{
  "schema": "cain.e91.eval.acceptance.v1",
  "slice": "session-1b-evaluation-harness",
  "base_commit": "2d444e70746447f3b8c70ff089e9ce0eb5f22109",
  "overall_status": "PARTIAL",
  "counts": {"PASS": 3, "PARTIAL": 3, "FAIL": 2, "BLOCKED": 2},
  "gates": [
    {"id": "E91-E01", "name": "Six verticals imported read-only", "status": "PASS", "evidence": "vendored snapshot under benchmarks/e91/eval/_vertical_snapshot; 6 vertical modules import and generate tasks"},
    {"id": "E91-E02", "name": "120 held-out tasks", "status": "PASS", "evidence": "HELD_OUT_TASK_MANIFEST.json count=120, 20 per vertical, split=heldout"},
    {"id": "E91-E03", "name": "Ten arms plus G0", "status": "PASS", "evidence": "1,440 held-out decisions + 240 transfer decisions in artifacts/e91/eval/rows.jsonl"},
    {"id": "E91-E04", "name": "Learning improves held-out performance", "status": "PARTIAL", "evidence": "G0 26/120 to G 120/120, +78.33 pp, p<1e-8; E and F also 120/120, so the gain is strategy selection from a hand-written expert, not model learning"},
    {"id": "E91-E05", "name": "Cross-domain transfer measured", "status": "PARTIAL", "evidence": "G_source 40/40 vs A 4/40, +90 pp; F also 40/40, so the number is not attributable to the signature abstraction"},
    {"id": "E91-E06", "name": "Collaboration benefit", "status": "FAIL", "evidence": "I 120/120 = F; 2 false accepts; 233 vs 155 tool calls. No measurable benefit on this suite"},
    {"id": "E91-E07", "name": "Real professional environments", "status": "FAIL", "evidence": "generated deterministic tasks; not real codebases, security targets, labs or customers"},
    {"id": "E91-E08", "name": "Independent validation", "status": "BLOCKED", "evidence": "same-project only; no external auditor or benchmark"},
    {"id": "E91-E09", "name": "ClawX public evidence", "status": "BLOCKED", "evidence": "not in this slice; owner approval and the other E91 sessions' bundle are required"}
  ],
  "limits": [
    "hand-written deterministic expert policies, not an LLM",
    "the suite is saturated by the expert; E=F=G=H=I=J=120/120",
    "no long-horizon, kill-switch or production test in this slice",
    "no model calls, no network, no paid service"
  ]
}
