call-transcript-reliability-gate

call-transcript-reliability-gate

Before you trust what the phone heard, check how it was written down.

Every verification skill in this repository - call-review, verity-verification-core, provenance-grade, exact-ref - starts from the same assumption: the transcript is ground truth. This skill audits that assumption. Speech-to-text systems hallucinate fluent text with no basis in the audio, and those hallucinations disproportionately appear as loops, caption-credit boilerplate, and even harmful phantom content. A downstream verdict built on a hallucinated confirmation is wrong with perfect confidence.

When To Use

• after any CALL-E call whose result will be written somewhere (booking, record update, payment) and before other transcript skills consume it
• when a summary contains values the transcript turns seem to repeat oddly, or boilerplate no phone caller would say
• before placing a number-critical call, to craft a goal that reduces transcription risk in the first place

When Not To Use

• to prove the provider hallucinated; text signals are advisory reasons to re-confirm, not verdicts about the audio
• during a call; this is strictly post-call transcript analysis plus pre-call goal crafting, because CALL-E exposes transcripts, not live audio
• as a replacement for word-level confidence scores; CALL-E does not expose them, and this skill says so
• to authorize any action; verdicts route work to humans, they never permit anything

Workflow

Gate a finished call

python3 scripts/transcript_reliability_gate.py analyze --transcript path/to/call-result.json

Reads the real get_call_run result shape ({status, result: {transcript}}) or the flat shape used by sibling skill fixtures. Emits a card:

• verdict: RELIABLE / SUSPECT / UNUSABLE
• evidence: turn index, masked span, matched rules
• signals_summary: counts per rule - loop_repetition (same 3+-word phrase repeated 3+ times in one turn), boilerplate_phantom (caption credits and video boilerplate that non-speech audio triggers), harm_violence / harm_extremism / harm_slur_prefix (documented hallucination harm categories - always routed to human review), non_english_insertion (script switch mid-call), empty_turn, no_callee_turns, single_turn_call, extreme_turn_length, empty_word_content
• fields_to_reconfirm: numbers and date words inside suspect turns, masked
• recommended_action: proceed_with_caution, reverify_key_fields, or do_not_act_on_transcript