{"id":"0b573959-a7df-4176-b4e1-f47ac60ec4be","arxiv_id":"2610.06142","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"OpenQASM round trips strip quantum high-level operations, silently disabling requested synthesis in all 360 delivered test conditions, and a record-and-verify mechanism restores it in 830 recovery runs.","lead":"Quantum compilers silently lose the ability to synthesize high-level operations like the QFT when a circuit crosses a representation change such as an OpenQASM 3 round trip, even though compilation succeeds and functional equivalence holds. This paper measures the loss across Qiskit, TKET, and Cirq and introduces Verified, which restores the lost synthesis by checking a recorded operation against the delivered circuit.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Verified's recovery envelope is exactly the ≤10-qubit regime: the fingerprint fast path never fires in the target scenario (it changed in all 830 recovery runs, §5.2.3, sentence b184), so any representation-change recovery above ten qubits is a guaranteed no-recovery and the dense O(4^n) matrix…","rationale":"I read the paper in good faith as establishing two things: (1) OpenQASM serialization routes silently destroy synthesis availability while compilation succeeds and functional-equivalence oracles pass (360 delivered conditions, b7, with careful version-pinned instrumentation and independent-looking controls); and (2) Verified can restore the request only when it can confirm record-to-delivered equivalence, which it does in all 830 runs (b10) by falling back from fingerprints to a dense-matrix operator-equivalence check bounded to ten qubits. The reader's weakest_assumption targeted exactly this: with the fingerprint changing in all 830 runs, any valid record above ten qubits is unrecoverable. My stress-test confirms and sharpens that: the fingerprint fast path does not merely happen to fail — it is structurally guaranteed to fail in the recovery scenario, because any representation change that motivates Verified also changes the region's serialized form. The above-ten-qubit branch is therefore a provable null set in the target use case, making the demonstrated 830-run success entirely a ≤10-qubit, dense-matrix result. This is an acknowledged limitation (b184, §7, conclusion), so I do not present it as a new finding; it is nevertheless load-bearing for the strength of the central recovery claim, since the paper's contribution is precisely a tool that 'restores requested synthesis' and the design guarantees that restoration cannot occur for the large circuits the quantum-ecosystem corpus itself contains (e.g., the 50-qubit QFT noted in §4.2.3). I do not find a separate load-bearing flaw in the 360-condition 'no effect' claim: it rests on measured dispatch instrumentation with an uninstrumented control (93 sampled cases, 144 paired PauliEvolution compilations), and its mechanism (name-based dispatch invoking a type-checking synthesis method) is internally consistent with Qiskit's documented HLS interface. The one weakness of that arm is that the replication package is withheld, which the reader already flagged as code_unavailable; my passes cannot do better than affirm that. The independence of the 830-run equivalence claim deserves a note: the paper's own matrix routine gates acceptance and is also used to assert post-reconstruction equivalence, with QCEC corroboration applied to the stale-record sets (§4.3.3) rather than the recovery population — but the acceptance comparison (record vs delivered region) and the output comparison (compiled reconstruction vs original) are distinct comparisons, and the paper's careful handling of final-layout swaps (§5.2.2, the four false rejections) shows awareness of the subtle cases, so I do not elevate this to a load-bearing attack. Net: the reader's CONDITIONAL verdict stands; my concrete test would settle whether the ten-qubit envelope is a hard design limit (as the paper's own text predicts) once the package is available.","tokens_in":23798,"tokens_out":7704,"duration_ms":53805,"concrete_test":"Once the replication package is released, run Verified on a 14-qubit QFT taken through a direct OpenQASM 3 round trip (a width the paper itself uses for study instances, §4.1), with a valid record created before the round trip. Prediction from the paper's own design: the fingerprint mismatches, no operator-equivalence check is permitted above ten qubits, and Verified returns the unchanged delivered circuit (recovery fails, request 'no effect'), exactly as the §5.2.3 'all 830 runs' fingerprint result implies. If Verified instead reconstructs the QFTGate and restores the requested synthesis at 14 qubits, the claimed envelope is wrong; if it fails, the ten-qubit-only scope of the 830-run recovery claim is confirmed as a hard design limit, not a coverage gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline recovery claim ('Verified restores requested synthesis in all 830 recovery runs', b10) is real but its demonstrated envelope is the ten-qubit regime by construction, not merely as an untested scaling limit. Verified's two-stage Verify (§4.3.1) accepts a record on a matching fingerprint (cheap, scalable, ~0.2 ms, §5.2.3) or, after a mismatch, on the bounded operator-equivalence check (dense O(4^n), stated in b98). Section 5.2.3 reports the fingerprint changes in all 830 recovery runs, all 24 MQT regions, and all 30 resynthesis cases, so the fast path never fires in the exact scenario Verified was built for: recovery after a representation change. Any real representation change (OpenQASM round trip, conversion, lowering) alters the QPY serialization, so the fingerprint provably always mismatches; hence every candidate above ten qubits is rejected by rule ('regions larger than ten qubits are accepted only on a matching fingerprint', §4.3.1) and recovery fails. The recovery population is also confined to ≤10 qubits by the equivalence filters (§4.3.2). The consequence is that Verified's entire demonstrated restoring capability rests on the same dense-matrix routine that also defines its ten-qubit bound, and extending the claim to the paper's own study widths (14- and 18-qubit QFT, §4.1) is impossible with the current design. This is not a mere acknowledged scope limit (the authors state the limit in b184 and the conclusion); it is a structural guarantee that Verified can only ever recover small circuits in its target use case, so the strong form of the central claim — that verified record reconstruction is a general remedy — is only evidenced at small width. This converges with the reader's weakest_assumption and reinforces the CONDITIONAL verdict rather than changing it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces and studies 'synthesis availability': whether a quantum compiler can still apply a requested synthesis method to a high-level operation (QFT, MCX, LinearFunction, PauliEvolution) after a representation change such as serialization, OpenQASM interchange, compiler conversion, or lowering. Through controlled experiments in Qiskit, TKET, and Cirq, a project corpus audit, captured test circuits, and an end-to-end MQT pipeline, the authors find that synthesis availability survives only when the high-level operation itself survives; after a direct OpenQASM 3 round trip, the requested synthesis method is invoked but returns no circuit in all 360 delivered conditions, while compilation succeeds silently (Section 5.1.2). They also introduce Verified, a record-and-verify mechanism that stores an operation record before a representation change and reconstructs the operation only after a check (fingerprint match or bounded operator-equivalence) confirms the record still matches the delivered circuit; Verified restores requested synthesis in all 830 recovery runs with median verification times of 2.0–638 ms and rejects every stale record in the evaluated sets.","tokens_in":24205,"tokens_out":10592,"duration_ms":66688,"significance":"The RQ1 finding is a robust, practically relevant result for quantum compiler and interchange design: a functionally equivalent circuit that passes all existing oracles (successful compilation, functional-equivalence checks) can silently lose its intended optimization, with measured two-qubit-count increases of 10–164% in captured MCX circuits and 37.2% in the eight-qubit MQT Grover pipeline. The empirical methodology is unusually careful: pinned compiler versions, ten routing seeds, frozen analysis protocols, instrumented HLS dispatch with uninstrumented controls, request selection on the QPY control before the representation change (avoiding outcome-based selection bias), and independent validity labels that require agreement with the QCEC tool. The RQ2 contribution (Verified) is sound within its demonstrated population, but its demonstrated envelope is exactly the ≤10-qubit regime, and that cap is structural rather than merely an untested scaling limit (see major comment 1); the authors acknowledge the cap in Section 7 and the conclusion but understate its structural necessity in the abstract and Section 5.2.1.","major_comments":[{"comment":"Verified's recovery envelope is structurally capped at ten qubits, not merely as an untested scaling limit, and the paper should state this explicitly. The fingerprint is the SHA-256 digest of the region's QPY serialization (§4.3.1), and the recovery scenario is precisely the case where the representation changed and the high-level operation was lost; any such change also alters the serialization, so a fingerprint match is impossible by construction in that scenario. Section 5.2.3 confirms the fingerprint changed in all 830 recovery runs, all 24 MQT regions, and all 30 resynthesis cases, so the fast path never fires in the scenario Verified was built for. Recovery therefore always falls to the bounded dense-matrix operator-equivalence check (O(4^n), stated in Section 4.1, sentence b98), and Section 4.3.1 step (2) rejects regions above ten qubits by rule. The consequence is that Verified's demonstrated restoring capability is confined to ≤10 qubits by construction, leaving the paper's own 14- and 18-qubit QFT study instances (Section 4.1) and the 50-qubit QFT found in the corpus (Section 4.2.3) outside its reach. The authors state the limit (b184, Section 7, conclusion) but do not spell out that it is a structural guarantee, not an untested scaling boundary; I recommend making that argument explicit and pairing the headline recovery claim with the ten-qubit qualifier.","section":"§4.3.1, §5.2.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The headline recovery claims — 'Verified restores requested synthesis in all 830 recovery runs' (abstract) and 'Original and Verified produce the requested synthesis in all 830 runs' (Section 5.2.1) — are stated without the ten-qubit qualifier, while the RQ1 study itself operates at widths up to 18 qubits (§4.1) and the corpus contains a 50-qubit QFT (§4.2.3). Given the structural cap in §4.3.1/§5.2.3, a reader could reasonably infer that Verified provides general-width recovery, which the design provably cannot deliver. I recommend that the abstract and Section 5.2.1 present Verified as a ≤10-qubit proof of concept whose verification path is, by construction, the dense-matrix operator-equivalence check, and explicitly note that the study's wider instances are unrecoverable with the current design rather than merely remarking on future work.","section":"Abstract and §5.2.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The shared-methodology caveat for the recovery-correctness claim should be stated more directly. Verified's Verify stage and the independent validity labels both rest on operator-equivalence checking via dense matrices: the labels use a separate implementation plus agreement with QCEC 3.10.0, which is a genuinely independent oracle and is a strength of the design, but every one of the 830 recovery runs is verified only by the in-tool bounded operator-equivalence check. Because that check is the same family of procedure that defines the ten-qubit cap, the safety of recovery in the demonstrated envelope is not independent of the envelope itself. This is not a circularity defect, but it should be acknowledged alongside b184 that the safety of even the ≤10-qubit recovery runs depends on the correctness of the dense-matrix equivalence check.","section":"§4.3.3","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The replication package [53] is stated to become publicly available only upon acceptance; given that every empirical claim rests on the pinned environment (Qiskit 2.5.2, pytket 2.18.1, Cirq 1.6.1, MQT revs), the reproducibility of the headline '360 conditions' and '830 runs' claims cannot yet be checked. I recommend providing the package with the revision rather than only at acceptance.","section":"Data Availability","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The cell label 'Refused' is used for delivery failures in Figure 2, while the text (Section 3.2) uses 'delivery failure'; please unify the terminology so the figure is self-consistent with the prose.","section":"Figure 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The text reports that 'none of the 180 delivered MCX and QFT conditions raises an exception or Python warning', while the source-audit sentence immediately after states that IntegerComparator's synthesis raises an exception after both OpenQASM 3 routes. These refer to different populations (the 180 delivered conditions versus the 37-instance family sweep), but the juxtaposition reads as inconsistent; one clarifying sentence on the population boundary would help.","section":"§5.1.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"For the main stale-record set, the text says 36 cases yield 28 delivered circuits but does not state which of the six transformation sequences produce the eight non-deliveries; a one-line breakdown (e.g., OpenQASM 3 import/export failures) would make the delivery accounting easier to follow.","section":"§4.3.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The MQT pipeline recovery paragraph states that 'reconstructing MCPhaseGate alone changes no value', which is consistent with MCPhaseGate having no registered synthesis method in Qiskit 2.5.2 (§4.3.2); consider adding one sentence in the results noting explicitly that this arm tests structural restoration of the nested definition, not synthesis restoration, so the 'no change' outcome is the expected one.","section":"§5.2.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision","confidential_remarks":"The RQ1 contribution — the silent loss of synthesis availability after OpenQASM round trips, with all existing oracles (successful compilation, functional equivalence) passing — is carefully established and should carry the paper; the methodology (pinned versions, ten routing seeds, frozen protocols, instrumented dispatch with uninstrumented controls, QCEC-agreed labels) is a model for this kind of measurement study. The RQ2 contribution is sound but its demonstrated envelope is structurally the ≤10-qubit regime because the fingerprint fast path cannot fire in the recovery scenario; the authors acknowledge the cap but the abstract overreaches slightly. I have recommended major revision on the strength of the reframing required, not because the empirical core is unsound. One caveat for the editor: the replication package is currently promised rather than available, and given that the claims are entirely empirical, availability at revision time would materially strengthen the review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Candid read: the central discovery is real and well supported. After a direct OpenQASM 3 round trip, requested synthesis fires, returns no circuit, compilation succeeds, and equivalence passes — all 360 delivered conditions. The HLS dispatch instrumentation (invoked vs. produced) makes that claim concrete rather than asserted, and the independent QCEC labels plus pinned compiler versions give the measurements real weight. Choosing beneficial requests on the QPY control before the representation change avoids outcome-selection bias, which they state openly. I agree the synthesis-availability concept is genuinely new relative to MorphQ/QDiff/IterTestQ, whose crash and equivalence oracles both pass exactly where this loss happens. The paper does its homework.\n\nSoft spots, in proportion. First, the replication package is withheld until acceptance. For a measurement study, that is a real hold: the 830-run and 360-condition numbers cannot currently be re-run. Second, the stress-test note lands. The fingerprint fast path never fires in the target scenario — it changes in all 830 recovery runs and all 24 MQT regions (§5.2.3) — so every real representation-change recovery above ten qubits is a guaranteed no-recovery by the paper's own rule (§4.3.1). The dense O(4^n) operator check both defines the ten-qubit bound and carries all demonstrated recovery weight. The authors acknowledge the limit (b98, conclusion), but it is structural, not merely an untested scale: the strong form of \"Verified restores synthesis\" is evidenced only at ≤10 qubits. Third, prevalence is not established — 31 of 38 MCX increases come from one repository (b250) — though they say so plainly, and the phenomenon itself is still proved.\n\nNone of this undermines the main discovery. The silent loss is measured, instrumented, and reproducible in principle; the cost consequences in the MQT pipeline (37.2% two-qubit increase) are real pipeline outputs, not fits. What it caps is the generality of the recovery tool's claim. CONDITIONAL is the right verdict: the artifact release plus an honest statement that Verified's envelope is the small-circuit regime would move it to accept.\n\nRecommendation: send to referees with the artifact condition front and center. This documents a failure mode in a path every quantum workflow touches, and the field should know about it even before Verified scales.","headline":"OpenQASM round trips silently kill synthesis availability with every oracle passing — the measurement work is solid, but Verified's recovery envelope is the ≤10-qubit regime by construction, not just as an untested scaling limit.","tokens_in":24830,"tokens_out":1896,"would_cite":true,"duration_ms":14731,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A quantum circuit can pass every functional-equivalence oracle and still silently lose its intended optimization when moved across an OpenQASM boundary, and a verified operation record can put that synthesis choice back.","keywords":["synthesis availability","quantum compilation","high-level synthesis","OpenQASM round trip","functional equivalence","operation records","verification","Qiskit"],"falsifier":"Take a single high-level operation, say a ten-qubit QFT, compile it with a non-default synthesis method, export it to OpenQASM 3, and import it back while instrumenting the HLS dispatch: if the requested method returns a synthesis circuit on the imported generic Gate, the central claim fails. A second check targets the recovery ceiling: run a record for a region larger than ten qubits through a representation change and have Verified successfully restore the request, which would contradict the stated fingerprint limit. A cheaper probe is quantitative: the paper predicts that after the direct OpenQASM 3 round trip the compiled two-qubit count equals the count produced from the imported gate definition alone, so any deviation would signal the synthesis path is not actually inert.","tokens_in":25221,"feed_emoji":"⚛️","tokens_out":4945,"duration_ms":34157,"temperature":0.7,"pith_summary":"The paper introduces synthesis availability, the property that a receiving compiler can still apply a chosen synthesis method directly to a high-level operation after a representation change, and argues that it is independent of functional equivalence. Using controlled experiments in Qiskit, TKET, and Cirq, it shows that OpenQASM interchange removes high-level operations: after a direct OpenQASM 3 round trip, the requested synthesis method is invoked in all 360 delivered conditions but returns no circuit, and compilation succeeds without error or warning. Because the loss is invisible to every existing correctness oracle, a developer can believe their gate-optimal synthesis was applied when it was not. The paper then offers Verified, a recovery mechanism that records each high-level operation before the boundary and reconstructs it only after checking that the record still matches the delivered circuit; Verified restores the requested synthesis in all 830 recovery runs with median verification times of 2.0 to 638 ms per operation family.","feed_headline":"OpenQASM round trips silently kill quantum synthesis choices","feed_subtitle":"360 of 360 synthesis requests became silent no-ops yet compilation passed; a verified record restores them.","key_machinery":"The central object is synthesis availability, defined as the receiving compiler being able to apply a synthesis method directly to a high-level operation in the delivered representation, and the paper's measure of success is whether the requested method both gets invoked and returns a synthesis circuit. The recovery machinery is the operation record, which stores an operation's type, parameters, and qubit binding together with a SHA-256 fingerprint of the region's QPY serialization, read by a three-stage procedure (Locate, Verify, Reconstruct). Verify uses the fingerprint as a fast path and falls back to a bounded operator-equivalence check that constructs dense matrices of size $O(4^n)$, applied only up to ten qubits, with regions above that size accepted only on an exact fingerprint match. The load-bearing distinction is between a valid record, where the region remains identifiable and reconstruction is functionally equivalent, and a stale record, where routing, optimization, or approximate recompilation has changed the region; Verified reconstructs only after the check passes, which is what prevents non-equivalent circuits, at the cost of also rejecting some valid records.","core_discovery":"The central claim is that preserving computation is not the same as preserving the compiler's ability to synthesize, and that a functionally equivalent circuit can silently lose its intended optimization. The paper demonstrates this with a ten-qubit QFT on a linear target: a direct OpenQASM 3 round trip keeps the computation but turns the QFTGate into a generic Gate named qft; Qiskit's dispatch still invokes the requested method qft.line, yet the method requires a QFTGate and returns no synthesis circuit for the generic object, so compilation proceeds from the imported gate definition and the median two-qubit count rises from 135 to 150. This is not a QFT quirk: 52 of the 55 registered synthesis entry points across 14 operation families check the delivered object's type and produce a circuit only for the expected type, so the silent no-op generalizes. Across the request study, the direct OpenQASM 3 round trip delivers 360 eligible conditions, and in every one the requested method is invoked, receives a generic Gate, returns nothing, and compilation still succeeds for all ten routing seeds with no exception or warning. The companion result is that Verified restores the requested synthesis in all 830 recovery runs, matching the original compilation exactly in two-qubit count and depth, while rejecting every stale record in the evaluated sets; unchecked reconstruction from a stale record produces a non-equivalent circuit in all 17 such cases tested.","pith_inferences":["The paper's audit found no HLSConfig request in the project corpus that crossed a representation change, so the 360-condition silent-no-op result is controlled-experiment evidence rather than a measured prevalence; how often real pipelines hit this pattern remains unquantified.","Verified's practical reach today is small circuits: the bounded equivalence check stops at ten qubits, and the SHA-256 fingerprint changes in all 830 recovery runs, so any region larger than ten qubits that undergoes any representation change is rejected outright rather than recovered.","A portable, non-dense equivalence oracle would be the enabling step the paper leaves open: storing operation records in OpenQASM 3 directives (listed as future work) would only help if the check itself scales past ten qubits.","A testable consequence for the broader toolchain is that any compiler reporting 'synthesis applied' without confirming the method produced a circuit can be audited by inspecting dispatch return values; the paper's finding that 52 of 55 entry points type-check suggests the failure mode is a design property, not a single-version bug."],"forward_implications":["Compiler developers should report whether a requested synthesis method is invoked, whether it produces a circuit, and whether that circuit is applied, because silent no-op synthesis is currently possible in Qiskit's HLS dispatch.","Interchange tools that cannot preserve a high-level operation should retain a verified operation record, since reconstruction from an unverified record can change the circuit's computation.","Researchers should report the representation supplied to the compiler along with workload and settings, because functionally equivalent representations can differ in synthesis availability and therefore in compiled cost.","Preserving the high-level operation keeps synthesis choices available but does not guarantee a cheaper circuit; the benefit of MCX synthesis depends on available clean ancillas and the benefit for adders on adder size."],"supporting_citations":[{"why":"Supplies the Qiskit high-level synthesis dispatch and the type-based applicability that makes a requested method a silent no-op for a generic Gate; it is the mechanism the whole study faults.","marker":"[38]"},{"why":"OpenQASM 3 is the interchange format whose direct round trip removes high-level operations, the route at the center of RQ1.","marker":"[7]"},{"why":"QPY serialization restores high-level operations, serving as the QPY control comparator and as the source of the SHA-256 fingerprint used in Verified.","marker":"[40]"},{"why":"The independent QCEC equivalence-checking tool used to label record validity, grounding the claim that Verified circuits are functionally equivalent.","marker":"[4]"},{"why":"SABRE layout and routing define the routing seeds and connectivities across conditions, and routing is one of the transformations that makes records stale.","marker":"[23]"},{"why":"Grounds the MCX synthesis opportunity that depends on available clean ancillas, the resource whose loss drives several reported gate-count increases.","marker":"[21]"},{"why":"MQT Bench supplies the Grover circuits whose OpenQASM 2 interchange raises the two-qubit count by 37.2 percent in the end-to-end pipeline.","marker":"[43]"},{"why":"MQT ProblemSolver is the pipeline whose downstream compilation is measured in the end-to-end study and whose Qiskit route loses MCX synthesis availability.","marker":"[44]"}],"fun_headline_variants":["OpenQASM round trips erase synthesis choices in every tested case","Synthesis choices vanish on OpenQASM round trips, Verified brings them back","Round-trip OpenQASM 3 made 360 of 360 synthesis requests no-ops","Functional equivalence is not enough when synthesis needs high-level structure","Verified restores synthesis in all 830 recovery runs after OpenQASM round trips"],"cache_read_input_tokens":0,"weakest_assumption_plain":"Verified's reconstruction guarantee rests on the bounded operator-equivalence check, which uses dense matrices and applies only up to ten qubits; above that size, any representation change alters the SHA-256 fingerprint, so no record for a larger region survives a changed representation, meaning Verified currently guarantees safety only for small circuits, and for those only if the shared equivalence check itself is correct.","fun_headline_variants_meta":{"raw":{"variants":["OpenQASM round trips erase synthesis choices in every tested case","Synthesis choices vanish on OpenQASM round trips, Verified brings them back","Round-trip OpenQASM 3 made 360 of 360 synthesis requests no-ops","Functional equivalence is not enough when synthesis needs high-level structure","Verified restores synthesis in all 830 recovery runs after OpenQASM round trips"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000324,"raw_usage":{"cost":0.00032382,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00032382,"upstream_inference_prompt_cost":0.00018102,"upstream_inference_completions_cost":0.0001428},"total_tokens":1803,"prompt_tokens":1293,"completion_tokens":510,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":421}},"tokens_in":1293,"tokens_out":510,"cost_source":"openrouter_usage","duration_ms":6087,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":421,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-06T16:46:21.929332+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"falsifier":"Take a single high-level operation, say a ten-qubit QFT, compile it with a non-default synthesis method, export it to OpenQASM 3, and import it back while instrumenting the HLS dispatch: if the requested method returns a synthesis circuit on the imported generic Gate, the central claim fails. A second check targets the recovery ceiling: run a record for a region larger than ten qubits through a representation change and have Verified successfully restore the request, which would contradict the stated fingerprint limit. A cheaper probe is quantitative: the paper predicts that after the direct OpenQASM 3 round trip the compiled two-qubit count equals the count produced from the imported gate definition alone, so any deviation would signal the synthesis path is not actually inert.","supporting_citations":[{"cited_title":"2026.Qiskit 2.5.2","cited_arxiv_id":null,"evidence_quote":"Supplies the Qiskit high-level synthesis dispatch and the type-based applicability that makes a requested method a silent no-op for a generic Gate; it is the mechanism the whole study faults."},{"cited_title":"Bishop, Steven Heidel, Colm A","cited_arxiv_id":null,"evidence_quote":"OpenQASM 3 is the interchange format whose direct round trip removes high-level operations, the route at the center of RQ1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"QPY serialization restores high-level operations, serving as the QPY control comparator and as the source of the SHA-256 fingerprint used in Verified."},{"cited_title":"Advanced equivalence checking for quantum circuits,","cited_arxiv_id":null,"evidence_quote":"The independent QCEC equivalence-checking tool used to label record validity, grounding the claim that Verified circuits are functionally equivalent."},{"cited_title":"Tackling the qubit mapping problem for NISQ-era quantum devices","cited_arxiv_id":null,"evidence_quote":"SABRE layout and routing define the routing seeds and connectivities across conditions, and routing is one of the transformations that makes records stale."},{"cited_title":"Quantum9, 1752 (May 2025).https://doi.org/10.22331/ q-2025-05-21-1752,http://dx.doi.org/10.22331/q-2025-05-21-1752","cited_arxiv_id":null,"evidence_quote":"Grounds the MCX synthesis opportunity that depends on available clean ancillas, the resource whose loss drives several reported gate-count increases."},{"cited_title":"Mqt bench: Benchmarking software and design automation tools for quantum computing,","cited_arxiv_id":null,"evidence_quote":"MQT Bench supplies the Grover circuits whose OpenQASM 2 interchange raises the two-qubit count by 37.2 percent in the end-to-end pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MQT ProblemSolver is the pipeline whose downstream compilation is measured in the end-to-end study and whose Qiskit route loses MCX synthesis availability."}],"review_version":1}