name: NullspanBench
description: >
  System One structured extraction, multi-numeric commercial contract disambiguation,
  calibrated abstention (p_null), cross-field invariant verification, and distributional
  quantile estimation benchmark (N = 500 field-level decisions across Open Structured,
  Sealed Multi-Numeric Contracts, Sealed Stanford SQuAD v2.0, and Adversarial Stress pools).

tasks:
  - id: capability
    config: default
    split: validation
  - id: intelligence
    config: default
    split: validation
  - id: calibration
    config: default
    split: validation
  - id: sealed_contracts
    config: default
    split: validation
  - id: ece
    config: default
    split: validation
    lower_is_better: true