When to use this file:
- MCP server reads this as context for the audio-classifier whenever a recording is dropped in a channel bound to this retreat (
@Funga autoingest audio here for berlin-26).
- This is a skeleton. Dates, venue, the Slack channel and the sessions table are placeholders until the Secretariat confirms the agenda (owner: Martin + Solveig, by ~20 Nov 2026). Keep
Session ID values stable once recordings start — the bot uses them as session_id in transcript frontmatter.
- The allocation-round mechanics this retreat runs are in portfolio/process.md; the proposals it decides on are under portfolio/rounds/round-5-nov-2026.md.
- Copied from
retreats/budapest-2026/overview.md — same structure, same classifier behaviour.
Collaborative End-of-Year Retreat 2026 — Audio Ingest Config
This file tells Funga how to classify audio recordings dropped in the retreat’s Slack channel. The bot uses the Sessions table (and the Synthesis Extras mapping further down) as context for a Haiku classification call against each upload’s caption + transcript head, then files the transcript to _transcripts/berlin-26/<date>-<short-id>.md with the matched session + team in frontmatter.
Curated synthesis lives under retreats/berlin-26/synthesis/<session>/ — five universal category files (Convergence, Divergence, Uncertainties & Risks, Impact Stories, Proposals & Opportunities), one Themes file, plus one file per session-specific Synthesis Extra (table at the bottom). The dashboard renders these at /retreats/berlin-26/; Funga’s own live pages are at funga.scac.openearth.dev/retreats/berlin-26.
Slack channel
TBC — create #retreat_26_recordings (or reuse the Budapest pattern) and bind it:
@Funga autoingest audio here for berlin-26
Bot will confirm. After binding, drop audio with a caption that hints at session + group — e.g. “Day 1 breathing in — ClientEarth proposal” or “decision day, finance table”.
How classification works
- Participant uploads an audio file to the bound channel with a short caption.
- Bot transcribes (Whisper) → calls Haiku with this whole file + caption + first 800 chars of transcript → gets back
{session_id, team_label, participants, confidence, reasoning}.
- Always files the transcript best-effort. High-confidence (>70%) classifications file silently with a confirmation. Low-confidence ones post “reply if I got this wrong” in the thread; user replies trigger reclassification.
The matching is semantic, not regex — vague, paraphrased, or partial captions are fine as long as they carry enough signal.
Sessions
Placeholder agenda derived from the allocation-round pattern in process.md. Replace titles, dates and times from the confirmed agenda; add or remove rows freely before day 1; keep IDs stable after.
| Session ID | Title | Date | Time | Description |
|---|
| context-setting | Context setting & agenda overview | Cross-cutting | varies | Daily morning framing — agenda walkthrough, where the portfolio stands, what today builds toward. Recordings aggregate across all days. |
| day1-breathing-in | Day 1 — Breathing in: proposal presentations | TBC | TBC | Each lead partner presents its Round 5 proposal (5 min + 5 min Q&A). Captions name the proposal or lead org. |
| day1-synergy-walk | Day 1 — Synergy walk: proposals against the existing portfolio | TBC | TBC | Small groups look at proposals alongside the positions they would amplify, enable, depend on or duplicate; uses the dashboard’s synergy and frames views. |
| day2-decisions | Day 2 — Making decisions | TBC | TBC | Working through the green / red / yellow vote: majority-green funding conversations, yellow conditions, red parking lot. Final 2–4 proposals with allocations. |
| day2-digital-infrastructure | Day 2 — Digital infrastructure: login, self-service, what to extend | TBC | TBC | Decisions the dashboard scope deferred to the retreat: who authenticates and how, partner self-service, mobile, and the 2027 extension list. |
| closing-reflection | Closing reflection | Cross-cutting | varies | End-of-day harvest. Recordings aggregate across days. |
Notes:
- Cross-cutting rows’ Date column says “Cross-cutting” so the bot understands the same session_id is reused across multiple days.
- Day-specific rows should use ISO dates once known so the bot can correlate “Tuesday” / “Day 2” caption hints back to the right session.
Session-specific synthesis sections in addition to the 5 universal categories. The synthesis pipeline reads this table, asks for the extra section by name, and writes it to retreats/berlin-26/synthesis/<session>/<kebab-name>.md.
| Session ID | Extra Section Name |
|---|
| day1-synergy-walk | Synergies Named Between Proposals and Positions |
| day2-decisions | Conditions Attached to Yellow Proposals |
| day2-digital-infrastructure | Decisions on Login and Self-Service |
Group labels — emergent, not predefined
Groups self-organise and re-form between sessions, so there is no fixed team enum. For each recording the bot extracts a free-form team_label from the caption + transcript + uploader. Cross-group synthesis is grouped by session_id, not by team.
Decisions taken at the retreat
Create retreats/berlin-26/decisions.md during the retreat (per the folder shape in retreats/README.md) — login strategy, mobile, allocation outcomes, the 2027 extension list.
Known Gaps
- Dates, venue, Slack channel and the real agenda — Secretariat (Solveig) to confirm; Martin to update this file.
- Whether the retreat format includes an indicative allocation round (Gina) — determines whether
day2-decisions is a decision or a signal-capture session.