{"id":"d0e28c81-cefb-456c-8a95-420ddcbe6c31","arxiv_id":"2507.00403","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A three-qubit Qiskit circuit encodes hand-set priors and conditional probabilities for network spikes, vulnerabilities, and false alarms, and statevector post-selection recovers the same Bayesian posteriors that went in.","lead":"The authors encode a small Bayesian network for intrusion detection into a three-qubit quantum circuit and read out joint and conditional probabilities by simulating the statevector. The demonstration is self-contained, but it only reproduces the hand-set probabilities it encodes, without real data, code, or a demonstrated quantum advantage.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed exact encoding of the Bayesian network by the RY/CRY recipe is unverified and plausibly overwrites or mis-encodes conditional structure; the reported 2-3% classical mismatch directly undermines the central inference claim.","rationale":"The reader's weakest assumption is that the circuit construction in Sections 4.1-4.3 exactly reproduces the user-defined Bayesian network. This is indeed the single most load-bearing condition for the paper's central claim: if the encoding is not exact, the extracted joint, marginal, and conditional probabilities are not the Bayesian inferences the paper says they are. My concern is more specific than the reader's summary but points to the same step. The preparation of X and Y as independent products in Section 4.1 restricts the representable parent joint, and the CRY recipe for CPT rows is not verified for exactness, especially for multi-row CPTs acting on the same target qubit. The paper's own fidelity statement is evidence against exactness: a statevector computation with symbolic post-selection should match a classical Bayes calculation to machine precision if both use the same model, so a 2-3% gap indicates either a different circuit or a different reference model. No code, parameters, or explicit circuit are provided, so the discrepancy cannot be resolved from the manuscript. I agree with the reader that REJECT is the appropriate verdict; since my reading does not alter that verdict, I set verdict_should_be to UNCHANGED. The proposed test would settle the technical question directly: if a faithful reconstruction matches classical Bayes exactly, the rejection would rest on other grounds such as missing reproducibility and internal figure inconsistencies; if it does not, the central claim fails on its own terms.","tokens_in":9815,"tokens_out":4959,"duration_ms":54748,"concrete_test":"Reproduce the Section 5 benchmark from the paper's own recipe: fix explicit priors P(X), P(Y) and a CPT P(FA|X,Y) matching the reported numbers, build the circuit with RY on X and Y followed by controlled rotations for each CPT row, simulate with Qiskit Statevector, and compare every extracted conditional P(Y=1|X=1), P(FA=1|X=1), P(FA=1|X=0), and the full 8-bitstring joint against classical Bayes computed from the same CPT. If any probability differs by more than 1e-10 (floating-point), the claimed exact encoding fails; if any differs by the reported 2-3%, it also fails. Then repeat with a model where X and Y are dependent, e.g., by adding a Y-to-X edge, using the same RY-marginals-then-CRY recipe; if the parent joint remains P(X)P(Y), the method cannot represent that intended network.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the claim in Sections 4.1-4.3 that the assembled RY/CRY circuit exactly realizes the user-defined Bayesian network. The paper never gives the full circuit or an algebraic proof; instead it relies on a recipe that is only correct under restrictive conditions. In Section 4.1, X and Y are prepared as independent marginals via RY, so the parent joint is forced to be P(X)P(Y). If the intended causal structure includes any correlation between X and Y, as suggested by the Section 5 phrase about a strong dependency between X and Y and by the scenario text saying a spike can be a side effect of a vulnerable system, then later CRY gates cannot repair the parent joint without changing the target state. Even in the independent-parent case, encoding each row of P(FA|X,Y) requires controlled rotations active on exactly one control configuration; the paper does not show whether open controls or multi-controlled decompositions were used, and it does not show that successive CRY rotations on the same target qubit do not overwrite previously encoded rows. The reported \"alignment within 2-3%\" in Section 5 is the decisive symptom: if the encoding were exact, statevector post-selection on a simulated pure state would agree with classical Bayes to floating-point precision. A 2-3% systematic discrepancy means either the circuit is not the claimed CPT or the classical reference does not match the model. Without the circuit diagram, gate ordering, or parameters, the central claim that the method yields joint, marginal, and conditional probabilities aligned with causal structure is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quantum Bayesian inference pipeline for an intrusion detection scenario. Three binary variables (network spike X, system vulnerability Y, and false alarm FA) are mapped to qubits; priors are encoded with RY rotations, conditional dependencies with controlled CRY rotations, and the joint distribution is extracted from a simulated statevector. Posteriors are computed by symbolic post-selection and renormalization. The authors report joint, marginal, and conditional distributions and claim agreement with classical Bayesian calculations within 2-3%, concluding that the method demonstrates feasible and interpretable quantum-native inference for security applications.","tokens_in":10110,"tokens_out":6031,"duration_ms":66755,"significance":"If the encoding were exact and the validation were consistent, this would be a clean pedagogical demonstration of state-preparation for a three-variable Bayesian network and of posterior extraction by post-selection. The paper correctly explains the standard RY/CRY construction and the relation between amplitudes and probabilities, and the symbolic post-selection idea is clearly stated. However, the manuscript offers no new algorithmic or theoretical insight beyond textbook state preparation, and the results as reported are internally contradictory. The lack of a full circuit specification, the unexplained 2-3% discrepancy in a noiseless statevector simulation, and the mutually inconsistent figures mean the central claim of correct and faithful Bayesian inference is not supported.","major_comments":[{"comment":"The reported results are mutually inconsistent. Figure 8 and Figure 9 state that P(X=0,Y=1)=0.59 is the dominant cell, whereas Figure 4 states that the 2-bit outcome '01' is 'nearly suppressed' and '00' occurs with highest frequency; if these figures refer to different variable pairs or different simulation runs, that is not disclosed. In addition, Figure 1 indicates P(FA=1|X=0) approaches 1, whereas Figure 10 gives P(FA=1)≈0.93 and P(X=0)≈0.67, which cannot be reconciled with a 'more balanced' conditional at X=1 unless P(FA=1|X=1) is also very high. Since the paper's conclusion that 'the inference engine correctly captured conditional causality' rests on these visualizations, the contradictions invalidate the reported validation.","section":"§5 and Figures 8–10"},{"comment":"Section 4.1 explicitly prepares X and Y as independent marginals with RY rotations and states that at this stage no entanglement is present; no later operation on the X or Y qubits is described anywhere in Section 4. Under that construction the final state must have P(X,Y)=P(X)P(Y) exactly. Section 5 nevertheless claims a 'strong dependency was encoded between X and Y', and Figure 8 reports P(X=0,Y=1)=0.59 and P(X=1,Y=1)=0.26. With the Figure 10 marginals P(X=0)≈0.67 and P(Y=1)>0.85, the product distribution would give values close to these reported cells, but not equal to them. The paper neither supplies the circuit that creates the extra dependency nor quantifies this discrepancy, and the text is therefore internally contradictory about the model's central structure.","section":"§4.1 and §5"},{"comment":"The paper reports that cross-validating simulated posteriors with classical conditional calculations confirms 'alignment within 2-3%' in all benchmark cases. In a noiseless statevector simulation, if the circuit exactly implemented the stated priors and CPT, the extracted probabilities should agree with classical Bayes to floating-point precision. A systematic 2-3% discrepancy implies either that the circuit does not implement the claimed Bayesian network or that the classical reference does not match the model. This directly undermines the abstract's claim that the method 'yields joint, marginal, and conditional probabilities aligned with causal structure'. The full circuit, the gate parameters, and the classical reference computation must be provided before this claim can be assessed.","section":"§5, 'Additional fidelity tests'"},{"comment":"The encoding of the CPT for P(FA|X,Y) is not specified to the level needed to verify correctness. The text says that multi-controlled gates 'are decomposed into Toffoli-based networks' and that 'direct implementation is often used', but it does not give the circuit, the control conditions for rows where a parent equals 0, the ordering of the controlled rotations, or an algebraic proof that the final target-qubit amplitude for each control configuration equals the square root of the corresponding CPT entry. Since this encoding is the load-bearing step of the method, the absence of a verifiable construction is a serious gap.","section":"§4.2–4.3"}],"minor_comments":[{"comment":"The title contains a typo: 'Suport' should be 'Support'.","section":"Title"},{"comment":"Figure 10 is referenced in the text before Figure 9, and the figure numbering appears out of order; the manuscript should renumber the figures consistently.","section":"§5"},{"comment":"The mathematical notation is sometimes corrupted in the rendering, e.g., the summation limits appear as '2n−1X' rather than a proper sum; the authors should ensure all equations typeset correctly.","section":"§2"},{"comment":"The statement that classical Bayesian updating 'irreversibly conditions on evidence' is conceptually misleading; classical Bayesian updating does not destroy the prior distribution, and conditioning is not an irreversible physical operation.","section":"§4.5"},{"comment":"The paper repeatedly reports qualitative behavior ('often holding the highest probability amplitude', 'consistently peaked above 0.6') without giving a table of the specific priors, CPT entries, rotation angles, and resulting probabilities for a single run; such a table would make the results reproducible and checkable.","section":"§5"}],"recommendation":"reject","confidential_remarks":"The manuscript is closer to a lab notebook than a completed research paper. The central encoding claim is unverifiable from the text, the reported experiments are internally inconsistent, and no code or parameter values are provided. The paper also does not engage with the existing quantum Bayesian inference literature beyond citing it. I would not encourage resubmission unless the authors can provide an executable implementation, a single consistent simulation run with all parameters, and a circuit diagram or algebraic proof of the encoding."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a RY/CRY encoding of a three-variable Bayesian network, simulated with statevector and read out by post-selection. That method is already in the refs [11,12], so there is no new algorithm. The paper is honest that it is exploratory and pedagogical, and the mathematical framework is standard and correct. The toy intrusion scenario is plausible, and the step-by-step construction is clear enough to follow.\n\nThe problems are load-bearing. First, the results contradict each other. Figure 4 says the 2-bit outcome '01' is nearly suppressed, while Figure 8 gives P(X=0,Y=1)=0.59 as the highest cell in the P(X,Y) heatmap. Those cannot both be true for the same circuit. Second, the text claims a strong dependency was encoded between X and Y, but the construction in Sections 4.1-4.3 only applies independent RY rotations to X and Y and then CRY gates to FA. X and Y remain independent under that circuit. The heatmap values are actually close to the product of the reported marginals, so the claim of encoded dependency is simply wrong. Third, the reported 2-3% divergence between statevector posteriors and classical Bayes is a red flag: with exact amplitudes and no measurement noise, agreement should be at floating-point precision unless the circuit does not match the CPT or the classical reference is computed from different numbers. The paper never gives the circuit diagram, gate ordering, or numerical parameters, so none of this can be checked. Fourth, the fidelity test is circular: they set the priors and CPTs, encode them, then compare the output to classical Bayes applied to those same hand-set inputs. That is self-consistency, not validation.\n\nOn the positive side, the paper is not confused about the basic quantum formalism, and the symbolic post-selection procedure is described accurately. If the authors cleaned up the contradictions, provided code and exact parameters, and repositioned the paper as a tutorial rather than a research result, it could have some value for students. As submitted, the internal inconsistencies break the central claim that the inference engine correctly captured conditional causality.\n\nI would not bring this to a reading group, and I would not cite it. For peer review: this should be desk-rejected. It does not deserve referee time as a research contribution. If the venue is a workshop or an educational track, a conditional accept after major revision might be defensible, but not as a serious research paper.","headline":"A textbook three-qubit Bayesian network encoding with no new technique, internally inconsistent results, and no code; the central claim of correct inference is unsupported.","tokens_in":739,"tokens_out":912,"would_cite":false,"duration_ms":41307,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a three-qubit circuit whose gates encode prior and conditional probabilities can reproduce the joint, marginal, and conditional distributions of a three-variable intrusion-detection Bayesian network, with posterior…","keywords":["quantum Bayesian inference","intrusion detection","statevector simulation","RY rotation encoding","CRY controlled rotations","posterior inference","symbolic post-selection","Bayesian network encoding"],"falsifier":"Run the described construction with explicit values for $P(X)$, $P(Y)$, and every entry of $P(FA|X,Y)$; enumerate the same network classically and compare each of the eight squared amplitudes to the product $P(X)P(Y)P(FA|X,Y)$. If any entry departs from the classical value by more than the reported 2–3 percent, or if the final marginal $P(X=1)$ differs from the originally encoded prior, the encoding is not faithful and the extracted posteriors are not genuine Bayesian updates.","tokens_in":9598,"feed_emoji":"🛡️","tokens_out":10172,"duration_ms":107617,"temperature":0.7,"pith_summary":"The paper tries to show that Bayesian inference for a three-variable intrusion-detection problem can be carried out inside a quantum circuit: prior probabilities are placed on qubits with $R_Y$ rotations, conditional probabilities are applied with controlled $R_Y$ rotations, and the full joint distribution is read from the squared amplitudes of a simulated statevector. Conditioning on evidence amounts to symbolic post-selection, filtering the consistent basis states and renormalizing, which the paper argues reproduces classical Bayes updates. The potential payoff is that the entire probability landscape stays visible and inspectable, with no measurement collapse and no hidden layer, so a security analyst could trace how evidence propagates from a traffic spike through vulnerability to a false alarm. If the construction is faithful, the same modular encoding could be reused for other Bayesian networks and other decision domains.","feed_headline":"Three qubits run Bayesian inference for intrusion detection","feed_subtitle":"RY and CRY rotations set priors and conditionals; post-selection on amplitudes yields posteriors.","key_machinery":"The carrying mechanism is the controlled $Y$-rotation encoding of a conditional probability table. A prior $p$ maps to the rotation angle $\\theta = 2\\arcsin(\\sqrt{p})$ so that the squared amplitude of $|1\\rangle$ equals $p$, and each conditional entry $P(\\text{child}|\\text{parents})$ becomes a $CR_Y$ gate whose control qubits are the parent variables. Because all gates are unitary and the circuit is never measured, the final statevector stores the entire joint distribution in its squared amplitudes. Symbolic post-selection, which filters basis states matching the evidence and renormalizes, plays the role of the Bayes update without collapsing the state; this is what allows many conditional queries to be answered from one simulation.","core_discovery":"At the center of the paper is a three-variable network: network spike $X$, system vulnerability $Y$, and false alarm $FA$. The marginals $P(X)$ and $P(Y)$ are encoded first with single-qubit $R_Y$ rotations using $\\theta = 2\\arcsin(\\sqrt{p})$, and the false-alarm qubit receives controlled rotations that encode the entries of $P(FA|X,Y)$. Simulating the circuit yields an eight-amplitude statevector, and the squared moduli are taken as the joint distribution $P(X,Y,FA)$. Conditioning on evidence such as $X=1$ is performed by summing squared amplitudes over matching basis states and renormalizing; the paper reports posterior values such as $P(Y=1|X=1)=0.68$ and $P(FA=1|Y=1)=0.74$, with the bitstring $111$ dominating the distribution. Fidelity checks against classical conditional calculations are reported to agree within 2–3 percent, which the paper presents as evidence that the circuit captures the intended conditional causality.","pith_inferences":["Editorial extension: the paper never verifies that the final marginals of $X$ and $Y$ match the priors it encoded, so the reader cannot rule out that later controlled rotations have overwritten part of the intended distribution; a direct comparison of initial and final marginals would settle this.","Editorial extension: the 2–3 percent fidelity gap implies the encoding is approximate even in the paper's own benchmark, so the central claim should be read as approximate Bayesian reproduction rather than exact amplitude encoding.","Editorial extension: the same pipeline is easy to test on a product-state CPT; if $FA$ is made independent of $X$ and $Y$, the final state should remain separable, and appearance of entanglement would mean the circuit introduces spurious causal dependence.","Editorial extension: because statevector simulation stores $2^n$ amplitudes, the method's practical value for larger networks depends on moving to real hardware or approximate sampling, neither of which the paper demonstrates."],"forward_implications":["Any Bayesian network with known conditional probability tables can be assembled from the same primitives, since every CPT entry becomes one controlled rotation and the joint distribution is recovered by squaring amplitudes.","Because conditioning is done symbolically on a single statevector, one simulation answers arbitrarily many conditional queries without re-running the circuit.","The reported concentration of probability mass on a few bitstrings, with $111$ dominant in high-risk settings, gives analysts a sparse, inspectable explanation of how evidence propagates through the network.","The stated 2–3 percent agreement with classical conditional calculations sets a concrete accuracy benchmark for future circuit-based Bayesian inference methods."],"supporting_citations":[],"fun_headline_variants":["Quantum Bayes on three qubits for intrusion detection","Three qubits compute posterior probabilities for network spikes","Post-selected RY rotations drive quantum intrusion inference","Quantum-native Bayesian inference for security monitoring","Small quantum circuit performs Bayesian intrusion analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the circuit's sequence of $R_Y$ rotations for priors followed by controlled rotations for conditionals exactly reproduces the analyst's Bayesian network, so that the squared amplitudes of the final statevector really are the intended joint distribution.","fun_headline_variants_meta":{"raw":{"variants":["Quantum Bayes on three qubits for intrusion detection","Three qubits compute posterior probabilities for network spikes","Post-selected RY rotations drive quantum intrusion inference","Quantum-native Bayesian inference for security monitoring","Small quantum circuit performs Bayesian intrusion analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1428,"prompt_tokens":821,"completion_tokens":607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":539}},"tokens_in":437,"tokens_out":607,"duration_ms":7208,"temperature":1.0,"reasoning_tokens":539,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:17:43.998203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the described construction with explicit values for $P(X)$, $P(Y)$, and every entry of $P(FA|X,Y)$; enumerate the same network classically and compare each of the eight squared amplitudes to the product $P(X)P(Y)P(FA|X,Y)$. If any entry departs from the classical value by more than the reported 2–3 percent, or if the final marginal $P(X=1)$ differs from the originally encoded prior, the encoding is not faithful and the extracted posteriors are not genuine Bayesian updates.","supporting_citations":[],"review_version":1}