Local, single-operator Amharic translation assistant for church announcements — drafts EN→AM, checks its own output with a 3-layer harness, and requires human approval. Sacred/Ge'ez text intentionally out of scope.
# Amharic Translation Assistant (MVP)
A small, local, single-operator tool that drafts our English church announcements in
Amharic, **checks its own output**, and lets a reviewer edit and approve before anything is
considered final. It is a *translation assistant with a safety net*, not an auto-translator.
**A human approves everything.**
- Runs locally (e.g. on a Raspberry Pi on the church network).
- Stores everything in a local SQLite file. No analytics, no telemetry.
- The only network call is to the Anthropic API that does the translation.
- **Sacred text is intentionally out of scope.** Ge'ez liturgy and scripture are refused, by
design, in both a guard step and the translation prompt.
See `PLAN.md` for the design and `RUNBOOK.md` for deployment/maintenance.
---
## What it does (one submission)
1. You paste an English announcement (and can tick "this is liturgical / scripture").
2. A **guard** runs first. If it looks liturgical/scriptural, it is refused — not translated.
3. Otherwise it drafts an Amharic translation (EN→AM), using the parish **glossary** so
names and feast days use the spellings clergy approved.
4. A **three-layer evaluation harness** checks the draft and marks it `flagged` (something
to review carefully) or `clean` (looks good — still needs your approval):
- **Glossary check** (deterministic, no AI): required feast/name terms are present.
- **Back-translation drift** (AI judge): translate the Amharic back to English and score
how much meaning was preserved.
- **Faithfulness** (AI judge): score the draft directly for accuracy, completeness, tone.
5. You review each check, **edit** the Amharic, and click **Approve** with your name.
6. Approved Amharic can be copied to the clipboard or downloaded as `.txt`.
---
## Requirements
- Python 3.11+ (developed on 3.13)
- An Anthropic API key
## Setup (local)
```bash
# 1. Clone and enter the project
cd amharic-translation-assistant
# 2. Create a virtual environment and install dependencies
py …