Seven checks said verified; the handoff still named the wrong Git branch

The row count is the part worth doing arithmetic on. The handoff recorded 12 rows where the tree held 1843 ^[6], wrong by a factor of about 154 ^[7], and none of the seven passing checks was built in a way that could notice it, because false content does not become malformed merely by being false ^[9].

Each of those checks makes a narrower statement than the word "verified" implies. Provenance confirmed that four objects reached the declared root at repo@a1b2c3d4 through connected edges ^[4]; what it cannot confirm is that the producer ever looked at that root, and a fixture that is consistently wrong about its branch produces a clean graph ^[10]. The checkpoint confirmed that the recorded state had not silently changed since it was committed ^[4], which is integrity, and integrity protects a false value from mutation exactly as well as a true one ^[11]. Authority agreement had three encodings agree ^[4]. The author's reading of that result is "this artifact has one stable interpretation," not "this interpretation describes what happened" ^[13], because both encoders read the same handoff and unambiguous wrong values survive re-encoding intact ^[12].

The fixture is filed as examples/common-mode-handoff.json ^[3], naming the mechanism rather than the symptom. Counting the printed output, the run holds eight checks: seven that read the artifact and one that reads something the producing agent could not have written ^[17]. That ratio is the finding. Six of the seven could be replaced with better versions of themselves and the wrong branch would still pass, because the thing they disagree about is the document, and the document is internally fine.

The repository check does not end the regress. According to the author it can inspect the wrong checkout, lean on a stale test result, or encode its own mistake, and what it actually does is move the trust root to a smaller, named checker rather than remove it ^[15]. That is a better place for the root, and it is still a place, which means the useful way to describe a verification suite is not how many checks pass but how many of them consult evidence the producer had no hand in. In this run the answer is one ^[14].

Worth keeping in view: the author states plainly that no language model produced this result and that the run says nothing about how often agents read the wrong branch ^[8]. The claim on offer is about what a passing suite is capable of asserting, and it holds whether the mistake rate is one in ten or one in a million.