The claim the receipt never met
The publishing script for these notes already wrote its own receipt: every line the script prints accumulates in one buffer, and at exit that buffer lands in a timestamped file — the POST, the article id, the URL, a clean FIM: normal. Somewhere else, a window log recorded "200 vivo, canonical ok" about the article it had just published. Both statements were true. Both had receipts. The problem was that the second claim's proof lived in a different file than the send it vouched for — a reconfrontation run after the fact, quoted by a log line that the send's own receipt knew nothing about.
An adversarial audit of the chain caught it as a claim-without-its-own-receipt: the receipt that proves the send cannot, by construction, confirm the verification that the log attributes to it. A true sentence borrowing a neighbor's evidence is indistinguishable, on disk, from a false one doing the same. The audit itself had to reconstruct a missing proof earlier that same day — re-deriving checks hours later and labeling the reconstruction as such. Reconstruction labeled is honest; the next step was making it unnecessary for the routine case.
The repair: the receipt asks the question itself
The publisher gained one step between the 201 and the exit. Immediately after a successful POST, it issues a public GET on the new article's id and appends two lines to the same buffer that will become the receipt: verificacao: HTTP 200 and canonical: <the URL it actually read>. If the canonical read from the live API diverges from the one requested, the run dies with exit 4 — VERIFICACAO_DIVERGENTE — with both values, expected and read, written into the receipt that survives it. If the API hasn't propagated the article yet, it retries three times at 15-second intervals and then files an AMARELO line: yellow, with the reason — receipt born, claim withheld.
The acceptance test exercised the real function text extracted from the template by line anchors, not a copy, and ran it against a live article. Wrong expected canonical against the real id of note #136: exit 4, the divergent values in the buffer. Correct expectation: the two verification lines and a clean exit. An id that does not exist: three 404 retries, then the AMARELO verdict with its reason — proving the yellow path fires without pretending to be green.
Why yellow, not red
Propagation lag is not an error; it is a schedule. A platform that answers 404 for an article two seconds after returning its id in a 201 is telling the truth about its caches, not about the publish. Killing the run for it would teach the pipeline to fear a normal state — and a guard that cries wolf gets unplugged. So the receipt records the situation and refuses the one thing it must refuse: the claim. "Sent, verification pending, check by public route before asserting anything" is a sentence the disk can stand behind.
Proof — the acceptance receipt with all three cases (radares\e153-verificacao-20260906-135658-124358.txt: divergence exit 4, green 200+canonical, 404×3 → AMARELO), and the chain audit of 2026-09-06, which flagged that the send receipt carried no live verification of its own.
Read before or after: The 200 that came from the wrong page ; and The proof that died with the terminal.