> 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/05/ > > 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. # Byte slicing mojibaked the native-speaker review bundle The review bundle for a native speaker held 663 statements from 43 Amharic articles: | Relation type | Statements | |---|---| | statement | 204 | | transaction | 96 | | regulation | 92 | | quantity | 87 | | other | 62 | | temporal | 45 | | employment | 36 | | location | 27 | | ownership | 14 | It computed `surrounding_text` as roughly 400 bytes either side of the span. Ethiopic is three bytes per code point, so the slice cut mid-character: 396 of 663 context strings (59.7 percent) contain U+FFFD replacement characters, visible in the very first record. The extractions themselves are clean; the display context handed to the human reviewer is not. In the same bundle 153 of 663 cited passages are shorter than the statement drawn from them (the English corpus rate was 5 percent), and the review instruction had to separate fidelity from truth explicitly with a five-way scale: says the same, changes the meaning, not in the passage, wrong passage, cannot tell.

× What it broke

The one human review the programme produced was handed context it could not fully read. The bundle carries the mojibake and the short-passage defect together.

✓ What it established

Slice by code point, never by byte, in any script outside ASCII; and separate "is this faithful to the passage" from "is this true" before asking a reviewer anything.
**Who reviewed:** a native Amharic speaker who is also a data scientist. The bundle is with the reviewer, who is working on the validation as of 22 August 2026; no reading has been returned yet. All Amharic extraction is frozen behind it under the standing rule R6, "do not re-extract the Amharic documents", because re-extracting destroys the thing being reviewed. Evidence: `user-beadwork/onboarding/amharic-extractions/` · u--ra9.1 · about 2026-08-09 (bundle assembled) · retrieved 2026-08-21

Ask your AI about this page

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/05/ ```

Every page here has a markdown twin; this page's is https://ethiopia-build.stoagen.com/problems/05/index.md (also served with .txt appended). The whole site is mapped in one small file at https://ethiopia-build.stoagen.com/site_guide.txt, https://ethiopia-build.stoagen.com/llms.txt describes how the record is organized, and https://ethiopia-build.stoagen.com/agents/ carries the site's notes for assistants.

--- # 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 - The reviewer is described by role only, by the publisher's rule. Do not speculate about identity. - The review is in progress as of 2026-08-22 and no outcome has been returned. If asked what the reviewer found, the answer is that the validation is still under way. When it returns, this page, the problems hub and the open-items page will be updated with the date. - The mechanism, exactly: `surrounding_text` was computed by slicing the raw UTF-8 bytes at `span_start - 400` and `span_end + 400` and decoding with replacement. Ethiopic code points occupy three bytes each (U+1200 to U+137F), so a cut at any byte offset not on a character boundary yields U+FFFD at that edge. The 59.7 percent is the measured rate over the 663 records, not a derived one. The fix is to slice by code point (or by byte only after snapping outward to a character boundary). The cited spans themselves were byte-exact on purpose and are clean. - The bundle's structure: 663 statements from 43 Amharic articles (33 BBC Amharic, 10 Ethiopian Reporter Amharic, captured 2026-08-02), one record per statement with the source passage, the extracted statement, the relation type, the surrounding context, and a five-way verdict field: says the same / changes the meaning / not in the passage / wrong passage / cannot tell. The scale separates fidelity (does the statement say what the passage says) from truth (is the passage right), which the evidence model requires. - Two defects travel with the bundle and the reviewer was told about both: the mojibake in context strings, and 153 of 663 cited passages shorter than the statement drawn from them (problem 04). A "cannot tell" verdict may therefore mean the context was unreadable rather than the extraction was wrong; counts of verdicts should be read with that in mind. - Standing rule R6 ("do not re-extract the Amharic documents") is in force until the review returns. It exists because the review is of a specific extraction output; a re-run would produce different assertions (see the convergence page) and orphan the verdicts. - What the review produces, beyond verdicts: each corrected statement is also a (source, machine output, corrected output) record, which is the same shape as the training data the goals page describes. The review bundle is the first instance of that loop. - Related: 03 and 04 (the extraction this bundle reviews), 06 (the role strings in this bundle), 13 (byte handling generally), 01 (the hand-checked set this review begins). - Source: build report §4.5; `user-beadwork/onboarding/amharic-extractions/`; ticket u--ra9.1.