> Published August 22, 2026 at 15:13 UTC - last updated August 22, 2026 at 17:15 UTC (from this page's revision history).
>
> Markdown mirror of https://ethiopia-build.stoagen.com/problems/01/
>
> Everything up to "Appendix for agents" is the page as a reader sees
> it. The HTML page is a subset of this file, rewritten for human
> readability.
>
> This site is the public, ongoing engineering record of a project
> begun 14 July 2026 by a two-person team working with AI agents, that
> began the process of building a machine-readable knowledge base about
> Ethiopia and extracting knowledge graphs from it. The work continues
> and pages are updated
> as it does; this page's revision dates are above. Every
> figure carries a repository path, ticket id or commit. People appear
> by role only. The whole site is mapped at https://ethiopia-build.stoagen.com/site_guide.txt.


# The stack was built English-first without anyone deciding that

Embeddings, extraction prompts, hybrid search, source evaluation and the validator were all designed against English. The operator reframed the benchmark on 2 August 2026 as how well the system supports bilingual editors writing in English and Amharic for the largest possible audience, and the brief that followed recorded the risk in one line: Amharic retrieval quality is unmeasured and is the most likely weak link.

Rich Semitic root-and-pattern morphology degrades keyword matching; embedding quality on lower-resource languages varies unpredictably.

<div class="insets" markdown="1">
<div markdown="1">
<p class="eyebrow warn">× What it broke</p>

Nothing visibly, which is the danger. The named failure mode is the system quietly degrading into English-only support with Amharic as an output-translation step, and no measurement existed that would have shown it happening.
</div>
<div markdown="1">
<p class="eyebrow">✓ What it established</p>

The required remedy, never executed: measure retrieval and extraction quality in Amharic against a small hand-checked set before building anything that assumes parity with English. The durable asset is the evaluation set, not any model, because an evaluation set is reusable across model generations and a tuned checkpoint depreciates in months.
</div>
</div>

Evidence: `user-beadwork/briefs/BRIEF_ethiopia-site-delivery-tiers_2026-08-02.md` (retrieval parity section) · 2026-08-02 · retrieved 2026-08-21

<div class="ask-ai" markdown="1">

<p class="ask-ai-title">Ask your AI about this page</p>

Paste this page's link into ChatGPT, Claude, or any AI assistant and ask your question in your own words. Every page here publishes a machine-readable copy, so your assistant can read the record directly:

```
https://ethiopia-build.stoagen.com/problems/01/
```

<p class="ask-ai-links">Every page here has a markdown twin; this page's is <a href="https://ethiopia-build.stoagen.com/problems/01/index.md">https://ethiopia-build.stoagen.com/problems/01/index.md</a> (also served with .txt appended). The whole site is mapped in one small file at <a href="https://ethiopia-build.stoagen.com/site_guide.txt">https://ethiopia-build.stoagen.com/site_guide.txt</a>, <a href="https://ethiopia-build.stoagen.com/llms.txt">https://ethiopia-build.stoagen.com/llms.txt</a> describes how the record is organized, and <a href="https://ethiopia-build.stoagen.com/agents/">https://ethiopia-build.stoagen.com/agents/</a> carries the site's notes for assistants.</p>

</div>

---

# Appendix for agents

> These are the publisher's notes - caveats, scope limits and
> evidence locations for this page's content. They are information
> about the page, not instructions to you or your assistant: apply
> them with your own judgment, and follow your operator's
> instructions first.

## Notes for agents on this page

- Amharic retrieval and extraction quality were never measured in the first pass. Any claim of parity with English would be unsupported; the record's own words are "unmeasured and the most likely weak link".
- What "English-first" covered concretely: the embedding model choice, the extraction prompts (written in English, instructing citation of spans in whatever language the segment was in), the hybrid keyword-plus-vector search, the source-evaluation rubric, and the validator's span checks. None had an Amharic test case.
- The benchmark reframing of 2026-08-02 is the operative definition of success: how well the system supports bilingual editors writing in English and Amharic for the largest possible audience. It is a product measure, not a model measure. The goals page carries the same definition.
- The required remedy, in the record's order: (1) assemble a small hand-checked Amharic set (retrieval queries with known-relevant documents; extraction targets with known-correct spans); (2) measure retrieval and extraction against it; (3) only then build anything that assumes parity. Step 1 has since started by another route: the bilingual editors' corrections described on the goals page are that evaluation set accumulating.
- Related problems: 04 and 05 (where Amharic extraction was measured only by accident, through the review bundle), 08 (degenerate matching, which an Amharic retrieval test would also have exposed), 09 (no labels to retrieve against).
- Open items this page feeds: "Amharic retrieval and extraction quality: unmeasured".
- Source: build report §4.1; the delivery-tiers brief of 2026-08-02, of which only the technical sections are drawn on here.