{
 "schema": "cain.e128b.preregistration.v1",
 "frozen_at": "2026-11-21T05:12:11Z",
 "title": "E128-B de novo depth on the same targets (no benchmark feedback)",
 "hypothesis": "Feeding the official failing test output and the previous patch back to the same model produces more resolves per dollar than de novo generation on the same tasks.",
 "selection": "35 unresolved instances with an available official test_output.txt and a previous patch; repository-priority order; no labels.",
 "arm": {
  "candidates": 2,
  "steps": 24,
  "workers": 4,
  "model": "claude-sonnet-5-5",
  "hard_cap_usd": 16.0
 },
 "primary_endpoint": "official resolves gained",
 "comparison": "measured against de novo breadth (~$0.15/task) on the same tail",
 "budget_usd": 20.0,
 "note": "test output is public benchmark feedback, never the hidden test patch or FAIL_TO_PASS list",
 "agi_status": "NOT ACHIEVED",
 "status": "REJECTED_FOR_BENCHMARK_CONTAMINATION",
 "decision": "The official evaluation test_output.txt contains the hidden FAIL_TO_PASS test names. Feeding it back to the model would leak the benchmark and invalidate any rank. Arm aborted before producing a usable patch; no result is reported.",
 "replacement": "E128-B: de novo depth-2 on the same targets, no test feedback (legitimate).",
 "contamination": "none; the model sees only the problem statement, the repository and retrieval hints"
}
