{"id":"0d0d096a-2d7f-4708-8033-aada112c3f50","arxiv_id":"2607.01583","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Logistic regression and a shallow MLP compiled as deterministic C modules inside a FlexRIC xApp achieve 1–25 µs inference and sub-4 ms end-to-end service latency on an OAI testbed, meeting the 10 ms Near-RT budget for >95% of projected loops.","lead":"Researchers embedded tiny AI classifiers as plain C code inside an O-RAN Near-RT RIC xApp and measured microsecond inference on a live open-source 5G testbed. The work shows AI compute is not the timing bottleneck and ships a laptop-scale dashboard so others can reproduce the stack.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the paper's own feasibility scoping.","rationale":"The reader's strongest claim matches the paper's actual contribution: measured, deterministic C-embedded inference inside a Near-RT xApp with explicit three-component latency instrumentation. The weakest assumption the reader flags (synthetic proxy + small RFsim stack) is real but already owned by the authors as a feasibility, not generalization, result. That does not create an internal load-bearing failure for the claim as written; it correctly bounds how far the claim may be extrapolated. Correctness risk on the timing numbers is low given monotonic-clock instrumentation and the clear dominance of E2 path over model compute. Novelty is engineering/systems rather than algorithmic, which the reader already scored mid-band. Verdict remains CONDITIONAL pending real artifact availability and continued feasibility scoping—exactly as the reader concluded. No stronger technical objection is warranted.","tokens_in":12722,"tokens_out":531,"duration_ms":8464,"concrete_test":"If artifacts are released: recompile the xApp with the exported C modules, re-run the 10-minute indication loop under the same RFsim load, and recompute the inference-only and end-to-end CDFs of Fig. 8; if the >95% under-10 ms fraction or the 1–5 / 10–25 µs inference bands do not reproduce, the timing claim weakens. Absent artifacts, this remains the single highest-value verification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is carefully scoped as measured embedding feasibility of lightweight deterministic C inference (LR 1–5 µs, MLP 10–25 µs; typical E2E <4 ms; >95% of projected loops under 10 ms) on a live OAI/FlexRIC path, not production generalization. The synthetic five-feature proxy and 2-gNB/2-UE RFsim+Docker stack are the weakest transfer points, but the manuscript repeatedly states this (Abstract, §III-C, §VI) and treats accuracy only as pipeline evidence. Inference is shown to be negligible relative to E2 transport; rare >100 ms spikes are attributed to commodity OS/Docker, not the models. No internal inconsistency or hidden assumption undermines the stated feasibility result under the reported conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper demonstrates that lightweight supervised models (logistic regression and a shallow two-hidden-layer MLP) can be exported as deterministic C inference modules, compiled into a FlexRIC xApp binary, and executed inline on a live OAI/FlexRIC Near-RT RIC path. On a structured five-feature synthetic proxy for network-state classification (MAC/RLC/PDCP/GTP latencies + UE count), offline supervised accuracy clusters in 0.88–0.90; measured inference is 1–5 µs (LR) and 10–25 µs (MLP), typical end-to-end service latency stays below 4 ms, and both models are reported to meet the 10 ms Near-RT budget for >95% of projected loop executions. The contribution is scoped as embedding/execution feasibility and pipeline reproducibility (including the released RIC Workbench), not production generalization or closed-loop control.","tokens_in":12963,"tokens_out":1122,"duration_ms":22827,"significance":"The identified gap—measured, in-process, deterministic C inference with three-component latency instrumentation on a live OAI/FlexRIC stack—is real and under-served relative to simulation or external-inference xApp work. The offline-to-online export path, monotonic-clock instrumentation, noise ablation, and six-model offline comparison are concrete systems evidence that AI compute is negligible versus E2 transport under the reported conditions. Releasing RIC Workbench as a single-binary commodity orchestration tool is a genuine reproducibility contribution. If the measurements hold as stated, the paper supplies a useful template and lower bound for Near-RT RIC AI embedding rather than a new learning algorithm.","major_comments":[{"comment":"§IV-E and Fig. 8 (left): the claim that both models satisfy the 10 ms Near-RT budget for >95% of loop executions is based on a “projected” end-to-end loop that includes action dispatch, while §VI and the design text state that closed-loop control is not implemented and the xApp only classifies. The primary, fully measured result is E2 arrival + inference (and the reported typical service latency <4 ms). The manuscript should (i) report the measured E2+inference CDF as the main result, (ii) state explicitly and quantitatively what latency was assumed for the projected action-dispatch term, and (iii) keep “projected closed-loop” clearly labeled so the central feasibility claim does not rest on an unmeasured component.","section":null},{"comment":"Abstract / §I / §VII vs. §III-C and §VI: the abstract and conclusion state that the results “demonstrate that lightweight AI can operate within Near-RT RIC timing constraints,” while the body correctly scopes the work as embedding feasibility on a rule-structured synthetic proxy and a 2-gNB/2-UE RFsim + Docker commodity stack. The transfer claim should be tightened in the abstract and conclusion to match the body (e.g., “under the reported OAI/FlexRIC path and proxy task”), so the headline does not outrun the experimental conditions that the limitations section already acknowledges.","section":null}],"minor_comments":[{"comment":"Table IV “C Budget” column mixes qualitative deployability (✓ / Partial / No) with latency ranges; define the criteria for Partial/No in the caption or text so the six-model comparison is reproducible.","section":null},{"comment":"Fig. 8 caption and §IV-E: clarify whether the rare >100 ms service spikes are included in the plotted CDFs and in the “>95% under 10 ms” statistic, or whether that statistic is conditioned on non-preempted samples.","section":null},{"comment":"§III-C: state the exact random seed and generation script version (or release them with RIC Workbench) so the 32k-sample synthetic set is bit-reproducible by others.","section":null},{"comment":"Table VI peak inference latencies (<2 ms / <3 ms) are orders of magnitude above the typical 1–25 µs ranges; briefly explain what produces those peaks (e.g., first-call, cache, or measurement artifact) so they are not read as model compute.","section":null},{"comment":"Related work (§II-C): a short explicit contrast table (Python/external vs. compiled C; presence/absence of per-indication monotonic timing) would sharpen the claimed gap versus ColO-RAN and other cited xApp demos.","section":null},{"comment":"Minor polish: consistent hyphenation of Near-RT / Near-Real-Time; “µs” vs “us” in axis labels; ensure Fig. 3 console screenshot remains legible in print.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Fit is appropriate for a systems/networking venue that values measured O-RAN prototypes and tooling. Novelty is primarily integration, measurement, and reproducibility rather than ML method; that is fine if the journal accepts systems contributions. No integrity red flags; the authors repeatedly self-limit the synthetic proxy and scale. I would not require live closed-loop actuation for acceptance if the projected component is cleanly separated as requested."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The load-bearing result is the measured path: logistic regression and a shallow MLP exported as static C arrays, compiled into a FlexRIC xApp on OAI, with inference at 1–5 µs and 10–25 µs and typical end-to-end service latency under 4 ms. Both clear the 10 ms Near-RT budget for >95% of projected loops. Inference is not the bottleneck; E2 transport is. That is useful, concrete systems evidence, not another offline RAN ML paper.\n\nWhat is new is the combination, not any single piece. Prior work already has AI xApps, ColO-RAN, FlexRIC, and closed-loop SDR demos. Here they eliminate external ML runtimes, instrument three latency components with monotonic clocks and CDFs, run a six-model offline sweep plus noise ablation on a five-feature proxy, and ship RIC Workbench so others can stand the stack up on a laptop. Claims are carefully scoped to embedding feasibility, not production generalization. The synthetic dataset is treated as a controlled proxy; accuracy (0.88–0.90) is pipeline evidence only. That honesty is a strength.\n\nSoft spots are real but proportional. The proxy is rule-structured with injected Gaussian noise; live multi-cell E2 paths and temporal models may change the latency mix and the accuracy story. Scale is 2 gNB / 2 UE RFsim on commodity Ubuntu/Docker; rare >100 ms spikes are OS/Docker, not the models. Closed-loop actuation is deferred. None of that undercuts the stated feasibility result under the reported conditions. Math is elementary linear algebra; citations cover the right O-RAN and xApp literature without padding.\n\nThis is for people building Near-RT RIC software who need numbers on where the time actually goes. It deserves a serious referee. I would engage with it, cite the latency numbers and the C-embedding pattern when discussing real-time AI placement in O-RAN, and bring it to reading group if we are talking systems measurement rather than new algorithms.","headline":"Solid systems measurement paper: C-compiled µs inference inside a live FlexRIC xApp, with honest feasibility scoping and a useful commodity testbed tool.","tokens_in":13599,"tokens_out":501,"would_cite":true,"duration_ms":5418,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Lightweight AI compiled into a Near-RT RIC xApp runs in microseconds and stays under the 10 ms control budget on a live testbed.","keywords":["O-RAN","Near-RT RIC","xApp","AI inference","network state classification","logistic regression","MLP","OpenAirInterface"],"falsifier":"Replace the synthetic proxy with live multi-cell E2 indications under realistic load, re-export the same C models (or a modestly larger one), and re-measure the end-to-end latency CDF: if more than a few percent of loops exceed 10 ms or inference ceases to be negligible relative to E2 transport, the central feasibility claim fails for that workload.","tokens_in":13624,"feed_emoji":"📡","tokens_out":993,"duration_ms":7820,"temperature":0.7,"pith_summary":"This paper asks whether simple AI can be put inside the Near-Real-Time RAN Intelligent Controller software loop without breaking the 10 ms to 1 s control timing that O-RAN requires. The authors build a network-state classification xApp on a live OpenAirInterface and FlexRIC stack, train logistic regression and a shallow multilayer perceptron offline on a structured synthetic proxy dataset of cross-layer latency and load features, export the trained weights as pure C arrays, and compile them into the xApp binary so there is no external machine-learning runtime. On the live testbed they measure inference at 1–5 µs for logistic regression and 10–25 µs for the MLP, with typical end-to-end service latency under 4 ms; both models meet the 10 ms Near-RT budget for more than 95 percent of projected loop executions. The result they establish is feasibility of deterministic embedded inference, not production-grade accuracy on real traffic. They also release a single-binary orchestration dashboard so others can reproduce the stack on ordinary hardware.","feed_headline":"AI inside a Near-RT RIC xApp runs in 1–25 µs","feed_subtitle":"Compiled C models keep end-to-end service under 4 ms and meet the 10 ms budget on a live testbed","key_machinery":"The offline-to-online C-export pipeline: trained logistic-regression and shallow-MLP parameters are written as static const float arrays in C headers, compiled into the xApp binary, and run inline on five-dimensional service-model feature vectors (MAC, RLC, PDCP, GTP latencies and UE count) with monotonic-clock instrumentation of service, inference, and end-to-end latency.","core_discovery":"Lightweight supervised models, exported as deterministic C inference modules and compiled into a Near-RT RIC xApp, can classify cross-layer network states on a live OAI/FlexRIC testbed with measured inference of 1–5 µs (logistic regression) or 10–25 µs (shallow MLP) and typical end-to-end service latency below 4 ms, satisfying the 10 ms Near-RT control budget for more than 95 percent of projected executions and showing that AI computation is not the bottleneck under this setup.","pith_inferences":["The same C-export pattern would likely keep inference negligible for other compact tabular models (small trees, linear SVMs) until feature dimension or depth forces dynamic allocation or larger matrix work.","If E2 transport remains the dominant cost, research effort on this path should shift from model compression toward message-path and scheduling optimizations.","A natural next measurement is multi-xApp co-residency: whether concurrent compiled inferences still leave headroom under the same 10 ms budget."],"forward_implications":["Near-RT xApp designers can treat simple compiled classifiers as a negligible-cost inline step rather than an external service.","Model choice for this class of state classifiers can prioritize auditability and C-deployability over marginal accuracy gains.","Reproducible commodity-hardware stacks become a practical path for measuring AI inside the RIC loop without dedicated testbeds.","Future closed-loop control can start from the measured classification signal once E2SM-RC actuation is added."],"fun_headline_variants":["AI inference runs 1–25 µs inside Near-RT RIC xApp","Compiled C models hit 1–5 µs AI in live O-RAN xApp","Near-RT RIC embeds AI classifying states under 25 µs","Live testbed AI meets 10 ms budget with <4 ms service","Deterministic C LR/MLP enable AI in OAI FlexRIC xApp"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that lightweight AI can operate inside Near-RT timing rests on a rule-structured five-feature synthetic proxy task and a small RF-simulation testbed; if real traffic or larger models change the latency mix, the feasibility result may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["AI inference runs 1–25 µs inside Near-RT RIC xApp","Compiled C models hit 1–5 µs AI in live O-RAN xApp","Near-RT RIC embeds AI classifying states under 25 µs","Live testbed AI meets 10 ms budget with <4 ms service","Deterministic C LR/MLP enable AI in OAI FlexRIC xApp"]},"model":"grok-4.5","effort":"low","cost_usd":0.005136,"raw_usage":{"total_tokens":1542,"prompt_tokens":933,"num_sources_used":0,"completion_tokens":104,"cost_in_usd_ticks":51360000,"prompt_tokens_details":{"text_tokens":933,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":505,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":933,"tokens_out":104,"duration_ms":4653,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T08:41:52.825391+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace the synthetic proxy with live multi-cell E2 indications under realistic load, re-export the same C models (or a modestly larger one), and re-measure the end-to-end latency CDF: if more than a few percent of loops exceed 10 ms or inference ceases to be negligible relative to E2 transport, the central feasibility claim fails for that workload.","supporting_citations":[],"review_version":2}