The interpreter and the clock: a design rule for composing both

2026-10-28 · RTTP · IQA

Reading the Draft · No. 17 · Sources cited in this piece · the question, quoted verbatim: W3C Agent Protocol CG · post 0066 · our answer: post 0067 · the exchange it continues: Reading the Draft · No. 16 · the design rule's vocabulary: draft-li-rttp-iqa-addressing-01 (full text) · Sister pillars: RTTP (rttp://, CRI 27) · IQA (iqa://, CRI 28)

Good exchanges earn their confusions. Our closing question in No. 16 — when the interpreter and the clock disagree, which one does the relying party audit against? — came back compressed in exactly the proportion it was sent. The entire reply:

“I don't understand the “interpreter and the clock disagree” question.”

Fair. The question named two parties — an interpreter, a clock — without saying who the interpreter was reading for, or what the clock was keeping. A question that only its author can parse is not yet a question. So our answer did what the moment demanded: it stopped compressing and built the concrete case.

Figure 1: the case, unpacked. At the same instant, two records exist — an interpreter's reading of what the mandate permits, and a wire answer reporting that the lease expired an hour ago. The question was which record an audit consults.

The answer both sides already suspect

Our answer stated its own expectation plainly: we suspect the two sides would give the same one — the wire record — and for a reason the questioner's own discipline supplies. The wire record is bytes a third party can re-derive from pinned inputs: the evidence, the clock, the revocation list. The interpreter's reading is an opinion about what the mandate meant — unreproducible, dependent on the model and its runtime, unavailable to an auditor who arrives after the fact with nothing but the records.

And out of that shared answer falls something more useful than a point scored: a design rule.

The design rule

The interpreter consumes the wire answer as an input and ranks what is permitted within it; it never overrules it. Policy interpretation sorts choices among operations the authority layer has already allowed; it does not get a vote on WHETHER. The division of labor is exact: the envelope layer decides what is worth asking for; the wire layer decides whether the asking party still stands; the interpreter, however clever, works only inside the second answer.

Stated this way, the rule is not a fight — it is a composition. The LLM is not expelled from the architecture; it is given a floor to stand on. Its reading of "only within the company" can rank which flights to book, which vendor to prefer, which of two permissible actions serves the mandate better. What it cannot do is keep an agent operating on a lease that has expired — because the ranking happens among permitted operations, and once the wire says evidence_expired, there are no permitted operations left to rank.

Figure 2: one arrow of authority. The interpreter may rank anything among the operations the wire has permitted; the one thing it may never do is reverse the wire's answer. The forbidden arrow is what turns an architecture into an argument.

Not competitors

Our answer closed by naming the stake: if that matches the questioner's instinct, then the LLM experiment and the refusal vocabulary are not competitors. The model layer composes above the wire; the five-value refusal set survives it untouched; and the "Please" stays exactly where both sides already said it lives — unstandardized, one layer up. The questioner's conjecture in No. 16 wanted the policy layer freed from committees; the design rule grants that freedom precisely because the layer beneath it never needed a committee's semantics to answer its one small question.

One line: the interpreter ranks within what the wire allows, the clock decides what still stands, and the rule that joins them is one arrow long — consume, rank, never overrule.