Extractor score against the sample A hand read, dubai-01 (12 runs)
Rule: analysis/scoring/RULE.md. Per-name rows: results.csv.

Variant T: brand_entity, prose_list, table_cell  (like-for-like with the hand read)
  recall, registered brokerages              70 / 77    90.9%   95% CI 82.4% to 95.5%
  recall, all named organisations            79 / 133   59.4%   95% CI 50.9% to 67.4%
  precision, including not_a_firm            79 / 92    85.9%   95% CI 77.3% to 91.6%
  precision, excluding not_a_firm            79 / 81    97.5%   95% CI 91.4% to 99.3%

Variant A: brand_entity, prose_list, table_cell, local_business, source
  recall, registered brokerages              70 / 77    90.9%   95% CI 82.4% to 95.5%
  recall, all named organisations            79 / 133   59.4%   95% CI 50.9% to 67.4%
  precision, including not_a_firm            79 / 231   34.2%   95% CI 28.4% to 40.5%
  precision, excluding not_a_firm            79 / 219   36.1%   95% CI 30.0% to 42.6%

Sensitivity (not a headline figure): ground-truth names resolved with channels {prose_list}
  registered-brokerage denominator: 77 with no channel, 79 with prose_list
  recall, registered brokerages, T           72 / 79    91.1%   95% CI 82.8% to 95.6%
  recall, registered brokerages, A           72 / 79    91.1%   95% CI 82.8% to 95.6%
