SoK: Semantic Decision Engines in Network Control Loops

SoK: Semantic Decision Engines in Network Control Loops Delong Li, Chen Li, Xu Wang, Haochen Gong, Rui Lang, and Guangsheng Yu School of Electrical, Mechanical and Biomedical Engineering University of Technology Sydney, Sydney, Australia

arXiv:2610.06425v1 [cs.NI] 5 Oct 2026

¹ OniReimu/SoK-JEV

Abstract—A semantic decision engine such as Jev can return a valid answer and still miss a network deadline, select an infeasible action or leave the service unverified. We systematize 139 paper families by decision interface, execution path and check ownership. Fifty families claim that their engine fits a control loop or time budget, but only four support the claim with matched measurement. Across all 139, four report deadline attainment. The gap concentrates where the decision has no deterministic computation step. Those 72 families make 22 of the claims, none supported, and name a coverage owner in only two. Bounded tests under one event model show that each gap can reverse an admission verdict. A decision that meets a 10 s budget for every isolated request meets it for none once decisions queue ahead of replayed execution times. The same engine passes one coverage check and fails another. We derive a minimum reporting record, design rules and a research agenda for admitting decision engines to control loops. Index Terms—Intent-based networking, network control, semantic decisions, Jev, systematization of knowledge, service orchestration, decision latency.

* OniReimu/SoK-JEV

operator

“Restore the route”

Select

Checks

Submit

candidate

State Feasibility Coverage

ACK

Generate

network intent

…

config.

state · candidates constraints

Compute

intent + context

semantic interface

admit

f(x)

predicate

reject/refresh controller

Verify execution

t0

tq

td

ta

tb

Arrival

Slot start

Accepted

ACK

Verified

queue

Q = tq − t0

G1: Format

decision

submit

D = td − tq Ta = ta − t0 Tb = tb − t0 G2: Task correctness

probe

timing sketch; not to scale

G3: Verified outcome

Figure 1. Three alternative interfaces inside a control path. The controller checks state, feasibility and coverage before admission. Compute evaluates f (x). Application programming interface (API) or configuration acknowledgement (ACK) and verified outcome are distinct endpoints. G1–G3 report format validity, task correctness and verified outcome.

1. Introduction A network controller receives a request to move a service, restore a route or change a radio policy. Before the action takes effect, the controller interprets the request, checks the available state and resources and admits work to an execution path. A quick answer helps only when this complete path meets the requested outcome and its deadline. For example, the near-real-time controller in an open radio access network (O-RAN) operates on typical 10 ms– 1 s timescales. State collection, communication, decision waiting, control and observation all consume that budget [1]. Language interfaces make such requests easier to express, but they also move part of the controller’s responsibility into the interpreter. How much moves depends on the interface. The interpreter may choose among alternatives that the controller supplies, construct fields or configurations that the request leaves out, or pass a formalized request to a rule or solver that computes the answer. In each case a wellformed output can still be wrong for the network, because a schema-valid answer can exceed a shared quota, come from an incomplete catalogue or rest on contradictory observations. Conversely, a justified rejection can be the correct

outcome even though no service is activated. Figure 1 places these decisions in the surrounding control path. We ask which network control loops can admit a semantic decision engine and which checks the controller must keep. The literature rarely measures the complete path that this question requires. Of the 139 paper families we review, 50 claim that their engine fits a control loop or time budget, and 4 report measurement matched to the claimed budget and boundary. Nine families report latency at the 95th percentile or above, and four report deadline attainment (§5). In the systematization, the 72 families whose decision has no deterministic computation step make 22 of the 50 loop claims, none with matched measurement. Only 23 of them name a feasibility owner and 2 a coverage owner, against 45 and 11 of the 67 families with such a step. We test whether the missing measurements and checks change an admission verdict, using bounded measurements in one event model. The service-orchestration study [2], the RAN study [3] and our application tasks become instances of one framework, and fixed-action interventions on transport and edge execution stacks vary only the decision delay.

Each test measures the consequence of one gap. For missing load evidence, a decision that meets a 10 s budget for every isolated request meets it for none at 4.992 arrivals/s once decisions queue ahead of measured execution times. For unnamed check ownership, an engine that owns the coverage check requests a new candidate for 45 of 48 contracts without one but escalates 4 of 120 infeasible routes. This systematization of knowledge (SoK) contributes: • A three-axis systematization of 139 families by decision interface, control and execution path, and check ownership, mapped to standardized network functions and recoded blind to measure its reproducibility. It shows that check ownership and load evidence concentrate in families with a deterministic computation step. • An evidence audit with independent coding and explicit denominators, which finds matched support for 4 of 50 loop claims and yields a minimum reporting record. • Bounded tests of the gaps that the systematization exposes. One event model with four diagnostic lenses (measurement boundary, workload summary, check placement and robustness scope) compares the service, RAN and application studies without pooling their effects. Fixedaction delay interventions on transport and edge stacks quantify endpoint shares and queue amplification. • Design rules for admitting decision engines and a research agenda, each derived from a gap in the systematization or the audit.

2. Background 2.1. Control Loops and Their Time Budgets O-RAN distinguishes non-real-time (non-RT) radio access network (RAN) intelligent controller (RIC) functions above approximately 1 s from near-real-time (near-RT) RIC functions at approximately 10 ms–1 s [1]. Management and orchestration can act on slower operational timescales. An interpreter shares these intervals with state collection, communication and execution. The architecture also describes faster device-internal real-time (RT) control. Distributed applications (dApps), proposed outside the standardized twotier RIC hierarchy, extend artificial intelligence and machine learning toward this faster regime [4]. Non-RT applications (rApps) supply policy or enrichment through the non-RT RIC’s A1 services, while nearRT applications (xApps) consume measurements and issue control through the near-RT RIC’s E2 services. E2 service models distinguish measurement reporting, including Key Performance Measurement (KPM), from control functions [5]. A control acknowledgement and a later KPM change establish different milestones, and service verification requires an outcome check.

2.2. Intent, Analytics and Closed-Loop Functions The 3rd Generation Partnership Project (3GPP) defines intent-driven management in Technical Specification

(TS) 28.312, including feasibility checks before fulfilment [6]. TS 23.288 describes the Network Data Analytics Function (NWDAF), while TS 28.104 covers Management Data Analytics (MDA), its functions and services [7], [8]. Analytics informs policy and orchestration, while installation and outcome verification occur downstream. The European Telecommunications Standards Institute (ETSI) separates analytics, intelligent decision, orchestration and control in its Zero-touch network and Service Management (ZSM) architecture [9]. Its closed-loop automation specification adds governance and coordination of closed loops [10]. TM Forum’s TMF921 interface provides intent and report resources between an intent owner and handler, and IG1253 frames their intent-management relationship [11], [12]. Requests for Comments (RFCs) 9315 and 9316 supply intent-based networking concepts and intent classification, respectively. Both are Informational documents from the Internet Research Task Force’s Network Management Research Group [13], [14].

2.3. Selection, Generation and Computation Selection returns a member of a supplied alternative set, which may include rejection, clarification or escalation. Typed-choice application programming interfaces (APIs), candidate-constrained language models and discriminative classifiers can implement this interface. Generation constructs content not already represented as a complete alternative, including extracted values, programs, configuration edits and plans. Schema or grammar constraints govern output form, whereas the task predicate determines semantic correctness [15], [16], [17]. Computation uses rules, solvers or verifiers to evaluate formalized conditions or synthesize a satisfying assignment. A workflow can use all three interfaces. Intent interpretation may generate a bounded contract, a scheduler may compute joint feasibility, and a selector may choose a certified action. The classification follows each step’s output responsibility, independent of the vendor and of whether the step uses a language model internally. Comparing implementations requires the same usable output, available information and downstream checks.

3. Review Method 3.1. Scope and Source Identification We reviewed semantic and intent-based decision making for network configuration, orchestration and diagnosis. Eligible studies identify a network decision output and range from explicitly described conceptual designs to deployed systems, with or without latency measurements and positive results. For mixed-domain studies, we extract the identifiable network tasks. We exclude pure resource prediction without a configuration, orchestration or diagnostic decision. Source identification began with 92 records cited in the manuscript and added three fixed arXiv title-and-abstract

92 cited

151 unique arXiv IDs

21 additions

records

159 hits − 8 repeats

targeted discovery

−14 cross-source overlaps

250 merged records after source deduplication bounded through 28 September 2026

1. Discovery + merge

250 records

145 full-text candidates

139 source/version records

133

title / abstract screen

134 retained + 11 boundary

137 eligible + 2 linked versions

reading groups

2. Screening + grouping

−79 background −26 out of scope

6 pairs grouped

−6 background

assistant-led

Researcher 1

133

independent coding

Compare answers

reading groups

Researcher 2

adjudicate scope

independent coding

132 families → 405 coded scopes 374 step definitions + 31 linked scopes 3. Independent coding + author adjudication

132 included families 138 source IDs · 137 bibliography entries

231 initially comparable + 174 aligned rechecks = 405 answer pairs

−1 reading group

Task and version scopes remain distinct

CPU forecasts only

18 evidence fields field-level reporting and agreement

111 set-aside records

Blind second screen

24 full-text checks

+7 families

139 analysed families

79 + 26 + 6 from stage 2

two model coders

7 eligible · 2 split

24 coded scopes

429 scopes · 393 steps

966 OpenAlex works

847 screened

100 sampled

98 coded · 77 eligible

−119 already in corpus

594 pass title / abstract

50 + 50 by stratum

coverage audit, not added

4. Verification + coverage

Figure 2. Corpus formation and independent coding. Records, versions, groups, families and scopes are distinct units. Screening was assistant-led. Two researchers independently coded the groups, then jointly adjudicated. The 231 initially comparable pairs and 174 aligned rechecks form two strata. Stage 4 adds the blind second screening of all set-aside records, which contributes seven families, and the OpenAlex coverage audit, whose sampled records stay outside the analysed set. Search and coding definitions appear in Appendix D, and agreement is reported in Table 25 and Appendix E.

queries, publisher-page discovery and one backward-citation round. Publisher discovery was limited to individual pages from the Institute of Electrical and Electronics Engineers (IEEE) and Association for Computing Machinery (ACM). The queries cover 1 January 2016 through 28 September 2026. The citation seeds are INTA [18] for configuration, Chat-Driven Optimal Management [19] for orchestration and Network Arena [20] for diagnosis. The three complete arXiv result sets contain 159 hits and 151 unique identifiers. Combining these with the starting bibliography and targeted additions produces 250 records. Figure 2 records duplicates, cross-source overlaps, exclusions and report grouping. Assistant-led retrieval, deduplication and screening identified 145 records for review of the relevant full-text sections. These checks retained 139 source/version records in 133 reading groups and moved six records to background material. Appendix D gives the search routes, complete queries, retrieval date, eligibility criteria and coding definitions. Two researchers independently coded the retained groups. Their review excluded one study that predicted central processing unit (CPU) consumption without an identifiable in-scope network decision. This screen yields 132 paper/report families and 138 source/version records, mapped to 137 bibliographic entries. Figure 2 distinguishes initial screening from independent human coding. A second screening of all exclusions adds seven families (§3.2), giving 139 families and 145 source/version records.

3.2. Screening Verification and Search Coverage Two independent coders re-screened all 111 records that the initial screen had set aside. These comprise 79 background and 26 out-of-scope records from title-and-abstract screening and the six records moved to background after full-text review. The coders were two language models from different families, one GPT and one Claude model. Each worked from the published eligibility criterion, blind to the original decision and to the other coder. Title-and-abstract agreement over include, background and exclude was Cohen’s κ = 0.661 [21], with Gwet’s agreement coefficient (AC1) [22] at 0.738. After one exchange of rationales, 24 records went to full-text review, where both coders judged seven eligible. Because the original reasons for these seven applied a narrower operational scope than the published criterion, the seven enter the corpus as new families. Two further records remained split after reconciliation and stay outside the corpus. The same two model coders applied the unchanged 18field codebook to the seven families. Before reconciliation they agreed on 114 of 126 family-level field judgments (90.5%), including all four timing-claim fields. A field whose remaining disagreement after reconciliation changed family-level positivity is coded unclear. Seven field judgments are coded this way, none of them in the four timingclaim fields. This agreement is kept separate from the researchers’ agreement in Table 25. A recall check ran the three query concepts over OpenAlex, a multi-publisher index, with the same date window. The 966 unique works include 119 records already in the corpus. The two coders screened the other 847 by title

and abstract with three-class κ = 0.741 (AC1 0.812) and κ = 0.828 for include versus not. Of these, 594 passed. Before checking access to any full text, we drew a stratified random sample of 100 records, 50 from languagemodel work and 50 from earlier intent networking. Full text was retrieved for 98 draws, and the two coders applied the full-text criterion and the four timing fields of the codebook. Before reconciliation they agreed on eligibility for 90.8% of the draws and on the timing fields for 60.2–89.8%. Of the 98 draws, 77 are eligible and 2 remain unresolved. The frame-weighted rates are 32.4% for explicit loop claims (95% interval 22.2–42.9%), 6.5% for matched support (1.3– 12.4%), 5.3% for p95 reporting (1.2–10.8%) and 2.6% for deadline attainment (0–6.9%). Each interval contains the corresponding corpus rate. Bounds that count the two unretrieved draws and every unresolved code as negative or as positive give 30.7–45.4% for loop claims and 6.1– 20.9% for matched support. Appendix D.2 gives the queries, strata, per-stratum counts and estimator. Records outside the primary search routes therefore report loop claims and matched support at rates consistent with the corpus. The 139 families remain the analyzed set, and their denominators count families from the primary routes.

3.3. Coding Units and Adjudication Each coding unit identifies a work, task, decision step and evidence scope. The final inventory contains 429 scope records, made up of 393 step definitions and 36 linked scopes. The two researchers coded 405 of them, and the two model coders coded the 24 scopes of the added families. Linked scopes retain distinct implementations, versions or aggregate measurements without automatically creating another step. Shared measurements are linked instead of being copied to each component of a pipeline. Studies that share a benchmark remain separate families. The codebook contains 18 fields in four groups (Table 24). Eight record timing, workload and comparison evidence such as tail latency, deadline attainment and a nonlearning comparator. Four record the control-loop claim, its measured support and the deployment context. Four record the three correctness gates and whether a scope distinguishes them, and two record public code and data pointers. Missing, unclear and not-applicable evidence retain distinct codes. Of the 405 scope pairs, 231 compare the initial answers and 174 compare independent rechecks made after the scope boundaries were aligned. Table 25 reports these strata separately and in combination. The combined estimate summarizes both stages before adjudication. Coverage estimates use consensus values, whereas agreement estimates use the independent answers. For single-choice fields, we report unweighted κ and AC1. AC1 uses the codebook’s category universe. The two multi-select fields use exact-set agreement and option-wise binary κ/AC1. Table 27 gives the option-wise coefficients with each researcher’s selection count, so that common nonselection stays visible.

After comparing the independent answers, the researchers clarified the load and stability definitions during adjudication (Appendix D.4). Operational arrival or traffic load is distinct from single-request complexity, and stability concerns behavior over operational time or load. This clarification changed 65 of 810 load/stability values.

3.4. Aggregation and Uncertainty The primary estimand is reporting coverage across all 139 families. A family is positive for an indicator when at least one scope has the target code. Missing, unclear and not-applicable records remain in this denominator. Table 3 also reports coverage over 429 scopes and over the 393 step definitions, and Table 28 gives the scope-level category counts. Aggregate measurements count at the scope where they are recorded. Descriptive 95% confidence intervals (CIs) use 10,000 bootstrap resamples of whole paper/report families with seed 42. Resampling whole families preserves the dependence among scopes of the same family. These intervals summarize variation among families in the assembled corpus under the adjudicated codes. Zero or complete observed coverage produces a degenerate empirical interval. The task-group figure uses overlapping screening labels. Configuration covers 66 families, orchestration 52 and diagnosis 38. These labels were assigned while preparing the reading groups and were not independently coded. Because multi-task families appear in more than one group, comparisons between groups are descriptive.

4. Systematization 4.1. Three Axes We locate each network task on three axes. The decision interface distinguishes selection, generation and deterministic computation. The control or execution path distinguishes offline preparation from request-driven operation and identifies the functional role consuming the output. Check ownership records who performs three checks. The observation check supplies or reconciles the observed state a decision uses. The feasibility check evaluates joint feasibility, and the coverage check detects absent candidate coverage. A single paper can contain several interfaces and measurement scopes. Table 23 maps all 139 families on the three axes. Where the reviewed evidence specifies them, check entries give the owning component and the type of evidence at the reported task boundary, from a proposed design or model self-check to deterministic validation, a finite execution test or formal verification. A check whose responsible component remains unclear in the reviewed evidence has unresolved ownership. Two model coders, one from the GPT family and one from the Claude family, recoded the three axes blind from full text under a frozen codebook. A third blind pass mapped the published cells into the same fields (Table 26). Over the

TABLE 1. R EPORTING EVIDENCE AND CHECK OWNERSHIP BY DECISION INTERFACE ACROSS THE 139 FAMILIES . ROWS PARTITION THE FAMILIES BY THE INTERFACE LETTERS IN TABLE 23: SELECTION (S), GENERATION (G), DETERMINISTIC COMPUTATION (C) AND UNSPECIFIED (U). C ELLS COUNT FAMILIES . R EPORTING COLUMNS USE THE EVIDENCE AUDIT, AND CHECK COLUMNS COUNT FAMILIES WHOSE PUBLISHED ENTRY NAMES AN OWNER .

Interface

Loop claim Check owned n Any Matched p95+ Deadline Load Feasib. Cover.

S only G only U only Mixed, no C

11 24 2 35

4 5 0 13

0 0 0 0

0 1 0 1

1 1 0 0

4 3 0 6

1 9 1 12

0 0 0 2

C only Mixed with C

19 48

3 25

0 4

2 5

0 2

11 24

14 31

1 10

All

139

50

4

9

4

48

68

13

129 families outside a 10-family pilot, the coders agree on the execution path with κ = 0.69 and on the generation and computation interfaces with κ = 0.72 and 0.67. Selection is less reproducible (κ = 0.50). Whether a family places an observation, feasibility or coverage check is reproduced with κ from 0.62 to 0.73, and the owner category of a shared check matches exactly in 65 to 68% of cases. Agreement with the published entries is lower for check presence (κ from 0.10 to 0.42). Both coders agree that 40 of the 129 families place an observation check and 43 a coverage check, where Table 23 records 9 and 11. Of those 9 observation checks, both coders find 6, and of the 11 coverage checks they find 7. Neither coder finds one of the observation checks and two of the coverage checks. The codebook excludes plain reading of current state, which several published observation entries include. The check columns therefore support no prevalence estimate, and the paper draws none from them.

4.2. Patterns Across the 139 Families Generation is the most common decision interface. It appears in 94 of the 139 families, selection in 76 and deterministic computation in 67, and 83 families combine two or more interfaces (Table 1). In the published entries, check ownership concentrates in families with a deterministic computation step. Families with deterministic computation name a feasibility owner in 45 of 67 cases and a coverage owner in 11. Families without it name a feasibility owner in 23 of 72 and a coverage owner in 2. Operational load is reported by 35 of the 67 and by 13 of the 72. In the evidence audit (§5), families without a computation step make 22 of the 50 loop claims, and none of the 22 is supported by matched measurement. All four matched claims come from families that combine computation with selection or generation. These counts are descriptive and overlap with task groups. The lenses in §7 test what the missing responsibilities change.

4.3. Functional Roles and Check Ownership Table 2 maps tasks to model outputs, controller responsibilities and comparison baselines. In an O-RAN workflow, an xApp or rApp can consume an interpreted objective, but measurements, policy communication and device control remain separate functions. In management, the intent handler decomposes objectives and an orchestrator binds them to resources. A software-defined networking (SDN) controller or policy function installs decisions. These roles can place the observation, feasibility and coverage checks in different components. The implemented path fixes where each check runs. Check ownership determines when validation can affect an action. NetComplete computes a completion of an operator’s sketch under modeled requirements [26], and VeriFlow evaluates forwarding invariants [27]. INTA generates configurations after retrieval [18]. In configuration-repair evaluation, a final scoring oracle evaluates the completed configuration, whereas an agent-feedback checker guides intermediate repairs [33]. A deployment guard checks actions before admission. Table 23 records the reported check responsibilities.

4.4. Candidate Coverage and Conditional Selection For decision instance i, let Ci be the permitted output set, Ai the acceptable output set under the task predicate and âi ∈ Ci the selection. The coverage indicator Ki = 1{Ci ∩ Ai ̸= ∅} is one exactly when an acceptable output is available. If Pr(Ki = 1) > 0, then Pr(âi ∈ Ai ) = Pr(Ki = 1) Pr(âi ∈ Ai | Ki = 1).

(1)

Valid rejection or clarification belongs in both sets when the task allows it, and multi-field answers require an acceptable complete output. Invalid interface returns are counted separately. Equation (1) separates availability of an acceptable response from selection quality conditional on availability. When the permitted output is an escalation request, a correct selection initiates catalogue construction or service recovery. Candidate preparation, construction of missing content, selection, execution and verification each have their own outcome. The tests in §7 examine how escalation success changes across tasks.

5. Literature Reporting Coverage Table 3 reports all 18 indicators over the 139 families with their denominators and family-level intervals, and Figure 3 compares reporting across overlapping task groups.

5.1. Timing Coverage and Operational Conditions Among 139 families, 92 (66.2%) identify a timing endpoint and 49 (35.3%) specify a metric denominator under the codebook. Tail and deadline evidence is much less common. Nine families report latency at the 95th percentile (p95)

TABLE 2. M ODEL ROLES AND OTHER WORKFLOW RESPONSIBILITIES FOR REPRESENTATIVE NETWORK TASKS . BASELINES REFLECT THE OUTPUT RESPONSIBILITY OF EACH COMPARISON . Task

Model role

Other workflow steps

Baselines

Intent [2], [23] Documents [18], [24] Configuration [25], [26] Properties [27], [28] Diagnosis [29], [30] Reuse [31], [32]

Classify attributes Rank or reject Select template or tool Interpret requirement Rank tests or actions Trigger reinterpretation

Values; binding; joint checks Retrieval; extraction Generation; validation Formalization; checking Hypotheses; probes; execution Refresh; recomputation

Classifier; structured LLM Retrieval score; re-ranker Constrained generator Symbolic verifier Probing; procedures; tool agents Cache with controller

or higher (6.5%, 95% interval 2.9–10.8%), and four report deadline attainment (2.9%, 0.7–5.8%). Operational input load is reported in 48 families (34.5%, 26.6–42.4%), queue or waiting evidence in six (4.3%, 1.4– 7.9%) and stability evidence in 31 (22.3%, 15.8–29.5%). The stability indicator records analytical, simulated or testbed examination of behavior over operational time or load. Single-request scaling is coded separately. An explicit non-learning comparator appears in 48 families (34.5%, 26.6–42.4%). Twenty-three families explicitly include network round trips in a relevant timing measurement (16.5%, 10.8–23.0%). The denominator is all 139 families, including those for which network round trips are inapplicable. Table 28 distinguishes explicit exclusion, unclear boundaries and inapplicability.

5.2. Correctness Gates and Reusable Artifacts Semantic checks are reported in 133 families (95.7%), format checks in 71 (51.1%) and final-state checks in 54 (38.8%). These counts can refer to different scopes within a family. Nine families distinguish all three gates in some scope (6.5%, 2.9–10.8%). Public code pointers appear in 34 families (24.5%) and public data pointers in 51 (36.7%). The indicators record links supplied in the paper or supplement.

5.3. Coding Reliability Combined pre-adjudication exact agreement is 99.8% for tail reporting and 98.3% for deadline reporting, but 27.9% for load, 39.3% for queue evidence and 30.6% for stability. For these three fields, the κ/AC1 pairs are 0.083/0.088, 0.041/0.281 and 0.075/0.143. Agreement is also low for network round-trip inclusion (42.5%) and final-state checks (42.2%). Coverage of these five fields therefore carries greater coding uncertainty than tail and deadline coverage. Appendix D.4 gives the load and stability definitions clarified during adjudication. Table 25 reports agreement before adjudication, and Appendix E gives the category counts and multi-select agreement.

5.4. A Minimum Reporting Record Each network decision step needs a reporting record that connects its input to the claimed outcome. Identify

the input, decision output, candidates and checks assigned to the model and controller. Define the measured events from arrival and queue entry through decision, execution acknowledgement and service verification, and mark unavailable events explicitly. Report each correctness gate with its own denominator and retain failure, timeout, rejection, escalation and fallback counts. Give the latency distribution and deadline attainment for the claimed control loop. Report them with the operational arrival or traffic load, decision concurrency, queueing or waiting evidence and the observation window. Include a non-learning comparator that uses the same information and constraints and undergoes the same execution checks. Identify deployment location and whether communication, retries and validation enter the timing boundary. Provide code and data pointers, and state which measurements are inapplicable.

6. Unified Evaluation Framework A decision engine must be evaluated at the point where its output changes a network workflow. Our framework defines common events, correctness gates and perturbations for service contracts, RAN policies and application-layer control tasks. Fixed-input comparisons examine semantic decisions. Load traces measure queueing and deadline attainment. Fixed-action interventions isolate decision delay in transport and edge execution stacks.

6.1. Timing and Decision Capacity An intent arrives at t0 , leaves the decision queue at tq and produces a decision accepted by the declared schema or controller checks at td . The execution interface acknowledges the action at ta , and a domain-specific probe verifies the requested result at tb . We record queueing, decision time and the two downstream elapsed times separately: Q = tq − t0 , Ta = ta − t0 ,

D = td − tq , Tb = tb − t0 .

(2) (3)

Semantic correctness is evaluated separately against the task predicate (Gate 2, §6.2). The meaning of ta follows the interface contract. In the edge and transport intervention, ta is an API or configuration acknowledgement and tb verifies the new workload or route. Table 4 maps these events and the milestones recorded by each execution interface.

TABLE 3. L ITERATURE REPORTING COVERAGE . C ELLS GIVE COUNT / DENOMINATOR , PERCENTAGE AND 95% PAPER - FAMILY BOOTSTRAP INTERVAL (10,000 DRAWS ). A POSITIVE FAMILY HAS AT LEAST ONE SCOPE WITH THE CODE . U NCLEAR MAY OVERLAP POSITIVE COVERAGE , AND ALL -NA MEANS EVERY SCOPE IS INAPPLICABLE . A LL REMAIN IN THE STATED DENOMINATORS . S TEP DEFINITIONS EXCLUDE THE 36 LINKED SCOPES . Field Families (139) Timing, workload, and comparison Named timing endpoint 92/139 (66.2%) [58.3, 74.1] Specified metric denominator 49/139 (35.3%) [27.3, 43.2] p95 or higher 9/139 (6.5%) [2.9, 10.8] Deadline attainment 4/139 (2.9%) [0.7, 5.8] Input load 48/139 (34.5%) [26.6, 42.4] Queue / waiting 6/139 (4.3%) [1.4, 7.9] Stability evidence 31/139 (22.3%) [15.8, 29.5] Non-learning comparison 48/139 (34.5%) [26.6, 42.4] Control-loop claim and deployment Explicit loop claim 50/139 (36.0%) [28.1, 43.9] Supported loop claim 4/139 (2.9%) [0.7, 5.8] Named execution location 86/139 (61.9%) [54.0, 69.8] Network round trip included 23/139 (16.5%) [10.8, 23.0] Validation responsibilities Format check reported 71/139 (51.1%) [42.4, 59.0] Semantic check reported 133/139 (95.7%) [92.1, 98.6] Final-state check reported 54/139 (38.8%) [30.9, 46.8] Three gates distinguished 9/139 (6.5%) [2.9, 10.8] Artifact pointers Public code pointer 34/139 (24.5%) [17.3, 31.7] Public data pointer 51/139 (36.7%) [28.8, 44.6]

All scopes (429)

Step definitions (393)

Unclear families

All-NA families

164/429 (38.2%) [31.9, 44.9] 86/429 (20.0%) [14.6, 26.1] 14/429 (3.3%) [1.2, 5.8] 7/429 (1.6%) [0.2, 3.6] 87/429 (20.3%) [15.0, 26.0] 11/429 (2.6%) [0.7, 4.9] 49/429 (11.4%) [7.6, 15.5] 68/429 (15.9%) [11.7, 20.3]

143/393 (36.4%) [30.1, 43.1] 69/393 (17.6%) [12.5, 23.4] 11/393 (2.8%) [1.1, 4.9] 5/393 (1.3%) [0.2, 2.8] 74/393 (18.8%) [13.5, 24.5] 8/393 (2.0%) [0.5, 3.9] 36/393 (9.2%) [6.0, 12.6] 57/393 (14.5%) [10.5, 18.8]

2 55 1 2 1 2 0 3

0 46 0 0 0 0 0 0

127/429 (29.6%) [22.0, 37.5] 7/429 (1.6%) [0.2, 3.6] 214/429 (49.9%) [41.8, 57.5] 39/429 (9.1%) [5.1, 13.9]

112/393 (28.5%) [21.1, 36.3] 5/393 (1.3%) [0.2, 2.8] 191/393 (48.6%) [40.5, 56.6] 31/393 (7.9%) [4.1, 12.8]

0 0 51 58

0 2 16 61

138/429 (32.2%) [25.8, 38.7] 355/429 (82.8%) [78.1, 87.1] 83/429 (19.3%) [14.8, 24.2] 12/429 (2.8%) [1.1, 5.0]

123/393 (31.3%) [25.1, 37.7] 323/393 (82.2%) [77.6, 86.6] 72/393 (18.3%) [14.1, 22.9] 11/393 (2.8%) [1.1, 4.8]

0 4 9 2

0 0 19 0

96/429 (22.4%) [15.4, 30.1] 140/429 (32.6%) [24.4, 40.9]

92/393 (23.4%) [16.1, 31.4] 129/393 (32.8%) [24.3, 41.5]

8 5

0 0

Configuration (n=66)

0

p95+

Deadline

Reporting indicator

(a) Timing evidence

50

0

Load

Queue

Stability

Reporting indicator

(b) Load and waiting

100

Families (%)

50

Diagnosis (n=38)

100

Families (%)

100

Families (%)

Families (%)

100

Orchestration (n=52)

50

0

Nonlearn. Roundtrip

3 gates

50

0

Reporting indicator

(c) Comparators and checks

Code

Data

Reporting indicator

(d) Code and data

Figure 3. Reporting coverage by overlapping task group: (a) tail and deadline timing, (b) load, waiting and stability, (c) non-learning comparators, round trips and three-gate separation, and (d) code and data pointers. Bars show positive-family percentages with descriptive 95% whole-family bootstrap intervals (10,000 draws). Task labels come from screening. Evidence fields were independently coded and adjudicated. Group sizes are in the legend.

For all arrivals A, let Te,i = te,i − t0,i be the elapsed time to endpoint e ∈ {a, b} for arrival i. Deadline attainment at budget B is 1 X pbe (B) = 1{te,i observed, Te,i ≤ B}. (4) |A|

medians are computed separately. We denote the 50th, 95th and 99th percentiles by p50, p95 and p99. The slot can remain occupied after td . The quantity S = trel −tq uses actual release time trel . The offered arrival rate λ, mean occupied time E[S] and number of decision slots c define the load indicator

Equation (4) counts an unreached endpoint as a miss. The record keeps its failure reason and the elapsed time to timeout or termination. Endpoint quantiles use the requests that reach that endpoint and are reported with the observed count and all-arrival attainment. Ratios D/Ta and D/Tb use the corresponding observed pairs. We report medians of these per-request ratios. Segment medians and total-duration

ρ = λE[S]/c.

i∈A

(5)

For an open workload admitted without shedding, ρ < 1 is a necessary capacity condition. Deadline attainment also depends on burstiness, service-time variation and buffering. Admission-controlled traces distinguish offered and accepted rates. In service admission, the offered-load estimate

TABLE 4. T IMING EVENTS AND MEASUREMENT BOUNDARIES IN THE COMMON EVENT MODEL . Boundary Client return

Decision / slot

Interface ACK

Verified result

RAN observation

What the record establishes Lret is observed call duration. Format validity reported separately. Service completion needs its own probe. td is the decision after stated checks. S spans slot acquisition to actual release. Includes waits, retries and timeouts. ta is Kubernetes API acknowledgement or FRRouting configuration submission. RAN uses base-station acknowledgement. tb is a new edge replica Ready plus a correct image response or required route within its round-trip time bound. tK is the first recorded KPM change. Records a measurement update. Service verification uses a separate tb .

uses service times of dispatched requests. Service times for rejected requests are unobserved. The indicator ρ is unavailable for fixed-input and serial runs without open arrivals. For client-return records, we report Pr(Lret ≤ B) and the joint probability of format-valid return within B . RAN records supply an acknowledgement and a KPM observation at tK . In some of these records the first KPM change precedes the acknowledgement, and the signed KPM-minusacknowledgement difference is then negative.

6.2. Three Gates with Explicit Denominators Gate 1 is schema or interface validity over all arrivals. Gate 2 is satisfaction of the task predicate over applicable arrivals, including minimum-cost selection when required by that task. Deliberately unsupported requests are shown separately. Gate 3 is a verified correct terminal outcome over all arrivals. It can be a successful requested service or a completed rejection or escalation specified by the protocol. Failures, timeouts and fallbacks stay in their denominators. A rejection meets Gate 2 when the task predicate requires it. Online evaluation reports first-decision correctness, direct execution success and final success after fallback. Each task’s native predicate specifies what must be correct. Service-contract tasks report service top-1 and full-contract exact match. RAN-policy tasks report target-cluster correctness on telemetry-dependent inputs and full-policy exact match.

6.3. Perturbations and Diagnostic Lenses Table 5 maps perturbations to their experimental conditions, diagnostic lenses and measured outcomes. Structural expansion varies output fields, searched telemetry rows or interacting requirements along each task’s axis. Catalogue changes separate covered selection from rejection or escalation when coverage disappears. Unsupported service requests provide an analogous condition. The checkownership intervention compares model self-check with a

public-gate workflow for coupled placement. It changes instructions and refresh behavior together. The executionpath intervention estimates delay effects within domains and compares paths descriptively across domains.

6.4. Questions and Comparison Families The questions were formulated after inspecting the measurement records and are exploratory. Each family sets the varied condition against a declared reference. For the timing lenses, the decision-delay family tests endpoint slopes against the practical band [0.9, 1.1], and the deadline and capacity family spans all 83 load cells with pointwise time-block intervals. For check placement, the observation family compares placement prose and the missing, stale, conflicting and resolved-conflict states with complete state. It also compares stale, noisy and contradictory RAN telemetry with fresh telemetry at 57 cells. The coverage family sets rekeyed, replaced, two-candidate and absent contracts against the base catalogue and absent paths against present ones. The workflow family holds three public-gate minus self-check contrasts and shared minus independent quota under both workflows for every method. For robustness, the structure family compares access-policy conditions and service-contract padding with the base condition, and fresh RAN conditions with 7, 21 and 57 cells with three cells on the telemetry-dependent subset. Contract field-count contrasts stay descriptive because requests differ. Each finding receives a three-part verdict. The effect is repeated, task-dependent, unresolved or descriptive. Breadth is multi-task, single-task or single-study. Ownership is SoKowned, borrowed-and-reanalysed or borrowed context. We call an effect repeated when multiple tasks manipulate the same construct and use comparable outcomes. Their within-task intervals must also agree in direction beyond a declared negligible-effect reference. Different constructs or outcomes yield a task-dependent verdict. We use reference ranges of ±3 percentage points for correctness and attainment-rate differences. The latency range is ±50 ms. Each task retains its own effect estimate.

6.5. Estimation and Dependence Fixed-input tables report native proportions with 95% Wilson intervals [34], alongside return-time p50/p95 and the count with observed timing. In result tables, n is the number of evaluated cases or arrivals in the stated condition. Within-task paired contrasts match input identities and pairing strata. Service-contract field-count and catalogue-size conditions with different requests use descriptive conditionlevel summaries. Valid pairs are resampled within their original task or wording strata using 10,000 bootstrap draws with seed 42. Latency contrasts use the median of per-input differences for pairs with observed timing. Paired correctness contrasts use the two-sided exact McNemar test [35], implemented as a binomial test on discordant pairs with null probability 1/2. Holm adjustment [36] is applied separately to each declared correctness-comparison

TABLE 5. P ERTURBATION FAMILIES , DIAGNOSTIC LENSES , EXPERIMENTAL CONDITIONS AND MEASURED OUTCOMES . O UTCOMES FOLLOW THE TASK PREDICATES AND TIMING ENDPOINTS . N/A DENOTES NO CORRESPONDING ARM . ACK DENOTES ACKNOWLEDGEMENT. Family / lens

Service contracts [2]

RAN policies [3]

Application tasks / intervention

Measured outcomes

Structure L4 Irrelevant text L4 Observations L3/L4 Covered catalogue L3/L4 Absent coverage L3 Arrivals L2 Check ownership L3 Execution path L1/L2

4, 6 or 8 output fields

Fresh with 3, 7, 21 or 57 telemetry cells

Padding up to 16,384 tokens

N/A

N/A

Fresh / stale / noisy / contradictory telemetry

Scale / unseen services partial replacement Unsupported requests as an analogy

N/A

4, 16 or 64 interacting requirements Irrelevant context and repeated requirements Complete / prose / missing / stale conflict / resolved conflict Stable IDs / rekeying replacement with two candidates Missing contract candidate or infeasible path

Open admission traces

Open intent-arrival traces

N/A

Task correctness and return time Task correctness and return time Correct handling of observations and return time Selection correctness and return time Correct rejection or escalation and return Queue timetime and slot load ρ

Cached interpretation and numerical scheduling Real optical character recognition (OCR)

N/A

Model self-check versus public-gate workflow Transport/edge physical paths and real queues with replay

N/A

A1/E2 ACKs and KPM observations

with native deadline attainment Task correctness and return time Stage and outcome times with queue amplification

TABLE 6. D IAGNOSTIC LENSES , EXPLORATORY QUESTIONS , ESTIMANDS , DECLARED COMPARISON FAMILIES WITH THEIR CONTRAST COUNTS , AND DATA SOURCES . E ACH FAMILY RETAINS ALL ELIGIBLE CONTRASTS DEFINED BEFORE ANALYSIS . E DGE AND RAN DENOTE THE SOURCE STUDIES [2], [3]. Lens

Question

Estimands

Comparison families

Data

L1 Measurement boundary L2 Workload summary

How does the endpoint change timing effects and admission verdicts? How do median, tail and deadline summaries change under queued work?

Decision delay (4 slope tests) Deadline and capacity (83 load cells)

SoK, RAN

L3 Check placement

What changes when observation, coverage or feasibility checks move within the workflow?

Stage times, endpoint slopes, decision fractions Return and queue quantiles, attainment, ρ, modeled completion Correctness, validity, return p50/p95, online outcomes

SoK, Edge, RAN

L4 Robustness scope How do structure, length, representation and catalogue change bound a result?

Correctness, validity, return p50/p95, paired changes

Observation (66), coverage (39), workflow (45) Structure (89)

family. The notation pHolm denotes the adjusted p-value. Bootstrap latency intervals remain pointwise. The six-case and 24-case online subsets are descriptive. The timing analysis stratifies results by task, deployment, hardware and collection period. Its intervals resample occupied 30 s time blocks, with 60 and 120 s sensitivity views. Cells with fewer than eight occupied blocks are marked as having weak temporal coverage. Intervals are unavailable when timestamps are missing. The 10 ms, 100 ms, 1 s and 10 s budgets span control-loop timescales. Per-cell p99 is descriptive. For RAN Qwen, the primary record is the attempt returning Hypertext Transfer Protocol (HTTP) status 200 per case. Timing measures that attempt’s duration.

6.6. Execution-Path Intervention The measurement-boundary lens combines a sparse physical arm with a load arm using execution-time replay. In each domain, 30 fixed-action scenes receive delays of 0, 0.1, 0.5 and 2 s. Ten physical repeats are nested within each scene and delay. The sparse arm has zero decision queueing time (tq = t0 ). It measures each endpoint’s slope against

SoK, Edge, RAN

SoK, Edge, RAN

actual decision delay and the associated decision fraction. Whole-scene bootstrap resampling retains nested repeats and delay conditions. Four mean-slope tests, two domains by two endpoints, form one Holm family. The sparse arm completes 1,200 verified executions in each domain. Both stacks use the same guest environment. The scene, action, input and baseline recovery criteria remain fixed across delay arms. Seed 42 randomizes their order within each repeat. Transport uses containerlab 0.79 and FRRouting (FRR) 10.5.1 on an eight-router ring. The ring runs Open Shortest Path First (OSPF) with full-mesh internal Border Gateway Protocol (BGP) sessions. Each scene fixes a source, service prefix and link-metric change selecting the required lowerdelay path. Configuration submission defines ta . The route, traceroute and round-trip time (RTT) criterion defines tb . Baseline recovery checks all eight link-state databases and waits six seconds of protocol quiet. It then verifies the forwarding information base (FIB) and ping. Native protocol timers are retained. Edge uses k3d 5.9 and k3s 1.35.5. Its 30 migration or scaling scenes span two worker nodes and five image

workloads. A real Pillow service returns a checked grayscale image. New Pod identities, readiness and the output pixel hash define verification. Kubernetes API acknowledgement defines ta . Link emulation, API proxy and probe overhead are included in the measured path. The load arm uses a real first-in, first-out (FIFO) decision queue with one slot per condition. The delay elapses inside that slot, and the slot is released afterwards. An isolated 100 ms baseline is measured 200 times. The baseline sets three arrival rates with baseline load ρ0 at 0.5, 0.8 and 0.95. Those rates remain fixed across delay conditions. Each Poisson arrival trace uses seed 42, a 60 s warm-up and twenty subsequent 10 s measurement blocks. The same arrivals and physical-residual draws are shared across delay conditions. Four controlled delays and 24 empirical timing strata from the workload-summary lens define 84 queues. The empirical conditions replay recorded durations by task, implementation, deployment and collection period. After the queue, the load arm samples measured zerodelay physical residuals, including verification failures, to estimate outcome times under those residual distributions. Time-block intervals describe the finite observation window, with adjacent-block sensitivity at 20 s. A two-hour drain limit bounds observation of unfinished queues. Endpoint attainment retains every measured arrival. The edge 2 s and transport 10 s targets are exploratory task budgets. A common sensitivity grid covers 0.5/1/2/5/10 s. We estimate delay effects within each domain and compare the resulting paths across topologies, actions, stacks and probes. The load arm completes all 84 real single-slot FIFO queues. It drains all 166,432 arrivals, including warm-up. The analysis uses the 126,588 post-warm-up arrivals in the twenty measurement blocks of each queue. The measured 100 ms baseline sets rates of approximately 4.992, 7.987 and 9.484 arrivals/s. Execution after the decision queue uses paired replay from each domain’s 300 zero-delay physical observations. Twenty-four empirical timing strata supply recorded decision durations for further replay conditions [2], [3].

7. Consequences of the Systematized Gaps We apply the event model in §6.1 to service contracts, RAN policies and constructed application-layer tasks. Edge and RAN measurements are reused from their source studies. Each lens tests one gap exposed by the systematization and the audit. The measurement-boundary lens (L1) and the workload-summary lens (L2) test the missing timing and load evidence (§5). The check-placement lens (L3) tests the consequence of check ownership that the published entries leave unnamed, most often in families without a computation step (§4.2). The robustness-scope lens (L4) bounds the scope of the other verdicts. The measurements show consequences and do not estimate their prevalence. Labels identify the measured model versions. Servicecontract and RAN-policy tasks use Jev-1.13.0, DeepSeekV4.1-Flash, GLM-5.3-Flash and Qwen3.8-Flash. Their taskspecific comparators are SemIf-Qwen3.5-4B, Qwen3.5-4B-

JSON, Laya and AnyJev-L0. Application-control tasks use Jev-1.13, DeepSeek-V4.1-Flash, Gemini-3.1-Flash-Lite, GLM-4.7-Flash and local Qwen3.5-4B. Rules, solvers and retrieval baselines remain task-specific. Bold tinted cells mark the best observed value in the direction specified by each caption, including ties. Grey brackets show intervals. N/A marks an unavailable or inapplicable measurement. Appendix A reports native condition matrices. Appendix B gives supporting results by lens.

7.1. L1 Measurement Boundary Where a decision is timed changes the benefit of a faster engine. The audit finds explicit loop claims in 50/139 families, and 4/139 support the claim with measurement matched to its stated budget and boundary (§5). A claim checked at the decision return can differ from the same claim checked at the acknowledgement or verified outcome. With the action held fixed, verified outcome time follows decision delay approximately one for one, and both slopes lie within the practical band [0.9, 1.1]. The mean scene slope of Tb against D is 1.00025 in transport, with a 95% scene interval of 0.99970–1.00080. The edge slope is 0.99096, with an interval of 0.95906–1.02186. Removing the twosecond delay saves 2.00070 s in transport (1.99954–2.00181) and 2.00924 s in edge (1.92801–2.08679). The share that the decision occupies depends on the endpoint. A 0.5 s decision takes 91.9% of the transport acknowledgement time Ta and 47.2% of the verified outcome time Tb . In edge, the same delay takes 98.9% of Ta and 22.8% of Tb (Table 14). The average per-scene relative saving from the two-second arm to zero delay is 78.1% in transport and 55.8% in edge. The RAN study [3] shows the same separation at the radio boundary, and the RAN values in this paragraph come from it. Median interpretation takes 0.286–2.352 s, against 15.5–19.1 ms for A1 and 2.4–3.1 ms for E2, so the decision dominates the acknowledged control path. At the radio outcome, ideal enforcement moves the affected-class SLA violation by 3.96 percentage points, and no hosted model shows a resolved violation increase over Jev at the base point. In that base-point experiment, interpreters that miss the near-RT budget at the acknowledgement show no resolved penalty at the service outcome. Figure 4 separates decision, acknowledgement and observed or verified outcomes. Its RAN panel contains all 210 arrivals of the RAN study, and its stacks use the 203 records with both acknowledgement and KPM observations. Six signed KPMminus-ACK gaps are negative. Verdict. Repeated effect, multi-task evidence and SoKowned results with borrowed-and-reanalysed context.

7.2. L2 Workload Summary A latency summary taken without load can admit an engine that the loaded loop rejects. Only 9/139 families report p95 or higher, and 4/139 report deadline attainment

-J

If

w

m

Q

Se

-F ny

w

A

Q

S

LM G

Je

v

0

Implementation (a) RAN observed KPM

Outcome / KPM

2

1

0

0

.1

.5

Injected D (s)

2

Transport

Edge

3

Median Tb (s)

2

Acknowledgement

Stage median (s)

Stage median (s)

4

D

Stage median (s)

Decision

2 1 0

(b) Transport physical path

0

.1

.5

Injected D (s)

3 2 1

2

0.0

(c) Edge physical path

0.5

1.0

Injected D (s)

2.0

(d) Sparse verified outcome

Figure 4. Stage medians and sparse verified-outcome medians. Panel (a) retains 30/30/28/29/27/29/30 endpoint pairs from 30 RAN arrivals per interpreter [3]. Transport and edge use 30 scenes with ten repeats per delay. Stacks sum stage medians. Outcome intervals resample scenes. KPM is an observation bound.

7.3. L3 Check Placement An engine that owns a check can pass it under one condition and fail it under another. The check columns of the systematization support no prevalence estimate for observation reconciliation, coverage escalation or joint feasibility in the controller (§4). Engine-owned observation reconciliation depends on the conflict type. Jev gives the correct placement action in 50/120 current-conflict cases and 11/120 resolved-conflict cases. The latter condition requires precedence between

ρ0 = 0.8 Edge

ρ0 = 0.5 Transport

ρ0 = 0.95

10

3

10

2

10

1

10

0

P(Tb ≤ B) (%)

100

Tb p95 (s)

(§5). Because the queue field has low coding agreement, the audit gives no reliable queue prevalence. The load arm, which queues real decisions ahead of measured execution residuals, shows the verdict change directly. A 0.5 s decision meets the 10 s transport budget for all 300 isolated requests. At 4.992 arrivals/s, where the 100 ms baseline fills half of the decision slot, the same delay raises ρ to 2.500. Attainment then falls to 0.0% and the median verified outcome reaches 233.749 s (Table 14). In edge at ρ0 = 0.95, a 100 ms decision cuts 2 s attainment from 76.0% in isolation to 18.6% (9.3–29.0). Table 7 contrasts medians with tails and deadlines. It keeps task-specific endpoints and marks each task family’s ownership. At 16 service arrivals/s, Jev has ρ = 1.084, queue p95 1.492 s and native completion 90.7% [2]. At 2 RAN intents/s, Jev has ρ = 0.149 and queue p95 2.13 ms. Qwen 3.8 reaches ρ = 1.085 and queue p95 20.30 s. AnyJev reaches ρ = 1.169 and queue p95 34.53 s. DeepSeek has ρ = 0.403 and queue p95 6.37 s [3]. At ρ0 = 0.95, the transport and edge replays share one decision queue, and its p95 is 2.568 s for the 100 ms arm. The 500 ms arm reaches ρ = 4.749 and queue p95 970.063 s. The 2 s arm reaches ρ = 18.977 and queue p95 4,623.773 s. These overloaded queues accumulate work during the arrival window and drain after arrivals stop. Their intervals describe this finite nonstationary window. Figure 5 shows how the same arrivals yield different outcome tails and deadline attainment when slot occupancy changes. Verdict. Task-dependent effect, multi-task evidence and SoK-owned plus borrowed-and-reanalysed data.

50

0 0

.5

Injected D (s)

2

(a) Outcome tail

0

.5

Injected D (s)

2

(b) Budget attainment

Figure 5. Finite-window queue amplification at fixed arrival rates. Colour identifies baseline load ρ0 . Line and marker identify domain. Outcome tails use measured execution residuals after real decision queues. Budget attainment uses all arrivals with B = 10 s for transport and 2 s for edge.

newer evidence and an older contradiction. Figure 11 and Table 18 in Appendix B give the details. Coverage checks are also task-dependent. Jev requests new contract candidates in 45/48 absent-candidate cases. It escalates only 4/120 infeasible forwarding instances. Figure 6 compares the two coverage checks. The public-gate condition is a bundled workflow redesign. Instructions, refresh actions and check ownership change together. Jev’s independent-quota correctness rises from 29/120 to 84/120. The paired gain is 45.8 pp with a 95% interval of 37.5–54.2. With shared quotas, the gain is 23.3 pp with an interval of 15.8–30.8. Shared quotas plus fault-domain separation give 32.5 pp with an interval of 25.0–40.0. The per-stage rule fails when shared quotas require joint accounting. Figure 7 reports the bundled contrasts. Verdict. Task-dependent effect, multi-task evidence and SoK-owned results with borrowed-and-reanalysed context.

7.4. L4 Robustness Scope Robustness scope bounds where the verdicts of the other three lenses hold. The audit did not code robustness to

TABLE 7. F ORMAT- VALID RETURN BY BUDGET WITH TASK - FAMILY OWNERSHIP. I N THE ACCESS - POLICY BASE CONDITION , MODEL ROWS SUMMARIZE THE SAME 81 BASE CASES FROM THE SECOND COLLECTION BATCH , SPREAD OVER ALL FOUR REQUEST FAMILIES , WHILE RULE ROWS SUMMARIZE ALL 120 BASE CASES . R ETURN QUANTILES ARE IN SECONDS . T HE P 99 IS DESCRIPTIVE . P ROBABILITY INTERVALS USE 30 S BLOCKS . A N ASTERISK MARKS FEWER THAN EIGHT OCCUPIED BLOCKS . H IGH - LOAD ρ USES 16 SERVICE ARRIVALS / S AND 2 RAN INTENTS / S . E DGE AND RAN VALUES ARE BORROWED AND REANALYSED [2], [3]. Return (s) Model Ownership n p50 / p95 / p99 Access policy, SoK-owned base condition Jev-1.13 SoK-owned 81 0.350 / 0.483 / 0.620 DeepSeek-V4.1-Flash SoK-owned 81 0.469 / 1.623 / 6.123 Gemini-3.1-Flash-Lite SoK-owned 81 0.923 / 1.144 / 1.371 GLM-4.7-Flash SoK-owned 81 0.812 / 1.504 / 7.866 Qwen3.5-4B SoK-owned 81 0.329 / 0.388 / 0.435 Full rule SoK-owned 120 0.000 / 0.000 / 0.000 Cheapest SoK-owned 120 0.000 / 0.000 / 0.000 First SoK-owned 120 0.000 / 0.000 / 0.000 Service contract, reanalysed from the Edge study [2] Jev-1.13.0 From A 300 0.274 / 0.381 / 0.536 SemIf-Qwen3.5-4B From A 300 0.124 / 0.296 / 0.323 Laya From A 300 0.055 / 0.222 / 0.242 DeepSeek-V4.1-Flash From A 300 0.378 / 1.054 / 4.933 GLM-5.3-Flash From A 300 0.646 / 2.685 / 3.363 Qwen3.8-Flash From A 300 1.774 / 3.668 / 5.774 Qwen3.5-4B-JSON From A 300 0.774 / 0.868 / 1.036 Rule From A 300 0.000 / 0.000 / 0.000 0.000 / 0.000 / 0.000 Oracle From A 300 RAN policy at 57 fresh cells, reanalysed from the RAN study [3] Jev-1.13.0 From B 300 0.275 / 0.376 / 0.429 SemIf-Qwen3.5-4B From B 300 0.431 / 0.524 / 0.556 AnyJev-L0 From B 300 0.403 / 0.422 / 0.428 DeepSeek-V4.1-Flash From B 300 0.520 / 0.812 / 0.984 GLM-5.3-Flash From B 300 1.128 / 2.111 / 4.195 Qwen3.8-Flash From B 300 2.220 / 2.945 / 4.124 Qwen3.5-4B-JSON From B 300 0.429 / 0.439 / 0.443

SemIf GLM 4.7 Task rules

DeepSeek DistilBERT frozen First entry

100

50

50

0

0 Seen

Condition

(a) Service churn 50%

0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 100.0 [N/A] 100.0 [N/A] 100.0 [N/A]

100.0 [100.0,100.0] 88.9 [81.2,95.4] 74.1 [64.9,82.9] 72.8 [62.3,82.7] 100.0 [100.0,100.0] 100.0 [N/A] 100.0 [N/A] 100.0 [N/A]

N/A N/A N/A N/A N/A N/A N/A N/A

0.0 [0.0,0.0] 0.3 [0.0,0.5]* 83.3 [N/A] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 100.0 [N/A] 100.0 [N/A]

100.0 [100.0,100.0] 100.0 [100.0,100.0]* 100.0 [N/A] 94.7 [88.8,98.7] 81.0 [71.3,89.4] 0.0 [0.0,0.0] 98.7 [97.2,99.7] 100.0 [N/A] 100.0 [N/A]

1.084 10.170 1.037 1.842 2.972 5.876 15.823 N/A 0.002

0.0 [0.0,0.0]* 0.0 [0.0,0.0]* 0.0 [0.0,0.0]* 0.0 [0.0,0.0]* 0.0 [0.0,0.0]* 0.0 [0.0,0.0] 0.0 [0.0,0.0]*

99.7 [99.0,100.0]* 100.0 [100.0,100.0]* 100.0 [100.0,100.0]* 99.0 [97.7,100.0]* 24.0 [18.5,28.8]* 0.0 [0.0,0.0] 100.0 [100.0,100.0]*

0.149 0.261 1.169 0.403 0.350 1.085 0.217

GLM 5.3 DistilBERT retrained

Qwen 3.8 MiniLM

Rekey

Condition

50

50

0

New ×2

Present

Absent

Contract

Condition

(b) Contract catalogue

Qwen-JSON Full rule

100

0 Stable

Unseen

High-load ρ

100

Correct (%)

Correct (%)

100

Correct (%)

Laya Qwen 3.5 Link-only

≤1s (%) [CI]

Correct (%)

Jev Gemini 3.1 Lexical rule

≤ 100 ms (%) [CI]

Route

Condition

(c) Route coverage

(d) Escalation

Figure 6. Covered selection and absent-coverage escalation. Panel (a) reports seen and unseen service top-1 with n = 136/134 from the Edge study [2]. Contract candidates use n = 48. Paths use n = 120. Bars retain native predicates with 95% Wilson intervals.

−50 J D G L Q F P U 1

Implementation

(a) Independent quota

G: Gemini 3.1 U: Ungated rule

50 0 −50 J D G L Q F P U 1

Implementation

(b) Shared quota

L: GLM 4.7 1: First entry

Q: Qwen 3.5

50

Δ correct (pp)

0

D: DeepSeek P: Per-stage rule

Δ correct (pp)

50

Δ correct (pp)

Δ correct (pp)

J: Jev F: Full rule

0 −50 J D G L Q F P U 1

Implementation

(c) Quota and separation

50 0 −50 J D G L Q F P U 1

Implementation

(d) Shared minus independent

Figure 7. Public-gate minus self-check correctness in panels (a)–(c). Panel (d) shows shared minus independent quota under public gating. Intervals resample 120 paired cases within worker-count and demand strata. The intervention jointly changes instructions, refresh actions and check ownership.

structure, length, representation or catalogue change and therefore provides no prevalence estimate for this lens. Each task grows along its own structural axis. Jev’s lowdensity contract exact matches are 277/300, 236/300 and 173/300 for 4, 6 and 8 fields [2]. Its access-policy counts are 107/120, 97/120 and 47/120 for 4, 16 and 64 requirements. In records from the RAN study [3], Jev selects the correct RAN target in 150 and 145 of 150 cases with 3 and 7 fresh cells, and in 135 and 136 of 150 with 21 and 57 cells. Our reanalysis gives a contrast between 57 and 3 cells of −9.33 pp, with a paired 95% interval of [−14.00, −5.32] and structure-family Holm-adjusted p = 0.00757. Padding and repetition keep the requirement set but change its presentation. Figure 8 shows the four robustness scopes. Its contract and padding components come from the Edge study [2] and its telemetry component from the RAN study [3], while the SoK owns the access-policy arm. Verdict. Task-dependent effect, multi-task evidence and SoK-owned plus borrowed-and-reanalysed data.

8. Design Rules Each rule starts from a gap in the systematization or the evidence audit, adds the consequence measured under its lens, and names a controller action to close it. The boundary and workload rules rest on audit prevalence. The checkplacement rules start from the ownership the published entries name and rest on measured consequences, because the check columns support no prevalence estimate (§4). The robustness rule starts from a field the audit does not code. Budget the Decision at the Required Outcome. Fifty of the 139 families claim that their engine fits a control loop or time budget, and four support the claim with measurement at the stated boundary. The endpoint determines the share of time the decision occupies. A 0.5 s decision takes 91.9% of the transport acknowledgement time and 47.2% of the verified outcome time (§7.1). In the RAN study, interpretation dominates the acknowledged control path, yet the basepoint experiment shows no resolved penalty at the service outcome [3]. State which endpoint the loop budget governs. Measure D/Ta , D/Tb and the offered load before replacing an engine with a faster one, since a sparse verified outcome shortens by about the decision waiting removed. Admit an Engine by Attainment at Offered Load. Nine families report latency at p95 or higher, and four report deadline attainment. An isolated-request summary can pass an engine that the loaded loop rejects. A 0.5 s transport decision meets the 10 s budget for all 300 isolated requests and for none at 4.992 arrivals/s when decisions queue ahead of measured execution times (§7.2). In edge at ρ0 = 0.95, a 100 ms decision lowers attainment of the 2 s budget from 76.0% to 18.6%. Specify the budget at the required endpoint, a decision concurrency limit and an overload policy. Report attainment at the offered load with the occupied decision slot and observation window. Reconcile Observation Versions Before the Decision. The published entries name an observation owner for 11 of

the 139 families (§4.2). Missing, stale, contradictory and superseded observations require different actions. Jev gives the correct placement action in 50/120 current-conflict cases and in 11/120 resolved-conflict cases, where newer evidence supersedes an older contradiction (§7.3). Keep version, precedence and unresolved-conflict state in the controller. Pass the decision a reconciled or explicitly unresolved state. Make Absent Coverage a Controller State. The published entries name a coverage owner for 13 families, and for 2 of the 72 without a computation step. Jev requests a new candidate for 45 of 48 contracts without one and escalates 4 of 120 infeasible forwarding instances (§7.3). The same engine passes one coverage check and fails the other. Implement an explicit no-feasible-candidate outcome in the catalogue service or the path-computation workflow. Test covered selection and absent-coverage escalation separately, including how the controller obtains or certifies alternatives. Compute Joint Feasibility in the Controller. The published entries name a feasibility owner for 45 of the 67 families with a computation step and for 23 of the 72 without one. The public-gate redesign changes instructions, refresh actions and check ownership together. It raises Jev’s independent-quota correctness from 29/120 to 84/120, a paired gain of 45.8 pp (37.5–54.2). The per-stage rule fails when shared quotas require joint accounting. Where the access-policy, placement and forwarding tasks define complete feasibility and cost predicates, the full rule satisfies every measured condition with sub-millisecond computation (Tables 8, 12 and 11). Assign formalized constraints to a scheduler or solver after the engine interprets the intent. For open-vocabulary or underspecified objectives, construct the contract that defines this computation before the decision. Report the Robustness Scope With Each Verdict. The evidence audit does not code the structural scope a verdict covers and gives no count for it. Each task grows along its own structural axis. Jev’s contract exact matches fall from 277/300 to 173/300 between four and eight fields [2]. Its access-policy counts fall from 107/120 to 47/120 between 4 and 64 requirements. Its RAN target selection loses 9.33 pp between 3 and 57 cells [3]. Padding and repetition keep the requirement set but change its presentation (§7.4). Prune or preaggregate inputs that cannot affect the decision. Evaluate output width, observation retrieval and constraint interaction separately, and state each verdict’s tested scope.

9. Limitations The evidence audit measures what the literature reports. A family coded without matched timing evidence may still operate within its claimed loop. The audit therefore establishes missing support for admission claims. The measured consequences in §7 show why that missing support matters. They come from the Edge and RAN source studies and the SoK-owned tasks. The surveyed network systems of other authors remain untested. Each experiment uses one random stream. Its intervals come from resampling within the run, by input, scene or

DeepSeek Full rule

100

50

50

0

0 4

6

Output fields (a) Contract fields

8

GLM 5.3 Field error (a)

Qwen 3.8 Δ accuracy (d)

50

0 3

7

21

Telemetry cells

57

(b) Telemetry cells

Qwen-JSON Δ return (d)

100

100

Correct (%)

Correct (%)

100

Correct (%)

AnyJev-L0 Qwen 3.5

4

16

Requirements

(c) Policy requirements

64

1.0

Δ return (s)

SemIf GLM 4.7

Δ accuracy (pp)

Jev Gemini 3.1

0

−100

0.5

0.0 C:J

C:S

P:J

P:D

Padded − base

(d) Padding effects

Figure 8. Structural growth and paired padding effects. Panels (a) and the contract component of (d) are borrowed and reanalysed from the Edge study [2]. Panel (b) is borrowed and reanalysed from the RAN study [3]. Panel (c) and the policy component of (d) are SoK-owned. Bands in panels (a)–(c) show 95% Wilson intervals. Panel (d) uses paired stratified-bootstrap intervals.

time block, and do not describe variation across independent replications. Because the synthesis questions and comparison families were formulated after the measurement records were inspected, their tests are exploratory. Holm adjustment controls error within each declared family only. The access-policy, placement and forwarding tasks are constructed application-layer tasks with complete predicates. They isolate selector behavior under controlled structure, observation and coverage. Their graph costs do not represent measured network delay. Because the public-gate contrast changes instructions, refresh actions and check ownership together, it estimates the effect of a workflow redesign rather than the effect of moving one check. Because the load arm samples measured physical residuals after a real FIFO decision queue, its post-queue actions do not run concurrently. Near saturation, its intervals describe the finite observation window rather than a steady state. Transport and edge differ in topology, action, stack and verification probe. Comparisons between the two domains are therefore descriptive. Radio closed-loop results are taken from the published summaries of the RAN study [3].

Dependency-Driven Refresh and Cross-Domain Conflict. Reusable interpretation needs explicit dependencies on catalogue, contract and policy state. A cached interpretation can skip a decision call while the request, contract version and relevant catalogue semantics stay unchanged, but dependent network state requires recomputed feasibility. Refreshing everything on each telemetry change repeats unaffected computations. Refreshing only the model can retain invalid downstream assumptions. Dependency tracking must identify what a state change invalidates. For cross-domain intents, arbitration must resolve competing resources, priorities and authority and verify the shared outcome. Replication Across Deployments. The next step tests how decision-slot occupancy and execution overlap interact across sites and implementations. Experiments should vary arrival processes, provider concurrency and concurrent physical actions while matching task responsibilities and recording the complete event path. Independent traffic realizations would separate workload variation from the physical-repeat variation measured in §7.1. These comparisons can determine how deployment changes the relation between faster decisions, queue tails and verified outcomes.

10. Open Problems 11. Conclusion Guard Complementarity. Parsers, configuration verifiers and execution probes check different properties. Evaluating their complementarity requires measuring the errors each guard detects, its false alarms and the violations that survive their combination. Configuration translation and routingstack semantics are part of this comparison. The three gates supply its events and denominators. Candidate Construction and Certification. Typed selection needs a candidate set that contains an acceptable output. How to certify this coverage under changing constraints, and when to synthesize a new candidate instead, remain open. Measuring the full cost of candidate construction and certification is also open. A correct escalation initiates catalogue construction or service recovery, and coverage restoration and verified recovery require separate evaluation.

We systematized 139 paper families on semantic decision engines in network control and coded each by decision interface, execution path and check ownership. Of these, 50 claimed that their engine fit a control loop or time budget, and 4 supported the claim at the stated boundary. Families with a deterministic computation step made 28 of the claims and all four supported ones. The 72 families without such a step made 22, none supported, and only 2 of them named a coverage owner. Bounded tests showed that the endpoint set the decision’s share of elapsed time and that load turned a decision that met every isolated deadline into one that met none. The same engine passed some checks and failed others. We concluded that a loop claim needed to state its endpoint, load and check ownership, and derived a reporting record, design rules and a research agenda for doing so.

References [1]

[2]

O-RAN ALLIANCE, “O-RAN architecture description,” European Telecommunications Standards Institute, ETSI Technical Specification TS 103 982 V14.0.0, Mar. 2026, o-RAN.WG1.TS.OADR004-v14.00. [Online]. Available: https://www.etsi.org/deliver/etsi ts/103900 103999/103982/14.00.00 60/ts 103982v140000p.pdf D. Li, X. Wang, H. Gong, R. Lang, and G. Yu, “Replacing large language models with Jev decision models for low-latency edge service orchestration,” arXiv:2609.22753, Sep. 2026. [Online]. Available: https://arxiv.org/abs/2609.22753

[3]

——, “Intent interpretation at RIC timescales: Jev decision models versus large language models in 6G Open RAN,” arXiv:2609.23136, Oct. 2026. [Online]. Available: https://arxiv.org/abs/2609.23136

[4]

S. D’Oro, M. Polese, L. Kundu, Y. Huang, S. Sangal, V. Goyal, A. Roy, R. Gangula, D. Shyy, A. Damnjanovic, and D. Knisely, “dApps for real-time RAN control: Use cases and requirements,” O-RAN next Generation Research Group, Contributed Research Report RR-2024-10, Oct. 2024. [Online]. Available: https://mediastorage.o-ran.org/ngrg-rr/nGRG-RR-2024-10-dApp% 20use%20cases%20and%20requirements.pdf

[5]

[6]

[7]

O-RAN ALLIANCE, “E2 interface: Service model,” European Telecommunications Standards Institute, ETSI Technical Specification TS 104 040 V4.0.0, Oct. 2024, o-RAN.WG3.E2SMR003-v04.00. [Online]. Available: https://www.etsi.org/deliver/etsi ts/104000 104099/104040/04.00.00 60/ts 104040v040000p.pdf 3GPP, “Management and orchestration; intent driven management services for mobile networks,” European Telecommunications Standards Institute, 3GPP Technical Specification TS 28.312 V19.5.0, Release 19, Apr. 2026. [Online]. Available: https://www.etsi.org/deliver/etsi ts/128300 128399/128312/ 19.05.00 60/ts 128312v190500p.pdf ——, “Architecture enhancements for 5G system (5GS) to support network data analytics services,” European Telecommunications Standards Institute, 3GPP Technical Specification TS 23.288 V19.8.0, Release 19, Oct. 2026. [Online]. Available: https://www.etsi.org/deliver/etsi ts/123200 123299/123288/19.08.00 60/ts 123288v190800p.pdf

[14]

C. Li, O. Havel, A. Olariu, P. Martinez-Julia, J. Nobre, and D. Lopez, “Intent classification,” RFC Editor, RFC 9316, Oct. 2022, iRTF NMRG, Informational. [Online]. Available: https://www.rfc-editor.org/info/rfc9316

[15]

S. Geng, H. Cooper, M. Moskal, S. Jenkins, J. Berman, N. Ranchin, R. West, E. Horvitz, and H. Nori, “JSONSchemaBench: A rigorous benchmark of structured outputs for language models,” arXiv preprint arXiv:2501.10868, 2025. [Online]. Available: https://arxiv.org/abs/2501.10868

[16]

L. Beurer-Kellner, M. Fischer, and M. Vechev, “Guiding LLMs the right way: Fast, non-invasive constrained generation,” arXiv preprint arXiv:2403.06988, 2024. [Online]. Available: https: //arxiv.org/abs/2403.06988

[17]

L. Li, Y. Dong, G. Wang, Z. Xu, A. Jiang, and T. Chen, “XGrammar2: Dynamic and efficient structured generation engine for agentic LLMs,” in Proceedings of the ACM Conference on AI and Agentic Systems. San Jose CA USA: ACM, May 2026, pp. 1009–1022.

[18]

Y. Wei, X. Xie, T. Hu, Y. Zuo, X. Chen, K. Chi, and Y. Cui, “INTA: Intent-based translation for network configuration with LLM agents,” in 2025 IEEE 33rd International Conference on Network Protocols (ICNP). Seoul, Korea, Republic of: IEEE, Sep. 2025, pp. 1–16.

[19]

Y. Miyaoka, M. Inoue, K. Urata, and S. Harada, “Chatdriven optimal management for virtual network services,” arXiv preprint arXiv:2512.24614, 2025. [Online]. Available: https: //arxiv.org/abs/2512.24614

[20]

Z. Wang, A. Cornacchia, A. Sacco, F. Galante, M. Canini, and D. Jiang, “A network arena for benchmarking AI agents on network troubleshooting,” arXiv preprint arXiv:2512.16381, 2025. [Online]. Available: https://arxiv.org/abs/2512.16381

[21]

J. Cohen, “A coefficient of agreement for nominal scales,” Educational and Psychological Measurement, vol. 20, no. 1, pp. 37–46, 1960.

[22]

K. L. Gwet, “Computing inter-rater reliability and its variance in the presence of high agreement,” British Journal of Mathematical and Statistical Psychology, vol. 61, no. 1, pp. 29–48, 2008.

[8]

——, “Management and orchestration; management data analytics (MDA),” European Telecommunications Standards Institute, 3GPP Technical Specification TS 28.104 V19.3.0, Release 19, Oct. 2025. [Online]. Available: https://www.etsi.org/deliver/etsi ts/128100 128199/128104/19.03.00 60/ts 128104v190300p.pdf

[23]

D. M. Manias, A. Chouman, and A. Shami, “Towards intent-based network management: Large language models for intent extraction in 5G core networks,” in 2024 20th International Conference on the Design of Reliable Communication Networks (DRCN). Montreal, QC, Canada: IEEE, May 2024, pp. 1–6.

[9]

ETSI, “Zero-touch network and service management (ZSM); reference architecture,” European Telecommunications Standards Institute, ETSI Group Specification GS ZSM 002 V1.1.1, Aug. 2019. [Online]. Available: https://www.etsi.org/deliver/etsi gs/ZSM/ 001 099/002/01.01.01 60/gs zsm002v010101p.pdf

[24]

S. Nam, N. Van Tu, and J. W.-K. Hong, “LLM-based AI agent for virtual network function deployment,” Journal of Network and Systems Management, vol. 34, no. 4, Oct. 2026, Art. no. 111.

[25]

C. Wang, M. Scazzariello, A. Farshin, S. Ferlin, D. Kostić, and M. Chiesa, “NetConfEval: Can LLMs facilitate network configuration?” Proceedings of the ACM on Networking, vol. 2, no. CoNEXT2, pp. 7:1–7:25, Jun. 2024, Art. no. 7.

[26]

A. El-Hassany, P. Tsankov, L. Vanbever, and M. Vechev, “NetComplete: Practical network-wide configuration synthesis with autocompletion,” in 15th USENIX Symposium on Networked Systems Design and Implementation (NSDI 18). Renton, WA: USENIX Association, Apr. 2018, pp. 579–594. [Online]. Available: https://www.usenix.org/conference/nsdi18/presentation/el-hassany

[27]

A. Khurshid, X. Zou, W. Zhou, M. Caesar, and P. B. Godfrey, “VeriFlow: Verifying network-wide invariants in real time,” in 10th USENIX Symposium on Networked Systems Design and Implementation (NSDI 13). Lombard, IL: USENIX Association, Apr. 2013, pp. 15–27. [Online]. Available: https://www.usenix.org/ conference/nsdi13/technical-sessions/presentation/khurshid

[28]

M. K. Hossain and W. Aljoby, “NetIntent: Leveraging large language models for end-to-end intent-based SDN automation,” IEEE Open Journal of the Communications Society, vol. 6, pp. 10 512–10 541, 2025.

[10]

[11]

——, “Zero-touch network and service management (ZSM); closed-loop automation; part 1: Enablers,” European Telecommunications Standards Institute, ETSI Group Specification GS ZSM 009-1 V1.1.1, Jun. 2021. [Online]. Available: https://www.etsi.org/deliver/etsi gs/ZSM/001 099/00901/01. 01.01 60/gs ZSM00901v010101p.pdf TM Forum, “TMF921 intent management API,” Open API specification, version 5.0.0, 2024, aPI version released October 2024; accessed October 3, 2026. [Online]. Available: https://www.tmforum.org/open-digital-architecture/open-apis/ intent-management-api-tmf921/v5.0

[12]

——, “Intent in autonomous networks,” TM Forum, Introductory Guide IG1253 V1.3.0, Aug. 2022. [Online]. Available: https://www.tmforum.org/resources/how-to-guide/ ig1253-intent-in-autonomous-networks-v1-3-0/

[13]

A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura, “Intentbased networking - concepts and definitions,” RFC Editor, RFC 9315, Oct. 2022, iRTF NMRG, Informational. [Online]. Available: https://www.rfc-editor.org/info/rfc9315

[29]

M. Natu and A. Sethi, “Active probing approach for fault localization in computer networks,” in 2006 4th IEEE/IFIP Workshop on Endto-End Monitoring Techniques and Services. Vancouver, Canada: IEEE, 2006, pp. 25–33.

[44]

L. Dinh, S. Cherrared, X. Huang, and F. Guillemin, “Towards endto-end network intent management with large language models,” arXiv preprint arXiv:2504.13589, 2025. [Online]. Available: https://arxiv.org/abs/2504.13589

[30]

J. Clark, S. M. R. Pial, Y. Su, and T. Xu, “Can Jev make SRE agents more reliable?” SREGym Blog, Sep. 2026, published Sep. 17, 2026. Accessed: Sep. 20, 2026. [Online]. Available: https://sregym.com/blog/jev-sregym-lite

[45]

K. Dzeparoska, A. Tizghadam, and A. Leon-Garcia, “Intent assurance using LLMs guided by intent drift,” arXiv preprint arXiv:2402.00715, 2024. [Online]. Available: https://arxiv.org/abs/ 2402.00715

[31]

F. Bang, “GPTCache: An open-source semantic cache for LLM applications enabling faster answers and cost savings,” in Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023). Singapore: Association for Computational Linguistics, Dec. 2023, pp. 212–218. [Online]. Available: https://aclanthology.org/2023.nlposs-1.24/

[46]

A. El-Hassany, P. Tsankov, L. Vanbever, and M. Vechev, “Networkwide configuration synthesis,” in Computer Aided Verification, ser. Lecture Notes in Computer Science, R. Majumdar and V. Kunčak, Eds. Cham: Springer International Publishing, 2017, vol. 10427, pp. 261–281. [Online]. Available: https: //doi.org/10.1007/978-3-319-63390-9 14

[32]

D. Li, X. Wang, H. Gong, R. Lang, and G. Yu, “Fast intentdriven service orchestration with Jev for 6G edge networks,” arXiv:2609.23136v1, Sep. 2026. [Online]. Available: https://arxiv. org/abs/2609.23136v1

[47]

K. Islam and R. N. Calheiros, “Intent Engine: Natural-language intent translation for intent-driven orchestration in the compute continuum,” Journal of Systems Architecture, vol. 179, Oct. 2026, Art. no. 103938.

[33]

R. Asadli, B. Hoffman, I. Protogeros, and L. Vanbever, “Evaluating agentic configuration repair for computer networks,” arXiv preprint arXiv:2606.06212, 2026. [Online]. Available: https: //arxiv.org/abs/2606.06212

[48]

[34]

E. B. Wilson, “Probable inference, the law of succession, and statistical inference,” Journal of the American Statistical Association, vol. 22, no. 158, pp. 209–212, 1927.

A. S. Jacobs, R. J. Pfitscher, R. H. Ribeiro, R. A. Ferreira, L. Z. Granville, W. Willinger, and S. G. Rao, “Hey, Lumi! using natural language for intent-based network management,” in 2021 USENIX Annual Technical Conference (USENIX ATC 21). Virtual Event: USENIX Association, Jul. 2021, pp. 625–639. [Online]. Available: https://www.usenix.org/conference/atc21/presentation/jacobs

[49]

S. Jha, R. Arora, Y. Watanabe, T. Yanagawa, Y. Chen, J. Clark, B. Bhavya, M. Verma, H. Kumar, H. Kitahara, N. Zheutlin, S. Takano, D. Pathak, F. George, X. Wu, B. O. Turkkan, G. Vanloo, M. Nidd, T. Dai, O. Chatterjee, P. Gupta, S. Samanta, P. Aggarwal, R. Lee, P. Murali, J.-w. Ahn, D. Kar, A. Rahane, C. Fonseca, A. Paradkar, Y. Deng, P. Moogi, P. Mohapatra, N. Abe, C. Narayanaswami, T. Xu, L. R. Varshney, R. Mahindru, A. Sailer, L. Shwartz, D. Sow, N. C. M. Fuller, and R. Puri, “ITBench: Evaluating AI agents across diverse real-world IT automation tasks,” arXiv preprint arXiv:2502.05352, 2025. [Online]. Available: https://arxiv.org/abs/2502.05352

[50]

E. Li and H. Du, “JAUNT: Joint alignment of user intent and network state for QoE-centric LLM tool routing,” arXiv preprint arXiv:2510.18550, 2025. [Online]. Available: https: //arxiv.org/abs/2510.18550

[51]

T. B. Townsend and D. M. Manias, “Advanced LLM-Enhanced Intent-Based 5G Network Management using Dynamic Semantic Routes,” arXiv preprint arXiv:2608.22644v1, 2026. [Online]. Available: https://arxiv.org/abs/2608.22644v1

[35]

Q. McNemar, “Note on the sampling error of the difference between correlated proportions or percentages,” Psychometrika, vol. 12, no. 2, pp. 153–157, 1947.

[36]

S. Holm, “A simple sequentially rejective multiple test procedure,” Scandinavian Journal of Statistics, vol. 6, no. 2, pp. 65–70, 1979. [Online]. Available: https://www.jstor.org/stable/4615733

[37]

Y. Miyaoka, M. Inoue, K. Urata, and S. Harada, “Chat-Driven Interface for Virtual Network Reallocation,” in ICC 2025 - IEEE International Conference on Communications. IEEE, 2025, pp. 1608–1613. [Online]. Available: https://doi.org/10.1109/icc52391. 2025.11160902

[38]

Z. Wang, A. Cornacchia, F. Galante, C. Centofanti, A. Sacco, and D. Jiang, “Towards a Playground to Democratize Experimentation and Benchmarking of AI Agents for Network Troubleshooting,” arXiv preprint arXiv:2507.01997v2, 2025. [Online]. Available: https://arxiv.org/abs/2507.01997v2

[39]

C. Wang, M. Scazzariello, A. Farshin, D. Kostic, and M. Chiesa, “Making Network Configuration Human Friendly,” arXiv preprint arXiv:2309.06342v1, 2023. [Online]. Available: https://arxiv.org/ abs/2309.06342v1

[52]

D. M. Manias, A. Chouman, and A. Shami, “Semantic routing for enhanced performance of LLM-assisted intent-based 5G core network management and orchestration,” arXiv preprint arXiv:2404.15869, 2024. [Online]. Available: https://arxiv.org/abs/ 2404.15869

[40]

A. Angi, A. Sacco, and G. Marchetto, “LLNet: An intent-driven approach to instructing softwarized network devices using a small language model,” IEEE Transactions on Network and Service Management, vol. 22, no. 4, pp. 3403–3418, Aug. 2025.

[53]

J. Mcnamara, D. Camps-Mur, M. Goodarzi, H. Frank, L. ChinchillaRomero, F. Cañellas, A. Fernández-Fernández, and S. Yan, “NLP powered intent based network management for private 5G networks,” IEEE Access, vol. 11, pp. 36 642–36 657, 2023.

[41]

R. Beckett, R. Mahajan, T. Millstein, J. Padhye, and D. Walker, “Don’t mind the gap: Bridging network-wide objectives and devicelevel configurations,” in Proceedings of the 2016 ACM SIGCOMM Conference. New York, NY, USA: ACM, Aug. 2016, pp. 328–341. [Online]. Available: https://doi.org/10.1145/2934872.2934909

[54]

R. Mondal, A. Tang, R. Beckett, T. Millstein, and G. Varghese, “What do LLMs need to synthesize correct router configurations?” in Proceedings of the 22nd ACM Workshop on Hot Topics in Networks. Cambridge MA USA: ACM, Nov. 2023, pp. 189–195.

[55]

[42]

D. Brodimas, A. Birbas, D. Kapolos, and S. Denazis, “Intent-based infrastructure and service orchestration using agentic-AI,” IEEE Open Journal of the Communications Society, vol. 6, pp. 7150– 7168, 2025.

Y. Njah, A. Leivadeas, J. Violos, and M. Falkner, “Toward intentbased network automation for smart environments: A healthcare 4.0 use case,” IEEE Access, vol. 11, pp. 136 565–136 576, 2023.

[56]

S. Ramanathan, Y. Zhang, M. Gawish, Y. Mundada, Z. Wang, S. Yun, E. Lippert, W. Taha, M. Yu, and J. Mirkovic, “Practical intent-driven routing configuration synthesis,” in 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). Boston, MA: USENIX Association, Apr. 2023, pp. 629–644. [Online]. Available: https://www.usenix.org/conference/ nsdi23/presentation/ramanathan

[43]

H. Chen, Y. Miao, L. Chen, H. Sun, H. Xu, L. Liu, G. Zhang, and W. Wang, “Software-defined network assimilation: Bridging the last mile towards centralized network configuration management with NAssim,” in Proceedings of the ACM SIGCOMM 2022 Conference. Amsterdam Netherlands: ACM, Aug. 2022, pp. 281–297.

[57]

T. Schneider, R. Birkner, and L. Vanbever, “Snowcap: Synthesizing network-wide configuration updates,” in Proceedings of the 2021 ACM SIGCOMM 2021 Conference. New York, NY, USA: ACM, Aug. 2021, pp. 33–49. [Online]. Available: https: //doi.org/10.1145/3452296.3472915

[58]

B. Tian, X. Zhang, E. Zhai, H. H. Liu, Q. Ye, C. Wang, X. Wu, Z. Ji, Y. Sang, M. Zhang, D. Yu, C. Tian, H. Zheng, and B. Y. Zhao, “Safely and automatically updating in-network ACL configurations with intent language,” in Proceedings of the ACM Special Interest Group on Data Communication. New York, NY, USA: ACM, Aug. 2019, pp. 214–226. [Online]. Available: https://doi.org/10.1145/3341302.3342088

[72]

G. Borraccini, S. Straullu, A. Giorgetti, R. Ambrosone, E. Virgillito, A. D’Amico, R. D’Ingillo, F. Aquilino, A. Nespola, N. Sambo, F. Cugini, and V. Curri, “Experimental Demonstration of Partially Disaggregated Optical Network Control Using the Physical Layer Digital Twin,” arXiv preprint arXiv:2212.11874v1, 2022. [Online]. Available: https://arxiv.org/abs/2212.11874v1

[73]

K. Mehmood, K. Kralevska, and D. Palma, “Knowledgebased Intent Modeling for Next Generation Cellular Networks,” arXiv preprint arXiv:2302.08544v2, 2023. [Online]. Available: https://arxiv.org/abs/2302.08544v2

[74]

F. Christou and A. Kirstädter, “Grooming Connectivity Intents in IP-Optical Networks Using Directed Acyclic Graphs,” arXiv preprint arXiv:2304.09711v1, 2023. [Online]. Available: https: //arxiv.org/abs/2304.09711v1

[75]

N. Van Tu, J.-H. Yoo, and J. W.-K. Hong, “Towards Intent-based Configuration for Network Function Virtualization using In-context Learning in Large Language Models,” in NOMS 2024-2024 IEEE Network Operations and Management Symposium. IEEE, 2024, pp. 1–8. [Online]. Available: https://doi.org/10.1109/noms59830. 2024.10575237

T. He, A. N. Toosi, N. Akbari, M. T. Islam, and M. A. Cheema, “An Intent-based Framework for Vehicular Edge Computing,” arXiv preprint arXiv:2304.09916v1, 2023. [Online]. Available: https://arxiv.org/abs/2304.09916v1

[76]

H. Zou, Q. Zhao, L. Bariah, M. Bennis, and M. Debbah, “Wireless Multi-Agent Generative AI: From Connected Intelligence to Collective Intelligence,” arXiv preprint arXiv:2307.02757v1, 2023. [Online]. Available: https://arxiv.org/abs/2307.02757v1

[61]

N. Tu, S. Nam, and J. W.-K. Hong, “Intent-based network configuration using large language models,” International Journal of Network Management, vol. 35, no. 1, Jan. 2025, Art. no. e2313.

[77]

[62]

D. Wu, X. Wang, Y. Qiao, Z. Wang, J. Jiang, S. Cui, and F. Wang, “NetLLM: Adapting large language models for networking,” in Proceedings of the ACM SIGCOMM 2024 Conference. Sydney NSW Australia: ACM, Aug. 2024, pp. 661–678.

I. Cinmere, K. Mehmood, K. Kralevska, and T. Mahmoodi, “DirectConflict Resolution in Intent-Driven Autonomous Networks,” arXiv preprint arXiv:2401.08341v1, 2024. [Online]. Available: https://arxiv.org/abs/2401.08341v1

[78]

S. Mostafa, M. S. Elbamby, M. K. Abdel-Aziz, and M. Bennis, “Intent Profiling and Translation Through Emergent Communication,” arXiv preprint arXiv:2402.02768v1, 2024. [Online]. Available: https://arxiv.org/abs/2402.02768v1

[79]

F. Li, H. Lang, J. Zhang, J. Shen, and X. Wang, “PreConfig: A Pretrained Model for Automating Network Configuration,” arXiv preprint arXiv:2403.09369v1, 2024. [Online]. Available: https://arxiv.org/abs/2403.09369v1

[80]

D. Comer and A. Rastegarnia, “OSDF: An Intentbased Software Defined Network Programming Framework,” arXiv preprint arXiv:1807.02205v1, 2018. [Online]. Available: https://arxiv.org/abs/1807.02205v1

C. Muonagor, M. Bensalem, and A. Jukan, “Predictive Intent Maintenance with Intent Drift Detection in Next Generation Network,” arXiv preprint arXiv:2404.15091v1, 2024. [Online]. Available: https://arxiv.org/abs/2404.15091v1

[81]

A. Chaudhari, A. Asthana, A. Kaluskar, D. Gedia, L. Karani, L. Perigo, R. Gandotra, and S. Gangwar, “VIVoNet: Visually-represented, Intent-based, Voice-assisted Networking,” arXiv preprint arXiv:1904.03228v1, 2019. [Online]. Available: https://arxiv.org/abs/1904.03228v1

K. Mehmood, K. Kralevska, and D. Palma, “Knowledge Graph Embedding in Intent-Based Networking,” arXiv preprint arXiv:2405.07850v1, 2024. [Online]. Available: https://arxiv.org/abs/2405.07850v1

[82]

M. A. Habib, P. E. Iturria Rivera, Y. Ozcan, M. Elsayed, M. Bavand, R. Gaigalas, and M. Erol-Kantarci, “LLM-Based Intent Processing and Network Optimization Using Attention-Based Hierarchical Reinforcement Learning,” arXiv preprint arXiv:2406.06059v2, 2024. [Online]. Available: https://arxiv.org/abs/2406.06059v2

[83]

E. Karakaya, O. Ercetin, H. Ozkan, M. Karaca, E. D. Biyar, and A. Palaios, “Online Learning for Autonomous Management of Intent-based 6G Networks,” arXiv preprint arXiv:2407.17767v1, 2024. [Online]. Available: https://arxiv.org/abs/2407.17767v1

[84]

O. G. Lira, O. M. Caicedo, and N. L. S. da Fonseca, “Large Language Models for Zero Touch Network Configuration Management,” arXiv preprint arXiv:2408.13298v1, 2024. [Online]. Available: https://arxiv.org/abs/2408.13298v1

[59]

[60]

[63]

[64]

[65]

[66]

K.-H. Tseng, N. Bogahawatta, Y. Ginige, K. Patel, K. Dakic, and S. Seneviratne, “FaulT-Bench: Towards benchmarking network troubleshooting LLM agents under unreliable user tickets,” arXiv preprint arXiv:2608.27021, Aug. 2026. [Online]. Available: https://arxiv.org/abs/2608.27021

Y. Zhou, J. Ruan, E. S. Wang, S. Fouladi, F. Y. Yan, K. Hsieh, and Z. Liu, “NetArena: Dynamic benchmarks for AI agents in network automation,” arXiv preprint arXiv:2506.03231, 2025. [Online]. Available: https://arxiv.org/abs/2506.03231 L. Zhu, J. Yu, Z. Chen, Y. Wu, Z. Jiang, Y. Xian, Y. Liu, J. Su, S. Zhou, X. Li, H. Liu, X. Liu, D. Zhang, C. Wu, and X. Chen, “OmniPlan: An adaptive framework for timely and near-optimal network planning optimization,” arXiv preprint arXiv:2606.18105, 2026. [Online]. Available: https://arxiv.org/abs/2606.18105

[67]

M. Riftadi, J. Oostenbrink, and F. Kuipers, “GP4P4: Enabling Self-Programming Networks,” arXiv preprint arXiv:1910.00967v1, 2019. [Online]. Available: https://arxiv.org/abs/1910.00967v1

[68]

A. S. Jacobs, R. J. Pfitscher, R. A. Ferreira, and L. Z. Granville, “Refining Network Intents for Self-Driving Networks,” arXiv preprint arXiv:2008.05509v1, 2020. [Online]. Available: https://arxiv.org/abs/2008.05509v1

[69]

M. Bensalem, J. Dizdarević, F. Carpio, and A. Jukan, “The Role of Intent-Based Networking in ICT Supply Chains,” arXiv preprint arXiv:2105.05179v1, 2021. [Online]. Available: https://arxiv.org/abs/2105.05179v1

[70]

K. Mehmood, H. V. K. Mendis, K. Kralevska, and P. E. Heegaard, “Intent-based Network Management and Orchestration for Smart Distribution Grids,” arXiv preprint arXiv:2105.05594v1, 2021. [Online]. Available: https://arxiv.org/abs/2105.05594v1

[85]

X. Jiang, A. Gember-Jacobson, and N. Feamster, “CAIP: Detecting Router Misconfigurations with Context-Aware Iterative Prompting of LLMs,” arXiv preprint arXiv:2411.14283v1, 2024. [Online]. Available: https://arxiv.org/abs/2411.14283v1

[71]

K. Mehmood, D. Palma, and K. Kralevska, “Mission-Critical Public Safety Networking: An Intent-Driven Service Orchestration Perspective,” arXiv preprint arXiv:2205.03932v1, 2022. [Online]. Available: https://arxiv.org/abs/2205.03932v1

[86]

Z. E. A. Kherroubi, M. Prakash, J.-P. Giacalone, and M. Baddeley, “Poster: Could Large Language Models Perform Network Management?” arXiv preprint arXiv:2411.16232v1, 2024. [Online]. Available: https://arxiv.org/abs/2411.16232v1

[87]

Y. Tang, U. C. Srinivasan, B. J. Scott, O. Umealor, D. Kevogo, and W. Guo, “End-to-End Edge AI Service Provisioning Framework in 6G ORAN,” arXiv preprint arXiv:2503.11933v1, 2025. [Online]. Available: https://arxiv.org/abs/2503.11933v1

[102] A. Soliman, A. Refaey, A. Erbad, and A. Mohamed, “IntAgent: NWDAF-Based Intent LLM Agent Towards Advanced Next Generation Networks,” arXiv preprint arXiv:2601.13114v1, 2026. [Online]. Available: https://arxiv.org/abs/2601.13114v1

[88]

S. Mostafa, M. K. Abdel-Aziz, M. S. Elbamby, and M. Bennis, “RAG-Enabled Intent Reasoning for ApplicationNetwork Interaction,” arXiv preprint arXiv:2505.09339v2, 2025. [Online]. Available: https://arxiv.org/abs/2505.09339v2

[103] M. K. Hossain and W. Aljoby, “LEAD-Drift: Real-time and Explainable Intent Drift Detection by Learning a Data-Driven Risk Score,” arXiv preprint arXiv:2602.13672v1, 2026. [Online]. Available: https://arxiv.org/abs/2602.13672v1

[89]

M. Elkael, M. Polese, R. Prasad, S. Maxenti, and T. Melodia, “ALLSTaR: Automated LLM-Driven Scheduler Generation and Testing for Intent-Based RAN,” arXiv preprint arXiv:2505.18389v4, 2025. [Online]. Available: https://arxiv.org/abs/2505.18389v4

[104] ——, “MILD: Multi-Intent Learning and Disambiguation for Proactive Failure Prediction in Intent-based Networking,” arXiv preprint arXiv:2602.14283v1, 2026. [Online]. Available: https: //arxiv.org/abs/2602.14283v1

[90]

S. Lian, J. Tong, J. Zhang, and L. Fu, “Intelligent Channel Allocation for IEEE 802.11be Multi-Link Operation: When MAB Meets LLM,” arXiv preprint arXiv:2506.04594v1, 2025. [Online]. Available: https://arxiv.org/abs/2506.04594v1

[105] W. L. Phone, B. El Boudani, T. Dagiuklas, and S. Ghosh, “Performance Comparison of IBN orchestration using LLM and SLMs,” arXiv preprint arXiv:2603.06647v1, 2026. [Online]. Available: https://arxiv.org/abs/2603.06647v1

[91]

Z. Huang, J. Robin, N. Herbaut, N. Ben Rabah, and B. Le Grand, “Toward an Intent-Based and Ontology-Driven Autonomic Security Response in Security Orchestration Automation and Response,” arXiv preprint arXiv:2507.12061v1, 2025. [Online]. Available: https://arxiv.org/abs/2507.12061v1

[106] F. A. Bimo, C.-K. Lai, Z.-Y. Yang, and R.-G. Cheng, “Contractbased Agentic Intent Framework for Network Slicing in O-RAN,” arXiv preprint arXiv:2603.01663v1, 2026. [Online]. Available: https://arxiv.org/abs/2603.01663v1

[92]

F. A. Bimo, M. A. C. Galdon, C.-K. Lai, R.-G. Cheng, and E. K. P. Chong, “Intent-Based Network for RAN Management with Large Language Models,” arXiv preprint arXiv:2507.14230v2, 2025. [Online]. Available: https://arxiv.org/abs/2507.14230v2

[93]

N. Gupta, D. Das, T. K, U. M. Natarajan, S. Ravindran, K. Sharma, J. Bapat, and D. Das, “A Novel Integrated Architecture for Intent Based Approach and Zero Touch Networks,” arXiv preprint arXiv:2509.21026v1, 2025. [Online]. Available: https://arxiv.org/abs/2509.21026v1

[94]

[95]

[96]

M. Xing, C. Tian, J. Zhang, L. Pan, P. Liu, Z. Yan, and Y. Yue, “Understanding Network Behaviors through Natural Language Question-Answering,” arXiv preprint arXiv:2510.21894v1, 2025. [Online]. Available: https://arxiv.org/abs/2510.21894v1 U. U. Izuazu, M. Bensalem, and A. Jukan, “A Secured IntentBased Networking (sIBN) with Data-Driven Time-Aware Intrusion Detection,” arXiv preprint arXiv:2511.05133v1, 2025. [Online]. Available: https://arxiv.org/abs/2511.05133v1 A. Majlesara, A. Majlesi, A. Mamaghani, A. Shokrani, and B. H. Khalaj, “5G Network Automation Using Local Large Language Models and Retrieval-Augmented Generation,” arXiv preprint arXiv:2511.21084v1, 2025. [Online]. Available: https: //arxiv.org/abs/2511.21084v1

[97]

S. Qin, J. Zeng, H. Guo, X. Li, J. Kang, and Q. Chen, “Efficient Asynchronous Federated Evaluation with Strategy Similarity Awareness for Intent-Based Networking in Industrial Internet of Things,” arXiv preprint arXiv:2512.20627v2, 2025. [Online]. Available: https://arxiv.org/abs/2512.20627v2

[98]

O. G. Lira, O. M. Caicedo, and N. L. S. Da Fonseca, “Network self-configuration based on fine-tuned small language models,” arXiv preprint arXiv:2512.02861, 2025. [Online]. Available: https://arxiv.org/abs/2512.02861

[99]

D. Vijay and V. Ethiraj, “Graph-Symbolic Policy Enforcement and Control (G-SPEC): A Neuro-Symbolic Framework for Safe Agentic AI in 5G Autonomous Networks,” arXiv preprint arXiv:2512.20275v1, 2025. [Online]. Available: https://arxiv.org/ abs/2512.20275v1

[107] M. K. Hossain and W. Aljoby, “AI-driven Intent-Based Networking Approach for Self-configuration of Next Generation Networks,” arXiv preprint arXiv:2603.23772v1, 2026. [Online]. Available: https://arxiv.org/abs/2603.23772v1 [108] A. Twabi, Y. Ding, and T. Kondo, “NetAgentBench: A StateCentric Benchmark for Evaluating Agentic Network Configuration,” arXiv preprint arXiv:2604.09678v1, 2026. [Online]. Available: https://arxiv.org/abs/2604.09678v1 [109] J. Menezes, L. Bitzki, D. Kreutz, G. Almeida, M. Pohlmann, and R. Mansilha, “A Reproducible Semantic Benchmark for Multivendor DSM-to-CLI Translation,” arXiv preprint arXiv:2606.20564v1, 2026. [Online]. Available: https://arxiv.org/abs/2606.20564v1 [110] R.-J. Reifert, A. A. Ahmad, H. Dahrouj, and A. Sezgin, “Agentic AI-Based Joint Computing and Networking via Mixture of Experts and Large Language Models,” arXiv preprint arXiv:2605.02911v1, 2026. [Online]. Available: https://arxiv.org/abs/2605.02911v1 [111] I. Zacarias, M. Grunewald, F. Gentzen, X. Masip-Bruin, and A. Jukan, “Enhancing Secure Intent-Based Networking with an Agentic AI: The EU Project MARE Approach,” arXiv preprint arXiv:2604.06856v1, 2026. [Online]. Available: https: //arxiv.org/abs/2604.06856v1 [112] Y. Li, “Validated Intent Compilation for Constrained Routing in LEO Mega-Constellations,” arXiv preprint arXiv:2604.07264v1, 2026. [Online]. Available: https://arxiv.org/abs/2604.07264v1 [113] I. Protogeros, R. Asadli, B. Hoffman, and L. Vanbever, “Benchmarking LLM-Driven Network Configuration Repair,” arXiv preprint arXiv:2604.22513v1, 2026. [Online]. Available: https://arxiv.org/abs/2604.22513v1 [114] J. Martins, L. Mokrushin, and M. Orlic, “TIO-SHACL: Comprehensive SHACL validation for TMF Intent Ontologies,” arXiv preprint arXiv:2604.27359v1, 2026. [Online]. Available: https://arxiv.org/abs/2604.27359v1 [115] J. Parra-Ullauri, T. A. Khan, D. McHugh, S. Kapoor, A. Duke, A. Hey, and A. Corston-Petrie, “Role-based agentic AI for intent-driven network and service orchestration,” arXiv preprint arXiv:2606.20580, 2026. [Online]. Available: https://arxiv.org/abs/ 2606.20580

[100] G. Jiang, K. Wang, X. Chen, and Y. Huang, “Agentic AI Empowered Intent-Based Networking for 6G,” arXiv preprint arXiv:2601.06640v1, 2026. [Online]. Available: https://arxiv.org/ abs/2601.06640v1

[116] R. Bao, Y. Sun, Z. Chen, F. Yang, M. Tao, N. Li, and W. Zhang, “E 3 -Agent: An Executable and Evolving Agent for Resource Management of Edge Generative Inference,” arXiv preprint arXiv:2605.27428v1, 2026. [Online]. Available: https: //arxiv.org/abs/2605.27428v1

[101] T. Ahmed, Y. Zhu, and S. Choudhury, “Vision Language Models for Optimization-Driven Intent Processing in Autonomous Networks,” arXiv preprint arXiv:2601.12744v1, 2026. [Online]. Available: https://arxiv.org/abs/2601.12744v1

[117] M. K. S. Barbosa and K. L. Dias, “AgentxGCore: Agentic AI for Next-Generation Mobile Core Network,” arXiv preprint arXiv:2606.00417v1, 2026. [Online]. Available: https://arxiv.org/ abs/2606.00417v1

[118] İ. E. Sarıdaş, O. Salan, A. Görçin, I. Hokelek, and H. A. Çırpan, “RAG-driven Multi-Agent LLM Framework with Task Decomposition for Beyond 5G Auto-Configuration,” arXiv preprint arXiv:2606.01222v1, 2026. [Online]. Available: https://arxiv.org/ abs/2606.01222v1 [119] T. Haikal, S. Ismail, and E. Hammad, “Bridging High-Level Intent and Network Execution: Detecting Violations and Intent Drift Through Low-Level Traffic Analysis,” arXiv preprint arXiv:2606.05076v1, 2026. [Online]. Available: https://arxiv.org/ abs/2606.05076v1 [120] J. Armstrong, “Privacy-Preserving Intent Fulfilment and Assurance for 6G RAN,” arXiv preprint arXiv:2607.08809v1, 2026. [Online]. Available: https://arxiv.org/abs/2607.08809v1 [121] K. Tholl, F. Rivest, M. El Mezouar, A. Taylor, and R. Al Mallah, “Distilling Knowledge from Large Language Models into Lightweight Reinforcement Learning Agents for Autonomous Cyber Operations,” arXiv preprint arXiv:2607.28826v1, 2026. [Online]. Available: https://arxiv.org/abs/2607.28826v1 [122] Z. Zhang, V. Aggarwal, and T. Lan, “Learning Not to Optimize: Physics-Informed Action-Space Reshaping for IntentBased Network Control,” arXiv preprint arXiv:2608.00908v1, 2026. [Online]. Available: https://arxiv.org/abs/2608.00908v1 [123] C. Liu, X. Xie, X. Chen, and Y. Cui, “NetConfArena: An Executable Benchmark for LLM Agents in Closed-Loop Network Configuration,” arXiv preprint arXiv:2608.23179v1, 2026. [Online]. Available: https://arxiv.org/abs/2608.23179v1 [124] A. A. Alsamarneh and O. Alhussein, “Uncertainty Signals for Network Intent Translation: Risk Ranking and Ambiguity Localization,” arXiv preprint arXiv:2609.04486v1, 2026. [Online]. Available: https://arxiv.org/abs/2609.04486v1 [125] N. Chakraborty, P. Djukic, and B. Kantarci, “On Identifying Adversarial Intent Injection in AI-Native 6G Networks,” arXiv preprint arXiv:2609.12144v1, 2026. [Online]. Available: https: //arxiv.org/abs/2609.12144v1

[133] M. Shokrnezhad and T. Taleb, “An Autonomous Network Orchestration Framework Integrating Large Language Models with Continual Reinforcement Learning,” arXiv preprint arXiv:2502.16198v1, 2025. [Online]. Available: https://arxiv.org/abs/2502.16198v1 [134] Y. Tang, M. Zou, Z. Nezami, S. A. R. Zaidi, and W. Guo, “KP-A: A Unified Network Knowledge Plane for Catalyzing Agentic Network Intelligence,” arXiv preprint arXiv:2507.08164v1, 2025. [Online]. Available: https://arxiv.org/abs/2507.08164v1 [135] L. Zhang, H. Zhao, B. Xu, H. Zhu, and X. Wang, “Agentic AI for SAGIN Resource Management: Semantic Awareness, Orchestration, and Optimization,” arXiv preprint arXiv:2603.16458v1, 2026. [Online]. Available: https://arxiv.org/abs/2603.16458v1 [136] J. Tong, F. Liu, L. Xv, S. Lu, K. Li, Y. Zhang, Y. Song, Z. Xue, and J. Zhang, “WirelessBench: A Tolerance-Aware LLM Agent Benchmark for Wireless Network Intelligence,” arXiv preprint arXiv:2603.21251v1, 2026. [Online]. Available: https://arxiv.org/abs/2603.21251v1 [137] J. Martins, L. Mokrushin, M. Orlic, and A. K. A, “Intentdriven 6G service orchestration: Grounded translation, validation, and decomposition,” arXiv preprint arXiv:2606.28348, Jun. 2026. [Online]. Available: https://arxiv.org/abs/2606.28348 [138] W. Liang, Y. Tian, C. Chen, and Z. Yu, “MOSS: Endto-End Dialog System Framework with Modular Supervision,” arXiv preprint arXiv:1909.05528v1, 2019. [Online]. Available: https://arxiv.org/abs/1909.05528v1 [139] F. Tang, X. Wang, X. Yuan, L. Luo, M. Zhao, T. Huang, and N. Kato, “MSADM: Large Language Model (LLM) Assisted End-to-End Network Health Management Based on MultiScale Semanticization,” arXiv preprint arXiv:2406.08305v4, 2024. [Online]. Available: https://arxiv.org/abs/2406.08305v4 [140] L. Tulczyjew, K. Jarrah, C. Abondo, D. Bennett, and N. Weill, “LLMcap: Large Language Model for Unsupervised PCAP Failure Detection,” arXiv preprint arXiv:2407.06085v1, 2024. [Online]. Available: https://arxiv.org/abs/2407.06085v1

[126] Y. Zhang, H. Hu, and G. Gu, “NetInspector: Measuring and Improving LLM Capabilities for Reliable Intent-Based Networking Policy Generation,” arXiv preprint arXiv:2609.21103v1, 2026. [Online]. Available: https://arxiv.org/abs/2609.21103v1

[141] T. Tan, F. Tang, L. Luo, X. Wang, Z. Li, and M. Zhao, “Adapting Network Information into Semantics for Generalizable and Plug-and-Play Multi-Scenario Network Diagnosis,” arXiv preprint arXiv:2501.16842v2, 2025. [Online]. Available: https: //arxiv.org/abs/2501.16842v2

[127] B. Wang, C. Wu, H. Zou, Y. Tian, L. Bariah, L. Wei, C. Huang, Y. Shen, Z. Zhang, and M. Debbah, “TelecomGPT-R1: Unified PostTraining for Reasoning Across Heterogeneous Telecom Tasks,” arXiv preprint arXiv:2609.25356v1, 2026. [Online]. Available: https://arxiv.org/abs/2609.25356v1

[142] C. Shi, B. Jalli, G. Macdonald, J. Zou, W. Lei, M. Jain, and J. Philip, “Leveraging Multi-Agent System (MAS) and FineTuned Small Language Models (SLMs) for Automated Telecom Network Troubleshooting,” arXiv preprint arXiv:2511.00651v2, 2025. [Online]. Available: https://arxiv.org/abs/2511.00651v2

[128] B. Ojaghi, R. Vilalta, and R. Muñoz, “From Intents to Algorithms: Verified Algorithm Discovery for Transport Networks,” arXiv preprint arXiv:2609.27386v1, 2026. [Online]. Available: https://arxiv.org/abs/2609.27386v1

[143] J. Tan, F. Bu, Y. Gao, D. Khanolkar, J. Mackay, B. Sobolev, L. Jin, and L. Zhang, “HYVE: Hybrid Views for LLM Context Engineering over Machine Data,” arXiv preprint arXiv:2604.05400v1, 2026. [Online]. Available: https://arxiv.org/abs/2604.05400v1

[129] A. Masini, S. Acharya, P. Bellavista, L. Foschini, and B. Kantarci, “Intent2Tc: Automated Intent-to-Traffic Control Translation with Language Models,” arXiv preprint arXiv:2609.31397v1, 2026. [Online]. Available: https://arxiv.org/abs/2609.31397v1

[144] N. P. Tran, B. Jaumard, K. Premkumar, and S. Memon, “Cross-Domain Query Translation for Network Troubleshooting: A Multi-Agent LLM Framework with Privacy Preservation and SelfReflection,” arXiv preprint arXiv:2604.13353v2, 2026. [Online]. Available: https://arxiv.org/abs/2604.13353v2

[130] T. Szyrkowiec, M. Santuari, M. Chamania, D. Siracusa, A. Autenrieth, V. Lopez, J. Cho, and W. Kellerer, “Automatic Intent-Based Secure Service Creation Through a Multilayer SDN Network Orchestration,” arXiv preprint arXiv:1803.03106v1, 2018. [Online]. Available: https://arxiv.org/abs/1803.03106v1

[145] K.-H. Tseng, N. Bogahawatta, Y. Ginige, K. Dekic, A. Sivanathan, and S. Seneviratne, “SADE: Symptom-Aware Diagnostic Escalation for LLM-Based Network Troubleshooting,” arXiv preprint arXiv:2605.04530v1, 2026. [Online]. Available: https://arxiv.org/abs/2605.04530v1

[131] V. Raisanen, M. Elbamby, and D. Petrov, “Cross-stakeholder service orchestration for B5G through capability provisioning,” arXiv preprint arXiv:2008.07162v1, 2020. [Online]. Available: https://arxiv.org/abs/2008.07162v1

[146] Z. Wu, M. Zhao, F. Tang, and N. Kato, “PropLLM: PropagationAware Scene Reconstruction for Network Fault Diagnosis,” arXiv preprint arXiv:2606.00582v1, 2026. [Online]. Available: https://arxiv.org/abs/2606.00582v1

[132] A. Abdallah, A. Albaseer, A. Celik, M. Abdallah, and A. M. Eltawil, “NetOrchLLM: Mastering Wireless Network Orchestration with Large Language Models,” arXiv preprint arXiv:2412.10107v1, 2024. [Online]. Available: https://arxiv.org/abs/2412.10107v1

[147] A. Mekrache, A. Ksentini, and C. Verikoukis, “Intent-Based Management of Next-Generation Networks: an LLM-Centric Approach,” IEEE Network, vol. 38, no. 5, pp. 29–36, 2024. [Online]. Available: https://doi.org/10.1109/mnet.2024.3420120

[148] N. Zheng, F. Li, Z. Li, Y. Yang, Y. Hao, C. Liu, and X. Wang, “ConfigTrans: Network Configuration Translation Based on Large Language Models and Constraint Solving,” in 2024 IEEE 32nd International Conference on Network Protocols (ICNP), 2024, pp. 1–12. [149] R. Han, J. Wang, H. Sun, Z. Jiang, Q. Qi, Z. Zhuang, Y. Zhang, and J. Liao, “Network CoPilot: Intent-Driven Network Configuration Updating for Service Guarantee,” in IEEE INFOCOM 2025 IEEE Conference on Computer Communications. IEEE, 2025, pp. 1–10. [Online]. Available: https://doi.org/10.1109/infocom55648. 2025.11044495 [150] J. V. A. Garcês, N. R. De Oliveira, J. A. C. Watanabe, R. M. Tanaka, D. O. De Arruda, B. T. Leite, C. P. Galdino, R. S. Couto, I. M. Moraes, D. S. V. de Medeiros, and D. M. F. Mattos, “Intent-Based Management for Open RAN: Intelligent Network Configuration Automation via Chatbot,” in 2024 IEEE 13th International Conference on Cloud Networking (CloudNet). IEEE, 2024, pp. 1–9. [Online]. Available: https: //doi.org/10.1109/cloudnet62863.2024.10815823 [151] K. Aykurt, A. Blenk, and W. Kellerer, “NetLLMBench: A Benchmark Framework for Large Language Models in Network Configuration Tasks,” in 2024 IEEE Conference on Network Function Virtualization and Software Defined Networks (NFV-SDN), 2024, pp. 1–6. [152] C. Wang, X. Zhang, R. Lu, X. Lin, X. Zeng, X. Zhang, Z. An, G. Wu, J. Gao, C. Tian, G. Chen, G. Liu, Y. Liao, T. Lin, D. Cai, and E. Zhai, “Towards LLM-Based Failure Localization in Production-Scale Networks,” in Proceedings of the ACM SIGCOMM 2025 Conference. ACM, 2025, pp. 496–511. [Online]. Available: https://doi.org/10.1145/3718958.3750505 [153] H. Wang, A. Abhashkumar, C. Lin, T. Zhang, X. Gu, N. Ma, C. Wu, S. Liu, W. Zhou, Y. Dong, W. Jiang, and Y. Wang, “NetAssistant: Dialogue Based Network Diagnosis in Data Center Networks,” in 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). USENIX Association, 2024, pp. 2011–2024. [Online]. Available: https: //www.usenix.org/conference/nsdi24/presentation/wang-haopei [154] Z. Wang, S. Lin, G. Yan, S. Ghorbani, M. Yu, J. Zhou, N. Hu, L. Baruah, S. Peters, S. Kamath, J. Yang, and Y. Zhang, “Intent-Driven Network Management with Multi-Agent LLMs: The Confucius Framework,” in Proceedings of the ACM SIGCOMM 2025 Conference. ACM, 2025, pp. 347–362. [Online]. Available: https://doi.org/10.1145/3718958.3750537 [155] Y. Wang, Y. Pang, Y. Liu, Y. Zhang, L. Zhang, M. Zhang, and D. Wang, “Graph Structure-Enhanced Large Language Model for Optical Network Fault Diagnosis: An Explainable Alarm Root Cause Localization Approach,” IEEE Internet of Things Journal, vol. 12, no. 15, pp. 31 493–31 510, 2025. [Online]. Available: https://doi.org/10.1109/jiot.2025.3573056 [156] F. Lotfi, H. Rajoli, and F. Afghah, “LLM-Augmented Deep Reinforcement Learning for Dynamic O-RAN Network Slicing,” in ICC 2025 - IEEE International Conference on Communications, 2025, pp. 3827–3832. [157] M. Asif, T. A. Khan, and W.-C. Song, “R-IBN: A reinforcement learning-based intent-driven framework for end-to-end service orchestration and optimization,” Computer Networks, vol. 270, Oct. 2025, Art. no. 111564. [158] J. Liu, L. Chen, D. Li, and Y. Miao, “CEGS: Configuration Example Generalizing Synthesizer,” in 22nd USENIX Symposium on Networked Systems Design and Implementation (NSDI 25). USENIX Association, 2025, pp. 1327–1347. [Online]. Available: https://www.usenix.org/conference/nsdi25/presentation/liu-jianmin [159] K. An, F. Yang, J. Lu, L. Li, Z. Ren, H. Huang, L. Wang, P. Zhao, Y. Kang, H. Ding, Q. Lin, S. Rajmohan, D. Zhang, and Q. Zhang, “Nissist: An Incident Mitigation Copilot based on Troubleshooting Guides,” arXiv preprint arXiv:2402.17531v2, 2024. [Online]. Available: https://arxiv.org/abs/2402.17531v2

[160] P. Hamadanian, B. Arzani, S. Fouladi, S. K. R. Kakarla, R. Fonseca, D. Billor, A. Cheema, E. Nkposong, and R. Chandra, “A Holistic View of AI-driven Network Incident Management,” in Proceedings of the 22nd ACM Workshop on Hot Topics in Networks. ACM, 2023, pp. 180–188. [Online]. Available: https://doi.org/10.1145/3626111.3628176 [161] P. R. B. Houssel, P. Singh, S. Layeghy, and M. Portmann, “Towards Explainable Network Intrusion Detection using Large Language Models,” arXiv preprint arXiv:2408.04342v1, 2024. [Online]. Available: https://arxiv.org/abs/2408.04342v1 [162] H. Rezaei, R. Taheri, and M. Shojafar, “FedLLMGuard: A federated large language model for anomaly detection in 5G networks,” Computer Networks, vol. 269, 2025, Art. no. 111473. [Online]. Available: https://doi.org/10.1016/j.comnet.2025.111473 [163] S. D’Oro, L. Bonati, M. Polese, and T. Melodia, “OrchestRAN: Network automation through orchestrated intelligence in the open RAN,” arXiv preprint arXiv:2201.05632, 2022. [Online]. Available: https://arxiv.org/abs/2201.05632 [164] M. Bensalem, J. Dizdarević, and A. Jukan, “Benchmarking Various ML Solutions in Complex Intent-Based Network Management Systems,” arXiv preprint arXiv:2111.07724v1, 2021. [Online]. Available: https://arxiv.org/abs/2111.07724v1 [165] A. Bekri, A. Abane, A. Battou, and S. Bensalem, “Bridging Language Models and Formal Methods for Intent-Driven Optical Network Design,” arXiv preprint arXiv:2509.22834v1, 2025. [Online]. Available: https://arxiv.org/abs/2509.22834v1 [166] R. Prasad, M. Polese, and T. Melodia, “BLINC: ContextSpecific Causal Learning for Automated RAN Configuration,” arXiv preprint arXiv:2604.27084v1, 2026. [Online]. Available: https://arxiv.org/abs/2604.27084v1 [167] S. Kampakis, F. Rovai, M. Charalambides, T. Mourouzis, and C. Hicks, “Proving the Utility of Large Language Models in Cybersecurity Simulations: A Comprehensive Examination,” arXiv preprint arXiv:2608.16422v1, 2026. [Online]. Available: https://arxiv.org/abs/2608.16422v1 [168] S. Ghosh, I. Mavromatis, K. Antonakoglou, and K. Katsaros, “Performance Evaluation of Intent-Based Networking Scenarios: A GitOps and Nephio Approach,” arXiv preprint arXiv:2509.13901v1, 2025. [Online]. Available: https://arxiv.org/abs/2509.13901v1

Appendix A. Complete Native Conditions The following matrices report each task’s native conditions, baselines and denominators. They supplement the lens-based analysis in §7. Bold tinted cells mark the largest displayed correct fraction in each condition and column, including ties. TABLE 8. P OLICY INPUT STRUCTURE . C HANGED /P RESERVED EACH CONTAIN 60 CASES , DISTINGUISHED BY WHETHER ALL 64 REQUIREMENTS CHANGE THE MINIMUM - COST FEASIBLE CHOICE . C ONDITIONS RETAIN A FEASIBLE CANDIDATE AND PAIR THE SAME CASES AND CANDIDATES . L ONG IRRELEVANT CONTEXT AND REPETITION RETAIN FOUR DISTINCT REQUIREMENTS . T OKENS GIVES OBSERVED RANGES AT MATCHED CHARACTER BUDGETS . B EST CORRECTNESS PER CONDITION IS IN BOLD . Correct decisions ↑ Method Changed Preserved Short input: 4 requirements, 6k characters Jev-1.13 52/60 55/60 DeepSeek-V4.1-Flash 38/60 43/60 Gemini-3.1-Flash-Lite 39/60 47/60 GLM-4.7-Flash 43/60 39/60 Qwen3.5-4B 13/60 9/60 Full rule 60/60 60/60 Cheapest 48/60 45/60 First entry 20/60 17/60 Long irrelevant context: 4 requirements, 24k characters Jev-1.13 52/60 56/60 DeepSeek-V4.1-Flash 38/60 37/60 Gemini-3.1-Flash-Lite 38/60 50/60 GLM-4.7-Flash 41/60 37/60 Qwen3.5-4B 9/60 10/60 Full rule 60/60 60/60 Cheapest 48/60 45/60 First entry 20/60 17/60 Repeated requirements: 4 requirements, 24k characters Jev-1.13 55/60 56/60 DeepSeek-V4.1-Flash 38/60 39/60 Gemini-3.1-Flash-Lite 40/60 48/60 GLM-4.7-Flash 43/60 33/60 Qwen3.5-4B 16/60 16/60 Full rule 60/60 60/60 Cheapest 48/60 45/60 First entry 20/60 17/60 16 distinct requirements: 16 requirements, 24k characters Jev-1.13 43/60 54/60 DeepSeek-V4.1-Flash 24/60 39/60 Gemini-3.1-Flash-Lite 29/60 47/60 GLM-4.7-Flash 26/60 37/60 Qwen3.5-4B 2/60 3/60 Full rule 60/60 60/60 Cheapest 28/60 45/60 First entry 18/60 17/60 64 distinct requirements: 64 requirements, 24k characters Jev-1.13 11/60 36/60 DeepSeek-V4.1-Flash 11/60 39/60 Gemini-3.1-Flash-Lite 32/60 47/60 GLM-4.7-Flash 3/60 42/60 Qwen3.5-4B 0/60 2/60 Full rule 60/60 60/60 Cheapest 0/60 45/60 First entry 13/60 17/60

All

Valid ↑ All

Response (s) ↓ Median

107/120 81/120 86/120 82/120 22/120 120/120 93/120 37/120

120/120 119/120 120/120 120/120 120/120 120/120 120/120 120/120

108/120 75/120 88/120 78/120 19/120 120/120 93/120 37/120

p95

Input Tokens

0.382 0.558 0.987 0.883 0.328 < .001 < .001 < .001

0.646 1.441 1.252 4.645 0.389 < .001 < .001 < .001

1699–2640 1262–1630 1396–2321 1174–1559 1322–2261 N/A N/A N/A

120/120 119/120 120/120 120/120 120/120 120/120 120/120 120/120

0.468 0.608 1.167 1.034 0.568 < .001 < .001 < .001

0.988 1.046 1.470 2.964 0.648 < .001 < .001 < .001

4469–5410 4264–4633 4397–5323 3944–4331 4092–5032 N/A N/A N/A

111/120 77/120 88/120 76/120 32/120 120/120 93/120 37/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.431 0.630 1.161 1.171 0.984 < .001 < .001 < .001

0.687 1.446 1.462 2.055 1.048 < .001 < .001 < .001

9186–9816 8628–9440 8815–9444 6928–7161 8801–9436 N/A N/A N/A

97/120 63/120 76/120 63/120 5/120 120/120 73/120 35/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.415 0.602 1.125 1.013 0.576 < .001 < .001 < .001

0.673 1.205 1.393 1.929 0.644 < .001 < .001 < .001

4553–5486 4329–4685 4476–5393 3992–4372 4176–5105 N/A N/A N/A

47/120 50/120 79/120 45/120 2/120 120/120 45/120 30/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.406 0.634 1.140 1.033 0.601 < .001 < .001 < .001

0.758 1.601 1.393 2.793 0.688 < .001 < .001 < .001

4880–5818 4580–4902 4782–5705 4173–4558 4501–5440 N/A N/A N/A

Appendix B. Supporting Results by Lens The supporting material follows the lens order. For L1, the tables give the borrowed radio context and the complete SoK-owned endpoint results. The full deadline curves and native load traces behind the L2 synthesis in §7.2 follow. Observation-state detail and the declared coverage and workflow comparison families support L3, and the L4 tables give the native structural conditions and paired length effects.

TABLE 9. O BSERVATION CONDITIONS . E ACH NODE - COUNT GROUP HAS 40 CASES ACROSS FOUR DEMAND FAMILIES , BALANCED BY SHARED - QUOTA BINDING . I NPUTS CONTAIN FOUR CANDIDATES AND 9,000 CHARACTERS . C ORRECT REQUIRES MINIMUM - COST FEASIBLE SELECTION OR JUSTIFIED REFRESH . T HE COMPLETE / SHARED CONDITION IS THE SAME MEASUREMENT REUSED IN TABLE 12. R ESPONSE TIME ENDS AT DECISION RETURN . B EST CORRECTNESS PER CONDITION IS IN BOLD . Method Complete / shared quota Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Ungated rule Equivalent prose Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Ungated rule Missing required report Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Ungated rule Stale required report Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Ungated rule Conflicting current reports Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Ungated rule Older conflict resolved Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Ungated rule

3 nodes

Correct decisions ↑ 5 nodes 8 nodes

All

Valid ↑ All

Response (s) ↓ Median

p95

10/40 10/40 17/40 0/40 14/40 40/40 40/40

2/40 4/40 10/40 1/40 5/40 40/40 40/40

2/40 0/40 9/40 0/40 1/40 40/40 40/40

14/120 14/120 36/120 1/120 20/120 120/120 120/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.364 0.499 1.027 0.884 0.247 < .001 < .001

0.619 0.754 1.584 8.373 0.352 < .001 < .001

10/40 8/40 14/40 0/40 15/40 40/40 40/40

2/40 3/40 12/40 0/40 8/40 40/40 40/40

1/40 0/40 11/40 1/40 3/40 40/40 40/40

13/120 11/120 37/120 1/120 26/120 120/120 120/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.355 0.477 0.986 0.783 0.247 < .001 < .001

0.585 0.859 1.236 6.672 0.340 < .001 < .001

40/40 40/40 40/40 39/40 0/40 40/40 0/40

40/40 40/40 40/40 40/40 0/40 40/40 0/40

40/40 40/40 40/40 40/40 0/40 40/40 0/40

120/120 120/120 120/120 119/120 0/120 120/120 0/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.353 0.477 0.962 0.776 0.250 < .001 < .001

0.583 0.738 1.257 3.059 0.331 < .001 < .001

39/40 40/40 40/40 40/40 1/40 40/40 0/40

40/40 40/40 40/40 40/40 2/40 40/40 0/40

40/40 40/40 40/40 40/40 6/40 40/40 0/40

119/120 120/120 120/120 120/120 9/120 120/120 0/120

119/120 120/120 120/120 120/120 120/120 120/120 120/120

0.353 0.494 0.963 0.855 0.250 < .001 < .001

0.594 0.811 1.249 18.322 0.329 < .001 < .001

10/40 40/40 5/40 40/40 0/40 40/40 0/40

20/40 40/40 9/40 40/40 0/40 40/40 0/40

20/40 40/40 17/40 40/40 2/40 40/40 0/40

50/120 120/120 31/120 120/120 2/120 120/120 0/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.355 0.499 0.986 0.812 0.248 < .001 < .001

0.606 0.897 1.397 12.777 0.330 < .001 < .001

9/40 1/40 14/40 0/40 2/40 40/40 40/40

1/40 0/40 10/40 1/40 0/40 40/40 40/40

1/40 1/40 7/40 0/40 0/40 40/40 40/40

11/120 2/120 31/120 1/120 2/120 120/120 120/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.374 0.507 1.015 0.861 0.249 < .001 < .001

0.590 0.813 1.276 6.439 0.334 < .001 < .001

TABLE 10. C ONTRACT INTERPRETATION ON 48 CONFIGURATIONS PER CONDITION . R EPLACEMENT INTRODUCES TWO INCOMING CANDIDATES . A BSENT REQUIRES NEW CANDIDATES . VALID AND MEDIAN RESPONSE REFER ONLY TO S TABLE . RULE IS A LEXICAL CONTRACT MATCHER . M INI LM RE - RANKS DESCRIPTIONS . B EST CORRECTNESS PER CONDITION IS IN BOLD . Method Stable Rekeyed A. Contract interpretation: 48 configurations per condition Jev-1.13 48/48 48/48 DeepSeek-V4.1-Flash 47/48 43/48 48/48 48/48 Gemini-3.1-Flash-Lite GLM-4.7-Flash 27/48 19/48 Qwen3.5-4B 26/48 26/48 Rule 24/48 24/48 MiniLM 18/48 18/48

Correct decisions ↑ Replacement

Absent

Valid

Stable only Median (s)

48/48 37/48 48/48 1/48 31/48 24/48 13/48

45/48 3/48 14/48 16/48 1/48 48/48 24/48

48/48 48/48 48/48 48/48 48/48 48/48 48/48

0.323 0.447 0.893 0.775 0.243 < 0.001 0.041

TABLE 11. H ELD - OUT ROUTE CATALOGUES , 20 NEW GRAPHS PER CONTRACT AND 120 PER CONDITION , WITH EIGHT CANDIDATES . L INK - ONLY CHOOSES THE CHEAPEST CONNECTED LISTED PATH WITHOUT THE OTHER CONTRACT CHECKS . C ORRECT REQUIRES MINIMUM - COST FEASIBLE SELECTION OR JUSTIFIED CATALOGUE REBUILDING , WHICH LOADS A PRECONSTRUCTED CERTIFIED CATALOGUE . A LL FIXED RESPONSES ARE VALID . R ESPONSE TIME ENDS AT DECISION RETURN . B EST CORRECTNESS PER CONDITION IS IN BOLD . Method Connectivity A. Feasible candidate present Jev-1.13 17/20 DeepSeek-V4.1-Flash 14/20 Gemini-3.1-Flash-Lite 15/20 GLM-4.7-Flash 1/20 Qwen3.5-4B 5/20 Full rule 20/20 Link-only 20/20 First entry 3/20 B. No feasible candidate listed Jev-1.13 0/20 DeepSeek-V4.1-Flash 3/20 Gemini-3.1-Flash-Lite 1/20 GLM-4.7-Flash 18/20 Qwen3.5-4B 0/20 Full rule 20/20 Link-only 20/20 First entry 0/20

Jev Qwen-JSON

Budget

Ordered

Overall

Response (s) ↓ Median p95

15/20 16/20 16/20 1/20 4/20 20/20 6/20 1/20

16/20 13/20 19/20 1/20 6/20 20/20 16/20 1/20

19/20 17/20 16/20 5/20 7/20 20/20 18/20 3/20

18/20 6/20 18/20 0/20 3/20 20/20 20/20 1/20

10/20 12/20 13/20 1/20 6/20 20/20 1/20 3/20

95/120 78/120 97/120 9/120 31/120 120/120 81/120 12/120

0.351 0.525 0.954 0.872 0.357 < .001 < .001 < .001

0.484 0.722 1.208 6.666 0.447 < .001 < .001 < .001

1/20 4/20 3/20 16/20 0/20 20/20 0/20 0/20

0/20 12/20 16/20 19/20 2/20 20/20 0/20 0/20

0/20 13/20 7/20 19/20 0/20 20/20 0/20 0/20

1/20 18/20 9/20 20/20 0/20 20/20 0/20 0/20

2/20 17/20 13/20 17/20 0/20 20/20 0/20 0/20

4/120 67/120 49/120 109/120 2/120 120/120 20/120 0/120

0.350 0.512 0.950 0.853 0.363 < .001 < .001 < .001

0.507 0.744 1.322 5.507 0.451 < .001 < .001 < .001

SemIf Gemini 3.1

AnyJev-L0 GLM 4.7

0

50

.1

1

Budget B (s)

Qwen 3.8 RAN (d)

50

0 .01

GLM 5.3 Service (d)

ρ = λE[S]/c

50

DeepSeek Qwen 3.5

100

Valid return (%)

100

Valid return (%)

Valid return (%)

100

Correct decisions ↑ Avoid Hops

Required

0

10

.01

(a) Access policy

.1

1

10

Budget B (s)

.01

(b) Service contracts

.1

1

10 10

−1

10

−2

10

Budget B (s)

0

.1

(c) RAN telemetry

1

10

Arrivals/s

(d) Decision-slot demand

Figure 9. Task-specific format-valid return within budget and shared decision-slot demand. Panels (b)–(d) reuse measurements from the Edge and RAN studies [2], [3]. Whiskers at 100 ms and 1 s are 30 s time-block intervals.

50

0 5

10

Requests/s

(a) Service completion

15

Laya Qwen-JSON

AnyJev-L0 Service (d)

DeepSeek RAN (d)

100

0

10

−1

10

−2

.1

.5

Intents/s

(b) RAN slot load

1

2

10

1

10

0

10

−1

10

−2

Native ≤ B (%)

10

SemIf Qwen 3.8

Queue p95 (s)

100

ρ = λE[S]/4

Correct on time (%)

Jev GLM 5.3

50

0 .1

.5

Intents/s

(c) RAN queue tail

1

2

10

−2

10

−1

10

0

Offered ρ

(d) Native attainment

Figure 10. Native load traces borrowed and reanalysed from the Edge and RAN studies [2], [3]. Service admission reports correct on-time modeled service. RAN control reports a scheduled installable policy.

TABLE 12. O BSERVATION GATES AND COUPLED QUOTAS . E ACH NODE - COUNT GROUP HAS 40 CASES ACROSS FOUR DEMAND FAMILIES , BALANCED BY SHARED - QUOTA BINDING . C ORRECT REQUIRES MINIMUM - COST FEASIBLE SELECTION OR JUSTIFIED REFRESH . P ER - STAGE RULE CHECKS CAPACITIES SEPARATELY, OMITTING COMBINED DEMAND ON A SHARED WORKER . C OMPLETE / SHARED REUSES TABLE 9. P UBLIC GATING PRESERVES STATE BYTES AND CANDIDATES WHILE CHANGING INSTRUCTIONS AND REMOVING REFRESH . Q UOTA CONDITIONS ARE PAIRED WITHIN CASE . Method 3 nodes Independent quota Jev-1.13 17/40 DeepSeek-V4.1-Flash 12/40 Gemini-3.1-Flash-Lite 18/40 GLM-4.7-Flash 0/40 Qwen3.5-4B 17/40 Full rule 40/40 Per-stage rule 40/40 Complete / shared quota Jev-1.13 10/40 DeepSeek-V4.1-Flash 10/40 Gemini-3.1-Flash-Lite 17/40 GLM-4.7-Flash 0/40 Qwen3.5-4B 14/40 Full rule 40/40 Per-stage rule 20/40 Shared quota + domain separation Jev-1.13 6/40 DeepSeek-V4.1-Flash 3/40 Gemini-3.1-Flash-Lite 14/40 GLM-4.7-Flash 0/40 Qwen3.5-4B 2/40 Full rule 40/40 Per-stage rule 40/40 Public gate / independent quota Jev-1.13 29/40 DeepSeek-V4.1-Flash 22/40 Gemini-3.1-Flash-Lite 27/40 GLM-4.7-Flash 26/40 Qwen3.5-4B 21/40 Full rule 40/40 Per-stage rule 40/40 Public gate / shared quota Jev-1.13 16/40 DeepSeek-V4.1-Flash 15/40 Gemini-3.1-Flash-Lite 17/40 GLM-4.7-Flash 17/40 Qwen3.5-4B 15/40 Full rule 40/40 Per-stage rule 20/40 Public gate / domain separation Jev-1.13 17/40 DeepSeek-V4.1-Flash 9/40 Gemini-3.1-Flash-Lite 16/40 GLM-4.7-Flash 9/40 Qwen3.5-4B 3/40 Full rule 40/40 Per-stage rule 40/40

Correct decisions ↑ 5 nodes 8 nodes

All

Valid ↑ All

Response (s) ↓ Median

p95

7/40 5/40 18/40 1/40 10/40 40/40 40/40

5/40 2/40 4/40 0/40 3/40 40/40 40/40

29/120 19/120 40/120 1/120 30/120 120/120 120/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.358 0.502 0.994 0.813 0.246 < .001 < .001

0.642 0.861 1.326 9.581 0.329 < .001 < .001

2/40 4/40 10/40 1/40 5/40 40/40 20/40

2/40 0/40 9/40 0/40 1/40 40/40 20/40

14/120 14/120 36/120 1/120 20/120 120/120 60/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.364 0.499 1.027 0.884 0.247 < .001 < .001

0.619 0.754 1.584 8.373 0.352 < .001 < .001

0/40 1/40 14/40 0/40 1/40 40/40 40/40

0/40 0/40 14/40 0/40 1/40 40/40 40/40

6/120 4/120 42/120 0/120 4/120 120/120 120/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.354 0.499 0.989 0.823 0.248 < .001 < .001

0.627 0.790 1.220 2.839 0.340 < .001 < .001

25/40 21/40 22/40 21/40 18/40 40/40 40/40

30/40 17/40 20/40 24/40 13/40 40/40 40/40

84/120 60/120 69/120 71/120 52/120 120/120 120/120

120/120 120/120 120/120 119/120 120/120 120/120 120/120

0.358 0.509 1.009 0.843 0.311 < .001 < .001

0.604 1.175 1.241 3.553 0.411 < .001 < .001

13/40 10/40 13/40 16/40 13/40 40/40 20/40

13/40 8/40 10/40 17/40 11/40 40/40 20/40

42/120 33/120 40/120 50/120 39/120 120/120 60/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.353 0.501 1.009 0.840 0.319 < .001 < .001

0.715 1.087 1.234 8.439 0.389 < .001 < .001

11/40 6/40 18/40 12/40 6/40 40/40 40/40

17/40 2/40 27/40 13/40 5/40 40/40 40/40

45/120 17/120 61/120 34/120 14/120 120/120 120/120

120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.375 0.506 1.021 0.808 0.322 < .001 < .001

0.585 0.999 1.276 3.075 0.401 < .001 < .001

TABLE 13. R ADIO - LOOP SIMULATION BORROWED FROM THE RAN STUDY [3]. L ATENCY- ONLY CONTRASTS USE 0.3 INTENTS / S , 30 KM / H AND FIVE UE S PER CELL . VALUES ARE CHALLENGER MINUS ANCHOR SLA VIOLATION WITH PAIRED TIME - BLOCK INTERVALS AND H OLM ADJUSTMENT ACROSS EIGHT TESTS . Challenger - anchor Endpoint Radio base point: latency-only arms, event-triggered mode DeepSeek - Jev Affected class DeepSeek - Jev Network-wide GLM 5.3 - Jev Affected class GLM 5.3 - Jev Network-wide Qwen 3.8 - Jev Affected class Qwen 3.8 - Jev Network-wide Qwen JSON - SemIf Affected class Qwen JSON - SemIf Network-wide

∆ violation pp [95% CI]

pHolm

Contrast direction

0.03 [-1.45,1.55] -0.66 [-2.14,0.46] -2.45 [-5.13,-0.63] -2.66 [-4.94,-1.03] -0.98 [-2.21,0.23] -0.66 [-2.44,0.50] 1.46 [-0.31,3.97] 1.12 [0.46,1.97]

1 1 0.0072 0.0007999 0.581 1 0.6232 0.0007999

Unresolved Unresolved Lower violation Lower violation Unresolved Unresolved Unresolved Higher violation

TABLE 14. F IXED - ACTION TIMING AND LOAD . C ONDITIONAL P 50/ P 95 ARE IN SECONDS . T HE TERM nv /n GIVES VERIFIED OUTCOMES OVER ALL ARRIVALS . ATTAINMENT USES ALL ARRIVALS WITH B = 10 S FOR TRANSPORT AND 2 S FOR EDGE . L OAD OUTCOMES USE MEASURED RESIDUALS AFTER REAL FIFO QUEUES . D (s) nv /n Ta p50 / p95 Tb p50 / p95 Transport–sparse physical execution 0 300/300 0.044 / 0.051 0.561 / 0.586 0.1 300/300 0.144 / 0.151 0.659 / 0.687 0.5 300/300 0.544 / 0.551 1.060 / 1.086 2 300/300 2.045 / 2.051 2.561 / 2.587 Transport–queue + execution replay, ρ0 = 0.5, λ = 4.992 /s 0 1016/1016 0.044 / 0.051 0.561 / 0.588 0.1 1016/1016 0.154 / 0.348 0.683 / 0.869 0.5 1016/1016 233.246 / 372.255 233.749 / 372.767 2 1016/1016 1405.524 / 2229.573 1406.028 / 2230.085 Transport–queue + execution replay, ρ0 = 0.8, λ = 7.987 /s 0 1576/1576 0.044 / 0.051 0.561 / 0.586 0.1 1576/1576 0.287 / 0.838 0.802 / 1.355 0.5 1576/1576 508.040 / 773.329 508.566 / 773.853 2 1576/1576 2509.895 / 3838.564 2510.421 / 3839.087 Transport–queue + execution replay, ρ0 = 0.95, λ = 9.484 /s 0 1929/1929 0.045 / 0.051 0.561 / 0.587 0.1 1929/1929 1.480 / 2.711 2.002 / 3.231 0.5 1929/1929 627.838 / 970.613 628.367 / 971.128 2 1929/1929 2981.335 / 4625.823 2981.864 / 4626.337 Edge–sparse physical execution 0 300/300 0.005 / 0.006 1.379 / 2.777 0.1 300/300 0.105 / 0.106 1.602 / 2.739 0.5 300/300 0.506 / 0.506 2.189 / 3.001 2 300/300 2.006 / 2.007 3.394 / 4.758 Edge–queue + execution replay, ρ0 = 0.5, λ = 4.992 /s 0 1016/1016 0.005 / 0.006 1.382 / 2.734 0.1 1016/1016 0.115 / 0.307 1.555 / 2.867 0.5 1016/1016 233.210 / 372.216 235.246 / 374.051 2 1016/1016 1405.488 / 2229.533 1407.557 / 2230.965 Edge–queue + execution replay, ρ0 = 0.8, λ = 7.987 /s 0 1576/1576 0.005 / 0.006 1.382 / 2.798 0.1 1576/1576 0.248 / 0.795 1.774 / 3.063 0.5 1576/1576 507.999 / 773.286 509.794 / 774.647 2 1576/1576 2509.854 / 3838.521 2512.032 / 3839.745 Edge–queue + execution replay, ρ0 = 0.95, λ = 9.484 /s 0 1929/1929 0.005 / 0.006 1.382 / 2.812 0.1 1929/1929 1.441 / 2.674 2.954 / 4.655 0.5 1929/1929 627.802 / 970.569 628.943 / 972.304 2 1929/1929 2981.299 / 4625.779 2982.439 / 4627.480

D/Ta

D/Tb

P (Tb ≤ B)

ρ

0.1 69.3 91.9 97.8

0.0 15.2 47.2 78.1

100.0 [100.0, 100.0] 100.0 [100.0, 100.0] 100.0 [100.0, 100.0] 100.0 [100.0, 100.0]

N/A N/A N/A N/A

0.1 65.3 0.2 0.1

0.0 14.7 0.2 0.1

100.0 [100.0, 100.0] 100.0 [100.0, 100.0] 0.0 [0.0, 0.0] 0.0 [0.0, 0.0]

0.000 0.502 2.500 9.987

0.1 35.0 0.1 0.1

0.0 12.6 0.1 0.1

100.0 [100.0, 100.0] 100.0 [100.0, 100.0] 0.0 [0.0, 0.0] 0.0 [0.0, 0.0]

0.000 0.804 3.999 15.981

0.1 6.8 0.1 0.1

0.0 5.0 0.1 0.1

100.0 [100.0, 100.0] 100.0 [100.0, 100.0] 0.0 [0.0, 0.0] 0.0 [0.0, 0.0]

0.000 0.955 4.749 18.977

1.2 94.9 98.9 99.7

0.0 6.2 22.8 58.9

85.0 [78.3, 90.0] 76.0 [66.3, 84.3] 29.7 [17.3, 43.0] 0.0 [0.0, 0.0]

N/A N/A N/A N/A

0.6 87.9 0.2 0.1

0.0 6.5 0.2 0.1

85.2 [83.6, 87.0] 71.9 [69.0, 75.0] 0.0 [0.0, 0.0] 0.0 [0.0, 0.0]

0.000 0.502 2.500 9.987

0.6 40.6 0.1 0.1

0.0 5.7 0.1 0.1

84.6 [82.6, 86.6] 61.2 [56.8, 65.6] 0.0 [0.0, 0.0] 0.0 [0.0, 0.0]

0.000 0.804 3.999 15.981

0.5 7.0 0.1 0.1

0.0 3.4 0.1 0.1

84.3 [82.8, 85.9] 18.6 [9.3, 29.0] 0.0 [0.0, 0.0] 0.0 [0.0, 0.0]

0.000 0.955 4.749 18.977

TABLE 15. S ERVICE ADMISSION AT 1 AND 2 ARRIVALS / S , BORROWED AND REANALYSED FROM THE E DGE STUDY [2]. O FFERED ρ ESTIMATES DISPATCHED SERVICE DEMAND . R ETURN AND NATIVE MODELED COMPLETION USE ALL ARRIVALS . Q UEUE QUANTILES CONDITION ON OBSERVED WAITS . Offered ρ Accepted ρ Implementation [CI] Service admission: 1 arrivals/s, B = 2 s. 300 arrivals per row Jev-1.13.0 0.071 [0.070,0.073] 0.071 SemIf-Qwen3.5-4B 0.170 [0.156,0.184] 0.170 Laya 0.028 [0.027,0.029] 0.028 DeepSeek-V4.1-Flash 0.171 [0.094,0.273] 0.165 GLM-5.3-Flash 0.217 [0.202,0.231] 0.217 Qwen3.8-Flash 0.338 [0.326,0.350] 0.338 Qwen3.5-4B-JSON 0.739 [0.506,0.984] 0.643 Fixed 0.3 s 0.077 [0.077,0.077] 0.077 Fixed 0.6 s 0.152 [0.152,0.152] 0.152 Fixed 1.2 s 0.302 [0.302,0.302] 0.302 Oracle 0.000 [0.000,0.000] 0.000 Service admission: 2 arrivals/s, B = 2 s. 300 arrivals per row Jev-1.13.0 0.139 [0.133,0.144] 0.138 SemIf-Qwen3.5-4B 1.164 [1.027,1.268] 0.966 Laya 0.049 [0.048,0.050] 0.049 DeepSeek-V4.1-Flash 0.217 [0.199,0.236] 0.217 GLM-5.3-Flash 0.490 [0.384,0.607] 0.489 Qwen3.8-Flash 0.733 [0.719,0.749] 0.727 Qwen3.5-4B-JSON 2.154 [2.114,2.196] 1.019 Fixed 0.3 s 0.154 [0.154,0.154] 0.154 Fixed 0.6 s 0.304 [0.304,0.304] 0.304 Fixed 1.2 s 0.604 [0.604,0.604] 0.603 Oracle 0.000 [0.000,0.000] 0.000

Queue (s) p50 / p95

Return ≤ B (%) [CI]

Native completion (%) [CI]

0.002 / 0.002 0.002 / 0.002 0.002 / 0.002 0.002 / 0.002 0.002 / 0.002 0.002 / 0.002 0.002 / 1.807 0.002 / 0.002 0.002 / 0.002 0.002 / 0.002 0.002 / 0.002

100.0 [100.0,100.0] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 92.3 [83.3,100.0] 93.3 [90.7,96.1] 95.0 [92.6,97.6] 38.0 [15.1,63.7] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 100.0 [100.0,100.0]

94.0 [90.7,96.8] 63.0 [56.6,69.9] 11.7 [7.3,15.7] 90.7 [81.7,97.9] 91.7 [88.7,94.8] 90.0 [87.5,92.7] 19.3 [6.8,33.4] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 100.0 [100.0,100.0]

0.002 / 0.002 1.074 / 1.910 0.002 / 0.002 0.002 / 0.002 0.002 / 0.574 0.002 / 1.439 1.658 / 1.943 0.002 / 0.002 0.002 / 0.002 0.002 / 0.449 0.002 / 0.002

100.0 [100.0,100.0] 11.0 [0.4,29.1] 100.0 [100.0,100.0] 99.3 [98.1,100.0] 85.3 [74.2,95.4] 73.0 [60.9,86.3] 0.7 [0.0,2.0] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 99.0 [97.8,100.0] 100.0 [100.0,100.0]

95.3 [92.4,98.2] 8.7 [0.3,24.5] 11.7 [7.9,16.8] 97.7 [95.8,99.4] 82.3 [70.0,93.4] 66.0 [54.0,79.3] 0.3 [0.0,1.0] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 98.3 [96.5,100.0] 100.0 [100.0,100.0]

TABLE 16. S ERVICE ADMISSION AT 4, 8 AND 16 ARRIVALS / S , BORROWED AND REANALYSED FROM THE E DGE STUDY [2]. C OLUMNS AND UNCERTAINTY FOLLOW TABLE 15. R ETURN AND NATIVE MODELED COMPLETION USE ALL ARRIVALS . O FFERED ρ ESTIMATES DISPATCHED SERVICE DEMAND . Offered ρ Accepted Implementation [CI] ρ Service admission: 4 arrivals/s, B = 2 s. 300 arrivals per row Jev-1.13.0 0.272 [0.268,0.277] 0.272 SemIf-Qwen3.5-4B 2.637 [2.602,2.667] 1.037 Laya 0.102 [0.097,0.106] 0.102 DeepSeek-V4.1-Flash 0.401 [0.399,0.406] 0.401 GLM-5.3-Flash 0.883 [0.806,1.018] 0.871 Qwen3.8-Flash 1.478 [1.396,1.738] 1.019 Qwen3.5-4B-JSON 4.203 [4.052,4.359] 1.050 Fixed 0.3 s 0.307 [0.307,0.307] 0.307 Fixed 0.6 s 0.607 [0.607,0.607] 0.607 Fixed 1.2 s 1.207 [1.207,1.208] 1.018 Oracle 0.001 [0.001,0.001] 0.001 Service admission: 8 arrivals/s, B = 2 s. 300 arrivals per row Jev-1.13.0 0.556 [0.554,0.560] 0.555 SemIf-Qwen3.5-4B 5.179 [5.150,5.303] 1.087 Laya 0.232 [0.217,0.235] 0.231 DeepSeek-V4.1-Flash 1.053 [0.938,1.084] 0.954 GLM-5.3-Flash 1.416 [1.349,1.432] 1.047 Qwen3.8-Flash 2.954 [2.919,2.963] 1.083 Qwen3.5-4B-JSON 8.424 [8.352,8.819] 1.094 Fixed 0.3 s 0.615 [0.615,0.615] 0.615 Fixed 0.6 s 1.215 [1.213,1.215] 1.044 Fixed 1.2 s 2.414 [2.413,2.415] 1.054 Oracle 0.001 [0.001,0.001] 0.001 Service admission: 16 arrivals/s, B = 2 s. 300 arrivals per row Jev-1.13.0 1.084 [N/A] 1.084 SemIf-Qwen3.5-4B 10.170 [N/A] 1.186 Laya 1.037 [N/A] 1.002 DeepSeek-V4.1-Flash 1.842 [N/A] 1.110 GLM-5.3-Flash 2.972 [N/A] 1.119 Qwen3.8-Flash 5.876 [N/A] 1.155 Qwen3.5-4B-JSON 15.823 [N/A] 1.212 Fixed 0.3 s 1.229 [N/A] 1.097 Fixed 0.6 s 2.428 [N/A] 1.116 Fixed 1.2 s 4.829 [N/A] 1.142 Oracle 0.002 [N/A] 0.002

Queue (s) p50 / p95

Return ≤ B (%) [CI]

Native completion (%) [CI]

0.002 / 0.002 1.812 / 1.986 0.002 / 0.002 0.002 / 0.005 0.349 / 1.508 1.649 / 1.942 1.776 / 1.960 0.002 / 0.002 0.002 / 0.234 1.480 / 1.964 0.002 / 0.002

99.7 [99.2,100.0] 1.0 [0.0,2.7] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 84.3 [62.7,90.3] 3.7 [2.7,4.7] 0.7 [0.0,1.8] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 15.3 [11.9,18.6] 100.0 [100.0,100.0]

93.7 [89.8,95.3] 1.0 [0.0,2.7] 11.7 [8.0,15.6] 98.3 [96.6,99.2] 80.7 [57.6,86.7] 2.7 [2.3,3.4] 0.3 [0.0,0.9] 100.0 [100.0,100.0] 100.0 [100.0,100.0] 11.7 [8.6,15.9] 100.0 [100.0,100.0]

0.002 / 0.078 1.914 / 1.981 0.002 / 0.002 0.717 / 1.858 1.782 / 1.977 1.862 / 1.984 1.901 / 1.990 0.002 / 0.125 1.683 / 1.968 1.855 / 1.987 0.002 / 0.002

100.0 [100.0,100.0] 1.0 [0.0,1.2] 100.0 [100.0,100.0] 65.3 [56.8,100.0] 9.3 [1.7,11.2] 1.3 [0.0,1.7] 0.7 [0.0,0.8] 100.0 [100.0,100.0] 21.3 [6.8,24.9] 1.7 [0.0,2.1] 100.0 [100.0,100.0]

94.3 [89.8,95.4] 1.0 [0.0,1.2] 11.7 [10.2,12.0] 60.3 [51.5,96.6] 7.7 [0.0,9.5] 1.3 [0.0,1.7] 0.3 [0.0,0.4] 100.0 [100.0,100.0] 19.0 [5.1,22.4] 1.7 [0.0,2.1] 100.0 [100.0,100.0]

0.603 / 1.492 1.953 / 1.996 0.434 / 1.966 1.839 / 1.982 1.924 / 1.996 1.933 / 1.987 1.936 / 1.994 1.445 / 1.954 1.880 / 1.988 1.905 / 1.983 0.002 / 0.002

98.3 [N/A] 1.0 [N/A] 75.3 [N/A] 10.3 [N/A] 2.3 [N/A] 0.3 [N/A] 0.3 [N/A] 57.3 [N/A] 5.3 [N/A] 1.3 [N/A] 100.0 [N/A]

90.7 [N/A] 1.0 [N/A] 9.3 [N/A] 9.3 [N/A] 2.3 [N/A] 0.3 [N/A] 0.3 [N/A] 50.7 [N/A] 5.0 [N/A] 1.3 [N/A] 100.0 [N/A]

TABLE 17. RAN- POLICY ARRIVAL LOAD BORROWED AND REANALYSED FROM THE RAN STUDY [3]. C OMPLETION IS AN INSTALLABLE POLICY WITHIN MODELED A1/E2 TIMING . R ETURN AND COMPLETION USE ALL ARRIVALS . Q UEUE QUANTILES CONDITION ON OBSERVED WAITS . Offered ρ Implementation [CI] RAN policy: 0.1 arrivals/s, B = 1 s. 300 arrivals per row Jev-1.13.0 0.008 [0.008,0.008] SemIf-Qwen3.5-4B 0.011 [0.007,0.019] AnyJev-L0 0.017 [0.017,0.018] DeepSeek-V4.1-Flash 0.020 [0.018,0.023] GLM-5.3-Flash 0.039 [0.035,0.044] Qwen3.8-Flash 0.059 [0.055,0.063] Qwen3.5-4B-JSON 0.010 [0.010,0.011] RAN policy: 0.5 arrivals/s, B = 1 s. 300 arrivals per row Jev-1.13.0 0.038 [0.037,0.039] SemIf-Qwen3.5-4B 0.040 [0.037,0.042] AnyJev-L0 0.119 [0.105,0.135] DeepSeek-V4.1-Flash 0.095 [0.072,0.134] GLM-5.3-Flash 0.248 [0.187,0.310] Qwen3.8-Flash 0.277 [0.267,0.288] Qwen3.5-4B-JSON 0.052 [0.052,0.052] RAN policy: 1 arrivals/s, B = 1 s. 300 arrivals per row Jev-1.13.0 0.074 [0.073,0.076] SemIf-Qwen3.5-4B 0.095 [0.080,0.113] AnyJev-L0 0.347 [0.291,0.404] DeepSeek-V4.1-Flash 0.221 [0.161,0.298] GLM-5.3-Flash 0.232 [0.206,0.256] Qwen3.8-Flash 0.524 [0.513,0.538] Qwen3.5-4B-JSON 0.106 [0.104,0.107] RAN policy: 2 arrivals/s, B = 1 s. 300 arrivals per row Jev-1.13.0 0.149 [0.146,0.152] SemIf-Qwen3.5-4B 0.261 [0.193,0.318] AnyJev-L0 1.169 [1.134,1.210] DeepSeek-V4.1-Flash 0.403 [0.290,0.599] GLM-5.3-Flash 0.350 [0.334,0.361] Qwen3.8-Flash 1.085 [1.065,1.111] Qwen3.5-4B-JSON 0.217 [0.215,0.220]

Accepted ρ

Queue (s) p50 / p95

Return ≤ B (%) [CI]

Native completion (%) [CI]

0.008 0.012 0.019 0.022 0.042 0.063 0.011

0.001 / 0.002 0.006 / 0.027 0.005 / 0.026 0.001 / 0.002 0.001 / 0.002 0.002 / 0.002 0.006 / 0.028

100.0 [100.0,100.0] 98.7 [95.7,100.0] 92.7 [88.6,96.4] 89.0 [85.1,92.8] 23.3 [17.7,29.5] 0.0 [0.0,0.0] 99.7 [98.9,100.0]

100.0 [100.0,100.0] 98.7 [95.7,100.0] 74.3 [69.0,79.7] 89.0 [85.1,92.8] 22.0 [16.4,28.0] 0.0 [0.0,0.0] 99.7 [98.9,100.0]

0.040 0.042 0.128 0.102 0.265 0.296 0.056

0.002 / 0.002 0.002 / 0.006 0.001 / 0.006 0.002 / 0.002 0.002 / 1.437 0.002 / 0.002 0.002 / 0.006

99.7 [98.9,100.0] 100.0 [100.0,100.0] 63.7 [54.0,74.2] 93.0 [87.3,97.2] 25.3 [14.6,37.2] 0.0 [0.0,0.0] 100.0 [100.0,100.0]

99.7 [98.9,100.0] 100.0 [100.0,100.0] 52.3 [43.5,61.6] 92.7 [86.9,96.9] 24.7 [14.1,36.4] 0.0 [0.0,0.0] 100.0 [100.0,100.0]

0.079 0.102 0.371 0.236 0.248 0.561 0.113

0.002 / 0.002 0.001 / 0.003 0.001 / 1.474 0.002 / 0.914 0.002 / 0.002 0.002 / 1.260 0.001 / 0.003

100.0 [100.0,100.0] 97.3 [93.2,100.0] 36.7 [27.4,46.3] 84.0 [70.2,94.5] 73.0 [61.4,85.8] 0.0 [0.0,0.0] 100.0 [100.0,100.0]

100.0 [100.0,100.0] 97.3 [93.2,100.0] 29.0 [21.1,38.0] 84.0 [70.2,94.5] 70.7 [59.5,83.6] 0.0 [0.0,0.0] 100.0 [100.0,100.0]

0.159 0.279 1.250 0.431 0.374 1.160 0.232

0.002 / 0.002 0.001 / 0.259 25.171 / 34.533 0.002 / 6.366 0.002 / 0.231 15.227 / 20.295 0.001 / 0.002

99.7 [98.8,100.0] 90.0 [81.3,100.0] 0.7 [0.0,1.8] 84.7 [63.2,98.7] 92.7 [90.1,95.7] 0.0 [0.0,0.0] 100.0 [100.0,100.0]

99.7 [98.8,100.0] 88.7 [79.5,100.0] 0.7 [0.0,1.8] 84.7 [63.2,98.7] 91.3 [87.8,95.2] 0.0 [0.0,0.0] 100.0 [100.0,100.0]

AnyJev-L0 GLM 4.7

DeepSeek Qwen 3.5

100

50

50

0

0 Fresh

Stale

Noisy Conflict

Condition

(a) RAN telemetry

GLM 5.3 Full rule

Qwen 3.8

100

Correct (%)

Correct (%)

100

Correct (%)

SemIf Gemini 3.1

100

Correct (%)

Jev Qwen-JSON

50

0 Missing

Condition

Stale

(b) Placement refresh

50

0 Current

Resolved

Condition

(c) Version conflicts

Complete

Condition

Prose

(d) State representation

Figure 11. Observation-state correctness with 95% Wilson intervals. Panel (a) reanalyses 150 RAN cases per condition [3]. Placement uses 120 cases per condition.

TABLE 18. O BSERVATION - STATE CORRECTNESS WITH 95% W ILSON INTERVALS . T HE RAN ROWS ARE BORROWED AND REANALYSED [3]. P LACEMENT ROWS ARE S O K- OWNED . VALID COUNTS AND RETURN P 50/ P 95 USE CONTRADICTORY TELEMETRY OR RESOLVED CONFLICT. RAN policy: fresh / stale / noisy / contradictory, 57 cells Model Fresh Jev-1.13.0 90.7 [84.9,94.4] SemIf-Qwen3.5-4B 14.0 [9.3,20.5] AnyJev-L0 28.0 [21.4,35.7] DeepSeek-V4.1-Flash 96.0 [91.5,98.2] GLM-5.3-Flash 96.7 [92.4,98.6] Qwen3.8-Flash 88.0 [81.8,92.3] Qwen3.5-4B-JSON 28.7 [22.0,36.4] Placement: missing / stale / conflict / resolved conflict Model Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Per-stage rule Ungated First

Missing 100.0 [96.9,100.0] 100.0 [96.9,100.0] 100.0 [96.9,100.0] 99.2 [95.4,99.9] 0.0 [0.0,3.1] 100.0 [96.9,100.0] 100.0 [96.9,100.0] 0.0 [0.0,3.1] 100.0 [96.9,100.0]

Stale 70.0 [62.2,76.8] 5.3 [2.7,10.2] 21.3 [15.5,28.6] 69.3 [61.5,76.2] 95.3 [90.7,97.7] 73.3 [65.7,79.8] 27.3 [20.8,35.0]

Noisy 88.7 [82.6,92.8] 14.7 [9.9,21.2] 24.0 [17.9,31.4] 97.3 [93.3,99.0] 96.7 [92.4,98.6] 90.7 [84.9,94.4] 27.3 [20.8,35.0]

Contradictory 63.3 [55.4,70.6] 16.7 [11.6,23.4] 24.7 [18.5,32.1] 66.7 [58.8,73.7] 85.3 [78.8,90.1] 62.0 [54.0,69.4] 30.0 [23.2,37.8]

Valid / n 150/150 150/150 150/150 150/150 137/150 150/150 150/150

Return (s) ↓ p50 / p95 0.292 / 0.408 0.601 / 0.713 0.492 / 0.504 0.683 / 1.013 1.851 / 2.975 2.445 / 7.214 0.478 / 0.486

Stale 99.2 [95.4,99.9] 100.0 [96.9,100.0] 100.0 [96.9,100.0] 100.0 [96.9,100.0] 7.5 [4.0,13.6] 100.0 [96.9,100.0] 100.0 [96.9,100.0] 0.0 [0.0,3.1] 100.0 [96.9,100.0]

Current conflict 41.7 [33.2,50.6] 100.0 [96.9,100.0] 25.8 [18.8,34.3] 100.0 [96.9,100.0] 1.7 [0.5,5.9] 100.0 [96.9,100.0] 100.0 [96.9,100.0] 0.0 [0.0,3.1] 100.0 [96.9,100.0]

Resolved conflict 9.2 [5.2,15.7] 1.7 [0.5,5.9] 25.8 [18.8,34.3] 0.8 [0.1,4.6] 1.7 [0.5,5.9] 100.0 [96.9,100.0] 50.0 [41.2,58.8] 100.0 [96.9,100.0] 25.8 [18.8,34.3]

Valid / n 120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120

Return (s) ↓ p50 / p95 0.374 / 0.590 0.507 / 0.813 1.015 / 1.276 0.861 / 6.439 0.249 / 0.334 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000

TABLE 19. C ONTRACT- CATALOGUE CORRECTNESS (%) WITH 95% W ILSON INTERVALS . 48 CASES PER CONDITION . T WO NEW CANDIDATES FORM THE REPLACEMENT ARM . VALID COUNTS AND RETURN P 50/ P 95 REFER TO ABSENT COVERAGE . N/A DENOTES AN UNMEASURED CONDITION . T HE RULE BASELINE USES LEXICAL MATCHING . Contract catalogue: stable / rekeyed / two new candidates / absent Model Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Lexical rule MiniLM First

Stable 100.0 [92.6,100.0] 97.9 [89.1,99.6] 100.0 [92.6,100.0] 56.2 [42.3,69.3] 54.2 [40.3,67.4] 50.0 [36.4,63.6] 37.5 [25.2,51.6] 22.9 [13.3,36.5]

Rekeyed 100.0 [92.6,100.0] 89.6 [77.8,95.5] 100.0 [92.6,100.0] 39.6 [27.0,53.7] 54.2 [40.3,67.4] 50.0 [36.4,63.6] 37.5 [25.2,51.6] 29.2 [18.2,43.2]

Two new 100.0 [92.6,100.0] 77.1 [63.5,86.7] 100.0 [92.6,100.0] 2.1 [0.4,10.9] 64.6 [50.4,76.6] 50.0 [36.4,63.6] 27.1 [16.6,41.0] N/A

Absent 93.8 [83.2,97.9] 6.2 [2.1,16.8] 29.2 [18.2,43.2] 33.3 [21.7,47.5] 2.1 [0.4,10.9] 100.0 [92.6,100.0] 50.0 [36.4,63.6] 0.0 [0.0,7.4]

Valid / n 48/48 48/48 48/48 48/48 48/48 48/48 48/48 48/48

Return (s) ↓ p50 / p95 0.318 / 0.511 0.413 / 2.696 0.934 / 1.600 0.825 / 2.943 0.243 / 0.331 0.000 / 0.000 0.041 / 0.046 N/A / N/A

TABLE 20. P UBLIC - GATE MINUS SELF - CHECK CORRECTNESS ( PP ). I NTERVALS ARE PAIRED STRATIFIED - BOOTSTRAP 95% ESTIMATES . pHolm USES THE 45- COMPARISON WORKFLOW FAMILY. VALIDITY AND RETURN TIME DESCRIBE THE PUBLIC - GATE ARM . I NSTRUCTIONS AND REFRESH ACTIONS ALSO CHANGE . B EST CORRECTNESS PER CONDITION IS IN BOLD . Method Self-check Independent quotas Jev-1.13 29/120 DeepSeek-V4.1-Flash 19/120 Gemini-3.1-Flash-Lite 40/120 GLM-4.7-Flash 1/120 Qwen3.5-4B 30/120 Full rule 120/120 Per-stage rule 120/120 Ungated 120/120 First 28/120 Shared quotas Jev-1.13 14/120 DeepSeek-V4.1-Flash 14/120 Gemini-3.1-Flash-Lite 36/120 GLM-4.7-Flash 1/120 Qwen3.5-4B 20/120 Full rule 120/120 Per-stage rule 60/120 Ungated 120/120 First 31/120 Shared quotas + fault-domain separation Jev-1.13 6/120 DeepSeek-V4.1-Flash 4/120 Gemini-3.1-Flash-Lite 42/120 GLM-4.7-Flash 0/120 Qwen3.5-4B 4/120 Full rule 120/120 Per-stage rule 120/120 Ungated 120/120 First 32/120

∆ pp [95% CI]

pHolm

Valid / n

Return p50 / p95 (s)

84/120 60/120 69/120 71/120 52/120 120/120 120/120 120/120 28/120

45.8 [37.5,54.2] 34.2 [25.0,43.3] 24.2 [14.2,34.2] 58.3 [49.2,66.7] 18.3 [10.8,26.7] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0]

−15

2.331 × 10 9.366 × 10−9 0.0008467 1.391 × 10−18 0.003618 1 1 1 1

120/120 120/120 120/120 119/120 120/120 120/120 120/120 120/120 120/120

0.358 / 0.604 0.509 / 1.175 1.009 / 1.241 0.843 / 3.553 0.311 / 0.411 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000

42/120 33/120 40/120 50/120 39/120 120/120 60/120 120/120 31/120

23.3 [15.8,30.8] 15.8 [8.3,23.3] 3.3 [-5.8,12.5] 40.8 [32.5,49.2] 15.8 [8.3,23.3] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0]

2.079 × 10−6 0.004852 1 1.457 × 10−13 0.009012 1 1 1 1

120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.353 / 0.715 0.501 / 1.087 1.009 / 1.234 0.840 / 8.439 0.319 / 0.389 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000

45/120 17/120 61/120 34/120 14/120 120/120 120/120 120/120 32/120

32.5 [25.0,40.0] 10.8 [5.8,15.8] 15.8 [5.8,25.8] 28.3 [20.8,35.8] 8.3 [3.3,13.3] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0]

1.455 × 10−10 0.007324 0.1404 4.54 × 10−9 0.1587 1 1 1 1

120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120

0.375 / 0.585 0.506 / 0.999 1.021 / 1.276 0.808 / 3.075 0.322 / 0.401 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000

Public gate

TABLE 21. S TRUCTURAL GROWTH ON THREE NATIVE AXES . C ONTRACT AND RAN ROWS ARE BORROWED AND REANALYSED [2], [3]. P OLICY ROWS ARE S O K- OWNED . C ORRECTNESS IS IN PERCENT WITH 95% W ILSON INTERVALS . C ONTRACT INPUTS DIFFER , SO THEIR ∆ IS N/A. Service contract: 4 / 6 / 8 fields, low constraint density 8 fields 57.7 [52.0,63.1] 13.3 [9.9,17.6] 1.0 [0.3,2.9] 90.0 [86.1,92.9] 80.3 [75.5,84.4] 82.3 [77.6,86.2] 30.0 [25.1,35.4] 0.0 [0.0,1.3] 100.0 [98.7,100.0]

N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A

∆last−first pp [CI] N/A N/A N/A N/A N/A N/A N/A N/A N/A

Valid / n 300/300 300/300 300/300 300/300 300/300 298/300 300/300 300/300 300/300

Return (s) ↓ p50 / p95 0.272 / 0.371 0.133 / 0.194 0.057 / 0.242 0.460 / 0.591 0.930 / 1.954 2.092 / 3.525 1.507 / 1.789 0.000 / 0.000 0.000 / 0.000

7 cells 96.7 [92.4,98.6] 12.7 [8.3,18.9] 36.7 [29.4,44.6] 98.7 [95.3,99.6] 99.3 [96.3,99.9] 97.3 [93.3,99.0] 47.3 [39.5,55.3]

21 cells 90.0 [84.2,93.8] 13.3 [8.8,19.7] 28.0 [21.4,35.7] 96.7 [92.4,98.6] 98.0 [94.3,99.3] 92.7 [87.3,95.9] 36.7 [29.4,44.6]

57 cells 90.7 [84.9,94.4] 14.0 [9.3,20.5] 28.0 [21.4,35.7] 96.0 [91.5,98.2] 96.7 [92.4,98.6] 88.0 [81.8,92.3] 28.7 [22.0,36.4]

∆last−first pp [CI] -9.3 [-14.0,-5.3] -10.7 [-18.0,-3.3] -31.3 [-40.0,-22.0] -2.0 [-4.7,0.0] -2.7 [-6.0,0.0] -11.3 [-16.7,-6.7] -34.7 [-43.3,-26.7]

Valid / n 150/150 150/150 150/150 150/150 150/150 150/150 150/150

Return (s) ↓ p50 / p95 0.281 / 0.378 0.432 / 0.524 0.404 / 0.422 0.522 / 0.825 1.234 / 2.248 2.213 / 3.053 0.432 / 0.440

16 requirements 80.8 [72.9,86.9] 52.5 [43.6,61.2] 63.3 [54.4,71.4] 52.5 [43.6,61.2] 4.2 [1.8,9.4] 100.0 [96.9,100.0] 60.8 [51.9,69.1] 29.2 [21.8,37.8]

64 requirements 39.2 [30.9,48.1] 41.7 [33.2,50.6] 65.8 [57.0,73.7] 37.5 [29.4,46.4] 1.7 [0.5,5.9] 100.0 [96.9,100.0] 37.5 [29.4,46.4] 25.0 [18.1,33.4]

N/A N/A N/A N/A N/A N/A N/A N/A N/A

∆last−first pp [CI] -50.0 [-59.2,-40.8] -25.8 [-37.5,-14.2] -5.8 [-16.7,5.0] -30.8 [-40.8,-20.8] -16.7 [-23.3,-10.0] 0.0 [0.0,0.0] -40.0 [-49.2,-31.7] -5.8 [-15.0,3.3]

Valid / n 120/120 120/120 120/120 120/120 120/120 120/120 120/120 120/120

Return (s) ↓ p50 / p95 0.406 / 0.758 0.634 / 1.601 1.140 / 1.393 1.033 / 2.793 0.601 / 0.688 0.000 / 0.000 0.000 / 0.000 0.000 / 0.000

Model 4 fields 6 fields Jev-1.13.0 92.3 [88.8,94.8] 78.7 [73.7,82.9] SemIf-Qwen3.5-4B 54.3 [48.7,59.9] 40.0 [34.6,45.6] Laya 25.3 [20.7,30.5] 7.7 [5.2,11.2] DeepSeek-V4.1-Flash 98.3 [96.2,99.3] 95.7 [92.7,97.5] GLM-5.3-Flash 89.7 [85.7,92.6] 96.3 [93.6,97.9] Qwen3.8-Flash 98.3 [96.2,99.3] 90.0 [86.1,92.9] Qwen3.5-4B-JSON 53.0 [47.4,58.6] 48.3 [42.7,54.0] Rule 4.0 [2.3,6.9] 0.3 [0.1,1.9] Oracle 100.0 [98.7,100.0] 100.0 [98.7,100.0] RAN policy: 3 / 7 / 21 / 57 cells, fresh, state-dependent target Model 3 cells Jev-1.13.0 100.0 [97.5,100.0] SemIf-Qwen3.5-4B 24.7 [18.5,32.1] AnyJev-L0 59.3 [51.3,66.9] DeepSeek-V4.1-Flash 98.0 [94.3,99.3] GLM-5.3-Flash 99.3 [96.3,99.9] Qwen3.8-Flash 99.3 [96.3,99.9] Qwen3.5-4B-JSON 63.3 [55.4,70.6] Access policy: 4 / 16 / 64 distinct requirements Model Jev-1.13 DeepSeek-V4.1-Flash Gemini-3.1-Flash-Lite GLM-4.7-Flash Qwen3.5-4B Full rule Cheapest First

4 requirements 89.2 [82.3,93.6] 67.5 [58.7,75.2] 71.7 [63.0,79.0] 68.3 [59.6,76.0] 18.3 [12.4,26.2] 100.0 [96.9,100.0] 77.5 [69.2,84.1] 30.8 [23.3,39.6]

TABLE 22. PAIRED PURE - LENGTH EFFECTS . C ONTRACT EFFECTS ARE BORROWED AND REANALYSED FROM THE E DGE STUDY [2]. P OLICY EFFECTS ARE S O K- OWNED . E XACT M C N EMAR P - VALUES USE THE COMPLETE 89- COMPARISON STRUCTURE FAMILY. T HE CONDITIONS RETAIN THE REQUIREMENT SET. Method Condition n Service-contract padding: 16384 tokens minus base Jev-1.13.0 Padding 300 SemIf-Qwen3.5-4B Padding 300 Laya Padding 300 DeepSeek-V4.1-Flash Padding 300 GLM-5.3-Flash Padding 300 Qwen3.8-Flash Padding 300 Qwen3.5-4B-JSON Padding 300 Rule Padding 300 Oracle Padding 300 Access policy: irrelevant context or repetition minus base Jev-1.13 Irrelevant text 120 Jev-1.13 Repetition 120 DeepSeek-V4.1-Flash Irrelevant text 120 DeepSeek-V4.1-Flash Repetition 120 Gemini-3.1-Flash-Lite Irrelevant text 120 Gemini-3.1-Flash-Lite Repetition 120 GLM-4.7-Flash Irrelevant text 120 GLM-4.7-Flash Repetition 120 Qwen3.5-4B Irrelevant text 120 Qwen3.5-4B Repetition 120 Full rule Irrelevant text 120 Full rule Repetition 120 Cheapest Irrelevant text 120 Cheapest Repetition 120 First Irrelevant text 120 First Repetition 120

∆ correctness pp [CI]

pHolm

Paired ∆ return (s) [CI]

-2.7 [-5.7,0.3] -15.7 [-21.3,-9.7] -9.3 [-13.0,-6.0] -1.7 [-3.3,-0.3] -4.7 [-7.3,-2.0] 0.0 [-2.3,2.0] -30.0 [-35.3,-24.7] 0.0 [0.0,0.0] 0.0 [0.0,0.0]

1 4.31 × 10−5 1.823 × 10−5 1 0.07742 1 2.117 × 10−21 1 1

0.149 [0.135,0.157] 1.034 [1.032,1.036] 0.082 [0.081,0.082] 0.259 [0.249,0.276] 0.883 [0.781,0.962] 1.191 [1.135,1.272] 0.781 [0.774,0.791] 0.002 [0.002,0.003] 0.000 [0.000,0.000]

0.8 [-1.7,3.3] 3.3 [-0.8,8.3] -5.0 [-10.8,0.8] -3.3 [-10.8,4.2] 1.7 [-6.7,9.2] 1.7 [-7.5,10.8] -3.3 [-8.3,0.8] -5.0 [-11.7,1.7] -2.5 [-8.3,3.3] 8.3 [0.8,15.8] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0] 0.0 [0.0,0.0]

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

0.044 [0.022,0.084] 0.048 [0.013,0.084] 0.030 [-0.011,0.096] 0.099 [0.011,0.124] 0.158 [0.118,0.219] 0.172 [0.148,0.221] 0.141 [0.053,0.204] 0.247 [0.136,0.314] 0.239 [0.235,0.243] 0.653 [0.647,0.662] 0.000 [0.000,0.000] 0.000 [0.000,0.000] 0.000 [0.000,0.000] 0.000 [0.000,0.000] 0.000 [0.000,0.000] 0.000 [0.000,0.000]

Appendix C. Three-Axis Literature Map Table 23 maps the 139 families to their decision interfaces, execution paths and check responsibilities. Each row combines the reviewed tasks and versions of a family. Checks are qualified by component, test or formal model. TABLE 23: Three-axis literature map of 139 paper families, ordered by the lowest reference number in each row. S: selection; G: generation; C: deterministic computation; U: interface unspecified. Multiple letters retain multiple interfaces. Check labels distinguish observation state, feasibility or related checks, and candidate coverage. Unlisted check types have unresolved ownership. NE marks rows with all three unresolved. Work [2]

Interface S/G/C

Task and execution path Intent interpretation to edge admission/execution

[18]

G

Retrieved manuals to multivendor command-line output Intent parameters to placement/routing solver

[19], [37]

S/G/C

[20], [38] [23] [24] [25], [39] [26] [28] [30]

G S/G G G C G/C S/G

[32]

S/G/C

[33]

G

Offline configuration repair with feedback

[40]

G/C

Intent JSON to confirmed switch rules

[41] [42] [43] [44] [45] [46]

C S/G S/C G S/G C

Policy compilation to BGP configurations Agent plans to cluster/service deployment Offline device-model and parameter mapping Business intent to RAN/core service configuration Intent fulfillment and drift-triggered repair Solver-based network configuration synthesis

[47] [48] [49] [50] [51], [52] [53]

G G/S/C G S S/G S

[54] [55] [56] [57] [58] [59] [60], [61] [62] [63] [64]

G G C C C G G/C S G G/S/C

State-grounded service-objective construction Confirmed intent to compiler/controller Network-fault diagnosis and mitigation in ITBench Per-request tool selection with latency prediction Static route selection and dynamic parameters Registered workflow/model to slice or network function virtualization (NFV) action Configuration translation/generation with feedback Intent entity extraction; proposed SDN activation BGP compilation, staging and activation Safe update ordering with convergence waits Access-control-list verification, repair and migration Tool-driven diagnosis and proposed remediation Confirmed intent translation to NFV deployment Learned bitrate and cluster-action selection Agent planning/repair with environment feedback Intent preferences to network-planning experts

[65] [66] [67] [68] [69] [70]

C G/C C G/C C U

Policies to SDN paths and conflict advice Confirmed voice intent to OpenFlow rules Behavior-guided search for P4 programs Confirmed intent to NFV service-chain deployment Conceptual access-policy parsing and refinement Conceptual smart-grid intent to network slices

Tool-driven anomaly localization and diagnosis Intent extraction; proposed core placement Deployment-workflow generation and twin testing Task-specific policy, API and routing generation Routing-sketch completion before activation Flow generation, installation and drift repair Agent diagnosis with ranked tests and evidence gates Versioned intent contracts to edge placement

Check responsibility State: Scheduler reads queues/workers; Feasibility: Scheduler: predicted feasibility; Coverage: Scheduler: reject if no eligible node Feasibility: INTA: syntax matching; LLM: semantic equivalence Feasibility: Integer linear program: resource/placement constraints; Coverage: Optimizer: joint request-subset admission NE NE NE Feasibility: LLM diagnosis; task-specific tests/verifiers Feasibility: NetComplete: sketch and routing constraints Feasibility: LLM: conflict detection; controller: resolution State: Planner reads fresh snapshot; review: current failure; Feasibility: Jev review: cause, restoration and durability State: Controller versions; scheduler reads queues; Feasibility: Execution layer: permissions; scheduler: placement Feasibility: Batfish/Config2Spec: final scoring; optional agent feedback Feasibility: Network Manager: installed-intent conflicts; Coverage: Intent Converter: devices/scenarios Feasibility: Propane: permitted paths/preferences NE Feasibility: NAssim: command-line hierarchy and device-model tests NE Feasibility: LLM validator: policy-tree review Feasibility: SyNET: modeled forwarding constraints; Coverage: Solver: solution if one exists in model NE Feasibility: Exploratory classifier: intent contradictions NE NE NE NE Feasibility: Campion: equivalence; Batfish: routing properties NE Feasibility: Aura: switch-emulation policy tests Feasibility: Snowcap: temporal safety over modeled states Feasibility: Jinjing: formal reachability/placement constraints State: Agent tests ticket premises against observations Feasibility: NFV-Intent/NI-MANO: capabilities and resource flavors NE NE Feasibility: Mixed-integer programs: constraints; fast experts: heuristics Feasibility: OSDF: rule-based conflict classification NE Feasibility: Search fitness: supplied examples/rules Feasibility: Intent Deployer: configuration-consistency check Feasibility: Proposed configurator: authorization conflicts Feasibility: Proposed orchestrator: SLA/profile checks

TABLE 23: Three-axis literature map (continued). S: selection; G: generation; C: computation; U: unspecified. Unlisted checks and NE denote unresolved ownership. Work [71] [72]

Interface C C

Task and execution path Controlled intent to simulated public-safety service Optical feasibility to lightpath deployment

[73] [74]

C C

[75]

C

Knowledge-based intent to simulated services Internet Protocol (IP)/optical intent compilation in simulation Constrained vehicular-edge service mapping

[76] [77] [78]

G C S

[79] [80] [81]

G S/C G/S

[82] [83] [84] [85] [86] [87] [88] [89] [90] [91] [92] [93] [94] [95] [96] [97] [98] [99]

G/S/C S S/G G S G/S G G/C S/C S/U S/G G/S G/C S S/G G/S/C S/G G/C

[100] [101]

S/G G/C

[102] [103] [104] [105] [106] [107] [108] [109]

G/C S/C S/C G/S G/C G/S/U G G/S

[110] [111] [112]

S/G S/G G/C/S

[113] [114]

S/G C

[115] [116]

G/C G/C

Distributed transmit-power proposals Modeled RAN tilt-conflict resolution Device-intent symbols to quality-of-service (QoS) slices Configuration generation and vendor translation Time-series monitoring of intent drift Entity/knowledge-graph completion to intent templates Intent validity to hierarchical RAN control Parent-bandit selection of joint service priorities Intent to iterative verified configuration Retrieved configuration to error reports Wireless-mesh action recommendation Intent/model selection to service and xApp creation Retrieved context to structured network intent Lua scheduler construction before slot execution Wi-Fi channel initialization followed by search Proposed intent choice; implemented enforcement History-based transmit-power strategy over O1 Intent translation to learned throughput control Fact graph and generated network queries Malicious-intent classification at ingress 5G request classification to command generation Strategies to predicted deployment approval Iterative configuration generation with verification Language-planned actions through graph admission RAN/core recommendations to slice provisioning Optimization-code generation and SDN demonstration Planned, timed slice-rate changes via tools Risk-based drift alerts and attribution Multi-intent risk alerts and root-cause attribution Agent plans/code to emulated-network deployment Intent contract to rApp policy and xApp control Policy activation; proposed ranked assurance repair Interactive routing commands to converged state Multivendor command-line generation and model judging Expert/weight selection to resource allocation Security intent completion to domain agents Offline intent compilation to certified low-Earth-orbit routing Configuration-context selection and repair Resource Description Framework intent representation/ontology validation Business intent to governed service orders Event-driven parameter updates to task routing

Check responsibility Feasibility: Proposed intent validator: resource/conflict checks Feasibility: Lightpath computation and routing/spectrum assignment: optical and wavelength constraints Feasibility: Network-layer validator: resource/model checks NE Feasibility: Priority-aware installation/location-aware mapping heuristics: node/link constraints NE Feasibility: Bargaining: feasible utility sets NE NE NE Feasibility: KG2E classifier: triple plausibility Feasibility: Prediction and QoS-drift validity rules Feasibility: Learned parent policy: modeled resource conflicts Feasibility: Batfish: behavior checks and repair feedback Feasibility: LLM: syntax/dependency diagnosis NE NE NE NE NE Coverage: Ontology: operable candidate techniques NE Feasibility: Translator: basic prior-intent conflict check NE NE NE Feasibility: Federated evaluators: predicted deployability Feasibility: Assigned verifier; implementation unspecified State: Shape constraints and graph queries reject stale entities; Feasibility: G-SPEC: next-state policy guards Feasibility: Specialists/orchestrator: model consistency checks Feasibility: Gurobi: generated-program constraints Feasibility: Feasibility/session tools: action constraints NE NE Feasibility: Senior model: policy comparison; emulation: tests Feasibility: Agent: intent; rApp: history-based feasibility estimate NE NE Feasibility: Model-judge panel: benchmark semantic scoring NE Coverage: Orchestrator: support check and alert Feasibility: Validator/certifier: supported hard-constraint fragments; Coverage: Certifier: fragment-bound accept/reject; unresolved abstain NE NE Feasibility: Assigned tools: feasibility; supervisor: admission State: Feedback model uses completed, available records; Coverage: Router: device/risk filter; routes with penalties if all risky

TABLE 23: Three-axis literature map (continued). S: selection; G: generation; C: computation; U: unspecified. Unlisted checks and NE denote unresolved ownership. Work [117]

Interface G

[118]

G

[119] [120] [121] [122]

C S/C S C/S

[123]

G

[124] [125] [126]

G/C S G/S/C

[127] [128] [129] [130] [131] [132] [133]

S/G G/C G/C C U S/G S/G

[134] [135]

G/S G/S/C

[136] [137]

G/S G/S/C

[138] [139] [140] [141] [142] [143] [144] [145] [146] [147] [148]

G/S S/G/C S/C G/C/S S/G G/S S/G S/C S/G G S/C

[149] [150]

G G/S/C

[151] [152] [153] [154]

G G/S G/S/C G/S

[155] [156] [157]

G/C/S S G/S/C

[158]

G/S/C

[159] [160] [161]

G/S G/S/U S/G

Task and execution path Core-network plans to user-plane-function/user reassignment 5G base-station configuration/repair to approved startup Offline port-policy compliance and drift monitoring Intent profiles to O1 assurance counters Distilled policy for isolation/restoration actions Candidate screening to network deployment

Check responsibility Feasibility: Executor agent: plan-feasibility assessment Feasibility: Verification agent: review; operator: startup approval NE NE NE Feasibility: Executor: exact residual check and repair/fallback; Coverage: Guard: no executable generated candidate; fallback NE

Interactive command-line changes scored in emulation Command generation with uncertainty-based review NE Security classification of intent windows NE Intent translation to tool-based policy inspection State: State manager: atomic, versioned snapshots; Feasibility: Tools: network predicates; LLM: final verdict Telecom fault classification with reasoning NE Numerical algorithm search to path allocation Feasibility: Trusted network functions: feasibility constraints Intent compilation to corrected traffic-shaping rules Feasibility: Fixed-template module: policy and tc constraints Multilayer SDN service setup and teardown NE Conceptual domain/capability/service orchestration NE Wireless tool selection and allocation generation NE Semantic objectives to sequential resource actions State: Controller updates state after each allocation; Feasibility: Executor: represented-state feasibility mask Network queries and service subscriptions NE Semantic rewards to reinforcement-learning NE placement/resource solving Tool-assisted wireless allocation and QoS decisions Feasibility: Agent: predicted-throughput QoS verdict Catalog discovery to service-profile decomposition Feasibility: Capability predicates: extracted requirements; Coverage: Catalog layer: supplied-profile match; greedy coverage Troubleshooting dialogue acts and responses NE Network-health diagnosis and proposed mitigation Feasibility: LLM prompts: evidence and action checks Packet reconstruction to failure/localization labels NE Network knowledge to diagnostic plans/reports Feasibility: Z3 component: rules; model: diagnostic report Delegated telecom diagnosis and action planning NE Runbook, API and diagnostic-branch assistance NE Clarification to cross-domain diagnostic queries NE Symptom probes to skill-based fault localization State: Diagnostic agent cross-checks IP/layer evidence Multimodal faults to root-cause/propagation chains NE Multidomain intent translation to service descriptors Feasibility: Proposed handlers: conflict/resource checks Parameter/command mapping to target Feasibility: Constraint-satisfaction solver: translation and view command-line trees consistency Graph-grounded routing-parameter updates NE Dialogue-based intent to Open Network Operating NE System installation Command generation and probes in emulation NE Production alert diagnosis with operator assistance NE Dialogue-selected workflows to health reports NE Workflow/domain-specific-language generation with Feasibility: Parsers, dry runs and invariant validators diagnostic tools Alarm propagation graphs to root-cause diagnosis NE LLM-assisted learning for O-RAN slice allocation NE Intent translation to learned service-chain placement Feasibility: Similarity/k-nearest-neighbour classifier: intent contradictions Example/template selection to formal synthesis Feasibility: CEGS: formal checks of encoded intent, topology and templates Clarified intent to guide-based mitigation actions NE Conceptual incident diagnosis and mitigation NE Intrusion classification with explanatory alerts NE

TABLE 23: Three-axis literature map (continued). S: selection; G: generation; C: computation; U: unspecified. Unlisted checks and NE denote unresolved ownership. Work [162] [163]

Interface S C/S

[164]

C/S/U

[165]

G/C

[166]

G/C

[167]

G

[168]

C

Task and execution path Check responsibility Federated traffic/log anomaly classification NE Request-driven model placement and xApp dispatch Feasibility: Binary integer linear program: resources, timescales, input-data reachability; Coverage: Optimizer: request admission from catalog models Intent policies; ML-hardware performance Feasibility: Policy configurator: conflict check; Coverage: Policy estimation configurator: technique availability; both unevaluated Offline intent to planned optical network design Feasibility: Grammar parser: structure; planner: deployment constraints; Coverage: Planner: reports infeasible requirements Offline causal priors to RAN power-control State: Bayesian network conditions on current measurements; configuration incremental updates Offline emulator-scenario and attack-agent Feasibility: Validator: schema/topology; NASim load-and-run test generation Declarative intents to GitOps/Nephio deployment State: GitOps operators: desired/runtime state drift

Appendix D. Search Strategy and Evidence Definitions D.1. Search and Eligibility The arXiv API was queried on 28 September 2026 with the following title-and-abstract searches, covering 1 January 2016–28 September 2026. Q1, Q2 and Q3 returned 115, 28 and 16 hits, all of which were retrieved. Duplicates within arXiv results were removed by identifier. Cross-source matches used exact normalized titles. Related reports were grouped by work while preserving their distinct tasks and versions. Q1 (115 hits). ((ti:"network configuration" OR abs:"network configuration") OR (ti:"router configuration" OR abs:"router configuration") OR (ti:"intent based networking" OR abs:"intent based networking") OR (ti:"network intent" OR abs:"network intent")) AND ((ti:"language model" OR abs:"language model") OR (ti:"language models" OR abs:"language models") OR (ti:"LLM" OR abs:"LLM") OR (ti:"LLMs" OR abs:"LLMs") OR (ti:"natural language" OR abs:"natural language") OR (ti:"intent based" OR abs:"intent based") OR (ti:"intent driven" OR abs:"intent driven")) AND submittedDate:[201601010000 TO 202609282359] Q2 (28 hits). ((ti:"network orchestration" OR abs:"network orchestration") OR (ti:"service orchestration" OR abs:"service orchestration") OR (ti:"network function deployment" OR abs:"network function deployment") OR (ti:"intent driven resource allocation" OR abs:"intent driven resource allocation")) AND ((ti:"language model" OR abs:"language model") OR (ti:"language models" OR abs:"language models") OR (ti:"LLM" OR abs:"LLM") OR (ti:"LLMs" OR abs:"LLMs") OR (ti:"natural language" OR abs:"natural language") OR (ti:"intent based" OR abs:"intent based") OR (ti:"intent driven" OR abs:"intent driven")) AND submittedDate:[201601010000 TO 202609282359] Q3 (16 hits). ((ti:"network troubleshooting" OR abs:"network troubleshooting") OR (ti:"network diagnosis" OR abs:"network diagnosis") OR (ti:"network fault diagnosis" OR abs:"network fault diagnosis") OR (ti:"network fault localization" OR abs:"network fault localization")) AND ((ti:"language model" OR abs:"language model") OR (ti:"language models" OR abs:"language models") OR (ti:"LLM" OR abs:"LLM") OR (ti:"LLMs" OR abs:"LLMs") OR (ti:"natural language" OR abs:"natural language") OR (ti:"intent based" OR abs:"intent based") OR (ti:"intent driven" OR abs:"intent driven")) AND submittedDate:[201601010000 TO 202609282359]

Publisher-page discovery and citation tracing. Targeted web searches used the following queries to locate individual IEEE Xplore and ACM Digital Library pages. "network configuration" ("LLM" OR "language models"); domains: ieeexplore.ieee.org, dl.acm.org "intent" "orchestration" "language models"; domains: ieeexplore.ieee.org, dl.acm.org "network troubleshooting" ("LLM" OR "language models"); domains: ieeexplore.ieee.org, dl.acm.org site:dl.acm.org "network troubleshooting" "language models" site:dl.acm.org "network configuration" "language models" site:dl.acm.org "intent" "service orchestration"

One backward-reference round examined 152 references from INTA (arXiv:2501.08760), Chat-Driven Optimal Management (arXiv:2512.24614) and A Network Arena for Benchmarking AI Agents on Network Troubleshooting (arXiv:2512.16381). These seeds represent configuration, orchestration and diagnosis. The 21 additional records combine targeted discovery, this citation round and one previously uncited source. Eligibility requires an identifiable network configuration, orchestration or diagnostic decision. Conceptual designs and studies without latency measurements remain eligible when their decision task is explicit. Mixed-domain studies contribute their network tasks. Title/abstract screening assigned 79 records to background and excluded 26 as out of scope. Relevant full-text review assigned six further records to background. Independent researcher review excluded one work concerned

only with CPU-consumption prediction. A blind second screening of all 111 set-aside records found seven eligible under this criterion, and these enter the corpus as additional families (§3.2). Figure 2 gives the resulting record, report and family counts.

D.2. Coverage Audit The OpenAlex search applied a title-and-abstract filter for publications dated from 1 January 2016 to 28 September 2026. Three queries paired a task concept with an interface concept. The configuration concept was “network configuration”, “router configuration”, “intent based networking” or “network intent”. The orchestration concept was “network orchestration”, “service orchestration”, “network function deployment” or “intent driven resource allocation”. The diagnosis concept was “network troubleshooting”, “network diagnosis”, “network fault diagnosis” or “network fault localization”. Each was combined by AND with “language model”, “language models”, LLM, LLMs, “natural language”, “intent based” or “intent driven”. The sampling frame is the 594 records that passed title-and-abstract screening. A fixed pattern over title and abstract assigned 267 records to the language-model stratum when they mention LLM, language model, GPT, generative AI, GenAI, agentic, agent, natural language, NLP or SLM. The other 327 form the earlier intent-networking stratum. Fifty records were drawn from each stratum with seed 42 before any access check. Full text was sought from open-access locations, then by arXiv identifier or a unique arXiv title match, then through a library request. In the language-model stratum, 49 draws were coded and 45 are eligible, with 15 explicit loop claims, 3 matched claims, 2 p95 reports and 1 deadline report. In the earlier stratum, 49 were coded, 32 are eligible and 2 remain unresolved, with 10, 2, 2 and 1 positives. Each coded draw carries the weight of its stratum frame size over the coded draws in that stratum. A weighted rate divides weighted positives by weighted eligible draws. Its interval comes from 2,000 bootstrap resamples of coded draws within each stratum, with seed 42. This estimate assumes that access and coding are missing at random within each stratum. The bounds in §3.2 drop that assumption by counting the two unretrieved draws and every unresolved code as negative or as positive.

D.3. Evidence Definitions Table 24 defines the 18 reporting indicators. “Not reported” means that the relevant methods, evaluation and appendices supply no explicit report. “Unclear” means that the available description does not resolve the category, and N/A denotes an inapplicable measurement or operation. A positive indicator satisfies the stated criterion. Explicit absence, missing reporting and inapplicability remain separate categories. A coding unit is a work–task–decision step at a specified measurement boundary. Different implementations or versions can contribute linked scopes of the same step. A shared aggregate measurement belongs to its measured scope. Inference location concerns decision execution, and code/data indicators concern public pointers supplied by the study. TABLE 24: Definitions and response categories for the 18 evidence fields. The criterion specifies positive coverage in Table 3. Timing endpoint and execution location permit multiple selections. All other fields are single-choice. Field Positive reporting criterion Timing, workload, and comparison Timing endpoint Explicit end event for a measured interval: model return, action dispatch, confirmed network effect, verified service result, or another named event.

Response categories

Metric denominator

Explicit population entering the metric: all attempts, successful requests, another subset, or different denominators for different metrics.

All; successful; other; mixed; unclear; N/A.

Tail latency

A latency quantile at p95 or higher.

Reported; not reported; unclear; N/A.

Deadline attainment

The fraction or count attaining an explicit deadline.

Reported; not reported; unclear; N/A.

Input load

Operational arrival rate, traffic load, concurrency or queue input, as defined below.

Reported; not reported; unclear; N/A.

Queue / waiting

Explicit queueing or waiting evidence for the decision or service path.

Reported; not reported; unclear; N/A.

Stability evidence

Timing or operating-state behavior over operational time or load, as defined below.

Reported; not reported; unclear; N/A.

Non-learning comparison

A rule, solver or traditional controller evaluated as an independent comparator.

Reported; not reported; unclear; N/A.

Model; dispatch; network; service; other; not reported; unclear; N/A.

TABLE 24: Evidence definitions (continued). Field

Positive reporting criterion

Control-loop claim and deployment Explicit loop claim An explicit control-loop class or time-budget claim.

Response categories RT/fast; near-RT; non-RT; other budget or multiple loops; not stated; unclear; N/A.

Supported loop claim

A measurement supports the claimed budget at the corresponding execution boundary.

Matched; mismatch; insufficient evidence; N/A.

Execution location

An explicit location of inference or decision computation.

Hosted API; local GPU; edge device; local CPU; other; unclear; N/A.

Network round trip

The reported decision latency explicitly includes its network communication.

Included; excluded; unclear; N/A.

Validation responsibilities Format check

Evaluation of output syntax or structural validity.

Reported; explicitly unchecked; not reported; unclear; N/A.

Semantic check

Evaluation of task meaning or decision correctness.

Reported; explicitly unchecked; not reported; unclear; N/A.

Final-state check

Evaluation of the resulting network or service state.

Reported; explicitly unchecked; not reported; unclear; N/A.

Three gates distinguished

Format, semantic and final-state checking are explicitly distinguished within the scope.

All three; partial; combined; unclear; N/A.

A public code address or resource is given in the paper or its supplement.

Public link; on request/private; explicitly not public; not reported; unclear; N/A.

A public data address or resource is given in the paper or its supplement.

Public link; on request/private; explicitly not public; not reported; unclear; N/A.

Artifact pointers Public code pointer Public data pointer

D.4. Load and Stability Definitions During adjudication, the researchers clarified two ambiguous field definitions after comparing their independent answers. Operational load includes request/service arrivals, concurrency, traffic offered to the system and queue input, with the decision or service boundary identified. Evidence can be measured, simulated or analytical. Training batch size, prompt length, topology size and single-instance computational complexity describe different quantities. Stability evidence concerns timing or operating-state behavior over operational time or load, including queue/control oscillation and delay under varying traffic. Both stable and unstable behavior qualify. Fixed-condition run-to-run dispersion, a single-request runtime curve and training convergence are distinct from this indicator. The clarification applies to final reporting coverage. The pre-adjudication agreement in Table 25 uses the earlier independent responses.

Appendix E. Coding Distributions and Agreement Table 28 gives the category counts over the 405 scopes coded by the two researchers. The seven families added by the second screening contribute the other 24 scopes (§3.2). Table 27 supplies the option-level reliability of the two multi-select fields, and Table 25 gives the field-level results. Table 26 reports the reproducibility of the three-axis systematization. The companion files provide scope-level coding records and extended descriptions of the literature map for data reuse. Experimental companion tables contain 497 task-condition cells, 239 matched contrasts and 151 online summaries. They also contain 580 timing strata, 83 native load cells and 131 physical-path summaries. The six-case and 24-case policy, repair, forwarding, placement and network records remain in the machine-readable supplement. The execution-path table contains 176 domain-condition rows, including 24 empirical timing strata at each load. JSON-encoded columns store intervals and native metadata. Missing values are coded as unavailable. Measurement metadata identify the inputs, deployment versions and collection periods for each result.

TABLE 25. P RE - ADJUDICATION CODING AGREEMENT: 231 INITIAL PAIRS AND 174 INDEPENDENT RECHECKS AFTER SCOPE ALIGNMENT. A IS EXACT AGREEMENT (%), AND κ AND AC1 USE THE ORIGINAL CODEBOOK CATEGORIES . M ULTI - SELECT A IS EXACT- SET AGREEMENT, AND ITS OPTION - WISE COEFFICIENTS APPEAR IN TABLE 27. C OMBINED POOLS THESE TWO PRE - ADJUDICATION STRATA . Initial comparable n = 231 A (%) κ AC1

Field

Timing, workload, and comparison Named timing endpoint 93.5 Specified metric denominator 91.8 p95 or higher 100.0 Deadline attainment 99.1 Input load 26.0 Queue / waiting 37.7 Stability evidence 26.4 Non-learning comparison 92.2 Control-loop claim and deployment Explicit loop claim 86.1 Supported loop claim 97.8 Named execution location 83.5 Network round trip included 45.0 Validation responsibilities Format check reported 82.7 Semantic check reported 85.7 Final-state check reported 37.2 Three gates distinguished 65.8 Artifact pointers Public code pointer 96.5 Public data pointer 86.6

Aligned recheck n = 174 A (%) κ

AC1

A (%)

Combined n = 405 κ

AC1

N/A 0.839 1.000 0.663 0.062 0.006 0.021 0.736

N/A 0.908 1.000 0.991 0.071 0.268 0.099 0.914

87.9 89.7 99.4 97.1 30.5 41.4 36.2 90.8

N/A 0.835 0.935 0.674 0.108 0.083 0.144 0.725

N/A 0.882 0.994 0.970 0.115 0.300 0.203 0.896

91.1 90.9 99.8 98.3 27.9 39.3 30.6 91.6

N/A 0.838 0.965 0.674 0.083 0.041 0.075 0.731

N/A 0.897 0.997 0.982 0.088 0.281 0.143 0.906

0.692 0.000 N/A 0.187

0.850 0.978 N/A 0.314

81.6 98.9 77.6 39.1

0.649 0.851 N/A 0.118

0.799 0.988 N/A 0.232

84.2 98.3 81.0 42.5

0.673 0.625 N/A 0.169

0.828 0.982 N/A 0.276

0.645 0.294 0.213 0.183

0.803 0.850 0.245 0.619

84.5 85.6 48.9 73.6

0.661 0.464 0.340 0.365

0.825 0.846 0.380 0.706

83.5 85.7 42.2 69.1

0.652 0.380 0.269 0.261

0.812 0.848 0.302 0.656

0.925 0.728

0.962 0.851

92.5 90.8

0.846 0.817

0.917 0.898

94.8 88.4

0.891 0.768

0.943 0.871

TABLE 26. R EPRODUCIBILITY OF THE THREE - AXIS SYSTEMATIZATION OVER THE 129 FAMILIES OUTSIDE THE 10- FAMILY PILOT. T WO MODEL CODERS (GPT AND C LAUDE ) RECODED EACH FAMILY BLIND FROM FULL TEXT UNDER A FROZEN CODEBOOK . A THIRD BLIND PASS MAPPED THE PUBLISHED CELLS INTO THE SAME FIELDS . A IS OBSERVED AGREEMENT. ROWS WITHOUT κ REPORT EXACT SET AGREEMENT. T HE LAST TWO COLUMNS GIVE EACH CODER ’ S AGREEMENT WITH THE PUBLISHED ENTRIES ( κ , OR EXACT AGREEMENT IN PERCENT ), OVER FAMILIES WHOSE PUBLISHED CELLS SETTLE THE FIELD ( PATH 103, RECEIVING ROLES 125). OWNER AND MECHANISM COMPARISONS WITH FEWER THAN 20 SHARED CHECKS ARE OMITTED .

Field

Coder vs coder Coder vs table n A (%) κ AC1 GPT Claude

Decision interface Selection (S) 129 76.7 0.50 Generation (G) 129 86.8 0.72 Computation (C) 129 82.9 0.67 Exact interface set 129 53.5 – Execution path Offline / request-driven / both 129 88.4 0.69 Exact receiving-role set 129 58.9 – Check present Observation 129 81.4 0.62 Feasibility 129 89.1 0.73 Coverage 129 86.8 0.72 Check owner and mechanism, where both coders find the check Observation owner 40 67.5 – Observation mechanism 40 70.0 – Feasibility owner 87 66.7 – Feasibility mechanism 87 81.6 – Coverage owner 43 65.1 – Coverage mechanism 43 81.4 –

0.58 0.39 0.75 0.76 0.66 0.57 – 46.5%

0.51 0.72 0.69 53.5%

0.87 0.47 – 43.2%

0.54 55.2%

0.64 0.82 0.75

0.13 0.42 0.10

0.12 0.42 0.17

– – – – – 43.4% – 63.2% – – – –

– – 47.2% 64.3% – –

TABLE 27. P RE - ADJUDICATION OPTION AGREEMENT FOR THE TWO MULTI - SELECT FIELDS OVER ALL 405 PAIRED SCOPES (231 INITIAL PAIRS AND 174 INDEPENDENT RECHECKS AFTER SCOPE ALIGNMENT ). C OUNTS GIVE THE NUMBER OF SELECTIONS BY EACH RESEARCHER . A COUNTS BOTH JOINT SELECTIONS AND JOINT NON - SELECTIONS . N/A DENOTES UNDEFINED κ FOR A CONSTANT CATEGORY. Option Timing endpoint Model return Dispatch Network effect Service result Other Not reported Unclear N/A Execution location Hosted API Local GPU Edge device Local CPU Other Unclear N/A

Researcher 1 selected

Researcher 2 selected

A (%)

κ

AC1

38 2 11 23 92 251 7 0

41 6 12 22 81 252 5 0

97.3 98.5 99.3 97.8 94.3 98.8 98.0 100.0

0.846 0.244 0.866 0.788 0.831 0.974 0.324 N/A

0.967 0.985 0.992 0.975 0.914 0.977 0.980 1.000

147 62 0 21 39 103 94

128 67 0 22 24 115 89

93.3 97.3 100.0 98.8 91.4 92.6 95.8

0.852 0.899 N/A 0.877 0.400 0.812 0.880

0.879 0.963 1.000 0.986 0.899 0.878 0.935

TABLE 28. A DJUDICATED CATEGORY COUNTS ACROSS THE 405 SCOPES CODED BY THE TWO RESEARCHERS . E ACH SINGLE - CHOICE ROW SUMS TO 405. T IMING ENDPOINT AND EXECUTION LOCATION ALLOW MULTIPLE SELECTIONS . N/A DENOTES INAPPLICABILITY, AND UNLISTED RESPONSE CATEGORIES HAVE ZERO COUNT. TABLE 3 REPORTS POSITIVE COVERAGE AND ITS FAMILY- CLUSTERED INTERVALS . Field Timing, workload, and comparison Named timing endpoint

Response category: number of scopes

Specified metric denominator

All attempts: 48; Successful: 5; Other: 16; Mixed: 7; Unclear: 76; N/A: 253

p95 or higher

Reported: 14; Not reported: 390; Unclear: 1

Model return: 39; Dispatch: 6; Network effect: 12; Service result: 23; Other: 86; Not reported: 253; Unclear: 2

Deadline attainment

Reported: 7; Not reported: 397; Unclear: 1

Input load

Reported: 81; Not reported: 324

Queue / waiting

Reported: 11; Not reported: 394

Stability evidence

Reported: 44; Not reported: 361

Non-learning comparison

Reported: 64; Not reported: 337; Unclear: 4

Control-loop claim and deployment Explicit loop claim

RT/fast: 54; Near-RT: 11; Non-RT: 36; Other: 19; Not stated: 285

Supported loop claim

Matched: 7; Mismatch: 1; Insufficient: 397

Named execution location

Hosted API: 143; Local GPU: 66; Local CPU: 21; Other: 29; Unclear: 105; N/A: 95

Network round trip included

Included: 35; Excluded: 5; Unclear: 91; N/A: 274

Validation responsibilities Format check reported

Reported: 135; Not reported: 266; N/A: 4

Semantic check reported

Reported: 344; Explicitly unchecked: 3; Not reported: 58

Final-state check reported

Reported: 75; Explicitly unchecked: 31; Not reported: 96; Unclear: 8; N/A: 195

Three gates distinguished

All three: 12; Partial: 262; Combined: 130; Unclear: 1

Artifact pointers Public code pointer

Public link: 96; On request/private: 3; Not public: 4; Not reported: 284; Unclear: 18

Public data pointer

Public link: 139; On request/private: 4; Not public: 1; Not reported: 248; Unclear: 13

Related documents

• SoK: Semantic Decision Engines in Network Control Loops#1215528
• Safety Signals to Verify NetOps Agents with Action-Level Granularity#1012360
• Αn empirical study to measure the use and impact of an imprecise information approach in the decision support system Entscheidungsnavi / Mendy Tönsfeuerborn, Rüdiger von Nitzsch#780384
• InfoOps Bench: A live information operations safety benchmark#414141
• DSS DESIGN--A SYSTEMIC VIEW OF DECISION SUPPORT#933670