One model, asked once, sounds certain about everything
Three judges read every candidate separately. When they land far apart, you get told, because that is the case a single pass gets wrong.
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 provider call times out, that judge does not vote and the remaining two are re-weighted. 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 raw persona scores are not returned.
See it in action
Join the waitlist. First 100 get Starter at 50% off for 3 months.
No credit card. No commitment.
Related
Full setup guide →
Quick start, scoring pipeline, full toolset
SignalPipe vs Clay →
Intent detection vs data enrichment
Buying-Intent Scoring API →
Point it at anything your scraper, bot or agent already reads. We never fetch on your behalf and never hold your keys.
Sarcasm and Rhetorical-Question Filtering →
Every buying phrase has a sarcastic twin. To a keyword they are the same string.
AI Buying Intent Detection →
Two pipelines — multilingual semantic scoring → urgency and specificity → sarcasm detection, then a 3-judge AI drafting swarm — surfaces real buyers in any language and rejects everything else.