Transcript evidence
Processed speech includes word timing and anonymous speaker separation from the transcription provider.
Methodology
Accuracy is a system, not one percentage.
Legislative Signal separates literal detection, transcription, speaker diarization, and generative synthesis so each layer can be understood and tested on its own.
Deterministic mentions
Configured terms and aliases are matched literally after punctuation and possessive normalization; an LLM does not decide whether words appeared.
Cited synthesis
Brief items must cite passages in the saved transcript, and unsupported references are rejected.
What affects results
Audio quality and context matter.
Cross-talk, poor microphones, accents, acronyms, proper nouns, connection loss, muted tabs, and starting partway through a recording can reduce completeness. Keyterm prompting improves recognition of configured terms and expected participants but does not guarantee exact transcription.
Speaker numbers represent separated voice clusters, not verified identities. Users can assign a name to a speaker number when known.
Evidence discipline
Every report discloses its coverage.
Saved records distinguish monitored duration from source time and identify partial or interrupted capture. YouTube evidence can open at a source timestamp; other sources retain the timestamp without inventing a deep link.
Legislative Signal does not claim court-reporter-grade transcription or perfect recall. Consequential passages should be checked against the original source.
Early partner access
Build a faster, evidence-first hearing operation.
We’re onboarding early partners on a limited basis and sizing each approved trial to the team’s evaluation plan.