{"id":"3ec938da-d4f1-4f73-94cb-e6ab308af232","arxiv_id":"2511.17505","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"An RCD + causal subgraph + KS/Z-score pipeline claims to recover the ordered KPI chain before RAN SLA breaches, but validates only one LTE case without ground truth.","lead":"This paper proposes an AI pipeline that finds likely causes and the ordering of events leading to network SLA breaches, tested on an LTE cell-load case. It combines a causal-discovery algorithm with deviation tests, but does not quantitatively validate that the identified sequences match real ground-truth faults.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ground-truth validation: Monte Carlo shows only RCD convergence, not that the recovered sequence is correct; the case study 'validation' reuses the same data, so the central claim 'pinpoints the correct trigger sequence' is untested.","rationale":"The central claim is empirical: the pipeline recovers the 'correct' causal intervention sequence. To support that, one needs a ground truth. The manuscript's Monte Carlo section is often read as validation, but it only establishes that RCD's selection frequency converges; convergence is a property of the estimator, not evidence that the estimator targets the true causes. The real-data case study uses histograms and time series from the same data to 'confirm' the sequence, which is circular. The causal invariance assumption (Section III.B) is a genuine additional concern, but it is a theoretical caveat; without a ground-truth experiment we cannot even assess whether the pipeline would work in the best case. Therefore the most load-bearing gap is the absence of any quantitative comparison to a known sequence. The reader flagged this in the rationale (Monte Carlo 'not agreement with ground truth') but listed hidden confounding as the weakest assumption; hence partial agreement. A synthetic ground-truth experiment would settle the issue decisively and is feasible. The verdict should remain CONDITIONAL because the lack of evidence is exactly what conditional acceptance requires the authors to remedy.","tokens_in":7541,"tokens_out":5807,"duration_ms":47262,"concrete_test":"Create a synthetic RAN dataset from a known structural causal model: e.g., 15-20 KPIs with a specified DAG, a planted intervention sequence of 3-5 anomalies preceding a binary SLA failure flag, and known hard/soft intervention types. Generate many Monte Carlo samples with realistic noise and optional hidden confounders. Run the full pipeline (RCD with g,n from the paper/Table I, KS/Z-score sequencing) and measure (1) precision/recall of identified root-cause KPIs against the true intervention set, and (2) order accuracy of the recovered sequence (e.g., normalized Kendall tau or edit distance) over 100+ seeds. If order accuracy is not significantly above chance or root-cause recall is low, the central claim is refuted; if it passes, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that the pipeline 'pinpoints the correct trigger sequence with high precision' (Abstract). The paper never compares the output to a known causal sequence. Section IV.A uses real LTE data, and Section IV.C 'validates' the identified sequence by inspecting histograms and time series (Figures 4-5) that were used to derive it; this is not an independent test. The Monte Carlo experiment (Section IV.E) only tracks the frequency P_{g,n} with which RCD names a KPI as a cause and shows that this frequency stabilizes as n grows (Figure 7). Convergence of an estimate to some value p does not imply p is the probability of a true cause; it only indicates internal algorithmic stability. Table I therefore cannot support 'correct trigger sequence'. Moreover, equations (3)-(4) select g* and n* from the 'prominent causal sources' identified by the same Monte Carlo runs, so the parameter choice is in-sample. A causal reading also depends on the unexamined invariance P(X|do(F_t=1))=P(X|F_t=1) and on causal sufficiency, but even granting those, the central empirical claim remains unvalidated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage AI/ML pipeline for root-cause and temporal-sequence analysis of SLA breaches in radio access networks. Stage one uses the RCD algorithm (from reference [11]) with a binary failure node F_t to select 'intervention variables' as root-cause candidates; stage two builds a causal subgraph among those variables from normal-state data; stage three applies Kolmogorov-Smirnov and Z-score deviation detection to order the anomaly events leading to the SLA breach. The method is demonstrated on a single LTE cell-load case with 60 KPIs at 15-second granularity, and a Monte Carlo experiment is used to choose the RCD parameters g and n and to estimate the probability that each KPI is a causal source. The central claim, stated in the abstract, is that the approach 'pinpoints the correct trigger sequence with high precision.'","tokens_in":7895,"tokens_out":3529,"duration_ms":32373,"significance":"The problem is practically important: operators need early, ordered indicators of impending SLA violations, and the paper attacks a real difficulty—high-resolution data are too voluminous for manual labeling. The proposed pipeline has attractive features: it avoids per-event human ground-truth labeling, uses a relatively small causal subgraph, combines causal discovery with simple deviation tests, and is claimed to be CPU-only and scalable. The comparison with PCMCI on the same data is a useful qualitative benchmark. However, the paper's central claim is not currently supported by the evidence. The case-study validation uses the same data for discovery and validation, and the Monte Carlo experiment shows only algorithmic convergence, not correctness against a known causal sequence. If the authors can provide a ground-truth validation (e.g., injected faults with known event order, or expert-annotated held-out events) and report quantitative accuracy, the contribution would be significant. As it stands, the paper is a promising systems description but not a validated causal-analysis result.","major_comments":[{"comment":"The abstract claims the approach 'pinpoints the correct trigger sequence with high precision,' but no experiment compares the output to a known or independently annotated causal sequence. In Section IV.C, Figures 4 and 5 are histograms and time series of the same KPIs used to derive the sequence in Figure 3; inspecting them after the fact is not an independent validation. The paper should either provide ground-truth events (e.g., fault-injection tests with a known order of KPI deviations, or held-out expert-labeled fault episodes) and report quantitative metrics (e.g., sequence edit distance, order accuracy, precision/recall of root-cause KPIs), or explicitly re-scope the claim to 'plausible sequence consistent with observed distributions.' Without this, the central claim is untested.","section":"Abstract; Section IV.C"},{"comment":"The Monte Carlo experiment demonstrates that the sample proportion P_{g,n} stabilizes as n grows, which is expected for any consistent estimator and does not imply that the limiting p is the probability of a true cause. The selection of g* and n* in Eqs. (3)-(4) uses the same 'prominent causal sources' that the Monte Carlo runs are supposed to validate, making the parameter choice in-sample. Additionally, Table I reports nonzero estimated probabilities for several KPIs with optimal n=0 (e.g., DL QPSK Success Rate p=0.16, MAC DL BLER p=0.50, DL 256QAM Fail Rate p=0.14); this is internally inconsistent if n is the number of experiments. Please clarify how these estimates are formed and how n=0 can support 'prominent causal sources.' At minimum, the distinction between 'internal stability' and 'ground-truth correctness' must be explicit.","section":"Section IV.E; Eqs. (2)-(4)"},{"comment":"The causal interpretation rests on the invariance assumption P(X|do(F_t=1)) = P(X|F_t=1) and on causal sufficiency of the observed KPI set. Neither assumption is defended or stress-tested. In a RAN, unmeasured confounders such as cell location, UE distribution, scheduler state, or external interference can affect both the failure flag and candidate KPIs; if such confounders exist, the parent set found by RCD and the subsequent 'causal intervention sequence' are not causal. The authors should provide a sensitivity analysis (e.g., simulate a hidden confounder and assess how the recovered sequence changes) or at least state the limitation prominently and avoid overclaiming causal language in the abstract and conclusion.","section":"Section III.B, 'Intervention-Based Invariance'"},{"comment":"The reported causal intervention sequence is sensitive to the hand-tuned alpha value: Figure 3 uses alpha=0.1 and Figure 6 uses alpha=0.05, and the sequences differ. The paper justifies the larger alpha by noise tolerance and mentions Bonferroni and FDR corrections, but no details of the correction or the sensitivity of the final ordering to alpha are given. Because the central product is the ordered sequence, the authors should report the sequence across a range of alpha values, quantify the disagreement, and justify the chosen threshold with a principled criterion instead of a post-hoc preference for detecting 'subtle but consistent shifts.'","section":"Section IV.D; Figures 3 and 6"}],"minor_comments":[{"comment":"Typo: 'interventionsons' should be 'interventions'; Figure 6 caption uses 'Casual' instead of 'Causal.'","section":"Section IV.D; Figure 6 caption"},{"comment":"The dataset description is thin: it says each SLA breach event yields 240 data points across 60 KPIs, but does not state how many breach events were used, how many cells, or how the normal/abnormal periods were selected. This information is needed to judge the representativeness of the case study.","section":"Section IV.A"},{"comment":"The phrasing 'RCD algorithm(s)' and the citation to [11] are vague; a reader cannot tell which steps of the Ψ-PC algorithm are modified and which are unchanged. A more precise algorithmic description or pseudocode would improve reproducibility.","section":"Section I; Section II"},{"comment":"Reference [14] is malformed (author/publisher fields are incomplete). Several references to causal-discovery software (e.g., PCMCI implementations) would benefit from URLs or version numbers. No code or data availability statement is provided; for a methods paper, this is a reproducibility concern.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper reads as an industry experience/systems description rather than a validated research contribution. The core issue is that the headline claim—'pinpoints the correct trigger sequence with high precision'—is not matched by any ground-truth evaluation. I would encourage the editor to ask for either a re-scoped abstract and conclusions or, preferably, an experiment with a known causal ground truth. The Monte Carlo analysis and PCMCI comparison are useful but do not fill that gap. The manuscript is likely fixable within its scope if the authors can add a validation benchmark or explicitly lower the strength of the causal claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take: The paper combines RCD, causal subgraph construction, and KS/Z-score deviation detection into a pipeline for root-cause sequence tracking in RAN. The engineering is coherent, and the application is new. But the central claim in the abstract — ‘pinpoints the correct trigger sequence with high precision’ — is not supported by the evidence. The only case study is a single LTE cell-load issue with 60 KPIs, and the ‘validation’ in Figures 4-5 reuses the same histograms and time series that generated the sequence. The Monte Carlo experiment shows that the RCD estimate P_{g,n} stabilizes as n grows, which is internal consistency, not agreement with a known causal chain. No ground-truth accuracy metric is reported anywhere.\n\nWhat the paper does well: It applies Root Cause Discovery to high-resolution RAN data, which I have not seen in the cited literature, and it proposes a concrete Monte Carlo protocol for choosing g and n. The authors openly discuss PCMCI's shortcomings and the sensitivity of the sequence to the CIS p-value, and they say they applied Bonferroni/FDR corrections. The pipeline is lightweight and could be a useful operational heuristic if properly validated. The writing is clear and the related work is fair.\n\nThe soft spots are real. First, validation: there is no external labeled dataset, no quantitative comparison with PCMCI, and the parameter selection for g* and n* is in-sample — equations (3)-(4) pick values based on the same Monte Carlo runs that define ‘prominent causal sources.’ Second, the causal reading depends on the invariance P(X|do(F_t=1)) = P(X|F_t=1), which is asserted without argument; an observed failure flag is not a do-intervention, and unmeasured confounders like cell location or scheduler state would break it. Third, the paper claims scalability to millions of points, but no scaling experiments are shown. Minor: Table I lists n=0 for several KPIs, which is confusing.\n\nWho should read it: network operations researchers and practitioners working on fault automation. It deserves a serious referee, because the problem is important and the pipeline is nontrivial, but the referee should demand ground-truth validation and released code/data. I would not cite the accuracy claim in my own work, but I would bring it to a reading group to discuss how far internal consistency can carry a causal claim.","headline":"Plausible RAN root-cause pipeline, but the headline causal claim is untested: validation is in-sample, and the Monte Carlo shows internal convergence, not correctness.","tokens_in":8339,"tokens_out":3083,"would_cite":false,"duration_ms":37928,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a three-module AI/ML pipeline that identifies root-cause KPIs and recovers the exact temporal order in which their deviations lead to an SLA breach.","keywords":["radio access networks","root cause analysis","causal discovery","intervention sequence","SLA breach prediction","deviation detection","Kolmogorov-Smirnov test","Monte Carlo simulation"],"falsifier":"Create a synthetic RAN-like dataset with a known intervention order and one hidden confounder that affects both the failure flag and two candidate KPIs; if the pipeline inserts the confounded KPI earlier in the sequence or reverses two adjacent events, the invariance assumption is violated. A second check is to collect historical data from many cells and verify that the predicted first deviation precedes the SLA breach by a consistent lead time.","tokens_in":7457,"feed_emoji":"📡","tokens_out":5704,"duration_ms":50402,"temperature":0.7,"pith_summary":"The paper aims to show that the causal chain behind an SLA breach can be extracted from high-resolution network data without manual labeling. It does this in three steps: find which KPIs have been forced into anomalous states, build the normal-state causal subgraph among them, and then use deviation detection to order the anomalies by their first-exit times. In the worked LTE cell-load case, the pipeline points to the leading indicators and the sequence that ends in downlink throughput falling below the SLA threshold. Monte Carlo experiments estimate how often each KPI appears as a causal source and give a rule for setting the algorithm's two main parameters. If the claim holds, operators could act on the first flagged KPI rather than respond to the breach itself.","feed_headline":"Causal AI pipeline orders the events that lead to SLA breaches","feed_subtitle":"Engineers see which KPIs deviate first and in what order, gaining time to act before subscribers feel the outage","key_machinery":"The central object is the causal intervention sequence: a temporal ordering of KPI deviations, produced by three components. RCD identifies intervention variables as the parents of the failure flag; causal subgraph construction restricts attention to the causal connections among those variables; and deviation detection (K-S and Z-score) assigns the order and direction of each deviation. The load-bearing identity is the intervention invariance P(X|do(F_t=1)) = P(X|F_t=1), which allows the pipeline to read observed conditional distributions after a failure as if they were produced by a controlled intervention on the failure flag.","core_discovery":"On the paper's own terms, the central claim is that root-cause discovery can yield not just a list of suspicious KPIs but an intervention sequence. The pipeline starts with a binary failure flag and uses the Root-Cause Discovery (RCD) procedure to identify variables whose dependence on the failure flag survives conditional independence tests, treating those variables as parents of the failure and therefore root-cause candidates. It then learns a causal subgraph from normal-state data and applies Kolmogorov-Smirnov and Z-score deviation detection to mark when each candidate KPI first leaves its normal distribution. The order of those first deviations is the causal intervention sequence, and t","pith_inferences":["A testable extension the authors do not run is a hidden-confounding stress test: inject a known intervention sequence into synthetic RAN data while adding an unmeasured common cause, and check whether the recovered sequence still matches. That would directly probe the invariance condition.","The Monte Carlo analysis validates convergence of the per-KPI causal-source probability, but not end-to-end sequence accuracy; pairing the binomial estimate with a measure of full-sequence recovery would sharpen the headline claim about precision.","If confirmed across other fault families (interference, handover failures, transport issues), the same three-module pattern could serve as a generic early-warning layer around any SLA-type threshold, with each threshold needing only a new labeling rule."],"forward_implications":["Operators can shrink the monitored KPI set to just the prominent causal sources identified by the Monte Carlo ranking, cutting monitoring cost and complexity.","Because the earliest deviation is flagged as the root cause, the sequence gives lead time to act before the SLA indicator crosses its threshold.","The pipeline runs on standard CPUs and samples only a subset of variables, so it can be deployed per cell at scale rather than in a central GPU cluster.","The Monte Carlo parameter selection provides a reproducible way to choose the algorithm's g and n and to rank KPIs by their probability of being a causal source.","The labeling is flexible: a low-resolution SLA rule can be defined simply, and the pipeline automatically maps it to high-resolution KPI deviations."],"fun_headline_variants":["AI decodes the causal order behind network SLA breaches","Causal AI tracks the precise sequence of fault triggers","Know the next fault: AI orders root-cause events","AI spots which KPIs fail first and in what order","Proactive network care: AI sequences fault indicators"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The causal interpretation stands on the assumption that the failure flag is the only common cause of the flagged KPIs; if an unmeasured factor such as subscriber distribution or scheduler state influences both the failure flag and the KPIs, the recovered 'intervention sequence' is only an association sequence.","fun_headline_variants_meta":{"raw":{"variants":["AI decodes the causal order behind network SLA breaches","Causal AI tracks the precise sequence of fault triggers","Know the next fault: AI orders root-cause events","AI spots which KPIs fail first and in what order","Proactive network care: AI sequences fault indicators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1105,"prompt_tokens":650,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":394,"completion_tokens_details":{"reasoning_tokens":393}},"tokens_in":394,"tokens_out":455,"duration_ms":46048,"temperature":1.0,"reasoning_tokens":393,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T12:44:50.510515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Create a synthetic RAN-like dataset with a known intervention order and one hidden confounder that affects both the failure flag and two candidate KPIs; if the pipeline inserts the confounded KPI earlier in the sequence or reverses two adjacent events, the invariance assumption is violated. A second check is to collect historical data from many cells and verify that the predicted first deviation precedes the SLA breach by a consistent lead time.","supporting_citations":[],"review_version":1}