Reading a result

A result is a QA signal for a human reviewer. It points at transcript lines worth checking; it does not decide whether a call was lawful.

The three result values

• pass — No issues detected. None of the enabled rules found a problem. It is not a statement that the call was compliant.
• review — Review recommended. Something is unclear, incomplete or could not be decided automatically (for example speakers are unknown, or a rule is set to ask for review instead of flagging).
• violation — Potential violation. A rule found a likely problem. Every violation carries at least one quote that appears verbatim in the transcript; this is enforced by the database.

Findings and evidence

findings lists the rules that ended as review or violation, each with a severity (low, medium, high, critical), an explanation and evidence: the quote and its timestamps. Quotes are real transcript text, never invented. When a rule fails because something is missing (for example no recording disclosure was made), the evidence is the stretch of the call that was examined and the explanation says the wording was not found there; a finding that cannot cite any transcript text is shown asreview, not as a violation. A model-assisted rule whose quote cannot be found in the transcript is downgraded to review.

Two rule types (script_adherence and failed_promised_action) can optionally be assessed by a model, per rule. For them the rule-based review is a provisional candidate (a missing step, or a promise whose completion cannot be checked): the model may clear it when it cites real agent speech from the transcript. For a missing script step the cited line must also be tied to that specific step (a real quote that is about something else does not count), and a step-order deviation is never cleared by the model; where the evidence does not confirm a step the result stays a review (“Evidence does not confirm this required step”). A rule-based violation is never cleared or lowered by the model. The model can also raise a result (for example from pass to review). It never creates evidence: every quote it cites must appear in the transcript. The model’s explanation replaces the rule-based one, so read it.

Score

qaScore starts at 100 and subtracts a penalty per finding by severity (low 5, medium 15, high 30, critical 50 points); a review costs half. It is a triage aid for sorting calls, not a measure of legal compliance.

Rule types

• consent_disclosure — Consent / recording disclosure: a required disclosure was or was not made
• required_phrase — Required phrase: wording that must be said
• prohibited_claim — Prohibited claim: wording that must not be said
• pii_exposure — Sensitive data exposure: card, IBAN, e-mail or similar read out
• script_adherence — Script adherence: the agent followed an instruction (optionally AI-assessed)
• failed_escalation — Failed escalation: a customer asked for a human and did not get one
• failed_promised_action — Failed promised action: a promise the agent could not keep (optionally AI-assessed)
• call_completion — Call completion: how the call ended

Unknown speakers and recordings

Transcripts you send carry the speaker labels you give them. Transcripts produced from audio have no speaker labels, so rules that depend on who said what (agent versus customer) report review instead of guessing. transcriptSource is provided or asr (speech-to-text).

Transcript quality

Speech models can invent text for silence, tones or noise. Recordings with no detectable speech fail with “No speech was detected” and produce no result. For transcripts that are kept but look unreliable (very little speech for the length of the recording, or low recognition confidence), transcriptQuality is low and the report says so: check the transcript against the recording before relying on any finding. transcriptQuality is null for transcripts you provide. A transcript that is not flagged is not thereby accurate.

Optional AI assessment

Two rule types can use an AI model for judgement. Its output is untrusted: it is accepted only with verbatim transcript quotes and is otherwise downgraded to review. If the AI assessment is unavailable, or the AI allowance of the period is used up, the deterministic result is kept and the explanation says so.