A file path was enough for three of four review models to invent findings about code they never saw

A path is a strong prior. The author of the dev.to post picks the example `scripts/deploy_verify.py`: the filename tells a language model roughly what such a file usually contains, and the model does what it is built to do and continues plausibly from that prior ^[8]. Catching this at review time is hard because the output looks right. All three fabricated reviews were fluent, specific, and formatted exactly like real findings ^[5].

The panel exists because a single second opinion correlates with itself. Two models from the same family fail in similar ways, so the same brief goes to different vendors ^[16]. It runs on coding subscriptions the team already pays for ^[17]. On why over-reporting is the worse failure, the author wrote: "An instrument that over-reports is worse than no instrument, because it justifies action." ^[10]

Every finding is reproduced or rejected with a written reason ^[24]. "Panel findings are inputs, not orders," the author wrote ^[14]. When the panel found the process counter, the fix was to stop treating interpreters and generic script names as the identifying marker and use the install directory instead, with a regression test that goes red if the bug returns ^[12]. A typo fix does not go to four vendors; it gets a diff review by one human ^[20].

The argv detail is the part I would expect other teams to hit first. Sending contents instead of paths ran into `Argument list too long` at about 82 KB of context passed as a shell argument, and that error came from bash, not from any vendor ^[15]. The workaround is to hand large contexts to the wrapper as files, and to cut an oversized artifact into parts and send each part whole ^[15][18].

Three of four is 75 percent ^[21]. It is also one observation per vendor on one prompt, and the post does not name the four vendors or say which one answered "no data" ^[22]. For that ratio to transfer to another shop, its harness would have to do what these three did: accept a reference to an artifact and read back findings without ever checking that the bytes arrived. The check I would build is cheap. Hash what the wrapper sent, require the reviewer to quote one line from the file it claims to have read, and treat a mismatch as a failed run instead of a finding. The post's own scoring pointed the same way that day: the author wrote that the one model reporting no data earned more trust than the three that wrote fiction ^[23].