{"id":"8bed305e-5044-4d87-826e-cfeedef29c61","arxiv_id":"2505.11706","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A frequency-based method ranks qubit link error rates from transpiled circuits, matching IBM's public calibration bins within two ranks for about 80% of links on two 127-qubit machines.","lead":"The paper proposes a way to infer which qubit connections on a cloud quantum computer are low-error by counting how often a compiler selects each connection in transpiled programs. A customer of a third-party quantum cloud could use it to check whether the vendor actually executed the job on the high-quality hardware it promised.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central untested premise is that link-selection frequency is driven by error rates rather than by graph-centrality from SWAP minimization; a permutation-control experiment would settle it.","rationale":"The reader's weakest assumption correctly identifies the decisive modeling premise: transpiler link choice must be dominated by the fidelity metric E, not by graph topology or SWAP minimization. My reading of Section III-A strengthens this concern because the paper explicitly describes fidelity as a final tie-breaker among mappings that already minimize gate count, depth, and SWAP count. For large random circuits on a fixed heavy-hex graph, this makes graph centrality a plausible confound that has not been excluded. The proposed permutation control directly tests whether selection frequency causally depends on ECR values; without such a control, the empirical agreement with public ECR data is ambiguous. The reader's other concerns (few circuits, lenient bin tolerance, lack of error bars) are valid but secondary, because they affect the precision of an otherwise sound mechanism. The paper should remain CONDITIONAL: the direction is worth pursuing, but the core mechanism needs a controlled demonstration before the central claim is accepted. My read agrees with the reader's weakest assumption, so no verdict change is needed.","tokens_in":6264,"tokens_out":6446,"duration_ms":74355,"concrete_test":"Permutation control: for IBM Sherbrooke, build a synthetic backend with the same coupling map, basis gates, single-qubit errors, and the same multiset of ECR values, but with the ECR values randomly permuted across edges. Transpile the same five 100-qubit circuits at the same optimization level and recompute link-frequency bin ranks. If the inferred bins track the permuted ground truth as well as the original bins tracked the original ground truth, the method reads error rates. If the inferred bins stay close to the original (unpermuted) ranking and ignore the permutation, link-selection frequency is dominated by topology/SWAP minimization rather than by ECR, and the central claim is refuted. A second variant, transpiling with all ECR values set equal, should leave the ranking essentially unchanged under the topology-only hypothesis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim treats the frequency with which a physical link is used by the transpiler as a proxy for that link's ECR error rate. That requires the transpiler's choices to be dominated by the fidelity metric E in Section III-A. The paper's own description says fidelity is applied only after mappings with the lowest gate count, circuit depth, and number of SWAP gates are found: \"Several possible logical-to-physical mappings are found... Now the transpiler calculates which of these has the highest fidelity.\" Qiskit's routing passes (Sabre) primarily minimize SWAP count rather than ECR-weighted path cost. On 100-qubit random circuits mapped to 127-qubit heavy-hex graphs, many edges are forced as topology bottlenecks and are selected by any routing strategy, regardless of their error rate. Link-selection frequency may therefore track betweenness centrality rather than hardware quality. This is not merely a statistical weakness: if the frequency ranking is topology-dominated, the reported 80-83.5% agreement could arise from spatial correlation between topology and ECR, or from the fact that the same IBM backend error data may have been supplied to the Qiskit transpiler that generated the circuits and then used as ground truth for the comparison. The paper reports no control isolating the causal role of ECR rates in link selection, so the central inference is untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a forensic method to estimate relative two-qubit (ECR) error rates of qubit links on IBM quantum backends from transpiled circuits alone. The authors extract the physical qubit topology from transpiled QASM using the technique of [8], count how frequently each physical link is selected across five random 100-qubit circuits transpiled on IBM Sherbrooke and IBM Brisbane, and group the 144 links into six equal-sized bins by selection frequency. They compare this bin ranking against a bin ranking derived from publicly reported ECR error rates, reporting that for 80% (Brisbane) and 83.5% (Sherbrooke) of links the predicted bin differs by at most two bins from the ECR-based bin. The paper concludes that link-selection frequency is a usable side-channel for coarse backend error-rate ranking.","tokens_in":6425,"tokens_out":2467,"duration_ms":27822,"significance":"If the central modeling premise holds, the method would provide a non-invasive, user-side forensic tool for validating whether a cloud quantum provider transpiled and executed a circuit on high-fidelity qubit links, which matters for third-party quantum cloud security and trust. The approach uses only the transpiled circuit artifact, which is a realistic deliverable in the threat model, and it builds on an existing topology-extraction method. However, the paper's significance is conditional: the current empirical support is thin (five circuits, no statistical error bars, no baseline comparison), and the central inference that selection frequency is driven by ECR error rates rather than by routing topology is not tested by any control experiment. Strengths of the paper are its clearly stated threat model and the explicit use of an external public benchmark for comparison, which avoids circular derivation of the ranking.","major_comments":[{"comment":"The central modeling premise is untested. The paper states that the transpiler first finds mappings with the lowest gate count, circuit depth, and number of SWAP gates, and only then chooses among them using the fidelity metric E in Eq. (1). If routing passes such as Sabre minimize SWAP count rather than ECR-weighted path cost, then link-selection frequency will be dominated by graph-topology bottlenecks (e.g., central heavy-hex edges) rather than by hardware error rates. The reported 80-83.5% agreement could then arise from spatial correlation between topology and ECR, or from the fact that the same backend error data supplied to the transpiler are later used as the ground truth for comparison. The paper provides no control experiment, such as transpiling with permuted or synthetic error rates, or comparing the frequency ranking against a topology-based centrality ranking, to isolate the causal role of ECR rates. This is load-bearing because the entire forensic claim rests on selection frequency being a proxy for link error rate.","section":"Section III-A, Eq. (1)"},{"comment":"The statistical support for the headline claim is insufficient. Only five random 100-qubit circuits are used, with no repeated transpilation runs, no reporting of variance across circuits or across transpiler random seeds, and no confidence intervals on the reported percentages. With 144 links divided into 6 bins of 24 links each, the agreement metric counting a difference of up to 2 bins as success has a high chance baseline: for two independent uniform random bin assignments, the probability of a bin difference of at most 2 is about 67%. The reported 80-83.5% is therefore not far above chance, and the paper does not compare against a null model or report per-bin confusion statistics. The claim that the method is 'valid' requires a baseline and an error analysis.","section":"Section III-B"},{"comment":"The temporal alignment between the transpiled circuits and the public ECR snapshot is not given. The paper does not state the dates on which the IBM Sherbrooke and IBM Brisbane backends were transpiled, the date of the ECR data used as ground truth, or the calibration schedule of the backends. Since IBM error rates drift over time, a mismatch of even a few days could materially change the ground-truth bin ranking, especially for links near bin boundaries. The absence of these dates makes the quantitative agreement figures impossible to reproduce or interpret.","section":"Section III-B"}],"minor_comments":[{"comment":"The paper does not describe the random circuit generation parameters (depth, gate types, random seed) or the transpiler optimization level used, which prevents reproduction of the experimental setup.","section":"Section III-B"},{"comment":"The definition of T2 time is grammatically incomplete: 'measures the time it takes to collapse from the superposition state to either |0⟩ or|1⟩' should describe the decay of coherence rather than a collapse event.","section":"Section II-A4"},{"comment":"The figures showing ECR error information and frequency-based rankings are not referenced in the body text with specific discussion; the captions also lack units and color-bar legends, making it hard to read the actual error values.","section":"Fig. 2 and Fig. 3"},{"comment":"The abstract and introduction state that public error information 'may not be accurate' while the method uses that same public information as ground truth for validation; this tension should be discussed explicitly.","section":"Abstract and Section I"},{"comment":"Reference [1] is a bare URL without a title or access date, and several references are arXiv preprints without version or DOI information; a consistent citation format is needed.","section":"References"},{"comment":"The contributions list contains a formatting artifact: item (i) starts with '(i)Fingerprinting' and the text says '(b) Creating trust' after '(i)', suggesting a missing item (a); the list should be cleaned up.","section":"Section I"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible and useful direction for quantum cloud forensics, but the current version is best described as an preliminary empirical study rather than a validated method. The main missing piece is a control experiment that separates topology-driven link selection from error-rate-driven link selection; without it, the central claim is not established. I would encourage the authors to add a permutation or synthetic-error experiment, use a larger and more varied circuit set with repeated transpilation, report confidence intervals and a baseline for the bin-agreement metric, and include the dates of both the transpilation and the public ECR snapshot. The paper's scope fits the security/forensics audience, but the bar for this venue should be higher than five circuits without statistical support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea is genuinely new: use the frequency with which a transpiler selects physical qubit links to infer their relative error rates, then compare that ranking to public ECR data. That is a cheap forensics technique, needing only the transpiled circuits, not execution results, and it addresses a real trust gap in quantum cloud services. The paper is also honest in framing this as an early step, and the external comparison against published ECR rates is the right check.\n\nThe problem is the evidence. Five random circuits per backend produce noisy frequency counts, and no error bars, repeated runs, or calibration dates are reported. The success metric is also too generous: on a 6-bin scale, allowing a difference of up to 2 bins means random guessing succeeds about 67% of the time, so the reported 80–83.5% agreement is only modestly above chance. We never learn how the differences are distributed, whether most are exact or mostly within that loose two-bin window, or how the binning choices affect the result.\n\nThe deeper concern, which the paper does not address, is the causal mechanism. The paper's own description of the transpiler says it first finds mappings that minimize gate count, depth, and SWAPs, and only then picks the highest fidelity among those. On 127-qubit heavy-hex hardware running 100-qubit random circuits, many links are forced as routing bottlenecks regardless of their error rate. Link-selection frequency could therefore track graph topology or betweenness centrality, and the observed agreement with ECR data might be spatial correlation rather than evidence that the transpiler is making error-aware choices. A simple control would be to permute the ECR error labels over the fixed topology and see if the frequency ranking follows the permuted labels, or to compare against a topology-only baseline. Without such a control, the central inference is not established.\n\nThe paper does have a clean structure and avoids circularity by comparing against an independent public benchmark. It is a legitimate working paper, but the headline agreement numbers need much stronger support before they can be believed. I would send it to a serious referee, but the referee should demand more circuits, a stricter accuracy metric, a topology-control experiment, and released code and data.\n\nBottom line: worth engaging with, not worth citing yet as evidence.","headline":"Plausible side-channel for backend qubit quality, but the evidence is too thin and the topology confound is untested.","tokens_in":7036,"tokens_out":2093,"would_cite":false,"duration_ms":23815,"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 user can rank a quantum backend's qubit links by error rate just by counting which links its transpiler picks, with agreement within two bins for up to 83.5% of links.","keywords":["quantum hardware forensics","transpiler side-channel","qubit error rates","ECR error rates","cloud quantum security","hardware fingerprinting","backend characterization","quantum cloud transparency"],"falsifier":"Transpile the same suite of circuits on a backend while forcing the mapping cost function to ignore error rates and minimize only gate count and depth; if link-selection frequencies still match the published error bins at the reported rates, the agreement is an artifact of topology rather than a side-channel on error rates, whereas a drop in agreement would confirm the proposed mechanism.","tokens_in":5971,"feed_emoji":"⚛️","tokens_out":6833,"duration_ms":63653,"temperature":0.7,"pith_summary":"Quantum cloud customers send their programs to a vendor and get back a transpiled circuit, but they cannot see which physical qubits were chosen or how error-prone those qubits are. This paper tries to make that hidden choice visible by treating the transpiler's link selections as a side-channel: because the mapping step favors low-error qubit links, counting how often each link appears across a suite of transpiled circuits should rank the links by quality. The authors group the 144 links of two 127-qubit backends into six bins by selection frequency and compare the bin ranks with the vendor's published two-qubit error rates. They report agreement within two bin ranks for up to 83.5% of links on one backend and 80% on the other. If this holds, a customer with no privileged information can coarsely fingerprint a backend's high-fidelity regions and check whether the vendor honored its allocation promises.","feed_headline":"Transpiled circuits leak qubit error-rate rankings","feed_subtitle":"On two 127-qubit backends, selection-frequency bins match published error bins for up to 83.5% of links.","key_machinery":"The load-bearing object is the transpiler's fidelity metric $E = (1-E_{g_1})^{n_1}(1-E_{g_2})^{n_2}\\cdots(1-E_{g_N})^{n_N}$, the product over all executed gates of one minus each gate's error rate; the transpiler chooses among lowest-gate-count mappings the one with the largest $E$. This turns each decision into a biased sample of low-error links. The paper's ranking mechanism is then the frequency with which each physical qubit link appears in the extracted topologies of many transpiled circuits, converted into six equal bins of 24 links and compared with the same binning applied to the vendor's published two-qubit (ECR) error rates. The mechanism carries the argument because selection frequency is observable from ordinary transpiled output, while the error rates themselves are the hidden quantity being estimated.","core_discovery":"The paper's central claim is that the relative error rates of a quantum backend's physical qubit links can be recovered from the transpiled circuits returned to the user, without direct access to calibration data. The transpiler is assumed to minimize aggregate gate error through a fidelity metric $E=(1-E_{g_1})^{n_1}(1-E_{g_2})^{n_2}\\cdots(1-E_{g_N})^{n_N}$, so the physical links it selects most often across many circuits are the least error-prone ones. Ranking 144 links into six equal bins of 24 by selection frequency, and binning the same links by published ECR (echoed cross-resonance) two-qubit error rates, the authors find that the two rankings agree within two bin ranks for roughly 83.5% of links on one 127-qubit backend and 80% on another. The paper therefore claims that link-selection frequency in ordinary transpiled programs is a usable, coarse side-channel for the backend's error-rate structure.","pith_inferences":["I infer that this frequency-based ranking could be automated as a per-job verification service: a client submits probe circuits, computes bins, and compares the vendor-chosen mapping against the client's own ranking before paying.","The observed agreement likely depends on the circuit generation procedure; five random 100-qubit circuits are one sample, and circuits concentrated on a small region of the coupling map should make only local links visible, which is a testable boundary condition.","Because the mapping cost function is set by the software stack, the frequency signal mixes hardware quality with transpiler policy; a change in compiler version could shift the bins even with unchanged hardware.","A finer-grained use of the same data would be to look not at aggregate link counts but at pairwise ordering of links within a bin, which the paper's equal-width bins intentionally throw away."],"forward_implications":["A user who submits several large random circuits to a cloud backend can reconstruct a coarse quality ranking of that backend's qubit links using only the returned transpiled circuits.","A suspicious mapping that avoids the link bins most favored by the transpiler on other runs is evidence that the vendor did not allocate the highest-fidelity qubits.","The side-channel works even when the provider hides calibration data or hardware identity, since the signal is carried by the transpiled program itself.","If the same method is run repeatedly, changes in link-selection frequencies would indicate re-calibration, hardware drift, or a change in the vendor's allocation policy."],"supporting_citations":[{"why":"Supplies the topology-extraction heuristic that turns each transpiled circuit into the set of physical qubit links used, the input to the frequency ranking.","marker":"[8]"},{"why":"Vendor resource pages from which the authors take the public ECR error rates used as ground truth for the bin comparison.","marker":"[1]"},{"why":"Provides the background characterization of quantum gate errors and noise that motivates taking two-qubit error rates as the quality signal.","marker":"[9]"},{"why":"Frames the multi-tenant allocation threat model that motivates wanting visibility into which qubits a vendor assigns.","marker":"[2]"}],"fun_headline_variants":["Transpiler selections leak qubit error-rate rankings","Quantum error rates inferred from transpiled circuits","Side-channel reveals quantum hardware error rates","Circuit transpilation exposes hidden backend quality","Estimate qubit error rates from transpiler choices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's load-bearing premise is that the transpiler's link choices are driven mainly by the links' error rates; if gate count, circuit depth, or swap minimization dominate routing, link-selection frequency will reflect circuit topology rather than hardware quality.","fun_headline_variants_meta":{"raw":{"variants":["Transpiler selections leak qubit error-rate rankings","Quantum error rates inferred from transpiled circuits","Side-channel reveals quantum hardware error rates","Circuit transpilation exposes hidden backend quality","Estimate qubit error rates from transpiler choices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1545,"prompt_tokens":1074,"completion_tokens":471,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":690,"completion_tokens_details":{"reasoning_tokens":403}},"tokens_in":690,"tokens_out":471,"duration_ms":5096,"temperature":1.0,"reasoning_tokens":403,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:49:06.694403+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Transpile the same suite of circuits on a backend while forcing the mapping cost function to ignore error rates and minimize only gate count and depth; if link-selection frequencies still match the published error bins at the reported rates, the agreement is an artifact of topology rather than a side-channel on error rates, whereas a drop in agreement would confirm the proposed mechanism.","supporting_citations":[{"cited_title":"Forensics of Transpiled Quantum Circuits","cited_arxiv_id":"2412.18939","evidence_quote":"Provides the background characterization of quantum gate errors and noise that motivates taking two-qubit error rates as the quality signal."}],"review_version":1}