KB: frameworks/ner-drought/2024-10-24.md (Claude draft, Opus-reviewed)

[chd-ds-kb-steward]

Auto-drafted frameworks/ner-drought/2024-10-24.md via headless Claude on the Max plan, then reviewed in place by Opus against the template + public sources (the same QA gate as the interactive ingest). The Opus review summary is below — use it to spot-check and merge. Worth confirming yourself: frontmatter conforms to the template, deployment/jobs match reality, source_repo/branch/sha, depends_on.

---

🤖 Opus pre-review

Verdict: Ready to merge — the drift bump is accurate; I fixed one real frontmatter error, one schema-conformance gap, and tightened two attributions.

Checked

• Repo at source_sha: 1073c22 (public, cloned ocha-dap/ds-aa-ner-drought @ iri-trend): 1073c22 is branch HEAD ("Add CERF RR column…", 2026-07-30). All three code_ref paths exist. Confirmed every repo claim on the page: obs_pct_slider value=15 / pct_sel value=35, ref_window = 10, start_eval_year 2001, trig_either combined RP accounting, Window 1/Window 2 labelling by decision month, full_trigger.md expanding window (df_stats_iri["year"] <= current_year) with rp_target = 3.5, thresholds taken as sorted(...)[k-1] (no interpolation). docs/static/index.html was re-exported at 1073c22 and still carries data-initial-value='35' / '15' — the "re-verified unchanged" wording is correct. main last moved 2025-03-28 (e6323b4 init), so "15mo+ stale" holds.
• Framework PDF via raw/frameworks/ner-drought/2024-10-24.txt (ReliefWeb landing page, HTTP 200): 35% threshold; W1 pairs Jan+Feb / Feb+Mar, W2 pairs Mar+Apr / Apr+May / May+Jun; obs SPI Jun 1–Jul 31 lowest 35%, activating W2 only if the forecast path hasn't; 71% detection / 28% false alarm (1998–2022); 38% of years, 10/26 (1998–2023); south of 17°N; CHIRP fallback at ≥20% stations missing on 7 Aug; Maproom updated by the 22nd Jan–Jun and by 10 Aug; 455,400 people / 65,100 households / 28 communes; "Sept agences … six secteurs"; WFP FFA land-rehabilitation cancelled from Window 2 (section "Actions dirigées par le PAM").
• Funding, recomputed from the "Financements pré-convenus" table: funding_rows are all stated (not imputed) — W1 FAO 2.5M + WFP 2.75M = $5.25M, W2 = $9.6M, total $14,850,000. Rows reproduce funding_by_agency and funding_by_sector exactly, and both dicts equal funding_by_source = prearranged_funding_usd. Co-financing: WFP $2.5M is a single regional-bureau fund (2023–2027 "Scaling-up drought anticipatory actions for food security in Africa", forecast-path only per fn. 20) printed on both window rows → counting once gives $4,099,659 ✓.
• 2022 activation: $9.5M CERF, Aug 2022, Jun–Jul rainfall "among the lowest in the past 30 years, particularly in the southwest" — Anticipation Hub, OCHA. The repo's new CERF panel gives the allocation code 22-RR-NER-54991 at $9,513,802.
• All 5 external URLs return 200; every template field present, no extras; all 11 body headings present; n_windows: 3 = 3 table rows; YAML gate passes.

Changed

• monitoring_period.months: [1,2,3,4,5,6,8] → [2,3,4,5,6,8], source: inferred → stated. The PDF states the trigger points verbatim — « chaque mois entre février et juin pour les prévisions, et en août pour le SPI » — and January cannot host a decision (a pair needs its second month). This also contradicted the page's own Trigger-windows table ("decision Feb or Mar") and was already flagged as a cross-page mismatch in 2026-06-03.md's monitoring_period.note; the two pages now agree. Note rewritten with the doc quote.
• Added the missing **Activation structure:** bullet under Trigger logic (present on 41 of 45 framework version pages; template requires it).
• Method: "(OCHA Certification model, top-35% frequency)" was presented as doc-sourced, but "Certification" appears nowhere in the PDF — it comes from the repo's Maproom export. Re-attributed.
• Activation note: added the CERF application code 22-RR-NER-54991 and the repo's exact $9,513,802 figure alongside the public "~$9.5M" (prearranged_or_released_usd left at 9500000, matching public reporting).
• Sources & repo completeness: one sentence recording what afef129 → 1073c22 actually changed (bad-year-rank correlation panel; CERF-RR column) and that no trigger logic or threshold moved.

Still verify

• valid_until: "2025" overrides the PDF's "pendant une période de deux ans" (→ 2026) on a team correction. Unverifiable from public sources; the override is captured as a [conflict], but it drives the derived "expired" state — worth a human confirming.
• status: superseded is set from a development successor (issue KB: frameworks past their validity period #91), which is not the usual endorsed→superseded transition. Defensible here since valid_until has passed, but it is a judgement call.
• funding_rows use window labels "Window 1" / "Window 2", while the Trigger-windows table names three rows (Window 2 (forecast path) / Window 2 (obs path…)). INGESTION rule (3) asks these to match; splitting the money by path would double-count, so I left it — flagging in case the DB loader's per-window cross-check is strict about names.
• cofinancing_sources follows the doc's explicit list (WFP, FAO, UNICEF, UNHCR). UNDP separately describes "un fonds additionnel en co-financement" from its climate-resilience project, with no amount — possibly a fifth co-financer. Left unchanged.
• cerf.un.org and reliefweb.int both block scripted fetches (403 / empty body), so the CERF allocation summary for 22-RR-NER-54991 could not be read directly; the $9.5M is corroborated by OCHA and Anticipation Hub instead.
• Out of scope but adjacent (not edited, per instructions): supersedes: 2022-08-04 points at a version page that does not exist in frameworks/ner-drought/; and 2026-06-03.md is still pinned at afef129 / last_synced: 2026-07-30, with its extra.version_label asserting a source_sha for this page that this PR changes.
• visibility: internal despite a fully public PDF source — consistent with the sibling 2026 page, but per docs/PRIVACY.md (classification follows the source) public would also be defensible.

Closes #455