Why voice previews change: replaceable partials, final returns and evidence scope
A partial is replaceable display, not another transcript fragment to append
The recognition interface defines onPartialText as complete replaceable strings; only returned final text enters original deliverable boundaries. The controller uses bufferingNewest(1) with sessionID, phase and cancellation guards. Redecoding/final results can correct previews. Multiple displays cannot be joined into extra quotations, approvals or executed actions.
Different from meters, segment joins and receipts
Meters show energy, long-note join handles recognition pieces, receipts concern handoff. Here replacement semantics and evidence scope within one preview sequence matter. No model or actual speech is processed.
| Check | Verified scope | Manual review | Not established |
|---|---|---|---|
| Interface | Complete replaceable onPartialText | Not an incremental-token append contract | Each update is a new sentence |
| Buffer | bufferingNewest(1) keeps latest | Not every intermediate update displayed | Full token timeline |
| Session/phase | sessionID/expectedPhase/cancellation | Actual recognition ownership | Old preview can join a new target |
| Length condition | Partial display uses prior-call length condition | Final return can correct | Text always grows without revision |
| Final pass | Full pass gets priority after stop | Actual return/error versus preview | Visible words guarantee final success |
| previewUnavailable | Preview-path failure can be separate | Actual final flow | Model corrupt/must delete |
| Later guards | Final still subject to voiceprint/target etc. | Text versus authority | Final is always delivered |
| Editorial retention | Final/processing/audio separately | Actual reviewed scope | Screenshot is verbatim certification |
Synthetic preview sequence
Display sample, then sample needs review, then final sample deferred. These are not three quotations to concatenate into additional decisions. This illustrates replacement, not actual model output/accuracy.
Retain real errors/cancellation/differences instead of copying old previews around final/handoff guards. bufferingNewest is not permanent retention of every intermediate event.
FAQ
Should every partial append to an archive?
No. Complete replacement semantics make append create duplicates/conflicts.
Does visible preview establish final success?
No. Recognition, filtering and delivery remain separate.
Does preview always grow monotonically?
No guarantee. Redecoding/final corrections and display thresholds do not certify stable meaning.
Does bufferingNewest retain every token?
No. It is neither a full event archive nor alignment.
Should previewUnavailable trigger model deletion?
Do not infer corruption. Source retains final handling separately without deleting for preview failure.
Is returned final necessarily inserted?
No. Voiceprint/target/permission/actual receipts still need review.
Reading and source
Meter distinction · Long-note joins · Callback ownership · Actual receipts · Start · Privacy
