[patnr]

Three changes to the MiniRes benchmark:

• A row for distance localization: ES-MDA with a Gaspari-Cohn taper of 300 m (the prior's range) on the field, centred on each datum's well; sor is left unlocalized. The radius is given in metres, and run.py places the tapers on each case's grid.
• The case: the prior of the log-permeability field is now smooth (PET's "exp" variogram, exp(-3 (h/r)^1.9), all but Gaussian) instead of spherical, and the oil viscosity is known. With an uncertain viscosity ratio, much of what the wells' data say went to vo and not to the field. sor remains as the scalar unknown.
• run.py --quick: the coarse case on its first 5 truths, in about a minute instead of twelve, printed next to the official scores on the same truths. The experiments are seeded, so any difference is due to the code.

All rows are rerun. Means over the truths:

Coarse (25 x 25, 20 truths)

	misfit	RMSE forecast WWCT	logK RMSE	logK corr	sor error
prior	99.2	0.242	0.780	0.030	0.031
ES-MDA (4 steps, α = 4)	1.71	0.092	0.763	0.299	0.036
adaptive localization (universal threshold)	20.6	0.153	0.758	0.210	0.026
adaptive localization (3σ threshold)	13.5	0.124	0.750	0.262	0.029
distance localization (Gaspari-Cohn, 300 m)	8.96	0.101	0.718	0.337	0.030

Fine (50 x 50, 8 truths)

	misfit	RMSE forecast WWCT	logK RMSE	logK corr	sor error
prior	99.0	0.247	0.817	-0.072	0.028
ES-MDA (4 steps, α = 4)	1.83	0.098	0.793	0.355	0.031
adaptive localization (universal threshold)	33.3	0.173	0.793	0.153	0.031
adaptive localization (3σ threshold)	16.4	0.158	0.761	0.290	0.034
distance localization (Gaspari-Cohn, 300 m)	12.4	0.112	0.732	0.402	0.034

Checked locally: the benchmark's smoke test, the distance-localization tests, and ruff.

🤖 Generated with Claude Code