Exam-style
A support team runs an assistant on Amazon Bedrock that answers customer questions from the company's own help articles. The prompt asks the model to finish each answer with a line reading *Confidence: N%*. The team proposes a review gate: anything at 95% or above goes straight to the customer, and anything below it goes to a human agent first. Is that a sound gate, and if not, what should the threshold be read from instead?
Reveal the answer
C. Not sound. The figure is text the model generated, not a calibrated probability, so a confidently phrased invention scores as high as a correct answer. Threshold instead on the Amazon Bedrock Guardrails contextual grounding check, show citations so a person can verify against the source, and validate the answer's shape in code
Two different things are called confidence scoring, and the AI Practitioner material names the term without separating them. One is a real number. Amazon Comprehend returns a score for each entity or sentiment it detects, and Amazon Textract returns one for every field it lifts off a scanned form, both produced by a classifier that was measured against labelled data. Thresholding those is ordinary engineering, and it is how a document pipeline decides which extractions a person checks. The other is a sentence. Confidence: 97% is a continuation the model produced because it was a likely thing to write after that answer, and it is not derived from anything the model checked. An assistant that has invented a returns window will report a high figure for the invented window, because the wrong answer reads as fluent as the right one. Sorting by that number sorts by fluency and sends the smooth mistakes straight to the customer. The alternatives that do carry a usable signal are the other items under hallucination detection methods and grounding techniques. Retrieval Augmented Generation [RAG] grounding puts the help articles in front of the model so there is a source to be right about, which the fourth option correctly values and then overstates: retrieval reduces invention and does not end it, since a model can still summarise a retrieved passage into something the passage does not say. The Amazon Bedrock Guardrails contextual grounding check scores the answer against the passages that were actually supplied and against the question asked, and intervenes below a threshold you set, so that is the number the gate should read. Citations let the agent open the article and see it. Output validation is the code check on the answer’s shape before it is sent: a policy reference that exists, an amount inside a plausible range, a date the system of record agrees with. The fifth option invents a restriction; fine-tuning has no bearing on whether a model can rate itself.
Q. The assistant ends each answer with Confidence: N%. Auto-approve at 95% and above, send the rest to an agent. Sound gate?
A. No. That figure is generated text, so it tracks fluency rather than truth. Gate on a Guardrails contextual grounding score instead, with citations and a code check on the answer.
Why? Confidence scoring means two things and only one of them is a number to act on. Amazon Comprehend and Amazon Textract return scores from a classifier measured against labelled data, and thresholding those is sound. A model stating its own certainty is writing a likely sentence, and an invented answer gets a high figure just as easily as a correct one. The hallucination detection methods that do give you something are Retrieval Augmented Generation [RAG] grounding so there is a source to be right about, the Amazon Bedrock Guardrails contextual grounding check scoring the answer against those retrieved passages, citations so a person can verify it, and output validation in code on the parts with a checkable shape. Retrieval narrows the failure without closing it, which is why the check sits after it.