absence bench budget

	A	B	C	D
1	AbsenceBench — Budget
2	Volunteer team, incl. in-house scenario construction & validation (Arabic/English/Japanese). No stipend line. Spend is compute + hosting.
3	Scope: 20 models x ~1,000 base scenarios x 3 languages x 5 repeats = ~300K runs/pass; ~3.5 effective passes (build iteration, ablations, model-update reruns) = ~1.05M runs. Compute derived from that; blue/yellow cells = your inputs.
4
5	Line item	Minimum	Ideal	What it covers / how it's derived
6	Compute — closed models (API)	$2,500	$5,000	~10 frontier models via API over the runs. Blended ~$2.5/M in, ~$13/M out with batch (50%) + caching. Minimum = 1 clean pass + light iteration; Ideal = full ~3.5 passes + ablations. Derived, not estimated.
7	Compute — open-weight (GPU)	$1,200	$3,500	~10 open-weight models on rented cloud GPUs ($0.50-$5/hr by model size). ~440 GPU-hrs at full scope. Minimum = core models 1 pass; Ideal = full model set, all passes.
8	Scenario construction (materials/tools)	$500	$2,000	NON-labor costs only (labor is volunteer): annotation tooling, translation-check references, storage. Ideal funds broader tooling for the multilingual set.
9	Eval red-teaming / hardening	$0	$2,500	Stress-testing the benchmark so it can't be gamed. Some via paid API calls to adversarial-probe the eval. Ideal-only.
10	Dissemination / hosting	$800	$2,000	Public leaderboard hosting, open harness + dataset release, write-up. Recurring hosting over the project.
11	Buffer / contingency	$1,000	$2,000	Price moves (API/GPU rates shift), extra reruns when models update mid-project. Kept explicit rather than hidden in other lines.
12	TOTAL	$6,000	$17,000	Minimum ships the benchmark across a core model set + all 3 languages. Ideal = full 20-model coverage, all passes, red-teaming.
13
14	Note on size:			This project is compute-light and labor-heavy; because the labor (scenario authoring/validation across 3 languages) is volunteer, it is not a budget cost. The ask reflects real inference/GPU/hosting spend, not padded compute. Under-asking with correct arithmetic is intentional.
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100