Frozen 4B model proposes; external entailment gate releases
Sep 24 İşcan: candidate generation ≠ release authority — 21/22 fabricated ready plans on unanswerable tasks all rejected; 0 false releases / 83 gated (U≤3.54%); ungated shell 9/18 false.
Mehmet İşcan’s Sep 24 paper is a preregistered measurement of the Simplex-shaped split agentic SDLC keeps reaching for: candidate generation separated from release authority. On a sealed 144-task sensor coordinate/polarity commissioning benchmark (96 answerable / 48 unanswerable, written by isolated agent contexts blind to the gate), a frozen local 4B Qwen3 is only a residual candidate generator when a deterministic parser cannot bind the requirement. Plans release solely when an external entailment gate derives both facts under a sealed grammar — no model confidence score in the decision.
The receipts separate the two roles. On parser-routed unanswerable tasks the model committed fabricated ready plans on 21 of 22; the acceptance layer rejected all of them. Among 83 gated releases at seed 0 there were 0 false releases (one-sided 95% Clopper–Pearson upper bound 0.0354, below a sealed 5% headline — reported as diagnostic under i.i.d.). A model-free replay reproduced the same 83 releases with zero model calls. Comparators bite: separate ungated shell 9 of 18 false; in-run ungated terminal 6 of 13 false. Trust-boundary caveat owned: with a gold user the endpoint is clean, but wrong-answer sweeps still release when the requirement leaves a fact open (169/431 pairings), and a post-seal campaign outside the benchmark recorded 1 false release in 146. Gate sensitivity and real users were not measured.
Why it matters for code review
Continuous factories that let the same agent declare ready and merge are measuring self-verification — the arm İşcan shows fails open on unanswerable work. Review overnight inventory as two artifacts: the candidate, and an independent acceptance decision with a fail-closed third state when evidence is missing. FOMO for idle agents is the wrong fear if release authority still sits inside the generator; steady verified momentum needs an external gate that can reject a confident ready plan.