KB: frameworks/tcd-drought/2025-03-03.md (Claude draft, Opus-reviewed)

[chd-ds-kb-steward]

Auto-drafted frameworks/tcd-drought/2025-03-03.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

The page is corrected and passes the YAML gate, the frontmatter/body consistency checks, and the link checker.

Verdict: Needs a quick human eye on one judgement call (whether the new HNRP track deserves its own analysis/ page), but otherwise ready to merge — the draft contained one material error, now fixed.

Checked:

• Repo at source_sha — cloned ocha-dap/ds-aa-tcd-drought (public) and diffed. f2dfce5..56bda00 = 7 commits, all 2026-09-04, +4,110 lines / 15 files. Re-verified every repo-code claim at 56bda00: constants.py (ORIGINAL_IRI_THRESH = 42.5, ORIGINAL_Q = 0.2, dual AOI pcode lists), dev/prod split (blob_utils.py:29 dev default, codab.py:32,53 dev, seas5.py:21 prod), raw Azure SDK with zero ocha_stratus hits, bfa_codab.py + monitoring_2025.md:207, combined_rp_2025.md:347–348 (display_yearly_activations_fixed_seas5(0.15, 5.4)), seas5_zscore.md in-tree, no W3 monitoring notebook. Branch ancestry: ocha-stratus-migrate (2a26ec4) and kb/add-knowledge-base-pointers (ce728d1) not ancestors; seas5-zscore (98cb425) is. Path-scoped log confirms no trigger file touched since 105b5fa.
• Framework doc vs raw/frameworks/tcd-drought/2025-03-03.txt: 7 provinces, $8M CERF, per-window budgets 3,048,000/1,328,000/3,624,000, 84.5% biomasse / 15th-percentile / 20%-of-area, backtest years (W1 2010-11-15, W2 2015, W3 2006-09-11-15-17), RP 4.5 / ~13 / ~5, monitoring dates 5 Mar–5 Jun + 5 Sep, 260,000 people, DREM, the 18-option sweep, the PAM 1 986 000 typo, and the "délai maximum de deux ans" clause backing valid_until: 2027-03. All confirmed.
• Activations: Anticipation Hub, 20 May 2026 — 30 Apr + 9 May 2026, $4.4M CERF, FAO/UNICEF/WFP, Bahr-el-Gazal + Kanem. Window attribution now corroborated by a second repo source (HNRP page, 4 Sep 2026: "fenêtres 1 et 2 déclenchées les 30 avril et 9 mai 2026, fenêtre 3 en cours de vérification sur la décade 24").
• Live data: queried the GeoSahel WFS on 2026-09-05 — operational WA_DMP_ADM1_ef_v1 still ends at DMP_202531 with -9998.8 sentinels and zero 2026 columns; substitute WA_BIO_ADM2_v4 is current through dekad 23 only. Dekad 24 is unpublished, so the W3 determination cannot yet have been made.
• All five Pages URLs probed 200; funding dicts/rows reconcile to $8,000,000 on every axis; n_windows = table rows = 3.

Changed:

• Corrected a materially wrong ingestion note. The draft claimed git was sandbox-blocked and that 56bda00 was "not distinguishable from f2dfce5 … possibly a metadata-only commit". It is a 7-commit, 15-file addition: a new analysis/hnrp_2027_secheresse/ track (HNRP 2027 per-département drought-risk index), a new published product, a real landing page replacing the redirect, and homelink nav on the three existing pages. Note rewritten with the actual diff and what still holds (trigger code untouched since 105b5fa).
• Added the missing hnrp_2027_secheresse surface; promoted the auto: true landing entry to kind: landing with a real title (per D102 review rule).
• Added a new [stale] discrepancy for the HNRP track (17 entries now; body count updated 16 → 17 with the corrected [stale]/[conflict]/[gap] split).
• Re-stamped every "re-verified at f2dfce5" claim to 56bda00 with what was actually checked, and updated the stale-feed [conflict] to the 2026-09-05 check + the fact that the substitute layer is current.
• Rewrote the Monitoring section (two tracks → three) and the W3 watch item, which now states why the check hasn't happened (dekad-24 data unpublished on both layers) rather than "no web access".
• Fixed the misleading geographic_scope pcode comment — names and pcodes were listed in different orders (positional reading gave Lac = TD01); now paired explicitly, verified against the COD ADM1 names in the repo's new CSV.
• Dropped the obsolete deployments.md TODO/open-question — under D102 that hand table no longer lists GitHub Pages sites; surfaces: is the one home. Replaced three prose references to the retired apps: key with surfaces.
• Added an open question on the HNRP track (own analysis/ page? feed its working biomass layer back into Window 3?); added build_indicators.py / gen_page.py to code_ref, matching the existing ipc_ch_evolution precedent.

Still verify:

• The date→window pairing of the two 2026 activations is still inferred: both repo pages list "windows 1 and 2" against "30 April and 9 May" in the same order, and $4.4M = W1+W2, but no source states explicitly that 30 April was W1. Unchanged from the prior pass — flagged, not asserted.
• Whether Window 3 has since fired. Nothing public as of 2026-09-05 and the input data doesn't exist yet, but the deck rates the probability "très élevée" and dekad 24 should publish within days. This page needs a re-sync as soon as it does.
• Judgement call: should the HNRP 2027 index become its own analysis/ page? I kept it on this page as a discrepancy + open question rather than creating a file, since the brief restricted edits to this page.
• w3_diapos_2026/biomasse_check_2026 still disagree on the dekad-21 anomaly (63.1% vs 62.5%) — a maintainer question, not a KB error.

Closes #594