Cluster — 40 nodes × 8 GPUs
fill = GPUs in use · ring = thermal state
Shared cooling + network domain — high-centrality tier
000
010
020
030
040
050
060
070
080
090
100
110
120
130
140
150
160
170
180
190
200
210
220
230
240
250
260
270
280
290
300
310
320
330
340
350
360
370
380
390
Events
— run starting
Governance λ—
Violated jobs Σ0 vs 0 ghost
Migrations/epoch0
00:0006:0012:0018:0024:00
dashed = static λ=0.65 ghost (tuned elsewhere) · red tick = shared-domain failure