Module verbatim 

Verbatim stages: text the model must say word for word, verified.

A regulated call has lines that must be read exactly: a disclosure, a consent statement, the terms of a payment. An instruction to “read this verbatim” is a request, not a guarantee. A verbatim stage turns it into a guarantee at the governance level.

• While the stage is active, the stack publishes the required text under VERBATIM_KEY.
• At the end of each model turn, the control lane looks for the text in the model’s output transcript (contained_similarity) and writes the verdict to verbatim_flag(step). The text may be surrounded by other words in the same turn, such as a greeting before a disclosure. It also emits LiveEvent::VerbatimChecked.
• The stage completes only once the flag is true. A paraphrase keeps the conversation in the stage, where the posture asks for the exact text again.

The check needs output transcription, since it reads what was actually said. Without a transcript nothing is verified, and the stage does not complete.

Structs

VerbatimRequirement

The published requirement of the active verbatim stage.

VerbatimVerdict

The outcome of one verbatim check.

Constants

VERBATIM_KEY

The state key under which the active verbatim requirement is published.

VERBATIM_MIN_SIMILARITY

How close (0–1, word level) the spoken text must be to the required text. Tolerates a transcription slip or two in a long passage, not a paraphrase.

Functions

check_turn

Check a finished model turn against the active verbatim requirement, if any, and record the verdict in state. heard is the turn’s output transcript; an empty transcript checks nothing.

contained_similarity

How closely expected appears somewhere in heard: the best similarity between expected and a run of consecutive words of heard about as long as it (within a tenth either way). A model turn that says the required text and something else, such as “Welcome to Bright Smile. This call may be recorded. How can I help?”, scores as the text alone.

similarity

Word-level similarity of heard to expected, from 0 to 1: one minus the word edit distance over the length of the longer text, after lowercasing and dropping punctuation.

verbatim_flag

The state key holding whether step’s text was said verbatim.