The pushback we hear, and where we actually stand.
Every answer here points at something measured elsewhere on this site: a benchmark, a fix tier, a published false-positive rate. Nothing below is a talking point invented for this page.
Having your AI review its own code is like grading your own homework: it might catch some things, but there’s no outside number telling you how well it actually works. Most in-loop cleanup is a model judging its own output, or another model’s, inside the same session, with no published accuracy number behind it. Truemend runs separately from that loop: local analysis, no model in the detection loop, checked against the whole repository rather than one agent’s working set. Its precision is public on this site, self-consistency at 100.0% and real-corpus false-positive rates included, wins and losses both. It ships a check for pull requests and a hook for the coding agent you already use, so an existing agent pipeline can read its findings as a signal rather than compete with it.
See: Methodology, Product
Three objections where the honest answer is not yet.
A page of objections where every answer lands is a page nobody believes. These are the three we would raise if we were sitting on your side of the table.
Correct, on the tier that matters. We hold every language to under 10% wrong-match on medium-and-above findings, and when we pointed the scanner at our own source in August it came back at 10.3%, just over. A September re-run reads 6.6%, but under a different judge, so we do not call that a pass. It is published on Methodology rather than smoothed over.
Also correct. Eleven languages carry a head-to-head figure. Ruby, Dart, PHP, C, C++, SQL, PowerShell, Terraform and Dockerfile carry none yet, and two more have only a first baseline with no comparable tool wired in. If your stack is in that list, the honest position is that detection ships and the measurement does not exist.
Because you can see that we changed it, and why. A false-positive figure we led with turned out to be inflated by a measurement setting, and two benchmark rows we published as losses later turned into wins. All four states are on the site with dates. The alternative is a number that only ever moves in the flattering direction, which tells you nothing.
Do not take an answer on this page as evidence.
Every one of these leads somewhere with numbers and dates attached. The fastest way to settle an argument about the tool is to point it at one repository yourself: detection is free, needs no account, and takes an afternoon.