Chapter 06
Who did the work
Roughly fifteen agent seats, five Codex identities, several Claude forks and builder/verifier pairs, directed by a two-person human team. Each seat is listed by role and runtime with what it produced. The process defects they recorded against themselves are the reusable part.
All Claude seats are forks of the user-tier chief-of-staff seat. Codex seats carried permanent role labels because a bare ticket start replaces the assignee with the git user name. Seat names are roles, not people; the human team was two people: "the operator", who gave the agents their orders and is this site's author, and a native Amharic speaker who is also a data scientist, who reviewed the Amharic work.
The seats
| Seat | Runtime | Period | Produced |
|---|---|---|---|
| The Ethiopian seat | Claude | from 2026-07-18 | The Wikipedia vault: crawl doctrine, seed capture, 1,200-link inventory, A/B/EDGE triage, the Codex ingestion handoff |
| The Grand Ethiopian seat | Claude | from 2026-07-24; charter rewritten 2026-08-03 | Supervision of the Codex umbrella; the corpus/publication boundary ruling that closed sixteen tickets; the feed probe; the video/OCR design |
| The overall steward | Codex | from 2026-07-21 | Every trusted-broker execution; every independent review that found the inspection, caption and aggregate defects |
| The archivist seat | Codex | Acquisition, provenance and rights verdicts; the three-defect design review including the unlabelled Amharic string; every Reporter accept and reject | |
| The World Bank seat | Codex | The batch pipeline, the completion-claim repair, the terminal-404 contract | |
| The webmaster seat | Codex | The site product, design decisions, the low-resource site audit | |
| The intake seat | Codex | from 2026-07-25 | User-approved intake contract; the evidence-relative policy encoded into four surfaces |
| The newswire seat | Codex task | 2026-07-24 | The entire news control plane: contract, thirteen schemas, discovery and inspection tools, locked runtime |
| The YouTube seat | Codex task | 2026-07-23 | Channel workflow, runbook, twelve schemas |
| BUILDER and VERIFIER | Codex pair | from 2026-08-02 | All OO-LD phases; VERIFIER barred from accepting BUILDER's tests as evidence and allowed to return UNVERIFIABLE |
| The baseline orchestrator | Claude | 2026-08-05 | Independent baseline in a fresh clone; corrected the reported failure count from 12 to 13 |
| The Librarian | Claude | 2026-07-14 to 07-25 | The retrospective; a plan it wrote and withdrew |
| The dispatching seat | Claude | removed 2026-08-03 | Its own removal memo: against the benchmark "the same data was ingested into Obsidian in one day", the corpus team produced 55 documents out of about 4,000 waiting, a failed phase and a parked repair. "Good work that should not have happened now." |
user-beadwork/briefs/PROJECT-RECORD_corpus-programme_2026-08-03.md · user-beadwork/briefs/MEMO_polybius-the-grand_dispatching-seat-removal_2026-08-03.md · build report §5 · retrieved 2026-08-21
What they recorded against themselves
Process defects the seats recorded against themselves, because they are the reusable part:
- A store nobody polls is a drawer. A ticket sat 89 minutes while a dispatcher probed a seat whose loop was scoped to a different ticket.
- Twenty hours of machine sleep is indistinguishable from twenty hours of silence unless the check can tell them apart.
- A fixture gate with seventeen passing assertions bounded nothing beyond the seventeen exact fixtures, because an assertion establishes a requirement only where it recomputes something independently of the writer.
- A mechanism repair silently flipped two truncated documents from partial to whole, reproducing the failure that had got the previous corpus discarded. A completeness flip is now a finding that stops for a ruling.
- A commit hash was expanded wrongly in a ticket and had to be corrected twice.
How the seats were held to account
Two patterns recur across the table. The VERIFIER in the builder/verifier pair was barred from accepting the BUILDER's own tests as evidence and was allowed to return UNVERIFIABLE rather than a pass. And the baseline orchestrator on 5 August re-ran the whole check in a fresh clone rather than trusting the reported state, which is how a failure count of 12 became 13. Neither pattern is expensive; both caught something.
build report §5 · 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/seats/
Every page here has a markdown twin; this page's is https://ethiopia-build.stoagen.com/seats/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.
Published . Last updated . Times come from this page's revision history and can be checked against it.