Handing pricing to a system is only fair if you can see why it decided what it did. No black box is a
rule Elastly is built around, not a slogan. Every recommendation carries its full reasoning, so you
can check the number before you trust it.
What travels with a recommendation
Everything below is computed by the engine, with no language model involved. It is the same data
whether you read it in the review queue or pull it from the Prices API.
Drivers below about 0.25pp are folded away, so the list stays to the steps that actually mattered.
How confidence is measured
Confidence is not a guess. It comes from three kinds of evidence, and each one counts for more the
more of the price it is actually responsible for.
Each kind saturates on its own: a little evidence lifts it quickly, and more keeps lifting it with
diminishing returns. The three are then weighted by how much each one moved this particular price. A
price driven mostly by a well-evidenced part of your setup reads as high confidence; a price leaning
on a signal with little behind it reads lower. By default a score at or above 0.66 is high and at or
above 0.33 is medium.
A guardrail holding the price does not lower confidence. The two are separate: confidence is about
evidence, the guardrail check is about limits. A price can be held at its floor and still be high
confidence.
The plain-language summary
On top of the computed reasoning, Elastly can add a short written “why this price.” This is the one
place a language model is involved, and it is optional.
When it is on, Elastly passes the facts it already computed to a small model (Claude Haiku) and asks
for one or two sentences. The model is held to a strict rule: it may use only the facts it was
given, it can never invent a number or a claim, and if a guardrail capped the price it has to name
it. It explains the math in plain words. It does not make the decision, and it never sees your data
beyond the single recommendation it is describing.
The summary is a convenience, not the source of truth. Turning it off changes nothing about how
prices are set: every driver, the guardrail check, and the confidence are all still there. The
wording is just terser. If the model is unavailable, the recommendation is unaffected and simply
arrives without the sentence.
Where you see it
Open the review queue and expand any recommendation. You get a “why this price” panel with the
drivers laid out, the guardrail check (flagged if a limit held the price), any competitor context,
the confidence, and the written summary when it is on. It is the same reasoning whether you are
approving one price or reviewing a batch, and the same reasoning the API returns on every response. 
