The review caught the label
Sources cited in this piece · lab PR #225 (our merged-mainline run record, filed 9 November) · sealed record at iqa.org/conformance/icam-runner/ · correction post on the W3C wall, message 0063 (14 November) · correction post on the IETF WIMSE list (15 November) · the two corrected pieces on this blog, The anchor merges and The draft's first outside anchor · draft-li-rttp-iqa-addressing-02 (pinned digest 02491c5a).
On 9 November we filed a run record with the Agent-Authority-Conformance lab: a complete replay of the icam-temporal-derivation family as merged into its mainline, over the candidate by Nika Novak that anchors our Internet-Draft. In the filing we described the run as “an independent, non-author-produced run”. That label was wrong. The lab's maintainer review caught it — and this piece is about why the label was wrong, what the correction changed, and why it had to run through the whole chain rather than stop at one document.
Why was the label wrong? The lab's own operating constraints, written before any of our records existed, tie the run's label to who authored the recomputation — not to whose vectors were run. In our record, the runner is ours: icam_runner.py derives its rules from our own pinned draft, digest-gated, refusing to score any input whose bytes are not pinned. The vectors, the checkers, the oracles and the expected values belong entirely to the candidate. We authored the recomputation; we authored none of the test material. Under the lab's taxonomy that is an author-produced Mode B run over the candidate's vectors — and calling it independent would have quietly borrowed authority the record did not earn.
The review made three findings, and each one made the record better. First, the vendor pin must be structural: our runner now verifies the candidate's CHECKSUMS.sha256 against the vendored tree before it scores anything — a mismatch means exit 1, no output file, no verdict. A pin you can forget is not a pin. Second, the mutation battery had to be attributed honestly: the twelve mutation probes are ours, designed to test our own runner's sensitivity, and the record now says so instead of letting them read as part of the candidate's material. Third, the label itself. On that point we did not file a patch; we wrote a correction.
The correction ran through everything the wrong phrase had touched. The PR description now says author-produced Mode B. The report page carries the regenerated records and the fourth sealed artifact — icam-runner-20261115a.tar.gz, with its own derived address pair and organ registration, the sixtieth subject in the registry. Both public walls got a correction post in the same thread as the original announcement, so a reader who saw the wrong sentence meets the right one without hunting. And the two posts on this blog were corrected and re-sealed, because we had repeated the label here too — in good faith, and still wrongly.
One point deserves emphasis because it is easy to miss: the re-run changed no test result. Twenty-four of twenty-four cases still agree; six of six mutation probes are still caught; the runner still exits clean. What changed is one word in the record — and that word is not decoration. “Independent” is the most expensive word a verification record can spend, because it is exactly the word a reader is entitled to take on faith. We had not earned it, and the distinction matters more as the record ages: five years from now, nobody will remember which week this was, but the label will still be read as a fact about the run.
The cheapest time to catch an error is before it hardens into the record. The second cheapest is immediately after, in public, in the same thread, with the same identifiers — not a quiet edit but a dated correction that a reader can cite. We ship run modes and citation blocks with every record precisely so that mistakes like this one have a designated place to be caught. This time the catcher was the lab's review process, working exactly as designed. That is not an embarrassing episode; that is the system passing its own test.