{"id":"12eac2cf-3dea-43af-a5d7-bbf0f4353159","arxiv_id":"2607.16089","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"SQUIRO enforces security posture as hard scheduling constraints and ranks quantum backends by coherence margin and queue pressure, reporting up to 51% cost and 63% energy savings on synthetic underloaded Kubernetes clusters.","lead":"SQUIRO is a Kubernetes scheduling framework that treats security (post-quantum crypto, trusted hardware, fresh attestation) as a hard gate and picks quantum backends by circuit depth and queue state rather than error rate alone. On synthetic underloaded clusters it reports up to 51% lower cost and 63% lower energy versus a greedy baseline, but no code or data are released and the energy model makes savings follow directly from node consolidation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 51%/63% savings are computed from a binary active-node cost model (Eqs. 19–20), so they measure consolidation, not measured energy/cost; they may not survive realistic power curves or always-on clusters.","rationale":"The reader's weakest assumption identifies exactly the load-bearing issue: cost and energy are modeled as linear functions of the active-node set, so the E2 savings quantify consolidation rather than independently measured infrastructure effects. I agree with that diagnosis and make it the primary stress-test point. The security-by-construction claim (Eq. 27) is definitional and is empirically supported by E1's CSR=1; the backend-selection experiment E4 is explicitly self-flagged as not outcome-validated; and the paper is candid about scope and limitations. The only place the abstract's central claim can fail is the quantitative cost/energy part, because the objective function guarantees that activating fewer nodes yields lower cost/energy, making E2 a consistency check of the model rather than a measurement of physical cost or energy. The concern is concrete and addressable with a counterfactual simulation, and it does not undermine the structural USM/SDM/security-mask contributions. Therefore the CONDITIONAL verdict stands unchanged; the authors should be asked to run the proposed metric check or otherwise reword the headline savings as active-node-count savings under a linear accounting model.","tokens_in":25188,"tokens_out":5705,"duration_ms":61117,"concrete_test":"Re-run E2 under two counterfactual cost/power models while keeping placements identical: (A) all nodes remain powered/leased for the whole accounting interval, so y_j=1 for every node regardless of use; (B) nodes can be switched off, but with realistic idle power P_idle>0, utilization-dependent power P_j(t)=P_idle+(P_max−P_idle)u_j(t), and fixed shutdown/startup energy and delay per transition. Compute total cost and energy from the actual schedule, not from Σ y_j. If CP-SAT-balanced's advantage over K8s-greedy is ≈0 in (A) and below the reported 20–63% in (B), the abstract's headline numbers are artifacts of the y_j metric and should be reworded as model-accounting savings; if the savings persist under both models, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on E2, whose metrics C(y)=Σ κ_j y_j and E(y)=T Σ p_j y_j (Eqs. 19–20) are linear in the binary active-node variable y_j. The objective in Eq. (21) minimizes αC+βE, so any feasible schedule activating fewer nodes scores better on both metrics by construction. In the underloaded E2 regime, CP-SAT-balanced uses 8.4 active nodes versus 29.6 for K8s-greedy, and the reported 51%/63% savings are the direct accounting consequence of that count difference. The paper never states whether inactive nodes are deprovisioned/powered off or remain in an always-on cluster; it also omits idle power, load-dependent power draw, and power-state transition costs. If inactive nodes remain powered on—common for on-premises or managed Kubernetes clusters with daemons, minimum-size constraints, or standing capacity—cost and energy are nearly independent of placement, and the headline savings vanish. If nodes can be released, the savings depend on the power/cost model and time horizon, which are not specified. The current evidence therefore supports \"global packing reduces active-node count under a linear accounting model,\" not \"reduces infrastructure cost/energy\" in the unqualified sense of the abstract. A secondary gap: the prototype uses a uniform admission reward (Section VI B), so \"without compromising admission priorities\" is not demonstrated for differentiated priorities.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SQUIRO, a Kubernetes scheduling framework built on a Unified Scheduling Model (USM) and a six-step Scheduler Design Methodology (SDM). The authors separate hard security feasibility (Eq. 27) from residual-risk optimisation (Eq. 34), add a circuit-aware quantum backend selector with a colocation hierarchy, and report four synthetic experiments: E1 (security-mask compliance), E2 (cost/energy consolidation vs. greedy placement), E3 (CP-SAT scalability), and E4 (backend-rank divergence). The paper is unusually candid: it states that λR(x) is inactive in the main prototype, that M(x) is unimplemented, that the demand score diverges from Eq. (33), and that E4 has no outcome validation.","tokens_in":1664,"tokens_out":1609,"duration_ms":69312,"significance":"If the claims are accepted, the main contribution is a clean architectural separation between non-negotiable security requirements and gradational security preferences, expressed as a hard constraint rather than a score term; the global-packing result also gives a useful bound on consolidation headroom in underloaded clusters. The hard-mask construction (x_ij ≤ a^s_ij) genuinely makes any feasible schedule compliant with the modelled checks by construction. The paper is also commendable for explicitly stating its own implementation gaps. However, the headline cost/energy savings are computed under a binary active-node accounting model and therefore support a consolidation-accounting claim, not a measured operational-energy or infrastructure-cost claim. The residual-risk E1 result is, as reported, a consistency check on the solver rather than evidence about security value. The overall framework is a plausible design blueprint, but the quantitative headline claims need to be reframed or supported with additional measurement/modeling before publication.","major_comments":[{"comment":"The 51% cost and 63% energy savings are not measured infrastructure cost/energy savings; they are algebraic consequences of the model C(y)=Σκ_j y_j and E(y)=TΣp_j y_j. Since objective (21) minimizes αC+βE, any schedule activating fewer nodes scores better on both metrics by construction. In E2 underloaded, CP-SAT-balanced uses 8.4 active nodes vs. 29.6, and the reported savings are the direct accounting effect of that difference. The paper never states whether inactive nodes are deprovisioned/powered off, nor models idle power, load-dependent power, or power-state transition costs. If inactive nodes remain powered on, the savings largely vanish. Please either rephrase the abstract and conclusions as consolidation gains under a linear active-node model, or add sensitivity analysis with realistic power/cost curves and explicit node-lifecycle assumptions.","section":"§VI B, Eqs. (19)–(20); §VII E2"},{"comment":"The E1 'hard+soft' comparison does not validate the security fit score. R(x) is defined as Σ(1−Φ_ij/100)x_ij, and the reported 'mean Φ_ij rises from 81.0 to 84.7' is simply the solver increasing its own objective. That is a tautology, not evidence that the residual-risk term improves placement quality. The experiment only shows that CP-SAT can optimize the function it is given. Please replace or supplement the comparison with a baseline such as random feasible placement, or with an independent held-out security metric, and soften the conclusion's claim of a 'measurable 3.7-point security fit-score gain'.","section":"§VI B residual-risk note; §VII E1; Eq. (34)"},{"comment":"The abstract's claim 'without compromising admission priorities' is not supported. Section VI B states that 'the current CP-SAT prototype uses a uniform admission reward, with π_i held constant across workloads,' and E2's admission row is reported for CP-SAT-performance, not for a scheduler with differentiated priorities. No experiment uses workload-specific priority weights, so the paper cannot claim that admission priorities are preserved. Please either implement and test differentiated priorities, or reframe the claim as 'without reducing aggregate admission rate'.","section":"§VI B; §VII E2; Abstract"}],"minor_comments":[{"comment":"Typo in connectivity class description: 'hea vy-hex' should be 'heavy-hex'. Also, the sentence describing the omitted closed-loop pair from Fig. 5(d) should be in the figure caption or experiment setup rather than only in the results paragraph.","section":"§VI C"},{"comment":"The demand-score implementation is described as an additive pass with an unspecified 'compression factor' that diverges from the closed-form Eq. (33). The paper acknowledges this, but for reproducibility the exact formula should be provided (or Eq. (33) marked as aspirational).","section":"§VI D, Eq. (33)"},{"comment":"The caption mixes schedulers: the plot is labeled 'CP-SAT-balanced over K8s-greedy' but the admission row refers to 'CP-SAT-performance'. Clarify which scheduler is used for each row to avoid confusion.","section":"Fig. 5(b)"},{"comment":"The unimplemented M(x) locality/disruption term is listed in the objective (21) but never used in any experiment. This is disclosed, but the paper would be clearer if the evaluation explicitly stated that all experiments set δ=0.","section":"§VI B"},{"comment":"E4 demonstrates divergence between SQUIRO's score and naive error-rate ranking, but it does not show that the selected backend yields better fidelity or time-to-solution. The conclusion's phrase 'ranking backends by raw error rate alone is insufficient' is stronger than the evidence; it should be softened to 'can diverge under coherence- and queue-limited conditions' unless outcome validation is added.","section":"§VII E4 and Conclusion"},{"comment":"The result files are available 'upon reasonable request' but the experimental tooling is not released. Given that all four quantitative experiments depend on synthetic instances, releasing the generator and solver configuration would substantially increase reproducibility.","section":"Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper's candid statements about its own limitations are a genuine strength and should be encouraged. The main barrier is not correctness of the hard-mask construction but the mismatch between the abstract/conclusion wording and what the experiments actually measure. I would like the authors to address the cost/energy accounting issue and the residual-risk circularity before publication. The paper fits the journal's scope; no citation or novelty concerns beyond the usual need to position against existing Kubernetes scheduling work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First: ignore the 51%/63% headline until you check what it is counting. It is counting active-node count in a cost model where cost and energy are linear in the binary y_j variable. The paper's real contribution is structural: separating hard security requirements from optimizable residual risk, and adding coherence margin to backend selection. That part is worth your time.\n\nThe USM/SDM formalism is mostly Pinedo with new names—the four-layer taxonomy is fine but not new math. The security mask x_ij ≤ a^s_ij (Eq. 27) is the genuine piece: a single failed check makes placement infeasible regardless of objective. The paper is unusually candid about its own limits—λR(x) is inactive in the main prototype, M(x) is unimplemented, the demand score in the prototype diverges from Eq. (33), and E4 is explicitly not outcome-validated. That honesty earns it credit.\n\nNow the soft spots. E2's cost C(y)=Σκ_j y_j and energy E(y)=TΣp_j y_j are linear in the active-node set. So any optimizer that packs onto fewer nodes wins both metrics by construction. The reported savings are the accounting consequence of 8.4 vs 29.6 active nodes. The paper never states whether inactive nodes are deprovisioned or remain powered on. If they stay powered—typical for on-prem or managed clusters—the economic and energy claims largely disappear. The baseline is also a custom greedy, not stock kube-scheduler bin-packing, so the comparison is weaker than the abstract implies. And the E1 '3.7-point Φij improvement' is R(x) measuring its own objective; not circular, but not independent evidence.\n\nThe structural claims hold up. The framework is a reasonable design for hybrid quantum-classical orchestration, and the coherence quality factor (Eq. 26) captures something aggregate error rates miss. But the quantitative evidence is thin: synthetic instances, no released code or data, key parameter values unstated, and the main reported savings rest on a cost model that may not correspond to real infrastructure.\n\nWho should read this: people working on security-aware orchestration or hybrid quantum-classical scheduling. I'd send it to peer review—the structural contribution and the authors' candor give reviewers a clear path—but I'd expect major revision to clarify what the cost/energy metrics actually measure, replace or justify the baseline, and release artifacts. I would not cite the 51%/63% numbers; I would cite the hard-mask/residual-risk separation if I wrote about secure scheduling.","headline":"The framework's structural security model and coherence-aware backend scoring are the real contributions; the 51%/63% cost/energy savings are a consolidation artifact of a linear active-node cost model, not measured cost or energy.","tokens_in":26249,"tokens_out":3176,"would_cite":true,"duration_ms":27737,"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":"A hard security mask that excludes any non-compliant node before optimization, combined with global packing, can make regulated hybrid quantum-classical workloads compliant by construction while cutting cost and energy on underloaded cluste","keywords":["scheduling","Kubernetes","quantum-classical","security posture","hard constraints","energy efficiency","backend selection","CP-SAT"],"falsifier":"Measure real per-node idle and dynamic power on a live cluster and replay the underloaded experiment: if measured energy savings are far below 63%, the linear active-node energy model is the culprit. Also feed the security mask a node whose attestation certificate is expired but still listed as fresh; if the scheduler accepts it, 'complete compliance' is not actually enforced.","tokens_in":24873,"feed_emoji":"🔐","tokens_out":4525,"duration_ms":40980,"temperature":0.7,"pith_summary":"The paper argues that hybrid quantum-classical scheduling should treat security requirements as hard, non-negotiable constraints and optimize cost and energy globally across all pending workloads, rather than filtering one pod at a time. Its core claim is that a 'security mask' that excludes any node failing a mandated check makes compliance a property of the scheduler itself, not a downstream audit. On synthetic Kubernetes clusters, the resulting global optimizer is reported to cut infrastructure cost by up to 51% and energy by up to 63% compared with a greedy Kubernetes-style scheduler in underloaded conditions. The paper also claims that quantum backend choice should be circuit-aware, considering coherence margin, queue pressure, and calibration freshness, and shows that this ranking diverges systematically from ranking by raw two-qubit error rate.","feed_headline":"Security-aware scheduler enforces compliance, cuts energy 63%","feed_subtitle":"Hard feasibility mask rejects non-compliant nodes before cost optimization; global packing does the rest.","key_machinery":"The load-bearing mechanism is the hard security mask, a binary feasibility filter that is computed before any objective is evaluated and that makes non-compliant workload-node pairs impossible to select. Around it, the formulation uses binary active-node variables y_j with linear cost and energy functions C(y)=Σ κ_j y_j and E(y)=T Σ p_j y_j, turning the scheduling problem into a global packing problem solved with a constraint-programming solver. The backend selector adds a second hard-mask stage and a weighted score whose coherence quality factor q_coh = max(0, 1 − circuit_time/min(T1,T2)) penalizes backends whose coherence budget is exceeded by the circuit.","core_discovery":"The central discovery is that a strict separation between mandatory security constraints and optimizable preferences can be encoded directly into a scheduling instance. SQUIRO's security posture framework computes a mask as_ij = 0 for any workload-node pair that fails a mandatory requirement (PQC support, attestation freshness, TEE presence, compliance floor), and the solver constraint x_ij ≤ as_ij makes any such pair infeasible regardless of objective value. With feasibility settled, the remaining problem is a global packing problem over binary active-node variables y_j, minimizing cost C(y)=Σ κ_j y_j and energy E(y)=T Σ p_j y_j while maximizing admission. On synthetic clusters this yields","pith_inferences":["If real energy consumption scales with utilization rather than with node activation, the reported 63% energy saving will shrink substantially; the underlying packing benefit may still be real but the headline figure is a property of the linear active-node model.","The 'compliance by construction' claim holds only to the extent that the security metadata feeding the mask (attestation freshness, PQC capability, TEE presence) is accurate and current at scheduling time; stale or spoofed evidence would break the guarantee.","A natural test is to replay the cost-energy experiment on a live Kubernetes cluster with metered energy and cost; the synthetic-instance assumption is the main reason the headline numbers might not reproduce.","The coherence-quality factor is a heuristic proxy (the paper says it does not model transpilation or error mitigation); a stronger test would be noise-model simulation of the top-ranked backend versus the naive-error-rate backend for actual circuit fidelity."],"forward_implications":["Any workload whose security requirements are expressed as hard constraints is guaranteed compliant placement whenever a feasible solution exists; no post-hoc check is needed.","Global consolidation savings exist mainly where there is idle capacity: 51% cost and 63% energy in underloaded clusters, 20–28% in balanced loads, and near zero in packed or overloaded regimes.","The solver's time grows from about 0.3s at 10 nodes to about 21s at 75 nodes, with 60% of 75-node runs exhausting a 30s budget, so the approach as implemented suits small or incrementally-scheduled clusters.","Backend selection that ignores coherence margin and queue pressure can pick a backend that is strictly worse for a given circuit; the two-stage selector changes ranking accordingly."],"fun_headline_variants":["Security by construction: scheduler cuts cost 51%, energy 63%","SQUIRO: Enforce security, then optimize — energy down 63%","Hard security, then smart packing: cost -51%, energy -63%","Quantum-classical scheduler: compliance enforced, then optimized","SQUIRO masks non-compliant nodes, then cuts energy 63%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported cost and energy savings are computed from linear functions that charge every active node a fixed per-hour cost and fixed power draw, so any scheduler that turns off more nodes scores better by construction; the real-world savings depend on whether idle nodes actually dominate energy and cost as assumed.","fun_headline_variants_meta":{"raw":{"variants":["Security by construction: scheduler cuts cost 51%, energy 63%","SQUIRO: Enforce security, then optimize — energy down 63%","Hard security, then smart packing: cost -51%, energy -63%","Quantum-classical scheduler: compliance enforced, then optimized","SQUIRO masks non-compliant nodes, then cuts energy 63%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":2814,"prompt_tokens":771,"completion_tokens":2043,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":1944}},"tokens_in":515,"tokens_out":2043,"duration_ms":13888,"temperature":1.0,"reasoning_tokens":1944,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:23:36.576370+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure real per-node idle and dynamic power on a live cluster and replay the underloaded experiment: if measured energy savings are far below 63%, the linear active-node energy model is the culprit. Also feed the security mask a node whose attestation certificate is expired but still listed as fresh; if the scheduler accepts it, 'complete compliance' is not actually enforced.","supporting_citations":[],"review_version":1}