Requirement-Bound Verified Commissioning: A Frozen Four-Billion-Parameter Local Model as a Candidate Generator under an External Acceptance Layer with Verification and Release Authority

What it claims

This preprint (arXiv:2609.30219v1, single author Mehmet İşcan, Yıldız Technical University) applies the "candidate generator + external acceptance layer" pattern to a specific, narrow task the summary does not name: sensor coordinate and polarity binding during mechatronic commissioning — deciding, from natural-language requirement text, which physical quantity and sign convention a measurement channel gets before a system is authorized for service. A reversed polarity can turn intended negative feedback into positive feedback and induce instability, which motivates verifying the binding rather than trusting the generator. The frozen 4B local model only proposes; a deterministic entailment gate under a sealed grammar ("grammar V1") holds release authority and uses no model-derived confidence score.

How it works

A requirement first goes to a deterministic parser. Only requirements the parser cannot support are routed to the frozen model. Any candidate is checked by the gate, which releases a plan only when both facts (coordinate and polarity) can be derived from the text under grammar V1. If a fact cannot be bound, it is requested once from the user, appended as a single canonical sentence ("template T1"), and the gate is re-applied (a two-turn protocol). The design deliberately separates three things: derivability-from-text vs. correctness-against-answer-key; gate decision vs. scoring; and computed entailment vs. assumed truth — the user-supplied fact is trusted, not validated.

flowchart TD
  R[Requirement text] --> P{Deterministic parser supports it?}
  P -->|yes| G[Entailment gate under grammar V1]
  P -->|no| M[Frozen 4B model generates candidate]
  M --> G
  G -->|both facts derivable| REL[Release plan]
  G -->|fact missing, eligible| Q[Ask user once, append via template T1]
  Q --> G
  G -->|not derivable| AB[Abstain]

Evidence

Evaluated once under a criterion sealed before benchmark construction, on 144 tasks written by isolated agent contexts with no access to the gate, grammar, or plan.

• Failure to abstain, then caught: fabricated "ready" plans were committed on 21 of 22 routed unanswerable tasks — and all were rejected by the gate. The same 83 releases were reproduced by a model-free replay (model calls swapped for a coordinate/polarity-free plan), so the model's contribution to delivered plans was limited.
• False release: 0 among the 83 benchmark releases; one-sided 95% Clopper–Pearson upper bound of 0.0354, below the sealed 5% threshold. But one false release was later recorded among 146 releases outside the benchmark at seed 0.
• User-answer trust: both facts bound from original text on only 13 of 96 answerable tasks; incorrect answers were released in 169 of 431 pairings on the rest, including failures where a coordinate had been explicitly excluded.

Limits

Text-level only — hardware, closed-loop behavior, and control performance were excluded. The gold-standard user always supplied correct answers in the sealed run; a deployable questioning policy was never tested because question eligibility was read from the sealed answer key. Gate sensitivity (rate at which a truly supported gold pair is classified ready) and real user behavior were not measured. Results hold under one grammar and one benchmark, and the authors present them as conditional measurements, not deployment rates — the out-of-benchmark false release shows the sealed bound does not carry over.