Issues

• BE-004 cancel order: a cross-feature task with a validate-before-mutate trap

  measurement
  UnityInFlow/agent-observatory-benchmarks#28 · Filter by author hermanngeorge15 opened on Sep 3, 2026

• benchmark.yaml names its own baseline commit, so tasks can compound without moving each other's baseline

  enhancementL2-enforced
  UnityInFlow/agent-observatory-benchmarks#27 · Filter by author hermanngeorge15 opened on Sep 3, 2026

• No task fails a plain agent: the discrimination has moved from the evaluator to the rubric

  measurement
  UnityInFlow/agent-observatory-benchmarks#26 · Filter by author hermanngeorge15 opened on Sep 3, 2026
  1

• README documents BE-001 and BE-002, and not the benchmark every B2 run was measured on

  documentation
  UnityInFlow/agent-observatory-benchmarks#25 · Filter by author hermanngeorge15 opened on Sep 3, 2026

• test-quality carries 25 of 100 and is decidable on two of five variants

  UnityInFlow/agent-observatory-benchmarks#22 · Filter by author hermanngeorge15 opened on Aug 28, 2026