PaperlayerView exact revision on arXiv

Source-linked AI summary

SoK: Semantic Decision Engines in Network Control Loops

Delong Li, Chen Li, Xu Wang, Haochen Gong, Rui Lang, Guangsheng Yu

arXiv:2610.06425v1cs.NIcs.AI

TL;DR

Network-control research rarely measures the complete path needed to determine whether semantic decision engines can be admitted to a control loop. This SoK systematizes 139 paper families and finds that only 4 of 50 loop or time-budget claims have matched measurement.

  • Problem

    The literature rarely measures the complete path needed to determine which network control loops can admit a semantic decision engine and which checks the controller must keep.

    Source: §1 · p. 1

  • Method

    The paper systematizes 139 network-decision paper families by decision interface, execution path and check ownership, and uses bounded tests to examine consequences of missing load evidence and unnamed check ownership.

    Source: §1 · p. 2, §3.1 · p. 2

  • Results

    Only 4 of 50 loop or time-budget claims had matched measurement; among 72 families without a deterministic computation step, 22 made claims, none supported, and 2 named a coverage owner.

    Source: §1 · p. 1

  • Takeaways & Limitations

    A loop-admission claim should state its endpoint, load and check ownership.

    Source: §11 · p. 14

    The measured consequences come from Edge and RAN source studies and SoK-owned tasks; surveyed network systems by other authors remain untested.

    Source: §9 · p. 13

Abstract

from arXiv · show

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.

1. Introduction

A semantic decision engine is admissible only when its complete control path meets the requested outcome and deadline, yet the literature rarely measures that path or assigns checks. This SoK maps those evidence gaps and tests how they affect admission verdicts.

Introduction

  • Language interfaces shift part of the controller’s responsibility to the interpreter, and a schema-valid answer can still violate shared quotas, rely on an incomplete catalogue, or rest on contradictory observations.A justified rejection can be the correct outcome even when no service is activated.Source: §1 · p. 1
  • Of 139 paper families, 50 claim control-loop or time-budget fit, but only 4 have matched measurement; just 4 report deadline attainment.Among the 72 families without a deterministic computation step, 22 make loop claims with no matched support; only 2 name a coverage owner, versus 11 of 67 families with such a step.Source: §1 · p. 1
  • The bounded tests unify service-orchestration [2], RAN [3], and application studies in one framework, while fixed-action interventions vary decision delay on transport and edge execution stacks.The comparisons use four diagnostic lenses and do not pool effects across the studies.Source: §1 · pp. 1–2
  • A 10 s budget met by every isolated request is met by none at 4.992 arrivals/s under queuing; coverage ownership prompts candidates for 45 of 48 contracts but escalates 4 of 120 infeasible routes.The tests use one bounded event model to show that missing load evidence and unnamed check ownership can each reverse an admission verdict.Source: §1 · p. 2
  • The SoK systematizes 139 families across decision interface, control and execution path, and check ownership, and contributes an evidence audit, a minimum reporting record, design rules, and a research agenda.The systematization is mapped to standardized network functions and recoded blind to assess reproducibility; the audit uses independent coding and explicit denominators.Source: §1 · p. 2

2. Background

The background situates network control decisions within layered time budgets and distinguishes control acknowledgements from verified outcomes. It also frames intent and analytics standards alongside three decision interfaces: selection, generation, and computation.

2.1. Control Loops and Their Time Budgets

  • O-RAN places non-RT and near-RT RIC functions at roughly >1 s and 10 ms–1 s, while control acknowledgement and later KPM change mark different milestones.An interpreter shares these intervals with state collection, communication, and execution; outcome checks are required for service verification.Source: §2.1 · p. 2

2.2. Intent, Analytics and Closed-Loop Functions

  • Standards and frameworks assign feasibility, analytics, orchestration, governance, intent handling, and intent concepts distinct roles across closed-loop management –.3GPP TS 28.312 defines feasibility checks before fulfilment; ETSI ZSM separates analytics, intelligent decision, orchestration, and control, while TM Forum and IRTF documents address intent interfaces and concepts.Source: §2.2 · p. 2

2.3. Selection, Generation and Computation

  • Selection chooses among supplied alternatives, generation constructs content beyond complete alternatives, and computation evaluates formal conditions or synthesizes satisfying assignments [15]–.A workflow may combine the interfaces, with classification based on each step’s output responsibility; fair implementation comparisons require equivalent usable outputs, information, and downstream checks.Source: §2.3 · p. 2

3. Review Method

The review defines an eligible network decision corpus and uses independent screening, coding, adjudication, and family-level aggregation to assess it. Its methods also verify screening coverage and specify how uncertainty and overlapping task groups are handled.

3.1. Scope and Source Identification

  • The review retained 139 source/version records in 133 reading groups, covering 132 families before a second screen added seven families for a final 139-family corpus.Eligible studies identify a network decision output; mixed-domain studies contribute identifiable network tasks, while pure resource prediction without a network decision is excluded. Six records moved to background after full-text review.Source: §3.1 · pp. 2–3

3.2. Screening Verification and Search Coverage

  • Two blinded language-model coders re-screened 111 set-aside records and judged seven eligible; they agreed on 114 of 126 family-level field judgments (90.5%).The seven families were added because the original exclusions applied a narrower operational scope than the published eligibility criterion; two records remained split after reconciliation and stayed outside the corpus.Source: §3.2 · p. 3
  • 32.4% of the frame-weighted sample had explicit loop claims (95% interval 22.2–42.9%); matched support was 6.5%, p95 reporting 5.3%, and deadline attainment 2.6%.Each interval contained the corresponding corpus rate. The stratified sample drew 100 records, retrieved full text for 98, and left two unresolved; bounds treating unretrieved and unresolved records as negative or positive were 30.7–45.4% for loop claims and 6.1–20.9% for matched support.Source: §3.2 · p. 4

3.3. Coding Units and Adjudication

  • The inventory contains 429 scope records—393 step definitions and 36 linked scopes—and an 18-field codebook separates timing, loop claims, correctness gates, and code/data pointers.Researchers coded 405 scopes and model coders coded the 24 scopes from added families; adjudication clarified load and stability definitions, changing 65 of 810 values. Missing, unclear, and not-applicable evidence remained distinct codes.Source: §3.3 · p. 4

3.4. Aggregation and Uncertainty

  • Coverage is estimated across all 139 families, with descriptive 95% confidence intervals from 10,000 family-level bootstrap resamples; task-group comparisons are descriptive because labels overlap.Coverage is also reported across 429 scopes and 393 step definitions. Configuration, orchestration, and diagnosis labels cover 66, 52, and 38 families, respectively, and multi-task families can appear in multiple groups.Source: §3.4 · p. 4

4. Systematization

The systematization maps 139 paper families by decision interface, execution path, and check ownership. It finds that loop claims without deterministic computation lack matched measurement, and separates acceptable-candidate availability from selection quality.

4.1. Three Axes

  • The three axes classify decision interfaces, offline or request-driven execution paths and consuming roles, and ownership of observation, feasibility, and coverage checks.Check entries record the responsible component and evidence type at the reported task boundary; unclear responsibility is marked unresolved.Source: §4.1 · p. 4
  • Published-entry agreement on check presence was low (κ from 0.10 to 0.42), so the check columns support no prevalence estimate.The two coders found 40 observation checks and 43 coverage checks among 129 families, while published entries recorded 9 and 11, respectively.Source: §4.1 · p. 5

4.2. Patterns Across the 139 Families

  • Generation appears in 94 of 139 families, selection in 76, deterministic computation in 67, and 83 combine two or more interfaces.Families with computation named a feasibility owner in 45 of 67 cases, compared with 23 of 72 without computation.Source: §4.2 · p. 5
  • 22 of 50 loop claims came from families without deterministic computation, and none had matched measurement; all four matched claims combined computation with selection or generation.Families without computation named a coverage owner in 2 of 72 cases, compared with 11 of 67 families with computation.Source: §4.2 · p. 5

4.3. Functional Roles and Check Ownership

  • The implemented path determines which component runs each check: NetComplete completes sketches, VeriFlow checks forwarding invariants, and repair checks may guide intermediate actions [33].O-RAN xApps or rApps, management orchestrators, and SDN controllers or policy functions can consume or install decisions, with checks assigned across components.Source: §4.3 · p. 5

4.4. Candidate Coverage and Conditional Selection

  • Overall success combines the chance that an acceptable candidate exists with selection quality conditional on availability; escalation requests can initiate catalogue construction or service recovery.Candidate preparation, construction of missing content, selection, execution, and verification each have separate outcomes.Source: §4.4 · p. 5

5. Literature Reporting Coverage

Across 139 families, reporting coverage varies across timing, correctness checks, and operational conditions; the section also proposes a per-decision reporting record linking inputs, outcomes, checks, and measured events.

5.1. Timing Coverage and Operational Conditions

  • Four of 139 families report deadline attainment (2.9%, 0.7–5.8%), while nine report p95 latency (6.5%, 2.9–10.8%).Timing endpoints appear in 92 families (66.2%) and metric denominators in 49 (35.3%); load is reported in 48 (34.5%, 26.6–42.4%), queue or waiting in six (4.3%, 1.4–7.9%), and stability in 31 (22.3%, 15.8–29.5%). An explicit non-learning comparator appears in 48 (34.5%, 26.6–42.4%) and network round trips in 23 (16.5%, 10.8–23.0%); Table 3 gives 18 indicators and Figure 3 compares overlapping task groups.Source: §5.1 · pp. 5–6, §5 · p. 5

5.2. Correctness Gates and Reusable Artifacts

  • Semantic checks appear in 133 families (95.7%), format checks in 71 (51.1%), and final-state checks in 54 (38.8%).Only 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 data pointers in 51 (36.7%).Source: §5.2 · p. 6

5.3. Coding Reliability

  • Pre-adjudication exact agreement was 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.Agreement was also 42.5% for network-round-trip inclusion and 42.2% for final-state checks; the authors note greater coding uncertainty for load, queue, stability, network-round-trip, and final-state fields than for tail and deadline coverage.Source: §5.3 · p. 6

5.4. A Minimum Reporting Record

  • The proposed reporting record connects each decision's input, output, candidates, assigned checks, measured events, correctness gates, and operational timing conditions.It calls for separate gate denominators and failure, timeout, rejection, escalation, and fallback counts; latency distributions and deadline attainment should be reported with load, concurrency, queueing, and observation window. It also specifies a matched non-learning comparator, timing boundaries for deployment and communication-related costs, code and data pointers, and inapplicable measurements.Source: §5.4 · p. 6

6. Unified Evaluation Framework

The framework evaluates decision engines where outputs change network workflows, separating semantic correctness, decision timing, execution acknowledgement, and verified outcomes. It combines fixed-input comparisons with load traces and fixed-action interventions across service, RAN, transport, and edge tasks.

6. Unified Evaluation Framework

  • The framework combines common events, correctness gates, and perturbations with fixed-input comparisons, load traces, and fixed-action interventions at the point decisions change network workflows.It covers service contracts, RAN policies, and application-layer control tasks; load traces measure queueing and deadline attainment, while fixed-action interventions isolate decision delay in transport and edge stacks.Source: §6 · p. 6

6.1. Timing and Decision Capacity

  • Timing records separate arrival, queue exit, accepted decision, execution acknowledgement, and verified outcome, measuring queueing, decision time, and downstream elapsed times.Deadline attainment counts unreached endpoints as misses; load is characterized by arrival rate, mean occupied time, and decision slots, with ρ < 1 necessary for open workloads admitted without shedding.Source: §6.1 · pp. 6–7

6.2. Three Gates with Explicit Denominators

  • Evaluation separates schema or interface validity, task-predicate satisfaction, and verified terminal outcomes, retaining failures, timeouts, and fallbacks in applicable denominators.Online measures distinguish first-decision correctness, direct execution success, and final success after fallback; native predicates include service top-1 and full-contract exact match, and RAN target-cluster correctness and full-policy exact match.Source: §6.2 · p. 8

6.3. Perturbations and Diagnostic Lenses

  • Perturbations vary structural complexity, catalogue coverage, unsupported requests, check ownership, and execution path to diagnose their effects on task outcomes and delay.The check-ownership comparison contrasts model self-check with a public-gate workflow for coupled placement; execution-path effects are estimated within domains and compared descriptively across domains.Source: §6.3 · p. 8

6.4. Questions and Comparison Families

  • Exploratory comparison families contrast each varied condition with a declared reference, and findings are classified by effect, breadth, and ownership.An effect is repeated only when comparable constructs and outcomes agree in direction beyond negligible-effect references of ±3 percentage points for correctness and attainment, or ±50 ms for latency.Source: §6.4 · p. 8

6.5. Estimation and Dependence

  • Estimation uses native proportions with 95% Wilson intervals, paired contrasts, stratified bootstrap resampling, and task-, deployment-, hardware-, and period-specific timing analyses.Paired correctness uses exact McNemar tests with Holm adjustment; timing intervals resample occupied 30 s blocks, with 60 s and 120 s sensitivity views, and mark cells with fewer than eight occupied blocks as weakly covered.Source: §6.5 · pp. 8–9

6.6. Execution-Path Intervention

  • The execution-path intervention uses 30 fixed-action scenes per domain with delays of 0, 0.1, 0.5, and 2 s, measuring endpoint slopes and decision fractions without queueing.Ten physical repeats are nested within each scene and delay; the sparse arm completes 1,200 verified executions in each domain under a shared guest environment and fixed scene, action, input, and recovery criteria.Source: §6.6 · p. 9
  • A one-slot FIFO load arm tests controlled and replayed decision delays across 84 queues, using Poisson arrivals, shared residual draws, and endpoint attainment over all measured arrivals.It analyzes 126,588 post-warm-up arrivals from 84 completed queues; rates set by the measured 100 ms baseline are approximately 4.992, 7.987, and 9.484 arrivals/s, while empirical replay uses 24 timing strata [2], [3].Source: §6.6 · p. 10

7. Consequences of the Systematized Gaps

Four bounded lenses test how measurement boundaries, workload summaries, check placement and robustness scope affect admission verdicts; these measurements show consequences, not their prevalence.

7.1. L1 Measurement Boundary

  • Only 4/139 families support an explicit loop claim with measurement matched to its stated budget and boundary; decision delay also changes how much of the endpoint time it occupies.With action fixed, verified outcome time follows decision delay approximately one for one: transport slope 1.00025 and edge slope 0.99096. A 0.5 s decision occupies 91.9% of transport acknowledgement time versus 47.2% of verified outcome time, and 98.9% versus 22.8% in edge.Source: §7.1 · p. 10

7.2. L2 Workload Summary

  • A 0.5 s decision met the 10 s transport budget for all 300 isolated requests, but at ρ = 2.500 attainment fell to 0.0% and median verified outcome time reached 233.749 s.The queued load arm uses real decisions ahead of measured execution residuals. In edge at ρ0 = 0.95, a 100 ms decision cut 2 s attainment from 76.0% in isolation to 18.6% (9.3–29.0); the finite-window queues drain after arrivals stop.Source: §7.2 · p. 11, Figure 5 · p. 11

7.3. L3 Check Placement

  • Jev reconciles observations correctly in 50/120 current-conflict cases but 11/120 resolved-conflict cases, and requests candidates in 45/48 absent-candidate cases while escalating only 4/120 infeasible paths.The two coverage checks produce different outcomes; the reported check columns do not support prevalence estimates for reconciliation, coverage escalation or joint feasibility.Source: §7.3 · p. 11, Figure 6 · p. 12
  • Under public gating, Jev’s independent-quota correctness rose from 29/120 to 84/120, but the per-stage rule failed when shared quotas required joint accounting.The bundled redesign changed instructions, refresh actions and check ownership together; paired gains were 45.8 pp with independent quotas, 23.3 pp with shared quotas, and 32.5 pp with shared quotas plus fault-domain separation.Source: §7.3 · p. 11, Figure 7 · p. 12

7.4. L4 Robustness Scope

  • Jev’s results vary with task structure: contract exact matches fell from 277/300 at 4 fields to 173/300 at 8, while RAN target selection varied with fresh-cell count [3].The contrast between 57 and 3 fresh cells was −9.33 pp (paired 95% interval [−14.00, −5.32], Holm-adjusted p = 0.00757); robustness structure, length, representation and catalogue change were not coded for prevalence.Source: §7.4 · pp. 11, 13

8. Design Rules

The design rules translate gaps in the audit and evidence into controller actions, from budgeting and admission under workload to explicit ownership of checks and robustness scope. They require decisions to be evaluated at the outcome and load the controller actually needs to serve.

Budget the Decision at the Required Outcome

  • 50 of 139 families claim a control-loop or time-budget fit, but only four measure it at the stated boundary, so budget the decision at the required outcome.A 0.5 s decision consumes 91.9% of transport acknowledgement time and 47.2% of verified-outcome time; measure D/Ta and D/Tb before replacing an engine with a faster one. In the RAN study, interpretation dominates the acknowledged path, but the base-point experiment shows no resolved service-outcome penalty [3].Source: §8 · p. 13

Admit an Engine by Attainment at Offered Load

  • A 0.5 s transport decision meets a 10 s budget for all 300 isolated requests but none at 4.992 arrivals/s when decisions queue ahead of execution.Nine families report p95-or-higher latency, but four report deadline attainment. At edge ρ0 = 0.95, a 100 ms decision lowers 2 s budget attainment from 76.0% to 18.6%; specify the endpoint, concurrency limit, overload policy, offered load, occupied slot, and observation window.Source: §8 · p. 13

Reconcile Observation Versions Before the Decision

  • 11 of 139 families name an observation owner, and Jev gives the correct placement action in 50/120 current-conflict cases versus 11/120 resolved-conflict cases.Because missing, stale, contradictory, and superseded observations require different actions, retain version, precedence, and unresolved-conflict state and pass a reconciled or explicitly unresolved state to the decision.Source: §8 · p. 13

Make Absent Coverage a Controller State

  • 13 of 139 families name a coverage owner, including only 2 of 72 without a computation step; Jev requests a new candidate for 45/48 contracts without one.Jev escalates 4/120 infeasible forwarding instances, and the same engine passes one coverage check but fails another; implement an explicit no-feasible-candidate outcome and test covered selection separately from absent-coverage escalation.Source: §8 · p. 13

Compute Joint Feasibility in the Controller

  • The public-gate redesign raises Jev’s independent-quota correctness from 29/120 to 84/120, a paired gain of 45.8 pp (37.5–54.2).Per-stage rules fail when shared quotas require joint accounting; where 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 intent interpretation, and construct a contract first for open-vocabulary or underspecified objectives.Source: §8 · p. 13

Report the Robustness Scope With Each Verdict

  • Jev’s contract exact matches fall from 277/300 to 173/300 between four and eight fields [2], while access-policy counts fall from 107/120 to 47/120 between 4 and 64 requirements.RAN target selection loses 9.33 pp between 3 and 57 cells [3]; padding and repetition also alter presentation without changing the requirement set. Prune or preaggregate irrelevant inputs and state tested scope, evaluating output width, observation retrieval, and constraint interaction separately.Source: §8 · p. 13

9. Limitations

The evidence audit identifies missing support for admission claims, not proof that unaudited systems fail to meet their claimed loops; surveyed systems remain untested. Experimental findings are limited by single-run resampling, exploratory comparisons, constructed tasks, and descriptive cross-domain evidence.

9. Limitations

  • The audit measures reported evidence: unaudited families may still meet claimed loops, and surveyed systems by other authors remain untested.It establishes missing support for admission claims; the measured consequences come from Edge and RAN source studies and SoK-owned tasks.Source: §9 · p. 13
  • Each experiment uses one random stream, so within-run resampling intervals do not capture variation across independent replications.Intervals resample by input, scene, or time block; comparisons formulated after inspecting measurement records are exploratory, and Holm adjustment controls error only within each declared family.Source: §9 · pp. 13–14
  • The access-policy, placement, and forwarding tasks use complete predicates and isolate selector behavior, but their graph costs are not measured network delay.The public-gate contrast changes instructions, refresh actions, and check ownership together, so it estimates a workflow redesign rather than moving one check.Source: §9 · p. 14
  • Near saturation, the load arm's intervals describe a finite observation window, not steady state, because post-queue actions do not run concurrently.The arm samples measured physical residuals after a real FIFO decision queue.Source: §9 · p. 14
  • Transport and edge comparisons are descriptive because topology, action, stack, and verification probe differ; radio closed-loop results use published RAN summaries [3].Source: §9 · p. 14

10. Open Problems

The open problems concern how to measure complementary safeguards, certify candidate coverage, track dependencies and resolve cross-domain conflicts, and test whether findings generalize across deployments.

10. Open Problems

  • Guard complementarity requires measuring each parser, configuration verifier, and execution probe’s detected errors, false alarms, and surviving violations, including configuration translation and routing semantics.The three gates provide the events and denominators for this comparison.Source: §10 · p. 14
  • Typed selection still needs methods to certify candidate coverage under changing constraints, decide when to synthesize candidates, and measure construction and certification costs.Escalation should initiate catalogue construction or service recovery; coverage restoration and verified recovery need separate evaluation.Source: §10 · p. 14

Dependency-Driven Refresh and Cross-Domain Conflict.

  • Dependency tracking must identify which catalogue, contract, policy, and network assumptions a state change invalidates, avoiding both needless full refreshes and stale downstream feasibility.Cached interpretations can skip decision calls only while requests, contract versions, and relevant catalogue semantics remain unchanged; dependent network feasibility must be recomputed. Cross-domain intents also require arbitration over resources, priorities, and authority, followed by verification of the shared outcome.Source: Dependency-Driven Refresh and… · p. 14
  • Replication across sites and implementations should vary arrivals, provider concurrency, and concurrent physical actions while matching task responsibilities and recording the complete event path.Independent traffic realizations can distinguish workload variation from the physical-repeat variation measured in §7.1, clarifying how deployment affects faster decisions, queue tails, and verified outcomes.Source: Dependency-Driven Refresh and… · p. 14

11. Conclusion · Appendix A. Complete Native Conditions

The paper systematizes semantic decision engines for network control and finds that claims of deadline fit often lack matched evidence, especially without deterministic computation. Its bounded tests and appendix matrices clarify what loop claims and task comparisons need to report.

11. Conclusion

  • Of 50 families claiming control-loop or time-budget fit, 4 supported the claim at the stated boundary; all four had a deterministic computation step.Families with that step made 28 claims. The 72 families without one made 22 claims, none supported, and only 2 named a coverage owner.Source: §11 · p. 14
  • In bounded tests, the endpoint determined a decision’s share of elapsed time, while load turned a decision meeting every isolated deadline into one meeting none.The results show that deadline evidence depends on endpoint and load conditions.Source: §11 · p. 14
  • The same engine passed some checks and failed others, showing that an engine’s evaluation depends on which check is applied.The conclusion identifies check ownership as part of what a loop claim should state.Source: §11 · p. 14
  • The study systematized 139 paper families by decision interface, execution path, and check ownership.Fifty families claimed that their engine fit a control loop or time budget.Source: §11 · p. 14
  • The authors derive a reporting record, design rules, and a research agenda, and conclude that loop claims should state endpoint, load, and check ownership.These reporting requirements follow from the observed effects of endpoint, load, and check choice.Source: §11 · p. 14

Appendix A. Complete Native Conditions

  • The appendix matrices report tasks’ native conditions, baselines, and denominators; Table 8 compares policy-input conditions using 60 changed and 60 preserved cases.The caption distinguishes whether all 64 requirements change the minimum-cost feasible choice; conditions retain a feasible candidate and pair the same cases and candidates.Source: Table 8 · p. 21, Appendix A · p. 21

Appendix B. Supporting Results by Lens … D.1. Search and Eligibility

The appendices provide lens-specific supporting results, map 139 paper families across three axes, and document the search and eligibility process used to build the corpus.

Appendix B. Supporting Results by Lens

  • Appendix B supplies lens-ordered supporting material: L1 endpoint results, L2 deadline and load traces, L3 observation and check comparisons, and L4 structural and length analyses.Source: Appendix B · p. 21

Appendix C. Three-Axis Literature Map

  • Table 23 maps 139 paper families by decision interface, execution path, and check responsibility, qualifying checks by component, test, or formal model.Source: Appendix C · p. 30, Table 23 · p. 30

D.1. Search and Eligibility

  • The arXiv search covered 1 January 2016–28 September 2026; Q1, Q2, and Q3 returned 115, 28, and 16 hits, all retrieved.Source: §D.1 · p. 33
  • Publisher-page searches and a backward-reference round supplemented arXiv discovery; the round examined 152 references from three seeds, and 21 additional records combined discovery sources.Targeted web searches sought IEEE Xplore and ACM Digital Library pages.Source: §D.1 · p. 33
  • Eligibility required an identifiable network configuration, orchestration, or diagnostic decision; conceptual studies without latency measures and mixed-domain network tasks remained eligible.Source: §D.1 · p. 33
  • Title and abstract screening assigned 79 records to background and excluded 26 as out of scope; full-text review assigned six more to background.Source: §D.1 · p. 33
  • A blind second screening of 111 set-aside records found seven eligible additional families; one work focused only on CPU-consumption prediction was excluded.Source: §D.1 · pp. 33–34

D.2. Coverage Audit … Appendix E. Coding Distributions and Agreement

The appendices specify how the literature audit was sampled and coded, define the evidence and operational-load criteria, and document coding agreement and reusable companion data. The coverage audit reports stratum-level coding and eligibility counts, while the agreement tables describe the scope of several reliability checks.

D.2. Coverage Audit

  • The 594 records passing title-and-abstract screening were split into 267 language-model and 327 earlier intent-networking records, with 50 draws per stratum using seed 42.Draws occurred before access checks; full texts were sought through open-access locations, arXiv identifiers or title matches, then library requests.Source: §D.2 · p. 34
  • 45 of 49 coded language-model draws were eligible, versus 32 of 49 in the earlier stratum; both strata had one deadline report.The language-model stratum had 15 explicit loop claims, 3 matched claims and 2 p95 reports; the earlier stratum had 10, 2 and 2, respectively. Weighted rates assume access and coding are missing at random within each stratum; bounds in §3.2 drop that assumption.Source: §D.2 · p. 34

D.3. Evidence Definitions

  • The 18 reporting indicators distinguish positive coverage from “not reported,” “unclear” and N/A, while coding units are work–task–decision steps at specified measurement boundaries.Implementations and versions may contribute linked scopes; shared aggregate measurements belong to their measured scope, and inference location concerns decision execution.Source: §D.3 · p. 34, Table 24 · pp. 34–35

D.4. Load and Stability Definitions

  • Operational load covers arrivals, concurrency, offered traffic and queue input at an identified decision or service boundary; stability concerns timing or operating-state behavior over time or load.Evidence may be measured, simulated or analytical; fixed-condition dispersion, single-request runtime curves and training convergence do not by themselves qualify as stability evidence.Source: §D.4 · p. 35

Appendix E. Coding Distributions and Agreement

  • Adjudicated category counts cover 405 scopes coded by two researchers, while seven families added in the second screening contribute another 24 scopes.Table 28’s single-choice rows sum to 405; timing endpoint and execution location allow multiple selections.Source: Appendix E · p. 35, Table 28 · p. 37
  • Companion files expose scope-level coding and literature-map descriptions, plus experimental tables spanning task-condition cells, matched contrasts, timing strata, load cells and physical-path summaries.They also include machine-readable policy, repair, forwarding, placement and network records; JSON columns store intervals and native metadata, and missing values are coded unavailable.Source: Appendix E · p. 35
  • Pre-adjudication agreement was assessed on 231 initial pairs and 174 independent rechecks after scope alignment, with exact-set agreement for multi-select fields.Table 25 reports field-level agreement, while Table 27 describes option-level agreement for the two multi-select fields; the passages provide no agreement values.Source: Appendix E · p. 35, Table 25 · p. 36, Table 27 · p. 37
  • The three-axis systematization was checked across 129 families outside the 10-family pilot using two blind model-coder passes and a third pass mapping published cells.Table 26 reports observed agreement and coder agreement with published entries where the published cells settle the field; some low-count comparisons are omitted.Source: Appendix E · p. 35, Table 26 · p. 36

Related papers

Loading 2610.06425v1…