A "disable interpretation" toggle on the voice-capture control. When enabled, live transcription is skipped entirely — only the audio recording happens. The resulting held item shows placeholder content (e.g. "Voice note — not transcribed") rather than an empty or garbled transcript, with Replay as the primary way to review what was said.
Operator requirement: a voice-only held item with no transcript cannot later be found by search or recall — there is no text content to match against. The held item must carry an Operator-typed comment field (not a transcript — a short label or note the Operator supplies) so the item can be located and categorized later when the recording is being evaluated. This is separate from the placeholder content; the placeholder marks the item as untranscribed, the comment is what makes it findable.
No conflict. Operator-authority: the Operator chooses upfront whether to attempt interpretation. "Only show what is available": the toggle is a real, working option, not a placeholder control. The comment requirement supports the seed's Memory framing — accumulates with provenance — provenance here includes the Operator's own labeling of content the system could not itself describe.
Any change to Web Speech's own recognition accuracy. The v0.5 failure-path fix (Replay on failed/garbled transcription) — this is a separate, additive control for when the Operator chooses not to attempt transcription at all.