[patnr]

Adds to the MiniRes benchmark:

• A row for ES: a single update, without localization.
• The CRPS (continuous ranked probability score) of the ensemble, for the forecast water cut, the log-permeability field and sor. Unlike the RMSE of the ensemble mean, it also scores the spread: an ensemble that is confidently wrong scores worse than one that is wrong but knows it.

RESULTS.md is not rerun here, so it has neither yet. A score missing from the official results shows as blank.

Tentative numbers, from run.py --quick (coarse case, 5 truths; means):

	misfit	CRPS forecast WWCT	logK RMSE	logK CRPS	sor CRPS	simulations
prior	97.4	0.103	0.772	0.437	0.0178
ES (one update)	14.4	0.060	0.856	0.494	0.0203	100
ES-MDA (4 steps, α = 4)	1.56	0.039	0.795	0.464	0.0269	250
adaptive localization (universal threshold)	16.1	0.039	0.748	0.421	0.0124	250
adaptive localization (3σ threshold)	11.8	0.031	0.741	0.416	0.0117	250
distance localization (Gaspari-Cohn, 300 m)	8.67	0.028	0.730	0.407	0.0165	250

The ES row's 100 simulations include the simulation of the posterior.

Checked locally: the benchmark's smoke test, and ruff.

🤖 Generated with Claude Code