Verification status
89 v2 map cases · built Oct 3, 2026 · staged data, not yet released
Map cases: base locations and unit counts
Each v2 case is one UN deployment map that the 2026 update coded and that no person had coded before. A case climbs four stages as independent readings are adjudicated into the data. A reading that is finished but not yet adjudicated shows as a dashed pip and does not count.
What each stage means
- single-LLM
- One full model reading of the map is in the data.
- double-LLM
- A second, independent full model reading was compared with the first, and the disagreements were adjudicated into the data.
- single-human
- A research assistant coded the map blind, and the disagreements with the model reading were adjudicated into the data.
- double-human
- Two research assistants coded the map independently, and their readings were reconciled into the data.
Re-counts of only the multi-unit claims are partial checks, so they never move a case up a stage. Practice maps never count.
How good is one model reading?
The same scorer the research assistants face, run on model readings of three months that people coded before 2026. Agreement counts every base and country pair in either the answer key or the reading, so extra pairs count against it. Strict counts the answer key's pairs read with the exact unit count.
| Map | Reader | Strict | Agreement | Note |
|---|---|---|---|---|
| UNMISS Feb 2020 | Claude Sonnet | 52 of 52 (100%) | 52 of 52 (100%) | used to build the rules |
| MONUSCO Feb 2019 | Claude Sonnet | 57 of 61 (93%) | 55 of 60 (92%) | used to build the rules |
| MONUSCO Jun 2019 | Claude Sonnet with an Opus re-count of multi-unit claims | 23 of 57 (40%) | 24 of 75 (32%) | not used to build the rules; key still being checked |
February 2020 and February 2019 were used to develop the coding rules, so they flatter the method. The June 2019 answer key is still being checked against its map, so treat that row as provisional.
Model record
Exact model versions were not stamped when these maps were coded, and the session records from that period are not on this machine. Each batch below shows the model family the project log names, or "not recorded" where it names none. From now on, every new model reading records its exact model version, date and source document.
| Read | Scope | Maps covered | Model | Generation | Dates | In the data |
|---|---|---|---|---|---|---|
| Base presence read | full | MINUSCA Jun 2020 to Jun 2024 (12 maps); MINUSMA Mar 2020 to Jun 2023 (15 maps); MONUSCO Feb 2020 to Feb 2022 (9 maps); UNMISS Jun 2020 to Oct 2024 (13 maps) | not recorded | not recorded | Jun 1 to 3, 2026 | yes |
| Unit count read | full | MINUSCA Jun 2020 to Jun 2024 (12 maps); MINUSMA Mar 2020 to Jun 2023 (15 maps); UNMISS Jun 2020 to Oct 2024 (13 maps) | Claude Sonnet | not recorded | Jun 10, 2026 | yes |
| Unit count read | full | MONUSCO Feb 2020 to Feb 2022 (9 maps) | Claude Sonnet | not recorded | Jun 14, 2026 | yes |
| Re-count of multi-unit claims | partial (claims with 2 or more units) | MONUSCO Feb 2020 to Feb 2022 (9 maps) | Claude Opus | not recorded | Jun 14, 2026 | yes |
| Unit count read | full (coded from scratch) | MONUSCO Jun 2022 to Sep 2025 (14 maps) | Claude Sonnet | not recorded | Jun 14, 2026 | yes |
| Re-count of multi-unit claims | partial (claims with 2 or more units) | MONUSCO Jun 2022 to Sep 2025 (14 maps) | Claude Opus | not recorded | Jun 14 to 15, 2026 | yes |
| Unit count read | full (coded from scratch) | MINUSCA Oct 2024 to Oct 2025 (4 maps); UNMISS Jan 2025 to Jul 2025 (4 maps) | Claude Sonnet | not recorded | Jun 14, 2026 | yes |
| Re-count of multi-unit claims | partial (claims with 2 or more units) | MINUSCA Oct 2024 to Feb 2025 (2 maps); UNMISS Jan 2025 to Jul 2025 (4 maps) | Claude Opus | not recorded | Jun 15, 2026 | yes |
| Re-count of multi-unit claims | partial (claims with 2 or more units) | MINUSCA Jun 2025 to Oct 2025 (2 maps) | not recorded | not recorded | Jun 15, 2026 | yes |
| Unit count read | full (coded from scratch) | UNISFA Oct 2018 to Sep 2025 (15 maps) | Claude Sonnet | not recorded | Jun 16, 2026 | yes |
| Unit count read | full (coded from scratch) | UNISFA Feb 2020 to May 2023 (3 maps) | Claude Opus | not recorded | Jun 20, 2026 | yes |
| Re-count of multi-unit claims | partial (claims with 2 or more units) | UNISFA Feb 2020 to Feb 2023 (6 maps) | Claude Opus | not recorded | Jun 20, 2026 | yes |
| Re-count of multi-unit claims | partial (claims with 2 or more units) | UNISFA Apr 2020 to Sep 2025 (9 maps) | not recorded | not recorded | Jun 21, 2026 | yes |
| Full blind re-read | full | UNISFA Oct 2018 to Sep 2019 (3 maps) | Claude Sonnet and Claude Opus (Oct 2018); Claude Opus (Apr 2019); Claude Sonnet (Sep 2019) | not recorded | Jul 2, 2026 | adjudicated, not yet applied |
Troop counts: the counts layer
Parser-extracted, not model-read
- Country by gender counts come from the UN's monthly contributor reports, read by parsers rather than by a model
- For the 24 country-major months from Oct 2022 to Dec 2024, an independent re-extraction with a different engine agreed on 6,681 of 6,681 cells
- For the 1999 to 2018 corrections, 59 of 59 spot checks against the source pages favored the parser over v1 (each page read by a Claude Sonnet agent)
| Year | Status | UN report format | What it means | Scale |
|---|---|---|---|---|
| 1999 | Not yet checked | Scanned image PDF | Image-only scans (no text layer). Not yet validated: needs OCR / Vision. | 2 mission-months, 1 with no source report |
| 2000 | Not yet checked | Scanned image PDF | Image-only scans (no text layer). Not yet validated: needs OCR / Vision. | 22 mission-months, 2 with no source report |
| 2001 | Checked against source | Cross-tab / scanned | May-Dec source-validated Jun 9 (batch 11): 29/29 disputed cells render-verified parser-right; v1 page-skip + uruguay-into-ukraine blend fixed. Jan-Apr are scanned images (OCR bucket, pending scope decision). | 22 mission-months, 43 rows corrected, 2 with no source report |
| 2002 | Checked against source | Cross-tab (2 variants) | Source-validated Jun 9 via two new cross-tab parsers (rotated + transposed). All 7 v1-diff flags source-verified from rendered pages; 425+ cells agree with v1 exactly. | 14 mission-months, 7 rows corrected, 10 with no source report |
| 2003 | Checked against source | "Members" single-col | Source-validated Jun 9 via the 2003 Members parser: 100% printed-total checksums, all 6 on-disk months. 51 corrections incl. v1's mis-coded may-2003 MONUC (China engineer contingent: 188 troops, v1 had 1). | 12 mission-months, 44 rows corrected, 13 with no source report |
| 2004 | Checked against source | "Members" single-col | Source-validated. Parser certified vs Vision; v1 errors corrected. | 47 mission-months, 82 rows corrected, 5 with no source report |
| 2005 | Checked against source | "Members" single-col | Source-validated incl. the 8 UNOCI months recovered this spring via the ONUCI alias (the UN labeled the mission ONUCI before late 2006). | 69 mission-months, 199 rows corrected |
| 2006 | Checked against source | "Members" single-col | Source-validated incl. the 7 UNOCI months (ONUCI alias) and dec-2006 country-major (4 missions) recovered this spring. | 59 mission-months, 103 rows corrected |
| 2007 | Checked against source | "Members" single-col | Source-validated. | 52 mission-months, 153 rows corrected |
| 2008 | Checked against source | By-mission | Source-validated (all months). | 69 mission-months, 72 rows corrected |
| 2009 | Checked against source | By-country (mission-major) | Source-validated. The 2 country-major months (Jan, Jul) were rebuilt clean this spring with a new parser (100% subtotal reconcile; 99 v1 errors corrected). | 72 mission-months, 189 rows corrected |
| 2010 | Checked against source | M / F columns | Source-validated. The apparent 2010 fails were triaged: 9 spurious pre-2014 MINUSCA entries + the MONUC→MONUSCO rename month (v1 correct as-is). | 81 mission-months, 289 rows corrected |
| 2011 | Checked against source | M / F columns | Source-validated. | 61 mission-months, 112 rows corrected |
| 2012 | Checked against source | M / F columns | Source-validated. Feb-2012 (the lone country-major M/F month) fully rebuilt: MONUSCO (batch 8) + the other 5 missions (batch 9, Jun 9). 2010-15 bucket closed. | 72 mission-months, 94 rows corrected |
| 2013 | Checked against source | M / F columns | Source-validated. | 78 mission-months, 116 rows corrected |
| 2014 | Checked against source | M / F columns | Source-validated. | 80 mission-months, 141 rows corrected, 6 with no source report |
| 2015 | Checked against source | M / F columns | Source-validated. | 84 mission-months, 191 rows corrected |
| 2016 | Checked against source | Total-For-Country | Source-validated, including the late-2016 months rebuilt clean this spring (FPU bug fixed). | 84 mission-months, 192 rows corrected |
| 2017 | Checked against source | Total-For-Country | Source-validated, including the rebuilt Total-For-Country months and jul-2017 country-major M/F (recovered this spring). | 77 mission-months, 263 rows corrected, 7 with no source report |
| 2018 | Checked against source | Total-For-Country | Source-validated: all of the previously-blank 2018 months were recovered + rebuilt clean this spring. | 72 mission-months, 35 rows corrected |
| 2019 | New extension data | Extension (MSR) | New data: built + cross-checked (referee2), staged. The temporal refresh. | 63 mission-months, 6 with no source report |
| 2020 | New extension data | Extension (MSR) | New data: built + cross-checked (referee2), staged. | 60 mission-months |
| 2021 | New extension data | Extension (MSR) | New data: built + cross-checked (referee2), staged. | 60 mission-months |
| 2022 | New extension data | Extension (MSR) | New data: built + cross-checked (referee2), staged. (Jul-2022 is a real missing UN report.) | 40 mission-months, 15 with no source report |
| 2023 | New extension data | Extension (MSR) | New data: built + cross-checked (referee2), staged. | 0 mission-months, 60 with no source report |
| 2024 | New extension data | Extension (MSR) | New data: built + cross-checked (referee2), staged. (MINUSMA closed Dec-2024.) | 0 mission-months, 59 with no source report |
| 2025 | New extension data | Extension (MSR) | NEW data through Sep-2025: built + cross-checked (referee2), staged. | 0 mission-months, 36 with no source report |