Budget

	A	B	C	D
1	RATES
2	Item	Value	Note
3	Tinker train: Qwen3.5-9B	$1.463	$ per 1M tokens
4	Tinker train: Qwen3.5-4B	$0.737	$ per 1M tokens
5	Tinker train: Qwen3.6-27B	$4.103	$ per 1M tokens
6	Tinker train: Qwen3.6-35B-A3B (MoE)	$1.177	$ per 1M tokens
7	RunPod: A100 80GB	$1.390	$ per GPU-hour (on-demand)
8	RunPod: H100 80GB	$2.990	$ per GPU-hour (on-demand)
9	Avg training tokens per run	0.8	millions of tokens (approx.)
10	Eval overhead (× training cost)	1.15	logprob + sampled evals (approx.)
11
12	Runs: core dissociation matrix (9B)	450	for 6 envs (approx.)
13	Runs: feature validation (9B)	200	predicted vs random vs control (approx.)
14	Runs: seed-sweep cross-check (4B)	250	approx.
15	Runs: scale, 27B	50	approx.
16	Runs: scale, MoE	50	approx.
17	Runs: base model (9B-Base)	50	approx.
18	GPU-hours (A100)	300	base-model forward passes + dev (approx.)
19	SAE training + large states (H100)	200	approx.
20	Stipend: hours/week	15	part-time
21	Stipend: weeks	17	~4 months
22	Stipend: $/hour, minimum	15	local minimum wage
23	Stipend: $/hour, ideal	30	modest research wage
24
25	BUDGET
26	Item	Minimum (USD)	Ideal (USD)	Note
27	Compute: fine-tuning
28	Core dissociation matrix (9B)	$606	$606	450 runs × 0.8M tok × $1.463 × 1.15
29	Feature-specific validation (9B)	$269	$269	200 runs × 9B rate
30	Seed-sweep cross-check (4B)	$170	$170	250 runs × 4B rate
31	Scale generalization: 27B	$189	$189	50 runs × 27B rate
32	Scale generalization: MoE	$54	$54	50 runs × MoE rate
33	Base-model (9B-Base)	$67	$67	50 runs × 9B rate
34	Subtotal: fine-tuning	$1,355	$1,355
35	Compute: scanning (RunPod, hidden-states)
36	Scan + baselines + development (A100)	$417	$417	300 GPU-hr × $1.39
37	SAE training + large-models (H100)	$598	$598	200 GPU-hr × $2.99
38	Subtotal: scanning	$1,015	$1,015
39	Frontier-model API
40	LLM-judge evaluation	$1,000	$1,000	EM / toxicity scoring
41	Dataset synthesis & filtering	$600	$600	trait-pair & EM datasets
42	SAE feature auto-interpretation	$400	$400	label Qwen SAE features
43	Subtotal: API	$2,000	$2,000
44	Personnel
45	Researcher stipend	$3,825	$7,650	15 hr/wk × 17 wk × ($15 min / $30 ideal)
46	Subtotal: personnel	$3,825	$7,650
47	Tooling
48	Experiment tracking + storage	$150	$150	W&B, artifacts
49	Dev-assistance API credits	$0	$1,000	ideal: coding/agent assistance
50	Subtotal: tooling	$150	$1,150
51
52	Contingency (10% of compute + API)	$440	$440	reruns, seed expansion (rounded)
53	GRAND TOTAL	$8,785	$13,610
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