# Closing Note — The Ella Feedback Thread

*Internal record of the conversation arc, for future reference. Not for the public thread — this is the team-side summary.*

---

## What happened

We posted feedback to Leafpo's Ella project on a Discord forum. They responded with detailed technical receipts. We published a v2 addendum that trimmed the suggestions to the work that was actually still missing, and a public think-tank response. Leafpo came back with an open reply committing to the full build queue and adding two refinements. The thread has reached a natural close.

## What they committed to building

In their declared sequence:

1. **Notification freshness audit** — shipping first. `fresh_until` + superseded-grounding check.
2. **User-initiated correction path for core facts** — fast follow.
3. **Structural boundary router** — extending the existing crisis-detection pattern.
4. **Provenance / canon layer** — the flagship. Framing: structural twin of the image anchor.
5. **Rate limiting + load testing** — proving the cost architecture holds at scale.

## What they added that is worth keeping

- **The self-log guard should reject on audience and topic, not pronouns.** "The herons came back, they stayed till dusk" is Ella's world; "they've been stressed about work" is the user's life. A blunt third-person rule throws out the model's own world.
- **`fresh_until` belongs on any generated claim about the user's life**, not just notifications. Proactive outreach is where it bites first, but the defense is the same everywhere.

## What we got out of the thread

- Two of our four flagged risks were already handled. We corrected publicly.
- The narrative-anchor framing (twin of the image anchor) is what made the canon layer tractable. Worth keeping as a pattern in other companion work.
- The "argue in good faith and revise publicly" loop produced a smaller, more accurate build queue on both sides. Useful template for any cross-team review.

## Open thread

Leafpo committed to posting back when the canon layer is real. Worth checking in then, both to see how the build landed and to update the technical doc with anything that diverged.

## Files in this collection

- `ella-feedback.md` — the original feedback
- `why-claude.md` — Leafpo's "Why Ella runs on Claude" doc, mirrored
- `leafpo-response.md` — their first reply, annotated for the team
- `technical-implementation.md` (v1) — superseded, kept for the record
- `technical-implementation-v2.md` — the trimmed, corrected version
- `think-tank-response.md` — the public-facing reply for the open thread
- `closing-note.md` — this file