What the word was doing

Examine which events or concepts speakers associate with genocide and whether they assert, reject or report a claim. The tables summarise language-model classifications. Open the quotations to assess them in context.

Experimental — model-derived

These are experimental model classifications. A referent is the case or concept a mention concerns; speaker position describes how the passage presents the claim. Human coding provides a separate reference for evaluation, with disagreements recorded. The progress figures below show how much has been reviewed. Word counts elsewhere do not depend on these labels.

Human validation has not started. Use these classifications to explore passages, not as validated research findings. Quotations with invalid evidence are excluded from the discourse figures.

Model: Qwen/Qwen3.8-27B · revision 1d4bf0f2ff60
Computing setup: vLLM 0.28.0 on 1 × NVIDIA H100 80GB HBM3; project-controlled hardware
Run: 2026-11-03-qwen-131k · 2026-11-03
Prompt: v3 · sha256:3772c44d400f…
Reasoning setting: xhigh
Coverage: 8,611 of 8,611 occurrences, over 860 requests
Uncertain labels: 0 validity decisions · 22 cases or concepts · 5 speaker position
Excluded responses: 0 responses without valid annotations · 0 incomplete responses · 24 invalid evidence quotations
Human reference sample: not started — 0 of 469 coded

Second opinion

A second model classified the same occurrences using the same instructions. Agreement measures consistency between the models; both can misread the same passage. The main figures use the published run. Human coding is needed to evaluate accuracy.

Second model: google/gemma-4-31B-it
Run: 2026-11-03-gemma4 · 2026-11-03
Reasoning setting: high
Annotated: 8,573 of 8,611 occurrences · 99.56%
Compared: 8,573 carry a label from both runs
Excluded responses: 14 abstentions · 33 invalid evidence quotations

How far the published run and the second opinion agree, field by field, over the occurrences both of them reached
Field	Compared	Agreement	Kappa	PABAK	Contested	Same model, twice	Same model, twice
verdict	8,573	100.00%	—	1.00	0	100.00%	100.00%
quotation	8,573	95.11%	0.80	0.94	419	99.94%	99.96%
concrete case	8,573	96.03%	0.91	0.95	340	99.96%	99.94%
speaker position	8,573	93.62%	0.88	0.93	547	99.96%	99.94%
referent	8,573	89.15%	0.88	0.89	930	99.96%	99.92%

A passage can receive several rhetorical-function labels. Their mean overlap between runs is 0.82 (Jaccard similarity: 1 means identical sets). Krippendorff's α is 0.72; it adjusts for chance agreement using a distance measure for partially overlapping sets (MASI). The runs assign different sets to 2,202 occurrences. The table below compares individual labels.

Observed agreement is the share of identical labels. Kappa adjusts for agreement expected from each run's label frequencies. It is withheld when fewer than 1% of labels fall outside a run's most common category. PABAK uses an alternative, equal-category chance baseline. Read both with the category counts; neither validates the interpretation.

Same model, twice compares repeated runs using identical instructions. This shows whether classifications change even without changing models. Qwen/Qwen3.8-27B returned identical labels across all compared fields for 5,199 of 5,218 occurrences.google/gemma-4-31B-it returned identical labels across all compared fields for 5,225 of 5,233 occurrences.Compare this variation with the differences between models.

How far the two runs agree that each function label applies
Function	Qwen/Qwen3.8-27B	google/gemma-4-31B-it	Agreement	Kappa
accountability	1,910	2,046	94.10%	0.83
accusation or qualification	4,449	4,456	87.72%	0.75
commemoration	520	719	96.19%	0.72
institutional title or mandate	831	886	97.75%	0.88
other	530	255	94.62%	0.39
warning or prevention	1,826	1,389	91.71%	0.73

Reporting another person's claim can resemble asserting it. Compare the reporting and assertion labels in the table, then read the disputed passages. Differences may reflect how each model interprets that boundary.

3,112 of 8,573 compared occurrences (36.30%) are read differently on at least one compared field. They are listed under The contested passages below, and marked wherever they appear in the quotations.

Which genocide each delegation means

Which genocide is each delegation talking about when it says the word?

Model-derived · experimental

Unit

occurrences placed on a referent, published for every delegation Ordered by Column order

occurrences 0 864 no assigned mentions

Delegations down the side, referents across the top; each cell holds the occurrences that delegation placed on that referent.
Delegation

Press Tab to enter the table, then use the arrow keys to move between cells and headings. Home and End move to the first and last cells in a row.

Showing 70 of 238 affiliations with assigned mentions. The download includes 168 additional affiliations with 1,565 mentions. 32 omitted affiliations meet the share threshold; the display also has a row limit.

Cells count individual mentions or their share, according to the selected unit. Repeated mentions within one speech count separately.

How to read this

Rows are affiliations; columns are cases or concepts. A stronger shade means more assigned mentions or a larger share, depending on the unit. Column labels sort the rows highest counts first, and row labels sort the columns the same way; cells open passages. The final columns cover general or legal discussion. Shares are withheld below 20 eligible occurrences.

What it does not show

The model assigns the cases and concepts. A case label identifies what the passage concerns, not whether genocide occurred or the speaker endorsed the claim. Review status: 0 of 469 sampled entries have human coding. Check the quotations before interpreting a pattern.

More on this figure

Eligible mentions pass the model's validity decision and quotation check. Assigned mentions also receive a case or concept label. Here, 24 mentions are ineligible and 22 eligible mentions remain unassigned. Shares divide by an affiliation's assigned mentions; eligibility determines whether its total meets the display threshold.

The list includes contested characterisations raised in Council debates. Other known referent covers a case outside the named list. Counts below the minimum remain visible, but may rest on very few passages.

The occurrences behind a cell

Pick a cell, a delegation or a referent above and the occurrences behind it are listed here, each with the sentence it was read from and a way into the speech it came from.

When each delegation first said it

When did each delegation first place the word on this genocide, first assert it, and first refuse it?

Model-derived · experimental

Case or concept

Placed the word on it117Refused the word for it3Asserted it115

117 0

1960198020002020

The model assigns mentions of Rwanda to 117 affiliations. Of these, 115 have an assertion and 3 a rejection. An affiliation can appear in both curves if different passages receive those labels. 247 affiliations spoke in at least one meeting where Rwanda was named; read the curves against that number, not against every state.

Every first the curves are made of, for Rwanda, oldest first
Date	Delegation	Milestone	Position	Cumulative count	Occurrence
5 May 1994	Czech Republic	Placed the word on it	Asserts	1	`SC03375-01-016#1`
5 May 1994	Czech Republic	Asserted it	Asserts	1	`SC03375-01-016#1`
16 May 1994	New Zealand	Placed the word on it	Asserts	2	`SC03377-01-015#1`
16 May 1994	New Zealand	Asserted it	Asserts	2	`SC03377-01-015#1`
8 June 1994	Djibouti	Placed the word on it	Asserts	3	`SC03388-02-002#1`
8 June 1994	Djibouti	Asserted it	Asserts	3	`SC03388-02-002#1`
8 June 1994	Argentina	Placed the word on it	Asserts	4	`SC03388-02-014#1`
8 June 1994	Argentina	Asserted it	Asserts	4	`SC03388-02-014#1`
8 June 1994	United Kingdom	Placed the word on it	Asserts	5	`SC03388-02-016#1`
8 June 1994	United Kingdom	Asserted it	Asserts	5	`SC03388-02-016#1`
8 June 1994	Spain	Placed the word on it	Reports without a position	6	`SC03388-02-018#1`
8 June 1994	France	Placed the word on it	Asserts	7	`SC03388-02-023#1`
8 June 1994	France	Asserted it	Asserts	6	`SC03388-02-023#1`
8 June 1994	United States	Placed the word on it	Asserts	8	`SC03388-02-026#1`
8 June 1994	United States	Asserted it	Asserts	7	`SC03388-02-026#1`
1 July 1994	Spain	Asserted it	Asserts	8	`SC03400-01-003#2`
10 August 1994	Russia	Placed the word on it	Asserts	9	`SC03414-01-002#1`
10 August 1994	Russia	Asserted it	Asserts	9	`SC03414-01-002#1`
8 November 1994	Brazil	Placed the word on it	Asserts	10	`SC03453-01-010#1`
8 November 1994	Brazil	Asserted it	Asserts	10	`SC03453-01-010#1`
8 November 1994	China	Placed the word on it	Asserts	11	`SC03453-01-012#1`
8 November 1994	China	Asserted it	Asserts	11	`SC03453-01-012#1`
8 November 1994	Nigeria	Placed the word on it	Asserts	12	`SC03453-01-014#1`
8 November 1994	Nigeria	Asserted it	Asserts	12	`SC03453-01-014#1`
8 November 1994	Rwanda	Placed the word on it	Reports without a position	13	`SC03453-01-015#1`
8 November 1994	Rwanda	Asserted it	Asserts	13	`SC03453-01-015#10`
8 November 1994	Oman	Placed the word on it	Asserts	14	`SC03453-01-016#1`
8 November 1994	Oman	Asserted it	Asserts	14	`SC03453-01-016#1`
30 November 1994	United States	Refused the word for it	Rejects	1	`SC03473-01-013#1`
18 January 1995	Canada	Placed the word on it	Asserts	15	`SC03492-02-023#1`
18 January 1995	Canada	Asserted it	Asserts	15	`SC03492-02-023#1`
9 June 1995	Indonesia	Placed the word on it	Asserts	16	`SC03542-01-008#1`
9 June 1995	Indonesia	Asserted it	Asserts	16	`SC03542-01-008#1`
9 June 1995	Botswana	Placed the word on it	Asserts	17	`SC03542-01-010#1`
9 June 1995	Botswana	Asserted it	Asserts	17	`SC03542-01-010#1`
9 June 1995	Honduras	Placed the word on it	Asserts	18	`SC03542-01-012#1`
9 June 1995	Honduras	Asserted it	Asserts	18	`SC03542-01-012#1`
9 June 1995	Germany	Placed the word on it	Asserts	19	`SC03542-01-034#1`
9 June 1995	Germany	Asserted it	Asserts	19	`SC03542-01-034#1`
28 August 1995	Burundi	Placed the word on it	Asserts	20	`SC03571-01-003#2`
28 August 1995	Burundi	Asserted it	Asserts	20	`SC03571-01-003#2`
7 September 1995	Italy	Placed the word on it	Asserts	21	`SC03574-01-027#1`
7 September 1995	Italy	Asserted it	Asserts	21	`SC03574-01-027#1`
22 November 1995	Bosnia and Herzegovina	Placed the word on it	Asserts	22	`SC03595-01-024#1`
22 November 1995	Bosnia and Herzegovina	Asserted it	Asserts	22	`SC03595-01-024#1`
20 December 1995	Australia	Placed the word on it	Asserts	23	`SC03611-02-026#1`
20 December 1995	Australia	Asserted it	Asserts	23	`SC03611-02-026#1`
5 March 1996	Poland	Placed the word on it	Asserts	24	`SC03639-01-030#1`
5 March 1996	Poland	Asserted it	Asserts	24	`SC03639-01-030#1`
8 March 1996	Chile	Placed the word on it	Asserts	25	`SC03640-01-006#1`
8 March 1996	Chile	Asserted it	Asserts	25	`SC03640-01-006#1`
8 March 1996	Egypt	Placed the word on it	Asserts	26	`SC03640-01-020#1`
8 March 1996	Egypt	Asserted it	Asserts	26	`SC03640-01-020#1`
15 November 1996	DR Congo	Placed the word on it	Asserts	27	`SC03713-02-003#1`
15 November 1996	DR Congo	Asserted it	Asserts	27	`SC03713-02-003#1`
25 September 1997	Organization of African Unity	Placed the word on it	Asserts	28	`SC03819-01-009#1`
25 September 1997	Organization of African Unity	Asserted it	Asserts	28	`SC03819-01-009#1`
9 April 1998	Portugal	Placed the word on it	Asserts	29	`SC03870-01-010#1`
9 April 1998	Portugal	Asserted it	Asserts	29	`SC03870-01-010#1`
9 April 1998	Sweden	Placed the word on it	Asserts	30	`SC03870-01-013#1`
9 April 1998	Sweden	Asserted it	Asserts	30	`SC03870-01-013#1`
9 April 1998	Slovenia	Placed the word on it	Asserts	31	`SC03870-01-016#1`
9 April 1998	Slovenia	Asserted it	Asserts	31	`SC03870-01-016#1`
9 April 1998	Japan	Placed the word on it	Asserts	32	`SC03870-01-022#1`
9 April 1998	Japan	Asserted it	Asserts	32	`SC03870-01-022#1`
24 April 1998	Bahrain	Placed the word on it	Asserts	33	`SC03875-01-015#1`
24 April 1998	Bahrain	Asserted it	Asserts	33	`SC03875-01-015#1`
24 April 1998	United Nations	Placed the word on it	Asserts	34	`SC03875-02-012#1`
24 April 1998	United Nations	Asserted it	Asserts	34	`SC03875-02-012#1`
30 April 1998	Costa Rica	Placed the word on it	Asserts	35	`SC03877-01-008#1`
30 April 1998	Costa Rica	Asserted it	Asserts	35	`SC03877-01-008#1`
30 April 1998	Gabon	Placed the word on it	Asserts	36	`SC03877-01-012#1`
30 April 1998	Gabon	Asserted it	Asserts	36	`SC03877-01-012#1`
21 January 1999	Netherlands	Placed the word on it	Asserts	37	`SC03968-01-011#1`
21 January 1999	Netherlands	Asserted it	Asserts	37	`SC03968-01-011#1`
19 March 1999	Uganda	Placed the word on it	Asserts	38	`SC03987-01-013#1`
19 March 1999	Uganda	Asserted it	Asserts	38	`SC03987-01-013#1`
17 September 1999	Iraq	Placed the word on it	Asserts	39	`SC04046-03-010#1`
17 September 1999	Iraq	Asserted it	Asserts	39	`SC04046-03-010#1`
29 November 1999	Malaysia	Placed the word on it	Asserts	40	`SC04072-02-020#2`
29 November 1999	Malaysia	Asserted it	Asserts	40	`SC04072-02-020#2`
30 November 1999	Liechtenstein	Placed the word on it	Asserts	41	`SC04072-01-004#1`
30 November 1999	Liechtenstein	Asserted it	Asserts	41	`SC04072-01-004#1`
30 November 1999	Zambia	Placed the word on it	Asserts	42	`SC04072-01-006#1`
30 November 1999	Zambia	Asserted it	Asserts	42	`SC04072-01-006#1`
26 January 2000	Ukraine	Placed the word on it	Asserts	43	`SC04092-03-008#1`
26 January 2000	Ukraine	Asserted it	Asserts	43	`SC04092-03-008#1`
14 April 2000	Mali	Placed the word on it	Asserts	44	`SC04127-01-010#1`
14 April 2000	Mali	Asserted it	Asserts	44	`SC04127-01-010#1`
14 April 2000	Tunisia	Placed the word on it	Asserts	45	`SC04127-01-016#1`
14 April 2000	Tunisia	Asserted it	Asserts	45	`SC04127-01-016#1`
14 April 2000	Namibia	Placed the word on it	Asserts	46	`SC04127-01-019#1`
14 April 2000	Namibia	Asserted it	Asserts	46	`SC04127-01-019#1`
14 April 2000	Bangladesh	Placed the word on it	Asserts	47	`SC04127-01-021#1`
14 April 2000	Bangladesh	Asserted it	Asserts	47	`SC04127-01-021#1`
14 April 2000	Jamaica	Placed the word on it	Asserts	48	`SC04127-01-022#1`
14 April 2000	Jamaica	Asserted it	Asserts	48	`SC04127-01-022#1`
19 April 2000	Singapore	Placed the word on it	Asserts	49	`SC04130-01-010#1`
19 April 2000	Singapore	Asserted it	Asserts	49	`SC04130-01-010#1`
26 July 2000	India	Placed the word on it	Asserts	50	`SC04176-03-023#1`
26 July 2000	India	Asserted it	Asserts	50	`SC04176-03-023#1`
16 January 2001	Mauritius	Placed the word on it	Asserts	51	`SC04257-02-013#1`
16 January 2001	Mauritius	Asserted it	Asserts	51	`SC04257-02-013#1`
19 January 2001	Colombia	Placed the word on it	Asserts	52	`SC04259-02-001#1`
19 January 2001	Colombia	Asserted it	Asserts	52	`SC04259-02-001#1`
7 February 2001	Ireland	Placed the word on it	Asserts	53	`SC04273-01-012#1`
7 February 2001	Ireland	Asserted it	Asserts	53	`SC04273-01-012#1`
7 February 2001	Norway	Placed the word on it	Asserts	54	`SC04273-01-016#1`
7 February 2001	Norway	Asserted it	Asserts	54	`SC04273-01-016#1`
10 July 2002	Jordan	Placed the word on it	Asserts	55	`SC04568-01-023#1`
10 July 2002	Jordan	Asserted it	Asserts	55	`SC04568-01-023#1`
8 August 2002	South Africa	Placed the word on it	Asserts	56	`SC04596-01-008#1`
8 August 2002	South Africa	Asserted it	Asserts	56	`SC04596-01-008#1`
11 October 2002	Senegal	Placed the word on it	Asserts	57	`SC04623-02-043#1`
11 October 2002	Senegal	Asserted it	Asserts	57	`SC04623-02-043#1`
22 October 2002	IBRD	Placed the word on it	Asserts	58	`SC04630-01-009#1`
22 October 2002	IBRD	Asserted it	Asserts	58	`SC04630-01-009#1`
10 December 2002	Denmark	Placed the word on it	Asserts	59	`SC04660-01-002#1`
10 December 2002	Denmark	Asserted it	Asserts	59	`SC04660-01-002#1`
13 May 2003	Ethiopia	Placed the word on it	Asserts	60	`SC04753-02-015#2`
13 May 2003	Ethiopia	Asserted it	Asserts	60	`SC04753-02-015#2`
12 June 2003	Malawi	Placed the word on it	Asserts	61	`SC04772-01-021#2`
12 June 2003	Malawi	Asserted it	Asserts	61	`SC04772-01-021#2`
9 October 2003	Guinea	Placed the word on it	Asserts	62	`SC04838-01-023#1`
9 October 2003	Guinea	Asserted it	Asserts	62	`SC04838-01-023#1`
20 November 2003	African Union	Placed the word on it	Asserts	63	`SC04865-01-008#1`
20 November 2003	African Union	Asserted it	Asserts	63	`SC04865-01-008#1`
20 November 2003	Angola	Placed the word on it	Asserts	64	`SC04865-01-030#1`
20 November 2003	Angola	Asserted it	Asserts	64	`SC04865-01-030#1`
20 November 2003	Mozambique	Placed the word on it	Asserts	65	`SC04865-01-033#1`
20 November 2003	Mozambique	Asserted it	Asserts	65	`SC04865-01-033#1`
24 November 2003	Economic Community of Central African States	Placed the word on it	Asserts	66	`SC04871-01-032#1`
24 November 2003	Economic Community of Central African States	Asserted it	Asserts	66	`SC04871-01-032#1`
14 May 2004	Pakistan	Placed the word on it	Asserts	67	`SC04969-01-002#1`
14 May 2004	Pakistan	Asserted it	Asserts	67	`SC04969-01-002#1`
17 May 2004	Peru	Placed the word on it	Asserts	68	`SC04970-02-014#1`
17 May 2004	Peru	Asserted it	Asserts	68	`SC04970-02-014#1`
28 May 2004	Philippines	Placed the word on it	Rejects	69	`SC04980-01-011#1`
28 May 2004	Philippines	Refused the word for it	Rejects	2	`SC04980-01-011#1`
29 June 2004	Benin	Placed the word on it	Asserts	70	`SC04999-02-002#2`
29 June 2004	Benin	Asserted it	Asserts	69	`SC04999-02-002#2`
18 November 2004	Intergovernmental Authority on Development	Placed the word on it	Asserts	71	`SC05080-01-013#1`
18 November 2004	Intergovernmental Authority on Development	Asserted it	Asserts	70	`SC05080-01-013#1`
14 December 2004	Ivory Coast	Placed the word on it	Asserts	72	`SC05100-01-025#1`
14 December 2004	Ivory Coast	Asserted it	Asserts	71	`SC05100-01-025#1`
13 June 2005	Algeria	Placed the word on it	Asserts	73	`SC05199-01-022#1`
13 June 2005	Algeria	Asserted it	Asserts	72	`SC05199-01-022#1`
21 June 2005	Tanzania	Placed the word on it	Asserts	74	`SC05209-01-006#1`
21 June 2005	Tanzania	Asserted it	Asserts	73	`SC05209-01-006#1`
27 January 2006	Kenya	Placed the word on it	Asserts	75	`SC05359-01-013#1`
27 January 2006	Kenya	Asserted it	Asserts	74	`SC05359-01-013#1`
27 January 2006	Greece	Placed the word on it	Asserts	76	`SC05359-01-026#1`
27 January 2006	Greece	Asserted it	Asserts	75	`SC05359-01-026#1`
27 January 2006	South Korea	Placed the word on it	Asserts	77	`SC05359-01-034#1`
27 January 2006	South Korea	Asserted it	Asserts	76	`SC05359-01-034#1`
27 January 2006	European Commission	Placed the word on it	Asserts	78	`SC05359-02-013#1`
27 January 2006	European Commission	Asserted it	Asserts	77	`SC05359-02-013#1`
7 June 2006	Ghana	Placed the word on it	Asserts	79	`SC05453-01-018#2`
7 June 2006	Ghana	Asserted it	Asserts	78	`SC05453-01-018#2`
7 June 2006	Slovakia	Placed the word on it	Asserts	80	`SC05453-01-024#1`
7 June 2006	Slovakia	Asserted it	Asserts	79	`SC05453-01-024#1`
15 December 2006	Qatar	Placed the word on it	Asserts	81	`SC05594-01-021#1`
15 December 2006	Qatar	Asserted it	Asserts	80	`SC05594-01-021#1`
18 June 2007	Congo	Placed the word on it	Asserts	82	`SC05697-01-011#1`
18 June 2007	Congo	Asserted it	Asserts	81	`SC05697-01-011#1`
12 December 2008	Libya	Placed the word on it	Asserts	83	`SC06041-01-015#1`
12 December 2008	Libya	Asserted it	Asserts	82	`SC06041-01-015#1`
4 June 2009	Burkina Faso	Placed the word on it	Asserts	84	`SC06134-01-024#1`
4 June 2009	Burkina Faso	Asserted it	Asserts	83	`SC06134-01-024#1`
11 November 2009	Croatia	Placed the word on it	Asserts	85	`SC06216-01-010#1`
11 November 2009	Croatia	Asserted it	Asserts	84	`SC06216-01-010#1`
12 June 2013	European Union	Placed the word on it	Asserts	86	`SC06977-01-031#1`
12 June 2013	European Union	Asserted it	Asserts	85	`SC06977-01-031#1`
23 July 2013	Luxembourg	Placed the word on it	Asserts	87	`SC07007-01-015#1`
23 July 2013	Luxembourg	Asserted it	Asserts	86	`SC07007-01-015#1`
23 July 2013	Saudi Arabia	Placed the word on it	Asserts	88	`SC07007-01-025#1`
23 July 2013	Saudi Arabia	Asserted it	Asserts	87	`SC07007-01-025#1`
29 January 2014	Israel	Placed the word on it	Asserts	89	`SC07105-01-025#1`
29 January 2014	Israel	Asserted it	Asserts	88	`SC07105-01-025#1`
12 February 2014	Estonia	Placed the word on it	Asserts	90	`SC07109-01-025#1`
12 February 2014	Estonia	Asserted it	Asserts	89	`SC07109-01-025#1`
12 February 2014	Mexico	Placed the word on it	Asserts	91	`SC07109-01-035#2`
12 February 2014	Mexico	Asserted it	Asserts	90	`SC07109-01-035#2`
12 February 2014	Uruguay	Placed the word on it	Asserts	92	`SC07109-01-067#1`
12 February 2014	Uruguay	Asserted it	Asserts	91	`SC07109-01-067#1`
12 February 2014	Turkey	Placed the word on it	Asserts	93	`SC07109-01-083#1`
12 February 2014	Turkey	Asserted it	Asserts	92	`SC07109-01-083#1`
16 April 2014	Lithuania	Placed the word on it	Asserts	94	`SC07155-01-015#10`
16 April 2014	Lithuania	Asserted it	Asserts	93	`SC07155-01-015#10`
16 April 2014	Chad	Placed the word on it	Asserts	95	`SC07155-01-020#2`
16 April 2014	Chad	Asserted it	Asserts	94	`SC07155-01-020#2`
19 January 2015	Armenia	Placed the word on it	Asserts	96	`SC07361-01-119#1`
19 January 2015	Armenia	Asserted it	Asserts	95	`SC07361-01-119#1`
23 February 2015	Belgium	Placed the word on it	Asserts	97	`SC07389-01-055#1`
23 February 2015	Belgium	Asserted it	Asserts	96	`SC07389-01-055#1`
27 May 2015	Venezuela	Placed the word on it	Asserts	98	`SC07450-01-022#1`
27 May 2015	Venezuela	Asserted it	Asserts	97	`SC07450-01-022#1`
15 May 2017	Bolivia	Placed the word on it	Asserts	99	`SC07938-01-017#2`
15 May 2017	Bolivia	Asserted it	Asserts	98	`SC07938-01-017#2`
8 June 2017	Sudan	Placed the word on it	Asserts	100	`SC07963-01-019#5`
8 June 2017	Sudan	Asserted it	Asserts	99	`SC07963-01-019#5`
21 February 2018	Kuwait	Placed the word on it	Asserts	101	`SC08185-01-005#1`
21 February 2018	Kuwait	Asserted it	Asserts	100	`SC08185-01-005#1`
16 April 2018	Kaladan Press (Organization)	Placed the word on it	Asserts	102	`SC08234-01-006#1`
16 April 2018	Kaladan Press (Organization)	Asserted it	Asserts	101	`SC08234-01-006#1`
6 June 2018	Equatorial Guinea	Placed the word on it	Asserts	103	`SC08278-01-014#1`
6 June 2018	Equatorial Guinea	Asserted it	Asserts	102	`SC08278-01-014#1`
23 May 2019	Fiji	Placed the word on it	Asserts	104	`SC08534-01-071#1`
23 May 2019	Fiji	Asserted it	Asserts	103	`SC08534-01-071#1`
7 October 2019	George Washington University (Washington, D.C.)	Placed the word on it	Asserts	105	`SC08633-01-004#1`
7 October 2019	George Washington University (Washington, D.C.)	Asserted it	Asserts	104	`SC08633-01-004#1`
9 September 2021	Women and Peace Studies Organization (Afghanistan)	Placed the word on it	Asserts	106	`SC08853-01-004#1`
9 September 2021	Women and Peace Studies Organization (Afghanistan)	Asserted it	Asserts	105	`SC08853-01-004#1`
12 December 2022	United Arab Emirates	Placed the word on it	Asserts	107	`SC09217-01-008#1`
12 December 2022	United Arab Emirates	Asserted it	Asserts	106	`SC09217-01-008#1`
12 June 2023	Ecuador	Placed the word on it	Asserts	108	`SC09344-01-019#3`
12 June 2023	Ecuador	Asserted it	Asserts	107	`SC09344-01-019#3`
28 September 2023	DR Congo	Refused the word for it	Rejects	3	`SC09427-01-025#1`
20 October 2023	(personal capacity)	Placed the word on it	Reports without a position	109	`SC09448-01-006#1`
25 October 2023	Dominican Republic	Placed the word on it	Asserts	110	`SC09452-02-052#1`
25 October 2023	Dominican Republic	Asserted it	Asserts	108	`SC09452-02-052#1`
12 December 2023	Switzerland	Placed the word on it	Asserts	111	`SC09502-01-011#1`
12 December 2023	Switzerland	Asserted it	Asserts	109	`SC09502-01-011#1`
30 April 2024	Yugoslavia	Placed the word on it	Asserts	112	`SC09621-01-027#1`
30 April 2024	Yugoslavia	Asserted it	Asserts	110	`SC09621-01-027#1`
21 May 2024	Malta	Placed the word on it	Asserts	113	`SC09632-01-017#1`
21 May 2024	Malta	Asserted it	Asserts	111	`SC09632-01-017#1`
22 May 2024	Maldives	Placed the word on it	Asserts	114	`SC09632-03-046#5`
22 May 2024	Maldives	Asserted it	Asserts	112	`SC09632-03-046#5`
28 May 2025	Sierra Leone	Placed the word on it	Asserts	115	`SC09923-01-015#1`
28 May 2025	Sierra Leone	Asserted it	Asserts	113	`SC09923-01-015#1`
15 April 2026	Latvia	Placed the word on it	Asserts	116	`SC10136-01-016#1`
15 April 2026	Latvia	Asserted it	Asserts	114	`SC10136-01-016#1`
17 June 2026	Somalia	Placed the word on it	Asserts	117	`SC10175-01-017#1`
17 June 2026	Somalia	Asserted it	Asserts	115	`SC10175-01-017#1`

Height is a count of delegations, not of occurrences: each is counted once, on the day it first crossed that line.

How to read this

The solid amber curve counts affiliations with an assertion recorded by that date; the dashed curve counts those with a rejection. The faint line, where different, counts any mention of the case. Each affiliation enters each curve once, at its first recorded occurrence. The table lists those first mentions.

What it does not show

These are first occurrences within this corpus, as classified by the model. An absent affiliation may not have spoken about the case. The curves do not track subsequent changes of position or establish when a state adopted a policy. Model disagreement can make a case unavailable.

More on this figure

The time axis stays fixed across cases; the vertical scale adjusts to each case's counts. Changing membership and speaking opportunities can affect the curves. An affiliation can appear in both assertion and rejection curves, and stays counted after its first mention.

Where model comparison is available, a case needs F1 of at least 0.8 between runs to be shown. This score measures overlap in the occurrences assigned to that case. A lower score makes the first recorded date too dependent on the model choice; it is not a direct estimate of accuracy.

Cases over time

Which situations is the word applied to in each year, and how often to none?

Model-derived · experimental

Drawing…

Bars count occurrences the model applied to a case; the line below is a share of all of them.

How to read this

Each bar stacks one year's occurrences by the case the model assigned, the leading cases in their own colour and the rest pooled. The line below is the share the model applied to no case: the Convention, a tribunal's or an office's name, a general formula. Select a bar to read that case in that year.

What it does not show

These are a model's labels, not yet checked against a human reading. Before 1992 a year holds few occurrences, so one wrong label moves it; the share is withheld below 20 occurrences.

More on this figure

The busiest year for cases is 2024, with 826 occurrences.

Historical referents such as the Holocaust are stacked with the cases, because a speaker applies the word to them as to a situation. A case the controlled list does not hold yet joins the pooled remainder.

The contested passages

Which passages did the two models read differently, and how?

Model-derived · experimental

Loading the annotations and the concordance for genocide…

One row per occurrence, not per disagreement: an occurrence the two runs split on three fields is one row, ranked above one they split on a single field.

Who rejects the word

When a delegation says genocide, is it making the claim or refusing it?

Model-derived · experimental

RejectsReports without a position †ConditionalAssertsUnclearNot applicableNo position (not a case)

A dagger flags categories whose boundaries may depend on the model, especially reporting a claim versus asserting it. Use the model-comparison table and quotations to examine that distinction.

Delegations with at least 20 eligible occurrences, ordered by how many of them the model labelled as rejecting the characterisation, then by name
Delegation	Eligible	Rejects	Positions
Yugoslavia	60	32	asserts 9, reports without a position 1, rejects 32, no position (not a case) 18
India	87	29	asserts 32, reports without a position 5, rejects 29, no position (not a case) 21
Sudan	92	25	asserts 44, reports without a position 1, rejects 25, conditional 1, no position (not a case) 21
Israel	176	18	asserts 127, reports without a position 2, rejects 18, conditional 5, no position (not a case) 24
United States	352	17	asserts 223, reports without a position 4, rejects 17, conditional 22, no position (not a case) 86
Russia	120	10	asserts 67, reports without a position 2, rejects 10, conditional 4, no position (not a case) 37
DR Congo	94	8	asserts 67, reports without a position 3, rejects 8, conditional 2, no position (not a case) 14
South Africa	99	7	asserts 43, reports without a position 1, rejects 7, conditional 11, no position (not a case) 37
Turkey	69	6	asserts 35, reports without a position 2, rejects 6, conditional 1, no position (not a case) 25
United Kingdom	176	6	asserts 98, reports without a position 1, rejects 6, conditional 8, no position (not a case) 63
Cyprus	27	5	asserts 15, rejects 5, no position (not a case) 7
Pakistan	157	5	asserts 92, reports without a position 2, rejects 5, conditional 20, no position (not a case) 38
Albania	65	4	asserts 39, rejects 4, conditional 1, no position (not a case) 21
China	42	4	asserts 11, reports without a position 2, rejects 4, conditional 3, no position (not a case) 22
Portugal	27	4	asserts 7, reports without a position 1, rejects 4, no position (not a case) 15
Azerbaijan	44	3	asserts 14, rejects 3, no position (not a case) 27
Burundi	85	3	asserts 42, reports without a position 4, rejects 3, conditional 14, no position (not a case) 22
Congo	21	2	asserts 9, reports without a position 1, rejects 2, conditional 1, no position (not a case) 8
Philippines	20	2	rejects 2, conditional 1, no position (not a case) 17
Slovakia	24	2	asserts 2, rejects 2, no position (not a case) 20
Syria	120	2	asserts 93, reports without a position 3, rejects 2, conditional 6, no position (not a case) 16
Ukraine	94	2	asserts 55, reports without a position 4, rejects 2, conditional 1, no position (not a case) 30, unclear 2
(personal capacity)	131	1	asserts 109, reports without a position 2, rejects 1, conditional 5, no position (not a case) 14
Angola	33	1	asserts 16, rejects 1, conditional 2, no position (not a case) 14
Gabon	22	1	asserts 10, rejects 1, no position (not a case) 11
Hungary	20	1	asserts 2, reports without a position 1, rejects 1, no position (not a case) 16
International Criminal Court	52	1	asserts 30, reports without a position 1, rejects 1, conditional 5, no position (not a case) 15
Ireland	35	1	asserts 14, rejects 1, conditional 1, no position (not a case) 19
Kenya	23	1	asserts 10, rejects 1, no position (not a case) 12
Libya	55	1	asserts 41, rejects 1, conditional 3, no position (not a case) 10
Liechtenstein	70	1	asserts 12, rejects 1, no position (not a case) 57
Lithuania	69	1	asserts 25, rejects 1, conditional 5, no position (not a case) 38
Norway	20	1	asserts 4, rejects 1, no position (not a case) 15
Qatar	29	1	asserts 12, reports without a position 1, rejects 1, conditional 4, no position (not a case) 11
Sweden	21	1	asserts 8, reports without a position 1, rejects 1, conditional 1, no position (not a case) 10
United Nations	604	1	asserts 340, reports without a position 11, rejects 1, conditional 44, no position (not a case) 208
Afghanistan	25	0	asserts 15, conditional 5, no position (not a case) 5
Algeria	80	0	asserts 65, conditional 7, no position (not a case) 8
Argentina	87	0	asserts 36, reports without a position 1, conditional 3, no position (not a case) 47
Armenia	97	0	asserts 32, conditional 3, no position (not a case) 62
Australia	41	0	asserts 17, reports without a position 1, conditional 4, no position (not a case) 19
Bangladesh	60	0	asserts 38, reports without a position 1, conditional 1, no position (not a case) 20
Belgium	35	0	asserts 6, reports without a position 2, conditional 1, no position (not a case) 26
Bolivia	21	0	asserts 12, no position (not a case) 9
Bosnia and Herzegovina	186	0	asserts 140, conditional 4, no position (not a case) 42
Botswana	23	0	asserts 11, reports without a position 1, no position (not a case) 11
Brazil	37	0	asserts 11, conditional 5, no position (not a case) 21
Canada	63	0	asserts 26, reports without a position 1, conditional 1, no position (not a case) 35
Chad	50	0	asserts 35, no position (not a case) 15
Chile	45	0	asserts 9, conditional 2, no position (not a case) 33, unclear 1
Colombia	35	0	asserts 24, reports without a position 2, conditional 1, no position (not a case) 7, unclear 1
Costa Rica	32	0	asserts 8, no position (not a case) 24
Croatia	73	0	asserts 44, conditional 1, no position (not a case) 28
Cuba	99	0	asserts 93, reports without a position 1, conditional 1, no position (not a case) 4
Czech Republic	33	0	asserts 22, reports without a position 1, conditional 1, no position (not a case) 9
Denmark	22	0	asserts 7, conditional 2, no position (not a case) 13
Ecuador	26	0	asserts 5, conditional 3, no position (not a case) 18
Egypt	31	0	asserts 20, no position (not a case) 11
Estonia	25	0	asserts 11, conditional 2, no position (not a case) 12
European Union	22	0	asserts 9, no position (not a case) 13
France	208	0	asserts 118, reports without a position 3, conditional 7, no position (not a case) 80
Germany	56	0	asserts 24, reports without a position 1, conditional 2, no position (not a case) 29
Ghana	41	0	asserts 13, reports without a position 1, conditional 2, no position (not a case) 25
High Representative for the Implementation of the Peace Agreement on Bosnia and Herzegovina	42	0	asserts 36, no position (not a case) 6
Indonesia	25	0	asserts 16, reports without a position 1, conditional 3, no position (not a case) 5
Iran	117	0	asserts 98, conditional 4, no position (not a case) 15
Iraq	68	0	asserts 53, reports without a position 4, conditional 4, no position (not a case) 7
Italy	29	0	asserts 4, conditional 2, no position (not a case) 23
Jordan	80	0	asserts 54, conditional 3, no position (not a case) 23
Kosovo	37	0	asserts 34, conditional 1, no position (not a case) 2
Kuwait	48	0	asserts 27, conditional 5, no position (not a case) 16
League of Arab States	39	0	asserts 37, conditional 1, no position (not a case) 1
Luxembourg	55	0	asserts 20, reports without a position 1, conditional 3, no position (not a case) 31
Malaysia	60	0	asserts 32, reports without a position 1, conditional 5, no position (not a case) 22
Maldives	29	0	asserts 20, conditional 1, no position (not a case) 8
Mauritania	36	0	asserts 34, conditional 1, no position (not a case) 1
Mexico	42	0	asserts 8, no position (not a case) 34
Morocco	26	0	asserts 13, no position (not a case) 13
Mozambique	28	0	asserts 8, conditional 3, no position (not a case) 17
Netherlands	58	0	asserts 22, reports without a position 1, conditional 7, no position (not a case) 28
New Zealand	66	0	asserts 39, reports without a position 1, conditional 5, no position (not a case) 21
Nicaragua	34	0	asserts 28, reports without a position 1, conditional 1, no position (not a case) 4
Nigeria	70	0	asserts 32, conditional 2, no position (not a case) 36
Palestine	242	0	asserts 194, reports without a position 1, conditional 10, no position (not a case) 37
Palestine Liberation Organization	53	0	asserts 46, conditional 4, no position (not a case) 3
Peru	43	0	asserts 5, reports without a position 1, no position (not a case) 37
Poland	30	0	asserts 8, conditional 3, no position (not a case) 19
Rwanda	1,095	0	asserts 933, reports without a position 1, conditional 12, no position (not a case) 149
Saudi Arabia	42	0	asserts 35, conditional 1, no position (not a case) 6
Senegal	29	0	asserts 14, conditional 2, no position (not a case) 13
Sierra Leone	83	0	asserts 27, conditional 6, no position (not a case) 50
Singapore	20	0	asserts 11, no position (not a case) 9
Slovenia	71	0	asserts 21, conditional 9, no position (not a case) 41
South Korea	30	0	asserts 7, conditional 1, no position (not a case) 22
Spain	41	0	asserts 15, reports without a position 1, conditional 1, no position (not a case) 24
Switzerland	37	0	asserts 9, conditional 9, no position (not a case) 19
Tanzania	33	0	asserts 16, reports without a position 1, conditional 2, no position (not a case) 14
Tunisia	46	0	asserts 35, conditional 1, no position (not a case) 10
Uganda	21	0	asserts 16, conditional 1, no position (not a case) 4
United Arab Emirates	26	0	asserts 14, conditional 1, no position (not a case) 11
Uruguay	36	0	asserts 3, conditional 4, no position (not a case) 29
Venezuela	69	0	asserts 43, reports without a position 2, conditional 2, no position (not a case) 22

All position counts, withheld delegations included

Delegation	Eligible	Asserts	Reports without a position	Rejects	Conditional	No position (not a case)	Unclear	Not applicable
Yugoslavia	60	9	1	32	0	18	0	0
India	87	32	5	29	0	21	0	0
Sudan	92	44	1	25	1	21	0	0
Israel	176	127	2	18	5	24	0	0
United States	352	223	4	17	22	86	0	0
Russia	120	67	2	10	4	37	0	0
DR Congo	94	67	3	8	2	14	0	0
South Africa	99	43	1	7	11	37	0	0
Turkey	69	35	2	6	1	25	0	0
United Kingdom	176	98	1	6	8	63	0	0
Cyprus	27	15	0	5	0	7	0	0
Pakistan	157	92	2	5	20	38	0	0
Albania	65	39	0	4	1	21	0	0
China	42	11	2	4	3	22	0	0
Portugal	27	7	1	4	0	15	0	0
Azerbaijan	44	14	0	3	0	27	0	0
Burundi	85	42	4	3	14	22	0	0
Congo	21	9	1	2	1	8	0	0
Philippines	20	0	0	2	1	17	0	0
Slovakia	24	2	0	2	0	20	0	0
Syria	120	93	3	2	6	16	0	0
Ukraine	94	55	4	2	1	30	2	0
(personal capacity)	131	109	2	1	5	14	0	0
Angola	33	16	0	1	2	14	0	0
Gabon	22	10	0	1	0	11	0	0
Hungary	20	2	1	1	0	16	0	0
International Criminal Court	52	30	1	1	5	15	0	0
Ireland	35	14	0	1	1	19	0	0
Kenya	23	10	0	1	0	12	0	0
Libya	55	41	0	1	3	10	0	0
Liechtenstein	70	12	0	1	0	57	0	0
Lithuania	69	25	0	1	5	38	0	0
Norway	20	4	0	1	0	15	0	0
Qatar	29	12	1	1	4	11	0	0
Sweden	21	8	1	1	1	10	0	0
United Nations	604	340	11	1	44	208	0	0
Afghanistan	25	15	0	0	5	5	0	0
Algeria	80	65	0	0	7	8	0	0
Argentina	87	36	1	0	3	47	0	0
Armenia	97	32	0	0	3	62	0	0
Australia	41	17	1	0	4	19	0	0
Bangladesh	60	38	1	0	1	20	0	0
Belgium	35	6	2	0	1	26	0	0
Bolivia	21	12	0	0	0	9	0	0
Bosnia and Herzegovina	186	140	0	0	4	42	0	0
Botswana	23	11	1	0	0	11	0	0
Brazil	37	11	0	0	5	21	0	0
Canada	63	26	1	0	1	35	0	0
Chad	50	35	0	0	0	15	0	0
Chile	45	9	0	0	2	33	1	0
Colombia	35	24	2	0	1	7	1	0
Costa Rica	32	8	0	0	0	24	0	0
Croatia	73	44	0	0	1	28	0	0
Cuba	99	93	1	0	1	4	0	0
Czech Republic	33	22	1	0	1	9	0	0
Denmark	22	7	0	0	2	13	0	0
Ecuador	26	5	0	0	3	18	0	0
Egypt	31	20	0	0	0	11	0	0
Estonia	25	11	0	0	2	12	0	0
European Union	22	9	0	0	0	13	0	0
France	208	118	3	0	7	80	0	0
Germany	56	24	1	0	2	29	0	0
Ghana	41	13	1	0	2	25	0	0
High Representative for the Implementation of the Peace Agreement on Bosnia and Herzegovina	42	36	0	0	0	6	0	0
Indonesia	25	16	1	0	3	5	0	0
Iran	117	98	0	0	4	15	0	0
Iraq	68	53	4	0	4	7	0	0
Italy	29	4	0	0	2	23	0	0
Jordan	80	54	0	0	3	23	0	0
Kosovo	37	34	0	0	1	2	0	0
Kuwait	48	27	0	0	5	16	0	0
League of Arab States	39	37	0	0	1	1	0	0
Luxembourg	55	20	1	0	3	31	0	0
Malaysia	60	32	1	0	5	22	0	0
Maldives	29	20	0	0	1	8	0	0
Mauritania	36	34	0	0	1	1	0	0
Mexico	42	8	0	0	0	34	0	0
Morocco	26	13	0	0	0	13	0	0
Mozambique	28	8	0	0	3	17	0	0
Netherlands	58	22	1	0	7	28	0	0
New Zealand	66	39	1	0	5	21	0	0
Nicaragua	34	28	1	0	1	4	0	0
Nigeria	70	32	0	0	2	36	0	0
Palestine	242	194	1	0	10	37	0	0
Palestine Liberation Organization	53	46	0	0	4	3	0	0
Peru	43	5	1	0	0	37	0	0
Poland	30	8	0	0	3	19	0	0
Rwanda	1,095	933	1	0	12	149	0	0
Saudi Arabia	42	35	0	0	1	6	0	0
Senegal	29	14	0	0	2	13	0	0
Sierra Leone	83	27	0	0	6	50	0	0
Singapore	20	11	0	0	0	9	0	0
Slovenia	71	21	0	0	9	41	0	0
South Korea	30	7	0	0	1	22	0	0
Spain	41	15	1	0	1	24	0	0
Switzerland	37	9	0	0	9	19	0	0
Tanzania	33	16	1	0	2	14	0	0
Tunisia	46	35	0	0	1	10	0	0
Uganda	21	16	0	0	1	4	0	0
United Arab Emirates	26	14	0	0	1	11	0	0
Uruguay	36	3	0	0	4	29	0	0
Venezuela	69	43	2	0	2	22	0	0
Fewer than 20 eligible mentions, so no band in the figure
Myanmar	9	0	0	4	0	5	0	0
Somalia	17	9	0	4	1	3	0	0
South Sudan	15	0	0	4	2	9	0	0
Greece	19	7	0	3	0	9	0	0
Latvia	13	2	0	3	2	6	0	0
Cambodia	10	8	0	1	0	1	0	0
Dominican Republic	6	5	0	1	0	0	0	0
Guinea	16	13	0	1	0	2	0	0
Honduras	13	6	0	1	3	3	0	0
Madagascar	2	0	0	1	1	0	0	0
Malta	18	5	0	1	3	9	0	0
Romania	11	0	0	1	1	9	0	0
African National Congress	11	10	0	0	1	0	0	0
African Union	7	4	0	0	0	3	0	0
All-Africa Conference of Churches	1	0	0	0	0	1	0	0
Andorra	1	0	0	0	0	1	0	0
Austria	14	1	0	0	0	13	0	0
Bahrain	13	6	0	0	0	7	0	0
Barbados	1	1	0	0	0	0	0	0
Belarus	1	1	0	0	0	0	0	0
Benin	13	6	0	0	1	6	0	0
Brunei	2	1	0	0	0	1	0	0
Bulgaria	10	4	0	0	2	4	0	0
Burkina Faso	7	3	0	0	0	4	0	0
Cameroon	7	5	0	0	0	2	0	0
Central African Republic	2	1	0	0	1	0	0	0
Centre for Strategic Communication and Information Security (Ukraine)	3	3	0	0	0	0	0	0
Columbia University (New York)	2	0	0	0	0	2	0	0
Comoros	2	1	0	0	1	0	0	0
Confluence Advisory (Organization)	3	1	0	0	2	0	0	0
Council of the European Union	1	1	0	0	0	0	0	0
Darfur Advocacy Group (Sudan)	5	2	0	0	3	0	0	0
Darfur Women Action Group	7	4	0	0	2	1	0	0
Djibouti	11	9	0	0	0	2	0	0
East Timor	4	3	0	0	0	1	0	0
Economic Community of Central African States	3	3	0	0	0	0	0	0
Equatorial Guinea	9	5	0	0	0	4	0	0
Ethiopia	16	7	0	0	4	5	0	0
European Commission	1	1	0	0	0	0	0	0
Facilitator of the Burundi Peace Process	1	0	0	0	0	0	1	0
Farida Global Organization	11	11	0	0	0	0	0	0
Fiji	5	2	0	0	0	3	0	0
Finland	9	0	0	0	0	9	0	0
Foundation for Human Rights (South Africa)	1	1	0	0	0	0	0	0
FRELIMO	1	1	0	0	0	0	0	0
Gambia	1	1	0	0	0	0	0	0
George Washington University (Washington, D.C.)	3	2	0	0	0	1	0	0
Georgia	12	6	0	0	0	6	0	0
German Democratic Republic	1	1	0	0	0	0	0	0
Grenada	3	3	0	0	0	0	0	0
Guatemala	14	1	1	0	2	10	0	0
Guinea-Bissau	1	1	0	0	0	0	0	0
Guyana	16	10	0	0	3	3	0	0
Haiti	5	2	0	0	1	2	0	0
Holy See	16	2	0	0	1	13	0	0
IBRD	1	1	0	0	0	0	0	0
Intergovernmental Authority on Development	2	2	0	0	0	0	0	0
International Center for Transitional Justice	1	0	0	0	0	1	0	0
International Committee of the Red Cross	4	0	0	0	0	4	0	0
Iraqi Al-Amal Association	1	1	0	0	0	0	0	0
Islamic Group at the United Nations	1	1	0	0	0	0	0	0
Ivory Coast	16	6	0	0	3	7	0	0
Jamaica	14	7	0	0	0	7	0	0
Japan	16	6	0	0	3	7	0	0
Kaladan Press (Organization)	1	1	0	0	0	0	0	0
Kazakhstan	4	0	0	0	1	3	0	0
Laos	1	1	0	0	0	0	0	0
Lebanon	10	8	0	0	1	1	0	0
Les Centres GHESKIO (Haiti)	1	1	0	0	0	0	0	0
Lesotho	1	1	0	0	0	0	0	0
Liberia	4	2	1	0	1	0	0	0
Malawi	2	1	0	0	0	1	0	0
Mali	19	13	0	0	0	6	0	0
Mauritius	7	4	0	0	0	3	0	0
McGill University (Montréal, Canada)	1	1	0	0	0	0	0	0
Medical Aid for Palestinians (Organization)	2	1	0	0	1	0	0	0
Moldova	1	0	0	0	0	1	0	0
Monaco	7	0	0	0	0	7	0	0
Mongolia	4	1	0	0	0	3	0	0
Montenegro	1	0	0	0	0	1	0	0
MPLA	8	8	0	0	0	0	0	0
Municipal Association of Women (Organization : Colombia)	1	0	0	0	0	1	0	0
Namibia	17	11	0	0	0	6	0	0
Nepal	1	0	0	0	0	1	0	0
New Development Bank	1	0	0	0	0	1	0	0
NGO Working Group on Women, Peace and Security	3	1	1	0	0	1	0	0
Niger	10	3	0	0	1	6	0	0
Nigeria; United Nations	1	1	0	0	0	0	0	0
North Korea	1	1	0	0	0	0	0	0
Norwegian Refugee Council	1	0	0	0	0	1	0	0
Oman	16	10	0	0	2	4	0	0
Organisation of Islamic Cooperation	1	1	0	0	0	0	0	0
Organization of African Unity	5	5	0	0	0	0	0	0
Organization of the Islamic Conference	10	7	0	0	1	2	0	0
Oxfam International	1	1	0	0	0	0	0	0
PAEMA	5	5	0	0	0	0	0	0
Palau	1	0	0	0	0	1	0	0
Pan Africanist Congress of Azania (South Africa)	7	6	0	0	0	1	0	0
Panama	18	6	0	0	0	12	0	0
Patriotic Front of Zimbabwe	2	2	0	0	0	0	0	0
POLISARIO Front	4	4	0	0	0	0	0	0
Regional Center for Human Rights (Ukraine)	1	0	0	0	0	1	0	0
Rossiya Segodnya (Russian Federation)	1	1	0	0	0	0	0	0
Samoa	1	0	0	0	0	1	0	0
Sao Tome and Principe	1	0	0	0	0	1	0	0
Sauti Ya Mama Mukongomani (Organization : Democratic Republic of the Congo)	1	1	0	0	0	0	0	0
secretariat of the OAU	2	2	0	0	0	0	0	0
South West Africa People’s Organization	1	1	0	0	0	0	0	0
Sri Lanka	2	0	0	0	0	2	0	0
St. Vincent and the Grenadines	8	3	0	0	0	5	0	0
State of Palestine	1	1	0	0	0	0	0	0
Strategic Initiative for Women in the Horn of Africa (Organization)	5	3	0	0	1	1	0	0
Sudan Family Planning Association	2	0	0	0	1	1	0	0
SWAPO	3	3	0	0	0	0	0	0
Swaziland	1	1	0	0	0	0	0	0
Thailand	2	0	0	0	0	2	0	0
Togo	2	0	0	0	1	1	0	0
TRIAL International (Organization : Switzerland)	3	3	0	0	0	0	0	0
Trinidad and Tobago	2	0	0	0	0	2	0	0
Ukrainian Trade Union of Law Workers	1	0	0	0	0	1	0	0
University of Oxford	3	0	0	0	0	3	0	0
US/Middle East Project (Organization : New York)	2	0	0	0	0	2	0	0
Vienna Center for Disarmament and Non-Proliferation	1	0	0	0	0	1	0	0
Vietnam	13	10	0	0	1	2	0	0
VOICE Amplified (Organization)	1	0	0	0	1	0	0	0
Women and Peace Studies Organization (Afghanistan)	1	1	0	0	0	0	0	0
Women's Initiatives for Gender Justice	1	0	0	0	0	1	0	0
Women's League of Burma	2	2	0	0	0	0	0	0
Women's Peace Network (Myanmar)	2	2	0	0	0	0	0	0
Yale University (New Haven, Conn.)	8	6	0	0	0	2	0	0
Yemen	11	11	0	0	0	0	0	0
Yemen Arab Republic	2	2	0	0	0	0	0	0
Yemen People's Republic	3	3	0	0	0	0	0	0
Zambia	16	16	0	0	0	0	0	0
Zan Times (Firm : Afghanistan)	1	0	0	0	1	0	0	0
Zimbabwe	14	10	0	0	0	4	0	0

Width is the share of that delegation's own eligible occurrences, not of the corpus.

How to read this

Each row is an affiliation with at least 20 eligible mentions. The band shows how the model labelled them. The two columns count those mentions and the ones labelled rejects. Rows are ordered by that second count, then by name. Hover over an affiliation for all position counts.

What it does not show

A count grows with how often a delegation spoke, and labels can misread negation or reported speech. No delegation is marked as unusual until human coding measures how often the rejects label is right. 136 affiliations under 20 eligible mentions appear only in the full table.

More on this figure

Asserts applies genocide to a case; rejects disputes that characterisation. Reports without a position attributes a claim without adopting it. Conditional makes its application conditional. Other categories cover abstract uses, uncertainty or inapplicable cases. Each eligible mention receives one position label. These are classifications of passages, not fixed delegation positions.

Until the label has been checked, the figure shows counts only: no share of rejections, no interval and no mark for an unusual delegation. These return only if rejects meets the standard in the project's evaluation plan. See classification and validation methods.

Who is named as the victim

Which peoples do speakers name as the victims of a genocide, decade by decade?

Model-derived · experimental

Victim groups down the side, decades across; each cell counts the occurrences naming that group.
Group	1940s	1950s	1960s	1970s	1980s	1990s	2000s	2010s	2020s	All
Palestinians			4	33	61	8	66	30	471	673
Tutsi						9	9	180	116	314
Yazidis								57	41	98
Bosniaks (Bosnian Muslims)						30	4	33	28	95
People of Bosnia and Herzegovina						55	1	3	1	60
Rwandans						16	16	18	6	56
Jews	1		8	9	7	2	3	3	10	43
Kosovo Albanians						8	6	10	17	41
Lebanese and Palestinians				2	24		4		5	35
Ukrainians									34	34
Africans			5	15	4	4				28
Armenians								15	12	27
Iraqis						7	6	10	4	27
Fur, Masalit and Zaghawa (Darfur)							4	11	11	26
Muslims in India	11	3	1						7	22
Victims and survivors (unspecified)				1		5	32	32	22	92
Civilians and populations (unspecified)			4	7	7	21	11	13	26	89
Women, children and men (unspecified)			1	1	6	3	14	11	10	46
Speaker's own people (unspecified)			3	9	5	1	2	1	19	40
Muslims (unspecified)	1				1	5				7

33 further peoples, named in 260 occurrences, are in the download.

Cells count occurrences whose victim group, as the AI model wrote it, falls in that group.

How to read this

Each row is a people the AI model named as the victim, grouped across spellings: "the Jewish people" and "Jews" count together. Rows are ranked over the whole period. Groups that name no particular people, such as civilians, sit under the rule, in italics.

What it does not show

An AI model wrote each victim label, and an AI-drafted table grouped them. The 650 rows of the grouping table were drafted by AI and none has yet been checked by a person. Read the quotations before relying on a count.

More on this figure

Across the whole period the people named most often is Palestinians: 673 occurrences, from 53 delegations.

A label the table does not yet group is left out of every row; the figure beside this one counts those labels. The table, annotations/lexicon/victim_groups.csv, keeps the place a speaker named in the original label: "Palestinians in Gaza" are counted as Palestinians.

Naming a people, or not

How often does a speaker name a particular people as the victim, and how often no one?

Model-derived · experimental

Drawing…

Each bar divides a decade's eligible occurrences into four kinds; the shares sum to 100%.

Which delegations name which people

Which delegations name which peoples as the victims of a genocide?

Model-derived · experimental

Delegations down the side, peoples across; each cell counts the occurrences in which that delegation names that people.
Delegation	Palestinians	Tutsi	Yazidis	Bosniaks (Bosnian Muslims)	People of Bosnia and Herzegovina	Rwandans	Jews	Kosovo Albanians	Lebanese and Palestinians	Ukrainians	All
Rwanda		210	2	3	1	14					234
Palestine	117										117
United Nations	2	32	10	9	1	6					67
Syria	51								2		65
Israel	1	1					29				57
Iran	43			3	5				4		55
Cuba	32								3		51
(personal capacity)	6		22					12		1	48
Pakistan	13			1	7		1	2	1		45
United States		6	9	12	1	1		1			44
Algeria	37			1	1						39
Iraq	15		5						1		35
Jordan	23	1					3		3		35
Libya	29										34
Palestine Liberation Organization	19						3		8		34
Bosnia and Herzegovina		1		20	9	1					31
Russia	7			1		3			3	1	31
Mauritania	26										28
Tunisia	26					1			1		28
Armenia			1								27

139 further delegations, with 734 occurrences, are in the download.

Cells count occurrences in which that delegation names that people, as grouped from the AI model's labels.

Instructions and categories

These are the instructions (the prompt) and category definitions used for this model run. They document how the model was asked to classify passages. The version and content identifier let you distinguish this run from later revisions.

Show the prompt (sha256:3772c44d400f…)

# Model annotation prompt — `genocide`

version: 3
date: 2026-09-07
codebook: `annotations/lexicon/CODEBOOK.md`, human schema version 3; model schema 3.1

Version 3 removes model self-reported confidence. Human confidence is unchanged.
Situation/modality coding remains a separate reviewed schema-4 proposal; it is
not introduced by this revision. Prompt v2 is archived byte-for-byte in
`prompts/v2.md`, so the Bayreuth run prepared with it remains readable.

The raw bytes of this file are hashed with SHA-256 into every run manifest and into every
row of every `annotations.jsonl`. Editing so much as a comma produces a different hash, and
a row can then no longer be matched to the text that produced it — so an edit here is a new
prompt version. Bump `version:` above, **move the old text to `prompts/v<n>.md` unchanged**,
and give the run that uses it a new run id. The archive is what lets an edit here be made at
all: `15_usage.py` resolves a run against every prompt this repository holds, so revising
the instrument costs a new run id and not the runs already paid for. The repository stores
and checks out this file with LF line endings (`.gitattributes`), so the hash does not depend
on the machine that computed it.

Two sections below are read by `scripts/lib/llm.py`, and only those two: the fenced block
under `## System` is the system message, verbatim; the fenced block under `## User template`
is the per-speech user message, with its `{placeholders}` filled in at run time. Everything
outside the two fences — this paragraph included — is documentation for a human and never
reaches the model.

The system block carries one placeholder, `{referents_table}`, rendered from
`annotations/lexicon/referents.csv` at run time as `id — label (years) — description`,
grouped by kind. The controlled list grows through reviewed additions to that file; the
prompt does not restate it, so adding a referent does not silently become a prompt edit.
What *would* change the meaning of a label — the field definitions, the boundary examples,
the cascade — lives here and is versioned here.

## What v2 changes, and on what evidence

v2 is the review of 1 September 2026, §4.2, carried out. Six changes:

1. **`stance` becomes two fields.** `speaker_position` records what the speaker does with the
   characterisation; `concrete_case` records whether the word is applied to a determinate
   case at all. v1's `neutral_legal_reference` was doing both jobs and competing with the
   `institutional_title_or_mandate` function and with three meta referents for the same
   decision — the largest referent confusion cells were exactly those pairs. Now the decision
   is made once and the other fields follow from it.
2. **`reports_without_position` is defined positively**, and a settled legal status is an
   assertion. v1's `attributes_or_reports` scored F1 0.37 across the two instruments, with
   445 occurrences read as *attributes* by one and *asserts* by the other — the single
   largest disagreement cell in the corpus.
3. **Ten worked examples, every one of them from the corpus**, quoted verbatim with the line
   id it was taken from, so a reader can go and check it. v1's examples were all invented,
   and the corpus's own recurrent formulae were absent from it.
4. **New fields the study's question needs**: `accused_actor`, `victim_group`,
   `own_state_accused`, `salience`, `referent_source` and a one-sentence `rationale`.
5. **Two rules the year ranges and the compound cases needed**, taken from the codebook so
   that a coder and a model read the same sentence.
6. **A false positive still needs a real quote.** v1's cascade said "all four fields, with no
   exception", and one model answered the *evidence* field with the literal string
   `not_applicable` three times.

Three of the new rules are stated as what the two committed runs already do, rather than as
a correction: both read the Special Adviser's title as a non-case mention in 176 of 176
occurrences, both code 98% of commemorations as assertions, and neither was ever told to.
Writing those down does not move them; it makes them testable, and it stops the next
instrument having to rediscover them.

The `distancing` rule is the opposite case. The grammatical-frames codebook
(`scripts/lib/node_frames.py`) finds 78 occurrences in which the label is marked as somebody
else's — *so-called*, *supposed*, scare quotes, *what she called "genocide"* — and the two
instruments part on them: 33 of Luna's are `attributes_or_reports`, 27 of Gemini's are
`rejects_or_denies`. That is §4.1's report/assert disagreement localised in one construction,
so one rule and one example can settle it.

**A provenance note about the two artefacts this prompt was written from.** The frames
artefact's `matched` column and the referent list's absorption table are both *projections*
over the two committed runs' text — a regex over a window, and a clustering of
`proposed_referent` strings. Neither is a re-coding, nothing in either committed run was
rewritten to produce them, and both are counted here as evidence of where the instruments
disagree and not as evidence of what the right label is. Only a run settles that, and no run
has yet been made against this prompt.

## System

```text
You are annotating occurrences of the word "genocide" in the verbatim records of the United
Nations Security Council, 1946-2024.

TASK BOUNDARY. You classify discourse. You never decide whether an underlying event legally
constitutes genocide, and no label you assign is a finding about what happened. The question
is always what the pattern found and what the speaker is doing with the word in this passage:
whether the matched string is really an instance of the expression and, where it is, how the
expression is being used here.

Read the whole speech before labelling anything. Then label each numbered occurrence
separately, from the passage around it, using the rest of the speech only to resolve what
that passage leaves open. Abstention is a correct answer, and a wrong label is worse than an
honest one: prefer "uncertain", "unclear" to a guess.

FIELDS

verdict — exactly one:
  true_positive   The highlighted span is a real instance of the expression. Example: "acts
                  of genocide".
  false_positive  The string does not instantiate the expression. Example: damaged OCR joins
                  unrelated fragments into something a tolerant pattern matched.
  uncertain       The transcription does not permit a defensible decision. Example: OCR
                  damage leaves the passage illegible.

quotation — exactly one:
  not_quoted              The current speaker uses the expression in their own formulation.
                          "We consider these acts genocide."
  direct_quotation        The expression occurs inside words explicitly quoted verbatim.
                          "The witness called it 'genocide'."
  attributed_or_reported  The speaker reports or paraphrases another actor's terminology.
                          "The Commission described the acts as genocide."
  unclear                 Attribution cannot be resolved from the available passage.
  not_applicable          False positives only.

  Quotation does not determine position. Naming the source of a characterization says where
  it came from; it says nothing about whether this speaker adopts it. A speaker may quote a
  claim in order to endorse it, in order to reject it, or in order to leave it standing
  unjudged, and those are three different values of speaker_position with the same value
  here.

concrete_case — exactly one. Answer this before speaker_position:
  yes             The word is applied, in this passage, to a determinate situation, event,
                  people or place: a case that could be named. It does not have to be named
                  in the passage — see referent_source — only determinate.
  no              The word is not applied to any case here. It names the crime as a legal
                  category, appears inside the title or mandate of an office, a treaty or a
                  tribunal, or stands in a general formula about prevention or atrocity.
  unclear         The passage does not let you tell which of the two it is.
  not_applicable  False positives only.

  This is the decision v1 made three times over and made differently each time. Make it once,
  here. Everything below follows from it: when concrete_case is "no", speaker_position is
  "no_position" and the referent is one of the three non-case referents; when it is "yes",
  speaker_position is one of the four real positions and the referent is a case.

  What matters is what the MATCHED WORD is doing, not what the sentence is about. "This
  crime is forbidden by the 1948 Convention on the Prevention and Punishment of the Crime of
  Genocide" is a sentence about a concrete crime in which the matched word occurs only
  inside a treaty's title: concrete_case is "no".

speaker_position — exactly one:
  asserts                   The speaker advances or endorses the characterization, or treats
                            it as established. "This is genocide." Also: a legal status the
                            speaker treats as settled — "those convicted of genocide by the
                            ICTR", "known genocidaires", "the ICTY decision finding him
                            guilty of genocide" — because a speaker who reasons from a
                            finding has adopted it.
  rejects                   The speaker disputes, denies or refuses the characterization.
                            "There was no genocide." Also: the speaker marks the label as
                            false while reporting it — see the distancing rule below.
  conditional               The speaker presents it as a possibility, a risk, a threshold or
                            a warning rather than as an established characterization.
                            "Unless prevented, this could become genocide."
  reports_without_position  Use this ONLY when both of these hold: the passage names the
                            actor doing the characterizing, AND the speaker neither adopts
                            nor rejects it. Both halves are required. If the characterizing
                            actor is not named, the speaker is speaking in their own voice,
                            which is one of the three above. If the speaker endorses what
                            they report, that is "asserts"; if they disown it, "rejects".
                            The value means "somebody said this and I am not saying whether
                            they were right", and nothing else.
  no_position               concrete_case is "no", so there is no characterization of a case
                            to take a position on. Use this value exactly when concrete_case
                            is "no", and never otherwise.
  unclear                   The evidence supports no reliable choice.
  not_applicable            False positives only.

  THE DISTANCING RULE. Ask what the hedge is scoping over.
  - Over the LABEL — scare quotes around the word, "so-called", "supposed", "purported",
    "what X calls genocide", "allegations of genocide" — the speaker is holding the
    characterization at arm's length as somebody else's and declining it: "rejects". Not
    "reports_without_position": marking a label as false is a position.
  - Over a PERSON OR AN ACT, with the category itself untouched — "alleged genocide
    financier", "those alleged to have committed genocide", "accused of genocide" — that is
    an unproven charge against an individual, made inside a category the speaker treats as
    real: "asserts". This is the ordinary language of an indictment, and it is not
    distancing at all.

function — one or more, as a JSON array of strings:
  accusation_or_qualification     Names or legally characterizes conduct.
  warning_or_prevention           Warns of risk, or urges prevention or protection.
  commemoration                   Remembers victims, anniversaries or historical events.
  accountability                  Concerns investigation, prosecution, punishment or
                                  impunity. A proceeding, an arrest, a sentence, a fugitive,
                                  an amnesty. NOT merely because a court or tribunal is
                                  named: "genocide denial by ICTR defence lawyers" is an
                                  accusation against the deniers and concerns no proceeding.
  institutional_title_or_mandate  Occurs in the name or mandate of an office, treaty,
                                  tribunal, adviser or institutional mechanism. This is a
                                  restatement of concrete_case "no", not a separate finding:
                                  a word inside a title is not applied to a case.
  other                           A clear function not represented above.
  unclear                         The function cannot be determined.
  not_applicable                  False positives only.

  "Prosecute those responsible for genocide so that it never happens again" may be
  ["accountability", "warning_or_prevention"]. Never repeat a label. "unclear" and
  "not_applicable" cannot be combined with any other label, so an array containing either of
  them has exactly one element. Prefer the smallest defensible set; do not infer purposes
  that are absent from the text.

  COMMEMORATION IS ASSERTION. A commemorative formula — an anniversary, a remembrance, "we
  honour the victims of the genocide" — treats the characterization as established, so its
  speaker_position is "asserts" even though nothing is being alleged against anyone in the
  passage. Add "accusation_or_qualification" only when the passage also characterizes
  someone's conduct.

referent — exactly one identifier from the controlled list below.

referent_source — exactly one. Where in the request you found the referent:
  passage  The passage around the occurrence names or unambiguously identifies it.
  speech   Another part of the speech does, and this passage does not.
  header   Only the header lines above the speech do — the speaker's country or
           organisation, the agenda item, the meeting.
  not_applicable  False positives only; and the value for referent "unclear".

  This field is a measurement of the instrument and not of the passage, so record what you
  actually used. A referent read off the header alone is a weaker assignment than one read
  off the passage, and the study needs to know how many there are.

accused_actor — free text, a few words. Who the passage says did it, or is charged with it,
  as the passage names them: "ex-FAR and Interahamwe", "Radislav Krstic", "the Government of
  the Sudan". Empty when the passage accuses no one.

victim_group — free text, a few words. The group the word is applied on behalf of, as the
  passage names them: "the Tutsi in Rwanda", "the Bosnian Muslims of Srebrenica". Empty when
  the passage names none.

  Both fields are copied from the passage, never inferred from the case. If the passage says
  "genocide in Rwanda" and names neither perpetrator nor victims, both are empty. That is
  data, not a failure.

own_state_accused — exactly one:
  yes             The speaker's own State or organisation is the one the passage accuses, or
                  the one it would accuse if the characterization stood. A denial by the
                  accused State is "yes": the accusation is what is being denied.
  no              Somebody else is accused.
  not_applicable  The passage accuses no one, this is a non-case mention, or the verdict is
                  false_positive.

salience — exactly one:
  passing         The word appears in a list, a formula or an aside, and the passage does not
                  argue about it. The four-crime responsibility-to-protect formula and the
                  triad "genocide, war crimes and crimes against humanity" are passing.
  substantive     The characterization is what the passage is doing: it is argued for,
                  argued against, dwelt on, or the subject of the sentence.
  not_applicable  False positives only.

rationale — one sentence, free text, always present. Why the position and the case decision
  are what they are, in your own words. Not a restatement of the labels: say what in the
  passage decided them. Where you hesitated between two values, name the other one. This is
  read by humans adjudicating disagreements and costs almost nothing to write.


THE FALSE-POSITIVE CASCADE

If verdict is "false_positive", then quotation, concrete_case, speaker_position, referent,
referent_source, own_state_accused, salience and every element of function are
"not_applicable"; accused_actor, victim_group and proposed_referent are empty strings. If
verdict is anything else, "not_applicable" may not appear in quotation, concrete_case,
speaker_position, function or referent: there it is reserved for false positives and means
nothing else. referent_source, own_state_accused and salience are the exceptions, because
each of them has a genuine "there is nothing here to record" answer that is not about the
verdict at all.

The cascade stops at the evidence. A false positive still needs a real, verbatim
evidence_quote and a real rationale: the claim "this match is not the word being used" is a
claim about a passage, and it is unreadable without the passage. Do not write the string
"not_applicable" into evidence_quote. A false positive whose quote cannot be found in the
speech is refused when the row is written.

REFERENT

Choose the referent the occurrence is ABOUT: the case, event, people or institution the word
is applied to in this passage. It is not every situation the speech mentions, and it is not
the speaker's own country unless the passage applies the word to it.

{referents_table}

The years in parentheses are DOCUMENTATION, NOT A CONSTRAINT. They say when the case is
usually invoked; they do not say which speeches may carry it. Where a passage names a period
a referent's years do not cover, code the referent anyway. A 2009 speech about Gaza takes the
Israel/Palestine referent because that is what the speech is about, not because 2009 falls in
a range.

A passage naming two cases is coded as the first one named, and the pair goes into
"proposed_referent" as the speaker gave it. "Rwanda and Srebrenica" is the Rwanda referent
with proposed_referent "Rwanda and Srebrenica". Where a passage names a case and something
more specific inside it — Khojaly inside Nagorno-Karabakh, the Yazidis inside ISIL's crimes —
that is not a compound: code the more specific identifier, and leave proposed_referent empty.

When concrete_case is "no", choose among the three non-case referents in this order, first
match winning:

  institutional title or mandate   The word is inside the name or mandate of an office, a
                                   body or a tribunal. "Special Adviser on the Prevention of
                                   Genocide".
  Convention and legal definition  The passage is about the 1948 Convention, the legal
                                   definition of the crime, or jurisdiction over it. "the
                                   Convention on the Prevention and Punishment of the Crime
                                   of Genocide", "the definition in article II".
  genocide in general              Everything else abstract: prevention doctrine, the
                                   four-crime formula, a general statement about the crime.

Using the three reserved identifiers:

  other           The referent is identifiable but is not on the list. Use it, and put a
                  short name for what you identified in "proposed_referent" — a place, a
                  conflict, a people or an institution, a few words at most. Never invent an
                  identifier; the list grows only through human review.
  unclear         The speech does not let you tell which referent is meant. Choosing this is
                  correct whenever it is true. Never guess.
  not_applicable  False positives only.

"proposed_referent" is a short free-text note about the referent. It is required when the
referent is exactly "other", it carries the pair when a compound is coded as the first case
named, it is empty on a false positive, and it is otherwise empty.

EVIDENCE

"evidence_quote" is the shortest verbatim passage of the speech that both contains the
matched word of this occurrence and supports every label you assigned to it. Usually a
clause or one sentence; two sentences when the attribution or the condition begins in the
one before. Never the whole speech.

The minimum is the clause containing the modal or the attribution — whatever it is that
makes your position label what it is. "could amount to genocide" rather than "genocide",
because the modal is what makes the position conditional; "what she called 'genocide'"
rather than "genocide", because the attribution is the whole of the reason. A bare "have been
convicted of genocide" cannot support a position label, and a quote too short to support one
is treated as unlocated evidence.

Copy it character for character out of the speech text supplied with the request: the same
spelling, punctuation, capitalisation and spacing, including any OCR damage. The records are
scanned, and they break words across lines: "the Secretary- General's warning" is what the
record says, and it is what you copy. No ellipses, no corrections, no added quotation marks,
no markup, nothing joined across a gap. The quote is located back in the speech
automatically, and a quote that cannot be found there is recorded as invalid evidence.

WORKED EXAMPLES

Ten occurrences from this corpus, quoted verbatim, with the line id each was taken from. They
are the cases the previous version of this prompt got least agreement on. Fields not shown
take the values the definitions above give them.

1. Commemoration. UNSC_2014_SPV.7105_spch0017#3, Rwanda, 29 January 2014.
   evidence_quote: "As we commemorate the twentieth anniversary of the 1994 genocide against
   the Tutsi in Rwanda"
   concrete_case yes | speaker_position asserts | quotation not_quoted |
   function ["commemoration"] | referent: Rwanda | referent_source passage |
   accused_actor "" | victim_group "the Tutsi in Rwanda" | own_state_accused not_applicable |
   salience substantive
   rationale: "Commemorating an anniversary treats the characterization as established, so
   the speaker asserts it even though the passage accuses no one."

2. The atrocity triad. UNSC_2005_SPV.5264_spch0022#1, European Union, 20 September 2005.
   evidence_quote: "our responsibility to protect populations from genocide, war crimes,
   ethnic cleansing and crimes against humanity"
   concrete_case no | speaker_position no_position | quotation not_quoted |
   function ["warning_or_prevention"] | referent: genocide in general |
   referent_source passage | accused_actor "" | victim_group "" |
   own_state_accused not_applicable | salience passing
   rationale: "The word names one of the four crimes in the responsibility-to-protect
   formula and is applied to no case, so there is no characterization to take a position on."

3. Perpetrator noun. UNSC_1999_SPV.3987_spch0015#2, United States, 19 March 1999, on the
   Democratic Republic of the Congo.
   evidence_quote: "to collaborate militarily with ex-FAR and interahamwe, known genocidaires"
   concrete_case yes | speaker_position asserts | quotation not_quoted |
   function ["accusation_or_qualification"] | referent: Rwanda | referent_source passage |
   accused_actor "ex-FAR and interahamwe" | victim_group "" | own_state_accused no |
   salience passing
   rationale: "The perpetrator noun applies the word to the killings in Rwanda and not to the
   Congolese war the debate is about, and 'known' treats the characterization as settled."

4. Denial and ideology. UNSC_2011_SPV.6545_spch0010#9, Rwanda, 6 June 2011.
   evidence_quote: "the ongoing scourge of genocide denial by some in the academic and legal
   professions"
   concrete_case yes | speaker_position asserts | quotation not_quoted |
   function ["accusation_or_qualification"] | referent: Rwanda | referent_source speech |
   accused_actor "academics and lawyers who deny it" | victim_group "" |
   own_state_accused no | salience substantive
   rationale: "Condemning denial asserts the genocide that is denied; the accusation is
   against the deniers, and no proceeding is at issue, so accountability is not added."

5. Own-state denial. UNSC_2017_SPV.7963_spch0019#6, Sudan, 8 June 2017.
   evidence_quote: "the report of the International Commission of Inquiry on Darfur, which
   was submitted to the Council in early 2005, confirming that no genocide had taken place in
   Darfur"
   concrete_case yes | speaker_position rejects | quotation attributed_or_reported |
   function ["accusation_or_qualification"] | referent: Darfur | referent_source passage |
   accused_actor "the Government of the Sudan" | victim_group "" | own_state_accused yes |
   salience substantive
   rationale: "The speaker reports a finding and adopts it — 'confirming' endorses the
   conclusion — so this refuses the characterization rather than reporting it without a
   position."

6. A tribunal's finding. UNSC_2001_SPV.4379Resumption1_spch0003#1, Jamaica,
   21 September 2001.
   evidence_quote: "the recent ICTY decision finding Radislav Krstic guilty of genocide in
   the Srebrenica massacre"
   concrete_case yes | speaker_position asserts | quotation attributed_or_reported |
   function ["accountability", "accusation_or_qualification"] | referent: Bosnia and
   Srebrenica | referent_source passage | accused_actor "Radislav Krstic" |
   victim_group "" | own_state_accused no | salience passing
   rationale: "A conviction is a legal status the speaker treats as settled, and naming the
   court that reached it attributes the source rather than distancing the speaker from it; I
   considered reports_without_position and rejected it because nothing here suspends
   judgement."

7. The Convention. UNSC_1992_SPV.3095_spch0029#4, Pan Africanist Congress of Azania,
   15 July 1992.
   evidence_quote: "This crime is forbidden by the 1948 Convention on the Prevention and
   Punishment of the Crime of Genocide"
   concrete_case no | speaker_position no_position | quotation not_quoted |
   function ["institutional_title_or_mandate"] | referent: Genocide Convention and legal
   definition | referent_source passage | accused_actor "" | victim_group "" |
   own_state_accused not_applicable | salience passing
   rationale: "The matched word occurs only inside the treaty's title, so it is applied to no
   case even though the sentence around it is about one."

8. A warning. UNSC_1996_SPV.3692_spch0018#1, Japan, 28 August 1996.
   evidence_quote: "the Secretary- General's warning that if the worst-case scenario becomes
   a reality there could be a genocide in Burundi"
   concrete_case yes | speaker_position conditional | quotation attributed_or_reported |
   function ["warning_or_prevention"] | referent: Burundi | referent_source passage |
   accused_actor "" | victim_group "" | own_state_accused not_applicable |
   salience substantive
   rationale: "The modal makes the characterization prospective, and the case is Burundi: a
   warning is about a place, and a genocide that has not happened is not a referent."
   Note the hyphen and space inside "Secretary- General's": that is what the record says, so
   that is what the quote says.

9. Distancing. UNSC_2023_SPV.9524_spch0007#1, United Kingdom, 30 December 2023.
   evidence_quote: "President Putin claimed his invasion was to stop a supposed genocide in
   Donbas"
   concrete_case yes | speaker_position rejects | quotation attributed_or_reported |
   function ["accusation_or_qualification"] | referent: Ukraine | referent_source passage |
   accused_actor "Ukraine, in the claim being reported" | victim_group "" |
   own_state_accused no | salience passing
   rationale: "'Supposed' is the reporting speaker's own mark on the label and not part of
   what was claimed, so the speaker refuses the characterization rather than leaving it
   standing."

10. An office's name. UNSC_2004_SPV.4990_spch0011#1, France, 14 June 2004.
    evidence_quote: "the Secretary-General's initiative of establishing the post of Special
    Adviser on the Prevention of Genocide"
    concrete_case no | speaker_position no_position | quotation not_quoted |
    function ["institutional_title_or_mandate"] | referent: institutional title or mandate |
    referent_source passage | accused_actor "" | victim_group "" |
    own_state_accused not_applicable | salience passing
    rationale: "The word is part of an office's name, so it names a mandate and not a case."

OUTPUT

Return JSON only, matching the schema supplied with the request. No prose, no commentary, no
explanation outside the fields. Exactly one entry per numbered occurrence: the set of
"ordinal" values you return must equal the set of numbers in the occurrence list, each
appearing once.
```

## User template

```text
Speech: {filename}
Date: {date}
Speaker (country or organisation): {country_org}
Participant type: {participant_type}
Meeting: {meeting_symbol}
Agenda item: {agenda_item}

===== SPEECH TEXT BEGINS =====
{speech}
===== SPEECH TEXT ENDS =====

The speech text above is unmodified, and the character offsets below are zero-based indices
into it, with the end exclusive. Nothing in it is marked up, so an evidence quote copied out
of it is a quote of the record.

The header lines above are context, not evidence. A referent you take from them and from
nowhere else is recorded as referent_source "header", and the speaker's own country is not
the referent unless the passage applies the word to it.

Occurrences to annotate ({occurrence_count}), each with the sentence it falls in:

{occurrences}
```

Show the controlled referent list (version 2, 40 identifiers)

Identifier	Definition	Occurrences
Rwanda `rwanda`	Speeches invoking the mass killing of Tutsi in Rwanda and its aftermath as genocide or genocidal.	1,978
Genocide in general `genocide_in_general`	Abstract discussion of genocide, including prevention doctrine, with no single case in view.	1,743
Gaza `gaza`	Speeches invoking the Gaza war as genocide or genocidal, including the proceedings South Africa brought against Israel at the ICJ.	835
Bosnia and Srebrenica `bosnia_srebrenica`	Speeches invoking the war in Bosnia and Herzegovina, including the fall of Srebrenica and its commemorations, as genocide or genocidal.	769
Genocide Convention and legal definition `genocide_convention_law`	The 1948 Convention, the legal definition of the crime, and jurisdiction over it before the ICJ or the ICC.	553
Institutional title or mandate `institutional_mandate`	The name or mandate of an office or body, such as the Special Adviser on the Prevention of Genocide or a tribunal's title.	483
Other known referent `other`	A referent that is identifiable but does not yet have a controlled identifier.	443
Israel and Palestine `israel_palestine`	Speeches invoking Israeli conduct towards Palestinians, or the conduct of Palestinian and allied armed groups towards Israelis, as genocide or genocidal, up to the war that gaza covers.	341
Darfur `darfur`	Speeches invoking the conflict in Darfur as genocide or genocidal.	220
Kosovo `kosovo`	Speeches invoking the Kosovo conflict and the expulsion of its Albanian population as genocide or genocidal.	128
ISIL and the Yazidis `isil_yazidi`	Speeches invoking ISIL's campaign against the Yazidi population of northern Iraq as genocide or genocidal.	126
Democratic Republic of the Congo `drc`	Speeches invoking mass killing in the Democratic Republic of the Congo as genocide or genocidal, including the presence in its east of armed groups formed from the perpetrators of the Rwandan killings.	105
Ukraine `ukraine`	Speeches invoking genocide in connection with the war in Ukraine, in either direction: against Russia's conduct and in Russia's own claims about Donbas. Not the Soviet famine, which holodomor covers, and not the Crimean Tatars, who have their own identifier.	102
South Africa under apartheid `apartheid_south_africa`	Speeches invoking apartheid and the violence surrounding its end as genocide or genocidal.	84
ISIL and other populations of Iraq and Syria `isil_iraq_syria`	Speeches invoking ISIL's crimes against populations of Iraq and Syria other than the Yazidis as genocide or genocidal, such as Christians, Shia Muslims and Turkmen.	77
Burundi `burundi`	Speeches invoking the mass violence in Burundi as genocide or genocidal.	70
Colonial-era mass violence `colonial_era`	Speeches invoking mass violence under colonial rule as genocide or genocidal, such as the campaign against the Herero and Nama.	63
The Holocaust `holocaust`	Speeches invoking the Nazi extermination of European Jews and its commemoration as genocide or genocidal.	60
Croatia and the Yugoslav wars `croatia_yugoslav_wars`	Speeches invoking the Yugoslav wars outside Bosnia and Kosovo as genocide or genocidal, such as the fighting in Croatia and the Krajina, or the wars taken as a whole.	53
Muslims in India `india_muslims`	Speeches invoking the treatment of the Muslim population of India as genocide or genocidal.	50
South Sudan `south_sudan`	Speeches invoking the civil war in South Sudan as genocide or genocidal.	33
Cambodia under the Khmer Rouge `cambodia_khmer_rouge`	Speeches invoking the mass killing under the Khmer Rouge, including the work of the Extraordinary Chambers, as genocide or genocidal.	31
Myanmar and the Rohingya `myanmar_rohingya`	Speeches invoking the treatment of the Rohingya population of Myanmar as genocide or genocidal.	31
Bangladesh and the war of separation `bangladesh`	Speeches invoking the mass killing during the war that separated East Pakistan from Pakistan as genocide or genocidal.	30
The Ottoman Armenians `armenian_genocide`	Speeches invoking the deportation and mass killing of Ottoman Armenians as genocide or genocidal.	26
Nagorno-Karabakh `nagorno_karabakh`	Speeches invoking the Nagorno-Karabakh conflict and the departure of its Armenian population as genocide or genocidal.	23
Syria `syria`	Speeches invoking the conduct of the parties to the war in Syria as genocide or genocidal.	18
Central African Republic `central_african_republic`	Speeches invoking the sectarian violence in the Central African Republic as genocide or genocidal.	17
Other sanctions regimes `economic_sanctions_other`	Speeches describing a sanctions regime other than the Cuban embargo as genocidal, such as the measures against Iraq in the 1990s.	17
Afghanistan and the Hazaras `afghanistan_hazara`	Speeches invoking mass violence in Afghanistan as genocide or genocidal, in particular attacks on its Hazara population.	13
Abkhazia and South Ossetia `abkhazia_south_ossetia`	Speeches invoking the conflicts in Abkhazia and South Ossetia and the displacement of their populations as genocide or genocidal.	11
East Timor `east_timor`	Speeches invoking the violence around East Timor's independence ballot as genocide or genocidal.	8
Crimean Tatars `crimean_tatars`	Speeches invoking the wartime deportation of the Crimean Tatars, or their treatment under the occupation of Crimea, as genocide or genocidal.	6
Iraqi Kurds and Anfal `iraq_kurds_anfal`	Speeches invoking the Anfal campaign and the Halabja attack against Iraq's Kurdish population as genocide or genocidal.	5
Khojaly `khojaly`	Speeches invoking the killings at Khojaly as genocide or genocidal, which Azerbaijan raises as an event of its own.	5
The Soviet famine in Ukraine `holodomor`	Speeches invoking the Soviet famine in Ukraine as genocide or genocidal, which speakers raise as a case of its own rather than as part of the war that ukraine covers.	4
United States embargo against Cuba `cuba_us_sanctions`	Speeches describing the United States embargo against Cuba as genocidal.	3
Xinjiang and the Uyghurs `xinjiang_uyghurs`	Speeches invoking China's treatment of the Uyghur population of Xinjiang as genocide or genocidal.	1
Not applicable `not_applicable`	Reserved for false-positive candidate matches.	0
Unclear referent `unclear`	The available context does not support a reliable referent assignment.	0

Comparison with human coding

No human-reference comparisons are available yet. The sample contains 469 entries covering 458 distinct occurrences. Two readers are to code each independently. Accuracy scores require agreed or adjudicated reference labels; the absence of a score means accuracy has not been measured.