{"id":"7d909a6a-1fb6-4cf6-9494-765c2667df47","arxiv_id":"2512.06484","paper_version":6,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Repurposing ancilla qubits for both magic-state cultivation and routing improves lattice-surgery schedule efficiency by 19-223% over dedicated-bus routing in simulations.","lead":"Quantum computers protected by surface codes need special 'magic' states for certain operations, and preparing them is slow. This paper proposes a scheduler that lets the same spare qubits both grow magic states and carry out routing, claiming faster and more compact schedules.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PureMagic's reported magic-prep speedup is a censored-statistic artifact; volume gains depend on an unverified free-restart, memoryless cultivation model.","rationale":"The paper's key evidence is Table II, where PureMagic's lower volumes and lower 'Avg. Magic Cultivation' times are generated by a simulator that cancels and restarts cultivation. I agree with the Reader that the assumption of free, memoryless restart is load-bearing. My sharper formulation: even if cancellation is free, the reported prep-time reduction is a censored statistic. With exponential cultivation and random preemption, completed attempts have mean 1/(μ+ν), while the true inter-completion time is ≥1/μ; so the 2.6–9.7x number does not indicate faster magic production. The actual scheduler benefit must come from having many more simultaneous cultivators and from the smaller N in the volume metric. Whether those benefits survive realistic interruption costs is untested. I therefore do not change the CONDITIONAL verdict. The abstract/full-text discrepancies (29 vs 17 circuits, missing DASCOT/FLASQ) and the redacted code URL are real artifacts that also support CONDITIONAL, but they are not the most load-bearing technical concern. The lambda=0.00227 issue is less severe than the Reader suggests: dividing the sampled exponential by code distance d=17 gives a mean of about 25.9 cycles, close to the stated 26. The decisive experiment is a throughput/restart-cost sensitivity analysis.","tokens_in":14918,"tokens_out":10250,"duration_ms":86126,"concrete_test":"Augment the PureMagic simulator to log every cultivation attempt (interrupted and completed) and compute end-to-end throughput R = completed magic states per ancilla-cycle for bus and PureMagic layouts. Then rerun the Table II experiments with (a) a one-cycle restart/re-initialization penalty after each interruption and (b) a cultivation distribution with a minimum duration (e.g., shifted exponential with 5-cycle minimum, consistent with Fig. 6's 14.5% ≤5 cycles). If PureMagic's efficiency improvements over bus routing move by more than ~10 percentage points under either change, the central volume claim depends on the free-restart/memoryless assumption; if R is unchanged while the completed-only mean drops, the average-prep-time claim is a censoring artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B models cultivation as exponential (Eq. 3) and Section IV asserts it can be terminated/restarted at any time. The central volume improvement over bus routing is computed under that model. But Table II's 'Avg. Magic Cultivation' for Pure Magic counts only completed cultivations: 'those terminated before cultivation is ready are not counted.' In a memoryless exponential with random preemption, completed attempts are length-biased: an attempt that survives to completion has mean 1/(μ+ν) < 1/μ, where ν is the interruption rate, so the reported 2.6–9.7x 'reduction' is a selection effect, not an increase in throughput. The expected time from one completion to the next, including restarts and routing occupation, is ≥1/μ, so per-ancilla magic-state production does not improve. The cycle counts in Table II would only be valid if canceling cultivation has zero overhead and the exponential model is exact; neither is established. Real cultivation has injection/cultivation/escape stages (Sec II-C) with likely minimum durations and re-initialization costs. If each interruption costs even one cycle, or the distribution has a minimum duration, per-ancilla throughput drops and the volume improvements (19–223%) shrink. Some gain will survive from the larger cultivator pool and reduced N, but the claimed magic-prep-time benefit is not a real speedup.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Pure Magic, a dynamic scheduler for surface-code lattice surgery that eliminates dedicated bus patches by repurposing all ancilla patches for both magic-state cultivation and routing. Cultivation is interrupted when a patch is needed for routing and restarted afterward; the authors argue this naturally truncates the long tail of probabilistic cultivation times. The scheduler greedily packs Steiner forests and is evaluated on 17 Benchpress circuits plus random circuits. The paper reports 19%–223% improvements in scheduling efficiency over a bus-routing baseline, 19%–80% reductions in logical qubits, and 2.6×–9.7× reductions in average magic-state preparation time. A comparison with Silva et al. benchmarks is also presented.","tokens_in":15172,"tokens_out":4261,"duration_ms":39620,"significance":"The core idea—using every ancilla patch as a dual-purpose cultivation/routing resource—is timely and potentially valuable for surface-code compilation with magic-state cultivation. Strengths include an open-source implementation, evaluation against an independent bus-routing baseline and Silva et al., and a useful sensitivity analysis of cultivation time and parallelism. However, the headline claim of reduced magic-state preparation time is built on a censored statistic and an unverified memoryless-restart model for cultivation; the abstract also advertises results (29 circuits, DASCOT/FLASQ comparisons) that are absent from the full text. If the volume improvements survive correction of these issues, the paper would be a meaningful contribution, but as written the central speedup claim is not established.","major_comments":[{"comment":"The abstract reports 29 benchmark circuits, a comparison against DASCOT with 'up to 21x' efficiency improvement, and near-optimal FLASQ bounds. The full text reports 17 benchmark circuits in Table I and contains no DASCOT or FLASQ evaluation. The abstract also promises a 'weight limit on Tableau transpilation' that is not discussed anywhere in the body. These claims are not supported by the manuscript and must be reconciled or removed.","section":"Abstract vs. Full Text"},{"comment":"The 'Avg. Magic Cultivation' values for Pure Magic (2.67–10.09 cycles vs. ~26 cycles for bus routing) are a length-biased statistic. The table note states that 'those terminated before cultivation is ready are not counted.' For an exponential distribution with preemption at rate ν, completed attempts have mean 1/(μ+ν), which is smaller than 1/μ, yet the memoryless property implies the expected time from an arbitrary restart to completion is still 1/μ. Thus the reported 2.6–9.7x reduction reflects selection bias, not an increase in per-ancilla throughput. The paper does not establish that interruption has zero overhead or that cultivation is memoryless (Eq. 3 is assumed, not derived). If restarts cost even one cycle or cultivation has a minimum duration, the throughput gain shrinks; the volume improvements in Table II may survive, but the claimed magic-state preparation speedup is not sup","section":"Section V-B, Table II"},{"comment":"The exponential model of Eq. (3) is fitted to MCMC simulations from Gidney et al., but the memoryless property is not verified. Real cultivation has injection, check-grow-stabilize, and escape stages (Section II-C), which likely impose a minimum duration and a re-initialization cost. The scheduler's 'cut the tail' benefit relies on the strong assumption that cultivation 'can be terminated and restarted at any time' with no lost progress (Section IV). This assumption is load-bearing for the reported time reductions and should be backed by evidence or explicitly stated as a worst-case approximation with sensitivity analysis.","section":"Section III-B and IV"}],"minor_comments":[{"comment":"Duplicate word in 'there are only only data and ancilla qubits.'","section":"Section I"},{"comment":"'Gidney at. al.' should be 'Gidney et al.'; also 'Figure 15a' in the text refers to the cited paper, not the current manuscript—consider clarifying.","section":"Section III-B"},{"comment":"The text says 'the right-most columns in Table I' but the average cultivation columns are in Table II.","section":"Section V-B"},{"comment":"The text states that for the 8-qubit circuit the 'overall efficiency improvement is 47%,' while Table II reports 59% for DNN 8. The discrepancy likely comes from different metrics (volume vs. scheduling efficiency), but the text should define which quantity is being reported and keep the numbers consistent.","section":"Section V-B / Figure 10"},{"comment":"In the paragraph beginning 'For 64 data qubits, the bus routing architecture has a total of 60 magic qubits,' the phrase 'there are only 2.3 magic state qubits ready each cycle' conflates a rate (qubits/cycle) with a count. Minor rewording would improve clarity.","section":"Section V-C"},{"comment":"The fitted constant α=216 is introduced without explanation. State how it is estimated.","section":"Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The gap between the arXiv abstract and the full text is substantial and should be addressed as a priority in revision. The censored-statistic issue is fixable by reporting throughput (e.g., completions per cycle per ancilla, including restarts) rather than mean completed-cultivation time. The authors' open-source code and honest comparison with Silva et al. are commendable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core idea here is genuinely new: instead of keeping bus qubits and magic patches in fixed roles, PureMagic lets every ancilla patch double as both a cultivation site and a routing resource, interrupting cultivation when the patch is needed for routing. That is a sensible design and the first I know to push it. The volume improvements over a bus-routing baseline (19-223% across 17 Benchpress circuits) are credible enough to take seriously, and the comparison to Silva et al. is a nice sanity check even if it only exercises the bus-routing mode.\n\nThe main soft spot is the magic-state preparation-time claim. The paper reports average cultivation times dropping from ~26 cycles to 2.67-10.09 cycles, and attributes this to cutting off long tails. But the table only counts cultivations that complete; interrupted attempts are excluded. Under their own exponential model with preemption and restart-from-scratch, the long-run throughput per ancilla is unchanged—the mean time between completions is still 1/λ. The low average is length-bias, not a real speedup. The aggregate throughput does improve because PureMagic has more cultivators (all ancillas vs. just the dedicated magic patches), and the volume gains from fewer qubits and shorter routes are plausible. But the headline \"magic prep time reduced\" should be reframed as a selection effect, or dropped.\n\nA second issue: the arXiv abstract claims 29 circuits, DASCOT/FLASQ comparisons, and 43-152% improvements, none of which appear in the full text, which reports 17 circuits and 19-223%. The metadata needs to be reconciled with the manuscript. This is the kind of thing that gets a paper desk-rejected if left as-is. Also, the model assumes cultivation can be interrupted at any time with zero overhead, and that cultivation times are memoryless. Those are strong assumptions; the paper notes them but doesn't test how sensitive the volume gains are to restart costs or non-exponential distributions. The dependency on a finite minimum cultivation time could erode some of the cycle improvements, though the qubit savings likely survive. (One concern from a quick read—λ=0.00227 vs. 26-cycle mean—resolves when you remember they divide by code distance d=17.)\n\nThe code URL is redacted, which is annoying for reproducibility but not a fatal issue in a preprint.\n\nOverall: the architecture idea is worth refereeing, but the paper needs a careful revision before its quantitative claims can be adopted. A referee should ask the authors to recompute the magic-prep metric as a throughput measure, fix the abstract, and do a sensitivity analysis on the cultivation model.","headline":"Novel dynamic scheduler that repurposes cultivation patches for routing, with a credible volume story but a misleading magic-prep-time metric and an abstract that overclaims; worth refereeing after revision.","tokens_in":15733,"tokens_out":4509,"would_cite":true,"duration_ms":43784,"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":"PureMagic reallocates magic-state cultivation qubits to routing on demand, cutting schedule volume by up to 223% and using up to 80% fewer logical qubits.","keywords":["lattice surgery","magic state cultivation","surface codes","dynamic scheduling","Steiner forest packing","ancilla reuse","fault-tolerant quantum computation","T states"],"falsifier":"Measure the actual cost and statistical memory of interrupting a cultivation run on a surface-code testbed, comparing the distribution of time-to-success for a restarted patch versus a fresh patch; even a one-cycle fixed restart overhead or any correlation between successive attempts will shift the reported average preparation times (2.7–10 cycles) and could erase the 4.5x reduction on low-parallelism circuits like DNN8.","tokens_in":14743,"feed_emoji":"⚛️","tokens_out":8726,"duration_ms":77669,"temperature":0.7,"pith_summary":"PureMagic is a dynamic scheduler for surface-code quantum computers that makes every ancilla patch serve both as a magic-state cultivator and as a possible routing path. When the scheduler needs to connect data qubits for a logical operation, it seizes any cultivating patches, interrupts their cultivation, and later resumes it. Because cultivation is probabilistic and has a long tail, these interruptions cut off the slowest attempts, dropping average magic-state preparation from about 26 cycles to as low as 2.7 cycles. On 17 benchmark circuits the scheduler reports 19% to 223% better scheduling efficiency (a volume proxy for space and time) while using 19% to 80% fewer logical qubits than traditional bus routing. If the underlying assumptions hold, this turns the randomness of magic-state production into a scheduling advantage, implying that architectures with dedicated bus qubits are not resource-optimal.","feed_headline":"Reuse magic-state qubits for routing, cut schedule volume by up to 223%","feed_subtitle":"PureMagic reuses cultivation qubits for routing, eliminating bus qubits and cutting logical-qubit count by up to 80%.","key_machinery":"The key mechanism is the dual-purpose ancilla patch: every non-data qubit in the layout is both a cultivation site for magic states and a candidate routing patch for lattice-surgery merges. The scheduler's greedy MINFIT algorithm packs approximately shortest Steiner trees for independent Pauli products each cycle, and when it seizes a cultivating patch, the cultivation is terminated (with the exponential renewal assumption) and restarted afterward. The second load-bearing piece is the exponential cultivation-time model (Eq. 3), which makes cancellation memoryless so that an interrupted cultivation is statistically a fresh start; this is what turns preemption into a tail-cutting device.","core_discovery":"The paper's central claim is that a lattice-surgery compiler need not dedicate any qubits to routing. In PureMagic scheduling, every non-data patch continuously attempts to cultivate a magic state; when the scheduler needs a path between data qubits, it seizes any convenient cultivating patches, cancels their attempts, and uses them as a Steiner tree for the Pauli product. Once the product finishes, the patches resume cultivation. Because cultivation is modeled as memoryless and restartable, this policy (1) removes idle ancillas entirely, (2) increases the number of concurrent cultivators, and (3) truncates the long tail of the cultivation-time distribution, cutting mean magic-state preparat","pith_inferences":["Inference: The same preemption principle could be applied to any stochastic resource-production process with a heavy tail and zero-cost restarts, not just magic-state cultivation, suggesting a general design rule for fault-tolerant architectures: treat resource generation as preemptible work.","Inference: If the memoryless-exponential assumption is replaced by a realistic cultivation model that has memory (e.g., stage-dependent success probabilities or a restart cost), the tail-cutting benefit may shrink; a concrete test would be to rerun the 17 benchmarks with a log-normal or phase-type distribution fitted to the same cultivation data and compare average preparation times.","Inference: The dual-purpose ancilla idea also implies a new hardware-level requirement: a qubit patch must support both cultivation and lattice-surgery merge/split with no extra fabrication overhead, and the transition between the two modes must be fast enough to avoid scheduling bubbles; if the mode switch costs more than a cycle, the volume gains degrade."],"forward_implications":["Surface-code compilers can eliminate dedicated bus patches, reducing logical-qubit count by up to 80% and shrinking the scheduler's space-time volume by up to 223% relative to bus routing.","The benefit grows with circuit parallelism: at average Pauli-product rates above about 10 per layer, the scheduler's relative gain rises because magic-state supply becomes the bottleneck and PureMagic supplies more cultivators.","The scheduler's gains persist even if cultivation becomes slower: the relative improvement over bus routing increases as the expected cultivation time grows, which matters for future high-fidelity T states.","Average magic-state preparation time falls from about 26 cycles to 2.7–10 cycles on the tested circuits because routing interrupts the longest cultivation attempts."],"fun_headline_variants":["No dedicated bus qubits: scheduler reuses ancillas for routing and magic-state prep","Scheduler repurposes every ancilla for routing and magic-state prep","Dynamic scheduler: no idle ancillas, 4.5x faster magic-state prep","Cut magic-state prep 4.5x by reusing ancillas for routing","No bus patches: dynamic cultivation scheduler cuts qubits up to 80%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire value proposition depends on magic-state cultivation being cancellable at any cycle with no lost progress and no restart overhead, so that an interrupted attempt is statistically a fresh exponential draw.","fun_headline_variants_meta":{"raw":{"variants":["No dedicated bus qubits: scheduler reuses ancillas for routing and magic-state prep","Scheduler repurposes every ancilla for routing and magic-state prep","Dynamic scheduler: no idle ancillas, 4.5x faster magic-state prep","Cut magic-state prep 4.5x by reusing ancillas for routing","No bus patches: dynamic cultivation scheduler cuts qubits up to 80%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001568,"raw_usage":{"total_tokens":6108,"prompt_tokens":766,"completion_tokens":5342,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":5247}},"tokens_in":510,"tokens_out":5342,"duration_ms":31629,"temperature":1.0,"reasoning_tokens":5247,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T18:09:23.489724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual cost and statistical memory of interrupting a cultivation run on a surface-code testbed, comparing the distribution of time-to-success for a restarted patch versus a fresh patch; even a one-cycle fixed restart overhead or any correlation between successive attempts will shift the reported average preparation times (2.7–10 cycles) and could erase the 4.5x reduction on low-parallelism circuits like DNN8.","supporting_citations":[],"review_version":2}