Ranking feasibility boundary under Gaussian weight perturbation

Rankings from pythia-70m-deduped step142000 (100-sample Pile set, first rankings), checked against the unembedding W + εG, Gij ∼ 𝒩(0, 1) (one fixed draw); x-axis is the RMSE of the perturbed weights (≈ ε). Boundary = first infeasible top-k (top-k order + top>bottom constraints, MOSEK LP); k = v = 50,304 means the full ranking is feasible. Weight std is 0.136.

2026-11-05T09:01:57.126296 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 1 0 − 1 0 1 0 − 8 1 0 − 6 1 0 − 4 1 0 − 2 R M S E   o f   p e r t u r b e d   w e i g h t s   + W ε G 1 0 2 1 0 3 1 0 4 1 0 5 F i r s t   i n f e a s i b l e   t o p - k v = 5 0 , 3 0 4 d = 5 1 2
εRMSEnmedianrangefrac full
1e-121e-12550,30416,384–50,3040.60
1e-111e-11550,30450,230–50,3040.80
3e-113e-11550,30428,672–50,3040.80
1e-101e-10550,30450,304–50,3041.00
3e-103e-10550,30450,239–50,3040.80
1e-091e-09550,30450,026–50,3040.80
3e-093e-09540,91915,536–50,3040.20
4e-094e-09532,51216,384–38,5180.00
5e-095e-09424,43212,286–32,9420.00
7e-097e-09424,02723,392–25,2520.00
1e-081e-08518,89617,193–19,9690.00
3e-083e-0859,3288,104–10,4420.00
1e-071e-0755,2484,784–5,6070.00
3e-073e-0753,0992,844–3,2970.00
1e-061e-0651,7651,705–1,9820.00
3e-063e-0651,1891,173–1,2190.00
1e-051e-055835821–8680.00
3e-053e-055669636–6750.00
0.00010.00015549543–5740.00
0.00030.00035472467–4840.00
0.0010.0015398384–4040.00
0.0030.0034366342–3730.00
0.010.015337279–3420.00
0.030.035232198–2550.00
0.10.14116111–1230.00

Data (CSV): summary, per ranking. Sweep still in progress. Updated 2026-11-05 09:01.