Transparent scoring, version by version.

Observable outcomes on the committed set

1. Give each of the ten domains weight 0.1.
2. Keep the frozen domain and difficulty quotas.
3. Keep missing and invalid tasks in completion accounting and block Official publication.
4. Classify complete synthetic fixtures as descriptive Synthetic Complete, never Official or ranking eligible.
5. Treat attributable agent, model, tool, timeout, budget, and wrong-artifact failures as valid zero scores.
6. Treat benchmark infrastructure failures as invalid and audit a rerun.

72 tasks · 10 equally weighted domains

Official, complete synthetic, provisional, or coverage-only

Attempted failure

Attributable failures are valid zero scores. Infrastructure failures are invalid and require an audited rerun.

Missing fixture or result

Missing and invalid tasks block Official. Synthetic Complete and Provisional output use descriptive observed domain means and fixed-fixture completion bounds without ranking eligibility.

Hidden fixture boundary

Hidden payloads stay sealed behind the published fixture-set commitment. Version, task counts, outcome states, and provenance remain public.

What the interval can and cannot say

The task-resampling interval uses finite_cluster_calibrated_percentile_sensitivity_v1 with a versioned 1.3 deviation correction calibrated for this fixed benchmark fixture. It is a fixed-fixture calibrated sensitivity interval, not a universal confidence interval for model capability.