Add Launa CSM-only evaluation results

[zomux]

Summary

Adds Launa to the clean English CSM results as a 2-part system. ITSM, HR, pooled scores, and perturbation results remain unavailable.

This is a draft submission for review of partial-domain coverage, incomplete-trial handling, and evaluation-version compatibility.

The only changed file is website/src/data/leaderboardStats.json. Evaluation result is documented below; supporting evidence is retained locally and can be provided on request.

Project website: Launa

Results

Metric	EVA-A	EVA-X	Scenarios
pass@1	48.70%	55.03%	50
Mean	64.19%	75.36%	50
pass@5	80.43%	95.65%	46
Estimated pass^5	29.09%	18.14%	46

There are 250 recorded trials across 50 scenarios. 244 trials passed simulation validation; six invalid trials affect four scenarios. Pass@1 and means first average valid trials within each scenario, then weight scenarios equally. Five-trial metrics use the 46 complete scenarios. Valid model failures are retained.

The original saved report averages valid trials directly (49.59% EVA-A / 55.33% EVA-X pass@1). This PR uses scenario weighting so point estimates and confidence intervals have the same sampling unit. All intervals are recomputed using EVA's bootstrap implementation: 2,000 scenario resamples, 95% percentile intervals, deterministic run-derived seed. Estimated pass^5 follows the website's per-scenario success-probability definition.

Maintainer guidance requested

Is a CSM-only entry acceptable with other domains and pooled results unavailable?

Thank you for maintaining EVA!

[fanny-riols]

Hi @zomux , thank you for your interest! We would actually prefer having submissions on the 3 data domains. Also, after some analysis, we concluded that we could use 3 trials only, so no need to run 5 trials. Finally, once you have all results (on the 3 domains), could you send us your output artifacts (all the output files created) so that we could do some sanity checks on your configs, etc., please?