision is a compact accountability hook. It names the balance the authorization point used. The held list names the unresolved calls it claims to know about. The settlement record names what was observed. Together they allow an auditor to ask whether a later decision had the right earlier obligations in view. The proposal does not itself set the spending cap, run the tool or prove that every record was delivered. Its security section says a coherent record speaks only for its own bytes.

That boundary deserves a public sentence because a signed audit file can look more authoritative than the process behind it. If the same operator controls what gets signed and what the reviewer receives, the cross-record check only sees the set handed over. A real consumer would need an independent way to retain the records or establish the observer and its key. The draft offers a data structure for checking; the organization adopting it would have to decide who may set a limit, reserve it, settle it and show the full series when challenged.

Original records