# ADR 0012: Preserve Corrections as Post-Run Knowledge

## Status

Accepted as a future product direction; not required for the current MVP.

## Context

Real-meeting review exposes corrections that should not require repeating
expensive processing. These include misspelled or recurring name variants,
people mentioned but omitted from the initial Meeting Context, and confirmed
`SPEAKER_XX -> participant_id` mappings. A destructive edit would lose useful
provenance, while rerunning Whisper or diarization would add cost without
improving a deterministic correction.

The product is also expected to provide both a detailed contextual protocol and
a short participant/distribution protocol. Both should eventually use the same
confirmed context and corrections.

## Decision

Original machine-generated artifacts remain immutable. Corrected or reviewed
artifacts are separate derivatives. Human corrections should evolve into
structured, meeting-specific knowledge rather than opaque destructive edits;
the correction schema is deliberately not defined by this ADR.

An early correction tool may be simple deterministic search and replace, for
example `Grossman` to `Herr Grossmann`. Applying such a confirmed correction to
a suitable editable artifact requires neither Whisper, diarization nor an LLM
call.

A later review stage may run after diarization or the complete initial run. It
may propose likely person/name matches, allow addition of previously omitted
mentioned people, and present anonymous speaker mappings for review. Every
suggestion is non-authoritative: anonymous labels remain anonymous until the
user explicitly confirms or corrects them, and `mentioned_only` people cannot
be mapped as speakers.

After confirmation, the product should offer two paths:

1. A fast path applies confirmed speaker, name or text corrections
   deterministically where that is semantically safe.
2. A quality path regenerates protocol output using the existing transcription,
   existing diarization, corrected Meeting Context and confirmed mappings or
   corrections. It reruns protocol generation only. Whisper and diarization run
   again only when separately requested or technically necessary.

The likely later workflow is therefore:

```text
Audio -> transcription -> optional diarization -> initial protocol
      -> name/person/speaker review -> human confirmation
      -> deterministic correction or protocol-only regeneration
      -> final reviewed detailed and/or distribution protocol
```

The exact ordering may evolve, and this review stage is not mandatory for the
current Streamlit MVP.

## Consequences

- Human knowledge can improve outputs without unnecessary upstream work.
- Provenance is retained because originals and corrected derivatives coexist.
- Correction data can later be reused consistently across detailed and short
  protocol views.
- Search/replace, correction storage, review UI, identity suggestions and
  protocol-only rerun controls remain future implementation work.