Working vault for a GSoC 2026 project to fine-tune Whisper on Greek municipal council speech, with LLM post-correction. Holds dataset exploration notes, decisions, specs, and the local review-UI prototype.

No training yet — dataset exploration comes first.

Start here:

• Current state
• Progress vs GSoC plan
• Agent instructions (AGENTS.md is a symlink to the same file)
• Project map
• Roadmap
• Decisions index
• Meetings index
• Logs index
• Exploration UI spec
• Local data model
• OpenCouncil meeting JSON notes
• GSoC proposal
• UI prototype

• CURRENT.md is the first file to read and should stay short.
• docs/decisions/ — accepted decisions and open questions, split by theme. See decisions index.
• docs/progress.md — where we are against the GSoC plan. The plan itself is in the proposal.
• docs/roadmap.md — phases and current direction.
• docs/meetings/ — normalized meeting notes.
• docs/specs/ — product and implementation specs.
• docs/logs/ — weekly digests only. See logs index for cadence rules.
• docs/reference/ — stable technical references.
• archive/ — superseded material, local only (gitignored).
• Data outputs live under data/; scripts live under scripts/.
• CLAUDE.md is the single source of truth for assistant instructions; AGENTS.md is a symlink to it for tools that read that filename.

• Full May 12 export: utterance-edits-may12-26.csv
• Historical sample (archived): archive/old-data/corrections-sample.csv
• Clean CSV: data/clean/corrections_clean.csv
• Rejected rows: data/reports/corrections_rejected.csv
• Summary JSON: data/reports/corrections_summary.json

Regenerate the cleaned data:

rtk python3 scripts/preprocess_corrections.py

The SvelteKit correction-review prototype lives in ui/. It ingests the May 12 CSV into local SQLite and supports review labels, timestamp adjustments, stats, and included-row export. Meeting JSON matching (utterance IDs, speaker context, surrounding transcript) is the next gap to close.