One model, asked once, sounds certain about everything
Three judges read every close call separately. When they land far apart, you get told, because that is the case a single pass gets wrong.
Updated 1 October 2026
A language model asked to score a post will produce a number. It will produce one for a real buyer, for somebody venting, and for a post it has no idea about, and all three will read as equally confident. The number that matters is the one you cannot see: how sure it actually was.
SignalPipe asks three times, separately. A skeptic, an analyst and an optimist each read the same post with no knowledge of the others. Their scores are fused into one verdict, and the spread between them is kept.
That spread is the useful part. On 255 scored posts in our own corpus the average spread was 0.478 on a 0 to 1 scale, which means mild disagreement is the normal state of three readers thinking independently. It is not news. Only the top decile is, so that is what gets flagged.
You will not find this in a keyword alert, and you will not get it from a single prompt. Both will hand you a confident answer on the posts where confidence is least deserved.
How it works
Three independent reads
Skeptic, analyst, optimist. Each sees the post and your product profile, none sees the others. Run concurrently, so the panel costs one round trip rather than three.
The split is reported, not averaged away
Fusing three numbers into one hides the thing worth knowing. The verdict carries a split flag when the judges landed far apart.
Tuned against real posts, not intuition
The flag threshold was set from the measured distribution of 255 production posts, not chosen because it sounded right. At the original setting it fired on three calls in four, which is the same as never firing.
A judge that fails abstains
If one judge’s call times out, that judge does not vote and the verdict comes from the two that answered. A lost call never silently becomes a lost opinion counted as agreement.
The numbers stay in
You get each judge as a stance, convinced or on the fence or unconvinced, plus whether they split. The judges’ raw scores are not returned.
See it in action
Starter is $29 a month for 3,000 judgements. The plugin and daemon stay free and open source.
Billed monthly. Cancel any time from your billing page.