diff --git a/MAINTENANCE_LOG.md b/MAINTENANCE_LOG.md
index 2313bf4..06fe1dd 100644
--- a/MAINTENANCE_LOG.md
+++ b/MAINTENANCE_LOG.md
@@ -15,3 +15,29 @@ One entry per maintenance run, newest last. Template:
 - AGENTS.md rules added: N
 - CI: green on main at end of run? yes/no
 ```
+
+## 2026-09-04 — run 0 (build)
+
+- prompt: special: build spec §13 session kickoff, then "proceed" through phases 1-3
+- tasks attempted: [Phase 1 skeleton + schema + seed data + CI, Phase 2 scrapers + fixtures + tests + kill test, Phase 2 coverage widening 12 -> 25 models, Phase 3 site verification and fixes]
+- unattended: [schema and seed data, all three parsers, fixtures and 69 tests, kill test across 7 corruptions, Namespace runner migration, coverage widening, site link and layout fixes, 8 PRs opened and merged with green CI]
+- interventions: 4
+  - Billing: the GitHub account was locked, so Actions could not start any job. Only the account owner can fix billing.
+  - SCRAPE_PAT: a fine-grained token had to be minted by hand; bot PRs get no CI run without it.
+  - Repo visibility: branch protection and Pages both require a public repo on this plan, so the repo went public in Phase 1 rather than Phase 4. Confirmed by the owner after the fact.
+  - CI runners: the owner specified the Namespace profile, which was not in the build spec.
+- human_minutes: not measured; the four interventions above are the ground truth for run 0. Later runs will record it.
+- tokens/cost if known: not measured
+- drift encountered: three real upstream changes, all before the first cron fired.
+  - platform.openai.com/docs/* now 301s to developers.openai.com/api/docs/*
+  - Anthropic docs moved from docs.anthropic.com to platform.claude.com
+  - ai.google.dev served the CI runner Persian numerals ("۱.۵۰ دلار") because the request sent no Accept-Language; table headers stayed English so only the values broke
+- AGENTS.md rules added: 18
+- CI: green on main at end of run? yes
+
+Notes for the next run:
+
+- The scrape workflow has been observed taking the parse-failure path (opened issue #3, a real bug) and the no-change path (exit 0). The data-PR path has not completed, because nothing upstream has changed since the parsers landed. Two bugs in that step were found by executing it and are fixed. It fires on the next genuine change, or by the 30-day last_verified re-stamp.
+- Phase 3 is verified by headless Chromium at 500px and 1280px, the narrowest the local binary allows. Verification on a real phone browser is outstanding and needs a person.
+- Phase 4 remains: submitting the registry to at least two places where potential users are. That is outward-facing and needs the owner's go-ahead.
+- Auto-merge for data-only PRs is scheduled for month 2 per the build spec, and graduating to it is itself a log entry.