[
 {
  "task": "dapt-intrusion-detection",
  "domain": "cybersecurity",
  "difficulty": "hard",
  "failed_check": "leak",
  "reason": "threat-detection/SKILL.md worked example uses this instance's exact expected metrics (packets_per_minute_max=2372, packets_per_minute_avg=262.9, both asserted in tests/test_outputs.py) and derives the has_dos_pattern=false answer",
  "replacement": "fix-druid-loophole-cve",
  "replacement_difficulty": "hard",
  "same_stratum": true
 },
 {
  "task": "earthquake-phase-association",
  "domain": "natural-science",
  "difficulty": "hard",
  "failed_check": "oracle",
  "reason": "environment image fails to build: unpinned setuptools 84 removed pkg_resources, seisbench import fails in Dockerfile step 6 (PhaseNet.from_pretrained); oracle cannot run",
  "replacement": "seismic-phase-picking",
  "replacement_difficulty": "hard",
  "same_stratum": true
 },
 {
  "task": "lab-unit-harmonization",
  "domain": "natural-science",
  "difficulty": "medium",
  "failed_check": "leak",
  "reason": "reference/ckd_lab_features.md reproduces all 62/62 verifier acceptance ranges (REFERENCE min/max in tests/test_outputs.py) exactly, while the prompt asks the agent to decide the physiological ranges itself",
  "replacement": "quantum-numerical-simulation",
  "replacement_difficulty": "medium",
  "same_stratum": true
 },
 {
  "task": "video-silence-remover",
  "domain": "media-content-production",
  "difficulty": "hard",
  "failed_check": "leak",
  "reason": "report-generator/pause-detector skills give a worked example for a 600 s video with opening 0-221 s, 249 s removed, 351 s compressed; ground truth is 0-236 s, 254 s, 346 s; copying the example passes the 20%-tolerance duration tests and the IoU opening match",
  "replacement": "mario-coin-counting",
  "replacement_difficulty": "medium",
  "same_stratum": false
 },
 {
  "task": "jax-computing-basics",
  "domain": "software-engineering",
  "difficulty": "medium",
  "failed_check": "leak",
  "reason": "jax_skills.py is the oracle: one function per sub-problem (reduce, map, logistic_grad, rnn_scan, jit) with the exact conventions the verifier expects (+-1 logistic labels, zero-init tanh RNN) that the prompt leaves unspecified; oracle comment says it inlines the skill's verbatim semantics",
  "replacement": "python-scala-translation",
  "replacement_difficulty": "medium",
  "same_stratum": true
 },
 {
  "task": "seismic-phase-picking",
  "domain": "natural-science",
  "difficulty": "hard",
  "failed_check": "oracle",
  "reason": "environment image fails to build: seisbench 0.10.2 imports pkg_resources, removed by unpinned setuptools 84 (same break as earthquake-phase-association); oracle cannot run",
  "replacement": "mars-clouds-clustering",
  "replacement_difficulty": "hard",
  "same_stratum": true
 }
]