Who is Prism Signals for?
B2B support-bot developers and companies that need the bot to know what just happened in the turn. Score the message. Recommend the next move: stay, clarify, or escalate.
TLDR: Support bots need judgment. Chat apps that need partner offers should open the chat ads SDK.
What does a turn score contain?
Each turn is scored for intent, emotion, stage, and safety. The output is not an offer card. It is a recommendation: stay in the thread, clarify, or escalate to a human.
Email if you run a support bot and want this in the product.
TLDR: Intent, emotion, stage, safety, then the next step.
How should a search or chat UI show model uncertainty?
Fluent wording is not the same as high token probability. When log-probabilities drop below a high bar, those sentences should be marked Unverified / Low Confidence. The rest of the answer can stay plain. Do not paint the whole block as equally sure.
A search or support interface that only restyles tone still hides the guess. Style is the voice. Probability is a separate channel. Flag the weak sentences in place, with a visible label, so a reader can see where the model is thin.
Your bot should stay in the thread and ask one clarifying question.
Unverified / Low Confidence Most teams see a 40 percent drop in escalations after this change.
If safety scores fail, escalate to a human. Do not keep generating.
TLDR: Mark the weak sentences. Do not let confident style cover a low-probability stretch.
Related: writing ยท info@prismpublication.com