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

1 Introduction

• An incorrect sensor coordinate supplies the wrong physical quantity, while reversed polarity can turn intended negative feedback positive and induce instability or oscillation depending on loop dynamics.Source: §1 · p. 2
• Prior work separates generation from acceptance, but leaves open reliable checking and the external decision mechanism’s false-acceptance frequency.Source: §1 · pp. 2–3
• The protocol routes parser-unsupported requirements to a frozen 4B local model, while a deterministic gate releases plans only when both facts follow from sealed grammar V1.When a fact cannot be textually bound, one canonical user answer is appended and the gate is applied again; model confidence does not authorize release.Source: §1 · p. 3
• Gold-equivalent candidates were generated on most routed answerable tasks, while nearly all fabricated ready plans on routed unanswerable tasks were rejected by the acceptance layer.The study measured generation and release decisions separately on the same parser-routed tasks.Source: §1 · p. 4
• The false-release endpoint used a criterion fixed before evidence collection and tasks withheld from method designers; a post-seal campaign observed false releases and was not pooled with the benchmark.The upper bound was reported as a diagnostic under the i.i.d. assumption.Source: §1 · p. 4
• Incorrect answers can be rejected only when the disputed fact is independently constrained by the requirement; deployable questioning, real user behavior, and gate sensitivity were not evaluated.Source: §1 · p. 4

2.1 Problem definition

• 144 tasks across 24 mechatronic compositions tested sensor-coordinate and polarity binding; modeling, estimator design, and controller design were excluded.The set comprised 96 answerable and 48 unanswerable tasks.Source: §2.1 · p. 6

2.2 Routing rule and entailment gate

• The parser routes unsupported requirements to the model, but every candidate passes through grammar V1, which releases only when the gate binds both facts without conflict.The gate uses evidence spans that must be nonempty contiguous substrings of the supplied text; it evaluates selected analysis units and binds a fact only when exactly one value is supported. Its operational grammar does not claim soundness for unrestricted natural-language inference.Source: §2.2 · pp. 7–8

2.3 Two-turn questioning protocol

• Turn 1 requires both a ready declaration and gate-derived readiness; eligible unreleased tasks receive one answer in template T1, after which the unchanged executor is rerun and the gate alone decides release.Eligibility depends on the gold answer key, excludes contradictory raw gate status, and prevents questions when no gold pair exists; the supplied fact is not itself validated.Source: §2.3 · pp. 8–9

2.4 Primary endpoint and statistical decision rule

• A release is false when its bound pair differs from the sealed gold pair or no gold pair is defined; the sealed rule uses two one-sided 95% Clopper–Pearson bounds at τ = 0.05 and τ = 0.10.Source: §2.4 · pp. 9–10

2.5 Call budgets, run integrity, and execution algorithms

• Each turn is a separate episode with separate ceilings; model calls and attempted non-submission tool operations are counted as ν(e) and χ(e).A tool charge applies whether an attempted operation is completed or rejected.Source: §2.5 · pp. 10–11
• Three kill rules check execution integrity: prohibit turn 2 after contradictory raw gate status, require one record per task–turn pair, and replay each plan through the sealed gate.Compliance was checked from the run records.Source: §2.5 · p. 11
• Parser-supported tasks receive a zero-call parser plan; otherwise the executor commits a model plan and applies the gate before any turn-1 release.The algorithm then permits a second turn only when eligible: it asks one question and evaluates a plan from the unchanged executor without a placement restriction.Source: §2.5 · p. 11

3.2 Primary endpoint and ungated comparisons

• 0 false releases among 83 plans yielded a one-sided 95% Clopper–Pearson upper bound of 0.0354, below the sealed 5% threshold under an i.i.d. diagnostic assumption.The result was assist-grade, requiring human review; none of the 48 tasks without ready gold status was released.Source: Table 4 · p. 15, §3.2 · pp. 15–16
• 9 of 18 ungated-shell deliveries were false, while the protocol correctly delivered plans for 83 of 96 answerable tasks versus 9 of 96 for the shell.The protocol used a three-call budget and the ungated-shell run a five-call budget; 74 answerable tasks were correctly delivered only by the protocol.Source: §3.2 · p. 16

3.3 Model contribution and model-free replay

• 21 of 22 routed unanswerable tasks received a fabricated ready plan from the model, and the acceptance layer rejected every such plan.Of 42 ready plans rejected at turn 1, 20 were equal to gold.Source: §3.3 · p. 17, Table 6 · p. 18
• The model-free replay released the same 83 tasks with 0 false releases and no model calls, versus 183 calls in the live run.It asked 80 questions rather than 76; the live run included 4 turn-1 deliveries without questions and 11 answerable tasks where only one fact was requested.Source: §3.3 · p. 17, Figure 5 · p. 20
• 68 releases and 0 false releases were observed in the development-labeled second-model arm, 15 fewer releases than the primary model’s 83.The second-model arm’s results were descriptive, not confirmatory.Source: §3.3 · p. 17

3.4 Protection against incorrect user facts

• 13 of 96 answerable tasks were text-bound for both facts, with none of 65 wrong-answer pairings released; among the remaining tasks, 169 of 431 incorrect pairings were released.For text-open tasks, the user’s answer was the sole source of an open fact, and the protocol trusted and released supplied answers whether correct or incorrect.Source: §3.4 · pp. 21–22, Figure 6 · p. 22
• 4 of 226 protocol-level pairings released a coordinate explicitly excluded by the requirement, while 2 of 129 protocol-level pairings released in a separate bound-cell classification.A distinct selected-stratum measurement released no tasks and failed its preregistered pass rule; the second selected set yielded only one observation.Source: §3.4 · p. 22

3.5 Preregistered predictions

• 1 false release among 146 plans in the subsequent confirmation campaign triggered its preregistered kill rule and paused the campaign.Four preregistered predictions were not met; the confirmation failure involved a second-clause anaphor whose polarity cue was not extracted by grammar V1.Source: §3.5 · p. 23
• 252 outside-benchmark tasks were included in a post-seal campaign whose results did not alter the primary tier verdict.The campaign was excluded from the sealed final look; its counts must not be pooled with Table 4 or its bounds compared across populations.Source: Table 8 · p. 24
• The campaign also examined incorrect ready decisions and clauses from which the polarity grammar extracted no cue.Gate-only counts cover tasks with incorrect ready status; the lexical sweep recognizes rise/fall patterns only with recognized head nouns and may undercount contradictions lacking that relation. Seed 17 was used only for sensitivity, not pooled with seed 0; when x = 1, the guarded label is inconclusive regardless of n.Source: Table 8 · p. 24

Candidate generation and acceptance

• The gate rejected all 21 fabricated ready plans committed on 22 routed unanswerable tasks, while release justifications remained traceable to supporting clauses.The discussion attributes rejection to textual support, rather than the model's commitment.Source: §4 · pp. 24–25

Model contribution

• The model made no calls on 60 of 83 releases, and a model-free replay reproduced the same 83 releases under grammar V1 with a gold user.The model's measured contribution was limited to process events involving questions or delivery without a question.Source: §4 · p. 24

False releases

• Zero false releases were observed among 83 seed-0 plans released with the gold user; the result is bounded to the evaluated setting.The separate ungated-shell run had 9 false deliveries among 18, and the in-run ungated terminal had 6 among 13.Source: §4 · pp. 25–26

Incorrect user answers

• Incorrect answers were released in 169 of 431 pairings when facts were left open, while none of 65 wrong-answer pairings was released when both facts came from text.Incorrect pairs were also released on two tasks with explicit coordinate exclusions.Source: §4 · p. 25

Assurance scope

• Textual entailment does not establish physical or formal safety, so closed-loop stability cannot be inferred from releasing a sign convention.The discussion says closed-loop testing must remain a separate commissioning activity.Source: §4 · p. 26

Grammar limitations

• Grammar V1's scope and lexicon mechanisms can permit release of excluded coordinates, and the study cannot separate these mechanisms from benchmark wording.The discussion identifies new unnegated clauses and exclusion wording outside the not/no/without family as mechanisms.Source: §4 · p. 26

Evaluation limitations

• Gate sensitivity and real user behavior were not measured, and the questioning policy was not evaluated using eligibility determined only from observable state.Positive coverage used the same grammar whose limitations were measured.Source: §4 · pp. 26–27

External and conclusion validity

• The benchmark results do not generalize beyond the single grammar, and one false release was later observed among 146 post-seal releases at seed 0.The single registered verdict was also based on seed 0, without pooled seeds or reported between-seed variance.Source: §4 · p. 27

Findings

• Fabricated ready plans were committed on 21 of 22 routed unanswerable tasks, and all were rejected; the same 83 releases were reproduced without model calls.The model-free replay used grammar V1 with a gold user.Source: §5 · p. 28

False-release result

• Zero false releases were observed among 83 seed-0 releases, with a one-sided 95% Clopper–Pearson upper bound of 0.0354 under the i.i.d. assumption.The bound was diagnostic and below the sealed 5% headline tier.Source: §5 · p. 28

Incorrect user answers

• Both facts were bound from original text on 13 of 96 answerable tasks, while incorrect answers were released in 169 of 431 pairings for facts left open.Incorrect pairs were also released on tasks with explicit coordinate exclusions.Source: §5 · p. 28

Scope and implications

• The protocol's benefits include traceable release justifications and a pre-evidence false-release bound, but it establishes no physical safety result.One correct delivery was lost relative to the ungated shell among 96 answerable tasks.Source: §5 · p. 28
• A post-seal campaign recorded one false release among 146 seed-0 releases, so the benchmark's zero-false-release observation cannot be extended beyond that benchmark.The identified mechanism was in grammar V1, which was not repaired.Source: §5 · p. 28

Future work

• Future evaluation should address coordinate-exclusion mechanisms, measure gate sensitivity independently of the grammar, and test partial, delayed, or refused answers.A deployable questioning policy should determine eligibility from observable state.Source: §5 · pp. 28–29

Data availability

• Research data, code, audit records, preregistration materials, and benchmark tasks are available from the corresponding author upon reasonable request, with sharing potentially constrained by file size, technical transfer, model licenses, or institutional policy.The materials have not been deposited in a public repository, DOI, or external archive.Source: Reproducibility and Data Availability · p. 29

Generative-AI use

• Open-weight models were studied and used to assist manuscript drafting and editing under the author's direction and review; experimental design decisions, preregistrations, and verdicts belong to the author.The study involved no human subjects or personal data.Source: Ethics and Generative-AI Use Disclosure · p. 29

• TÜBİTAK supported the work through its 1001 programme (project 225M316), whose meta-learning code was integrated into the test harness and used during agent design.Source: Funding · p. 29

• Appendix A covers sealing, adjudication, and search scope; B identifies comparison arms; C lists tasks; D reports predictions; E gives configurations and values; F compares prior work; G collects symbols.Source: APPENDICES · p. 33

A Sealing, archiving, and record integrity

• The release criterion and experimental plan were sealed before data and benchmark construction; a single-look rule limited evaluation to 24 compositions and 96 answerable tasks.The seals were archived internally, but no independent third-party registration was made.Source: A Sealing, archiving, and record… · p. 33
• The main experiment recorded 183 model attempts, met its sealed attempt and tool-charge limits, and recorded no kill-rule counts; model-free bindability cleared the stage-validity floor in all stages.Integrity replay agreement was 38/38, 112/112, and 70/70 stage records.Source: A Sealing, archiving, and record… · p. 33
• Neither selected-stratum instrument had a valid source and configuration record, so its measurements require that limitation to be considered.Adjudication addressed all 14 aggregate-report findings and recalculated 64 claims.Source: A Sealing, archiving, and record… · p. 33
• Across repeated stage seals, 139 of 144 method items matched the initial seal in each selected-stratum instrument, while the model-free replay found 0 mismatches across 1584 field comparisons.Source: A Sealing, archiving, and record… · p. 34
• Ready status was reproduced for both signs on both tasks, including T-4 despite contradictory parser status; all four incorrect-pair releases were among 57 confirmed substitutions.The remaining 23 substitutions lacked records, and the replay's 80 tasks differed from the live protocol's 76 questioned tasks.Source: A Sealing, archiving, and record… · p. 34
• The sealed T1 template joins two fixed clauses with a semicolon, substitutes the coordinate in both clauses and the polarity in the sign field, and appends the result to the requirement.Grammar V1 names the first sealed requirement-grammar version, including its lexicon and rules.Source: A Sealing, archiving, and record… · p. 34

A.1 Literature-search record

• The recorded term-level search covered 88 full-text sources, 1077 local-database records, and three external searches; it did not rule out equivalent concepts expressed differently.Source: §A.1 · p. 34

• In the ungated run, 18 plans were delivered, of which 9 were false and 9 correct; all 9 false deliveries came through the model-calling branch.All four deterministic-branch deliveries were correct.Source: B Comparison arms and controls · p. 34
• Gate specificity was measured on 4645 pairings, including 4230 across-composition cases: 3754 missing-fact and 476 contradictory outcomes.The remaining 415 pairings were within compositions.Source: B Comparison arms and controls · p. 34
• Of 496 wrong-answer pairings, 65 from 13 text-bound tasks had 0 releases, while the remaining 431 pairings came from 83 Topen tasks.The one-sided 95% Clopper–Pearson upper bound for the 65 pairings was 0.0450.Source: B Comparison arms and controls · p. 34
• The text-bound set was not defined to equal the text-only bindability-ceiling set; they coincided because the gold pair was text-bound on every task in that stratum.The ceiling required ready status and a gold-equal pair, whereas the plan-free call required only binding both facts.Source: B Comparison arms and controls · p. 34
• Table A1 distinguishes comparison arms by population, intervention, and reporting location; quantities are not pooled across changed conditions or seeds.The caption cautions that no common total or single verdict is implied.Source: Table A1 · p. 35

C Named tasks

• T-8 and T-10 were unreleased tasks, and T-9 retained ready status after a fact was removed.T-10 was identified as the second unreleased task.Source: C Named tasks · p. 36
• T-13’s fabricated plan was stopped before release; T-11 and T-12 were released without a question.A separate entry records no plan at seed 17 and release without a question at seed 0.Source: C Named tasks · p. 36
• The examples include released gold-matching pairs for liquid level +1 and sink temperature −1, alongside a development-experiment task outside the protected benchmark.The sink-temperature release is repeated as a separate result entry.Source: C Named tasks · p. 36
• T-4 released brake pressure instead of the gold vehicle-speed pair, and T-5 released runner speed instead of penstock flow.Both requirements contained wording that excluded the incorrectly released coordinate.Source: C Named tasks · p. 36
• T-14’s exclusion wording fell outside the grammar’s recognized not/no/without family, while T-15 tested nut deflection instead of gold carriage velocity.In T-15, the incorrect coordinate was independently supported by the original requirement.Source: C Named tasks · p. 36

D Preregistered predictions and verdicts

• Table A3 records each closing-arm prediction, measurement, and verdict, including failures; P2b and false-refusal results were measured after campaign closure.The single verdict under the primary sealed criterion is not replaced by the table.Source: Table A3 · p. 37, D Preregistered predictions and verdicts · p. 37
• 0 false releases were recorded among 68 releases (Rsm), meeting that predicate, while the 70–83 release prediction was not met.Both outcomes are reported as separate registered predictions.Source: D Preregistered predictions and verdicts · p. 37
• The release set was contained in primary-model seed-0 releases, with 0 false releases.This is a separate reported result, not a replacement for the sealed criterion verdict.Source: D Preregistered predictions and verdicts · p. 37, Table A3 · p. 37
• The subset measured at all three stages had 0 false releases and met its predicate.Source: D Preregistered predictions and verdicts · p. 37

E Model configurations and recorded numerical values

• The second model recorded 8, 37, and 23 releases by stage, with task identities checked individually against the sealed method and stage gold seal.Source: E Model configurations and recorded… · p. 37
• 7 of 22 routed unanswerable tasks had no committed plan from the second model; 17 of 47 routed tasks had no committed plan, versus none for the primary model.The second-model 17 episodes comprised 10 answerable and 7 unanswerable tasks.Source: E Model configurations and recorded… · p. 37
• All inference settings other than model identity and fingerprint were identical; no exact parameter count was recorded for the primary model.The second arm is labeled as development throughout its table column.Source: Table A4 · p. 38
• Table A5’s rates retain source-record precision and must be read with their arm and seed, not as exact fractions.The table introduces no new calculation.Source: Table A5 · p. 38
• The hardware supplement uses dated records obtained separately before the main runs; continuity of hardware use could not be established through persistent machine identity.Reported system memory is server-reported total memory, not nominal installed capacity.Source: Table A6 · p. 39

F Positioning relative to prior work and retained controls

• The literature positioning compares decision authority, evidence type, and scope without implying a ranking under a common performance metric.The shared architectural feature is separation of candidate generation from release authority.Source: Table A7 · p. 40

G Symbols

• Evidence notation distinguishes ordered evidence-span sequences, matched clauses, and the ordered analysis units evaluated separately by the gate.When no clause matches a span, the analysis-unit sequence uses the singleton span.Source: G Symbols · p. 41
• Plan notation distinguishes candidate plans, asserted coordinate–polarity pairs, unasserted facts, and the statuses ready, missing-fact, contradictory, and unsupported.Unsupported is restricted to plan and gate statuses.Source: G Symbols · p. 41
• Gate notation names the entailment relation, resolved coordinate and polarity values, evidence validity, and binding of both facts without plan disagreement.The notation also defines an order-preserving operator that retains the first occurrence of each identical sequence element.Source: G Symbols · pp. 41–42
• Protocol symbols distinguish routing branches, turn-specific statuses and release indicators, user-supplied facts, canonical answer templates, and episode call counts.Source: G Symbols · p. 41
• Endpoint and decision notation covers terminal release, false release, release and false-release counts, one-sided 95% Clopper–Pearson bounds, and supported, refuted, or inconclusive labels.Guard notation marks a concentration-guard indicator; composition symbols count tasks and releases by composition.Source: G Symbols · pp. 41–42
• Task-set notation distinguishes answerable tasks never released, text-only bound tasks, their complement, the released set, and the bindability ceiling.The symbol table groups notation by object and says definitions appear below equations or in associated sections.Source: Table A8 · p. 40, G Symbols · p. 42