{"id":"7061d092-2fd7-46f6-920d-62ad1fe0a41a","arxiv_id":"2607.18273","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":10,"one_line_summary":"A per-attempt linear SVR ensemble predicts CoAP retransmission timeouts on-device, and the authors report 97.25% packet delivery with a 768-byte model in simulation.","lead":"prCoAP replaces the CoAP protocol's fixed retransmission timers with a tiny per-attempt support-vector regression model trained on simulator data. It reports higher packet delivery than standard CoAP, but the evidence is weakened by in-simulation training, missing ML baselines, and unsupported numeric claims.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on a training label y that is never given an independent definition (Sec. II-B: 'assigned for that attempt in the simulator'); if y is just the simulator's existing RTO, the SVR is not learning a better timeout policy, and the abstract's 97.25% PDR / 37% goodput gain conflict","rationale":"The most load-bearing condition for the central claim is that the SVR is trained toward a target representing better timeout behavior. The paper never defines such a target: Sec. II-B says only that y is 'assigned for that attempt in the simulator,' which is circular unless the simulator is using an independent oracle or an explicit optimality criterion. If y is just the heuristic RTO already in use, the learned model is not a new policy and the comparison against CoAP/CoCoA is not a fair test of prCoAP. This is a correctness risk, not merely a disagreement with consensus. It is also independently supported by the paper's own limitation statement in the Conclusions and by the internal inconsistency between the abstract's claimed 97.25% PDR / 37% goodput gain and the lower goodput and higher latency shown in Tables III and IV. The reader's conditional verdict is appropriate: the design is plausible, the per-attempt R2 table shows internal fit, and the code is promised, so the issue is addressable. My concern focuses on the label definition, which is the part of the reader's weakest assumption that I find most decisive; I only partially agree because the reader also emphasizes the missing real-hardware validation, which I view as secondary to the label problem.","tokens_in":7823,"tokens_out":5678,"duration_ms":67957,"concrete_test":"Checkout the released GitHub repository; locate where the label y in Sec. II-B is computed. Add a diagnostic that records, for each training sample, the RTO that baseline CoAP/CoCoA used at that attempt. Compute the correlation/scatter between y and that baseline RTO. Then retrain the per-attempt SVR ensemble with y set to that baseline RTO (same features, same caps, same 80/20 split) and re-run the evaluation for Table III / Fig. 2. If correlation is ~1, or if the baseline-labeled model matches prCoAP's PDR/goodput, the supervisory signal carries no independent information and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II-B defines the regression target only as 'the attempt-level target RTO in milliseconds assigned for that attempt in the simulator after applying the attempt cap S_a.' No equation or optimality criterion specifies how y is generated; it could be the RTO already produced by the baseline CoAP/CoCoA logic in the same simulation. If so, Eq. (7) is trained to imitate the status quo, and the subsequent evaluation in the same simulator cannot separate learned timeout skill from simulator artifacts. The paper's own conclusion concedes 'training and most evaluation rely on simulator-generated data' and lists real-hardware evaluation as future work, despite the abstract claiming FIT IoT-LAB validation. The performance evidence is also internally inconsistent: Table III shows SVR Linear and SVR Linear+BO goodput below CoAP/CoCoA/CoCoA+ at every client count (e.g., 8326.8 vs 8734.7 at 150 clients), and Table IV shows higher latency for all ML variants, while the abstract/conclusion assert 97.25% PDR and a 37% goodput gain under burst traffic. No table in the paper reports 97.25%.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes prCoAP, a lightweight per-attempt linear Support Vector Regression ensemble that predicts a CoAP Retransmission Timeout directly from five node-observable features, together with a calibrated Random Forest classifier that early-abandons likely-failing transactions from the third retransmission onward. The authors evaluate three SVR variants against CoAP, CoCoA, and CoCoA+ in a discrete-event IEEE 802.15.4 simulator and report per-attempt R^2 values, goodput, latency, energy, and sensitivity results. The abstract claims 97.25% PDR and validation on FIT IoT-LAB, and the conclusion claims a 37% goodput gain under burst traffic. The body, however, does not report these numbers, the regression label y is not independently defined, and the claimed testbed validation is contradicted by the paper's own conclusion listing real-hardware evaluation as future work.","tokens_in":8211,"tokens_out":4746,"duration_ms":50097,"significance":"If the empirical claims were supported, the contribution would be useful: an on-device RTO predictor with a 768-byte parameter footprint and O(d) inference is a plausible fit for low-end IoT devices, and an attempt-stratified linear ensemble is a sensible way to handle the wide RTO range. The authors also open-source their simulations, which supports reproducibility. However, the central evidence as presented is not reliable: the training target is undefined, the headline performance numbers are absent from the body and partly contradicted by the tables, and the claimed FIT IoT-LAB validation does not appear anywhere in the manuscript. The scientific value of the paper therefore cannot be assessed until these load-bearing issues are resolved.","major_comments":[{"comment":"The regression label y is never independently defined. The text states only that y is 'the attempt-level target RTO in milliseconds assigned for that attempt in the simulator after applying the attempt cap S_a.' No equation, optimization criterion, or generator description specifies how this target is produced. If y is simply the RTO already generated by the baseline CoAP/CoCoA logic in the same simulator, then Eq. (7) is trained to imitate the status quo and the subsequent comparison cannot establish that the SVR learns a better timeout policy. This is a load-bearing gap: the central claim of the paper depends on y being a meaningful supervisory signal rather than an artifact of the simulator.","section":"Section II-B, Eqs. (6)-(7)"},{"comment":"The abstract claims 97.25% PDR and the conclusion claims a 37% goodput gain under burst traffic, but no table or figure in the body reports either number. Table III shows SVR (Linear) goodput below CoAP, CoCoA, and CoCoA+ at every client count (e.g., 8326.8 vs. 8734.7 B/s at 150 clients), and Table IV shows the SVR variants have higher mean latency than all baselines. The PDR figures are presented without numeric values, and the 'Deadline miss rate' figure is not supported by a table. The headline performance claims are thus unsupported by, and partially contradicted by, the quantitative results that are reported.","section":"Abstract and Section III-B, Tables III-IV"},{"comment":"The abstract states that the approach is 'validate[d] against the FIT IoT-LAB testbed,' and Section I lists 'real-hardware experiments on FIT IoT-LAB' as a contribution. Section III contains no FIT IoT-LAB results, methodology, or hardware details. Section IV's conclusion explicitly says that 'training and most evaluation rely on simulator-generated data' and that 'in the future, we will ... evaluate on physical hardware at scale.' This is a direct internal contradiction: the claimed testbed validation is absent from the evaluation and is stated to be future work. The authors must either provide the testbed experiments or remove the validation claim.","section":"Abstract / Section I / Section IV"},{"comment":"The evaluation does not establish generalization outside the training distribution. Both training and evaluation data are drawn from the same 300-scenario simulator sweep, with an 80/20 split stratified by attempt and channel regime. This makes the reported held-out R^2 and system-level metrics in-distribution interpolation, not evidence that the learned policy would transfer to real hardware or to channel dynamics not represented in the simulator. The paper's own limitation statement acknowledges this, but the abstract's 'outperforming standard CoAP under the evaluated conditions' is framed as a validated result. A cross-distribution or hardware evaluation, or at minimum a clearly specified held-out scenario set, is needed before the central claim can be accepted.","section":"Section II-B / Section III"}],"minor_comments":[{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the quantitative results. Given the small differences between some protocols, it is impossible to judge whether the reported differences are meaningful.","section":"Tables III-V"},{"comment":"The classifier's ROC-AUC of 0.7186 and 0.957 precision are reported as single numbers without cross-validation details or a confusion matrix. The 'residual false-drop probability of 0.01' injected at inference time is also not specified as an implementation mechanism or a sensitivity parameter; please clarify how this is applied and whether it affects the reported PDR/latency results.","section":"Section II-B, RF drop classifier"},{"comment":"Attempt 5+ has only 615 samples and R^2 = 0.2856. The text mentions this, but the abstract's general 'per-attempt linear SVR ensemble' claim would benefit from a more explicit caveat that the late-attempt sub-models are essentially untrained.","section":"Table V"},{"comment":"The success-rate decay factor of 0.90 per 2 s idle interval and the window W = 10 are introduced without a sensitivity analysis. Given that SR is a feature for both the SVR and the drop threshold tau(SR), a brief sensitivity discussion would help.","section":"Section II-A"},{"comment":"Several figures (Figs. 2, 3, 5, 8) are discussed qualitatively but do not report the underlying numeric values in the text or tables. Adding the plotted values in tables or captions would strengthen the reproducibility of the paper.","section":"Section III-B, Figures"}],"recommendation":"reject","confidential_remarks":"The manuscript's central claims are not supported by the body: the key performance numbers are missing, the training label is undefined, and the claimed FIT IoT-LAB validation is contradicted by the conclusion. These are not merely presentational issues; they undermine the paper's main assertions. Should the authors be invited to resubmit, they would need to provide a concrete definition of the training target, add the missing testbed results or remove the claim, reconcile the headline numbers with the tables, and evaluate outside the training distribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This one has a sensible design and a set of claims that outrun the reported data. If you read the body tables, the proposed SVR variants show worse goodput and higher latency than standard CoAP at every client count. The abstract's 97.25% PDR and 37% goodput gain are nowhere in the evaluation section. That, plus an undefined training label, makes the central result unverifiable as written.\n\nWhat's genuinely new is the specific combination: a per-attempt linear SVR ensemble, a log-domain target to handle the skewed RTO distribution, and a calibrated random forest that aborts likely-failed transactions after the third attempt. That's a sensible application of standard ML to CoAP congestion control, and the authors are transparent about the 768-byte footprint and the modest RF AUC (0.7186). The per-attempt R2 table is useful, and including a kernel SVR as an accuracy upper bound is a good check.\n\nThe soft spots are the ones that matter. First, the regression target y is defined only as \"the attempt-level target RTO ... assigned for that attempt in the simulator.\" There is no equation or criterion. If that target is just the simulator's existing RTO computed by the baseline logic, then the SVR is trained to imitate the status quo, and the later comparison collapses. Second, training and evaluation share the same simulator distribution; the contribution list and abstract claim FIT IoT-LAB validation, but the manuscript contains no testbed results, and the conclusion then admits \"training and most evaluation rely on simulator-generated data\" and lists real-hardware evaluation as future work. That is a direct contradiction in scope. Third, the headline PDR/goodput numbers are not merely missing; they are inconsistent with the body tables. That has to be fixed before any referee can take the performance claim seriously.\n\nThe underlying idea is not bad. The paper would be a reasonable candidate for peer review if the authors define the label, show either real-hardware data or remove those claims, and reconcile the abstract with the body. Send it to a serious referee with a request for major revision; it is not ready to be accepted, but it should not be desk-rejected without the chance to fix these issues.","headline":"Sensible design, unsupported headline numbers: the paper's body tables contradict its abstract, and the training label is undefined.","tokens_in":8644,"tokens_out":2979,"would_cite":false,"duration_ms":29905,"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":"Replacing CoAP's heuristic retransmission timers with a per-attempt linear SVR ensemble improves packet delivery on constrained IoT nodes, reaching 97.25% PDR under simulated lossy links.","keywords":["CoAP","congestion control","retransmission timeout","support vector regression","random forest","Internet of Things","constrained devices"],"falsifier":"Run the same prCoAP implementation on a real low-power wireless testbed with bursty link errors, comparing its packet delivery ratio, goodput, and energy per packet against standard CoAP and CoCoA+; if prCoAP does not outperform standard CoAP or if its PDR/goodput advantage disappears, the central claim is falsified, since the simulator-to-hardware transfer is the load-bearing assumption.","tokens_in":7722,"feed_emoji":"📡","tokens_out":2917,"duration_ms":31566,"temperature":0.7,"pith_summary":"The paper tries to show that CoAP congestion control can be made adaptive without heavyweight computation by replacing its fixed retransmission timeout (RTO) heuristics with an on-device machine learning predictor. It proposes a per-attempt ensemble of linear support vector regressors that predicts an absolute RTO from five sender-observable features, plus a random-forest drop classifier that aborts likely-failing transactions early. The authors report 97.25% packet delivery ratio, outperforming standard CoAP under the evaluated simulator conditions, while keeping the model footprint at about 768 bytes and inference to a single dot product. They also report that the more accurate kernel SVR variant consumes 84.6% more energy without a system-level benefit, arguing that linear models are the right trade-off for constrained devices.","feed_headline":"768-byte SVR model lifts CoAP packet delivery to 97%","feed_subtitle":"Per-attempt linear regression replaces binary exponential backoff, cutting energy waste on low-end IoT nodes.","key_machinery":"The core mechanism is a per-attempt linear SVR ensemble: for each retransmission attempt index (0 through 6), a separate linear support vector regressor predicts the log-transformed RTO from a five-feature vector (smoothed RTT, RTT variance, attempt index, inter-arrival time, short-term success rate). Features are standardized, the target is log-transformed, and predictions are bounded by BEB-aligned caps per attempt, which narrows the target range each learner sees. Inference is a single dot product over five features, costing O(d) and about 768 bytes of stored parameters. A separate calibrated random-forest classifier, activated from the third retransmission, outputs a calibrated failure p","core_discovery":"The central claim is that CoAP's heuristic RTO selection—random initial timeout, binary exponential backoff, and RTT-based variants like CoCoA or CoCoA+—can be replaced by direct, per-attempt prediction of an absolute RTO using a linear SVR ensemble. The model is trained offline on simulator-generated data, stratified by attempt index, and uses a log-domain target transform to handle the skewed timeout distribution. Inference is a single dot product over five standardized features, enabling on-device deployment within memory and energy budgets. A calibrated random-forest classifier, active from the third retransmission attempt, provides early transaction termination with a success-rate-adapt","pith_inferences":["The training labels are the simulator's own assigned target RTOs, so the model is effectively learning to reproduce a simulator's timeout policy; whether that policy is truly optimal or even representative of real hardware remains untested, since the claimed real-hardware validation is absent from the body.","The reported 97.25% PDR and 37% goodput gain under burst traffic are simulator results conditioned on specific channel models and traffic patterns; a reader should expect these numbers to shift on real wireless links with different burst statistics and hardware timing.","A natural testable extension would be to compare prCoAP against CoCoA+ on physical hardware under controlled burst and idle conditions, measuring not just PDR but also energy per delivered packet and tail latency, to see whether the early-drop classifier and SVR predictions actually reduce abort-related waste.","The adaptive threshold formula (0.75/0.88/0.92 based on success rate) is a hand-tuned heuristic layered on top of the ML components; a more principled approach might tune this threshold jointly with the regression or replace it with a cost-sensitive decision rule."],"forward_implications":["If the linear SVR ensemble works as claimed, CoAP stacks on low-end microcontrollers can adopt on-device ML without needing cloud offload or specialized hardware.","Per-attempt model decomposition with BEB-aligned caps appears to be a viable way to make simple linear predictors competitive with nonlinear ones for skewed regression targets.","The calibrated early-drop policy could reduce channel occupancy and energy waste in lossy IoT networks without significantly hurting delivery, if the reported precision of 0.957 on the timeout class holds.","The conclusion that kernel SVR's 84.6% energy overhead is not justified suggests a general design principle: regression accuracy gains must be weighed against deployment cost on constrained devices."],"fun_headline_variants":["Linear SVR replaces CoAP backoff, hits 97% delivery","On-device SVR predicts RTO, boosts CoAP to 97% PDR","Per-attempt SVR beats BEB in CoAP congestion control","CoAP gets adaptive RTO via lightweight SVR ensemble","Predictive RTO with SVR lifts IoT packet delivery"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's central claim rests on the assumption that the simulator-assigned 'target RTO' used as the training label is a valid supervisory signal for good timeout behavior, and that a model trained on a 300-scenario simulator sweep will transfer to real hardware—yet the claimed real-hardware validation is not reported.","fun_headline_variants_meta":{"raw":{"variants":["Linear SVR replaces CoAP backoff, hits 97% delivery","On-device SVR predicts RTO, boosts CoAP to 97% PDR","Per-attempt SVR beats BEB in CoAP congestion control","CoAP gets adaptive RTO via lightweight SVR ensemble","Predictive RTO with SVR lifts IoT packet delivery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2789,"prompt_tokens":791,"completion_tokens":1998,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":1919}},"tokens_in":535,"tokens_out":1998,"duration_ms":13232,"temperature":1.0,"reasoning_tokens":1919,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T11:25:16.639741+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same prCoAP implementation on a real low-power wireless testbed with bursty link errors, comparing its packet delivery ratio, goodput, and energy per packet against standard CoAP and CoCoA+; if prCoAP does not outperform standard CoAP or if its PDR/goodput advantage disappears, the central claim is falsified, since the simulator-to-hardware transfer is the load-bearing assumption.","supporting_citations":[],"review_version":1}