{"id":"ce327ce4-ddf3-41d8-9805-6c8fac79db7e","arxiv_id":"2412.18625","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A random splitting scheme lets the single-atom R1 structure solution method run without user-chosen partial models and reach correct small molecule structures.","lead":"This paper adds a lottery step to the single-atom R1 crystal structure solver: each cycle randomly splits the current partial model into a small and a large child, expands both, and keeps the one that reaches a lower approximate R1. The author reports that this removes the need for a user to choose the next partial structure, and that the scheme solved four benchmark structures and many additional datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The lottery's acceptance gate is the same approximate R1 used as the search objective, yet the paper never shows that lower sR1 minima track structural correctness; a wrong-model minimum could propagate confidently.","rationale":"The reader's weakest_assumption identified exactly the load-bearing vulnerability: the minimum approximate R1 is used both as the search objective and as the acceptance gate, with no independent demonstration that it correlates with structural correctness. My stress-test corroborates this concern and sharpens it by noting that the paper's own Section 2 describes the sR1 as lacking any implicit relation to a partial-structure electron density, and that external validation occurs only for selected final models after manual cleanup. This is not an internal contradiction, and the reported four cases are genuine demonstrations that the method can work. However, the stochastic claim 'can drive the sR1 calculation toward correct structure solutions' is underdetermined by the evidence as presented: a wrong-model attractor with lower approximate R1 would be faithfully propagated by the lottery. The proposed test would settle whether the acceptance criterion is trustworthy. Since the reader's verdict is already CONDITIONAL and this concern is the stated condition, no verdict adjustment is needed.","tokens_in":28152,"tokens_out":3180,"duration_ms":25933,"concrete_test":"Using the supplied code on sample 1, instrument each lottery cycle to record both the child/parent minimum approximate R1 and the fraction of atoms within 0.5 Å of the SHELXT-derived P1 correct model, using the matching algorithm in Section S1. If any accepted model has a lower sR1 minimum than its parent but a lower or equal correct-atom fraction, or if any rejected model has a higher correct-atom fraction, the acceptance gate is demonstrably not tracking correctness. A sharper variant: from a fixed parent state, generate 100 random splits and compute the Spearman correlation between Δ minimum sR1 and Δ correct-atom fraction; a non-positive correlation would falsify the paper's central mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the minimum approximate R1 used to accept or reject child models is a faithful proxy for structural correctness. Section 4 states this assumption explicitly: 'the minimum approximate R1 ever reached is a good measure of how good the resulting model is.' Section 2, however, concedes that the sR1 is an ad hoc approximation and is 'not even implicitly related to an electron density of some partial structure.' The only external correctness checks against SHELXT-derived models are applied to final models of the four showcased samples, often after manual ghost deletion, atom-type correction, and bond-length-guided completion, not to the intermediate models on which each lottery decision is made. The reported runs are therefore consistent with a different mechanism: the search greedily descends a numerical landscape that may contain wrong-model minima, while luck or manual intervention, rather than the R1 gate, may be responsible for reaching the correct basin. Since every accepted step is gated by the same quantity being minimized, the evidence does not establish that 'lucky' fluctuations are fluctuations toward the true structure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a \"lottery mode\" for the single-atom R1 (sR1) crystal structure solution method. In each cycle, a current partial model is randomly split into a small child (1-20 atoms) and a large child; each child is expanded by sR1 cycles, and the model with the lower minimum approximate R1 is carried forward. The stated goal is to remove the need for a user to select fragments or delete ghost atoms between cycles, making the calculation \"care-free.\" The author reports that the lottery mode solved four benchmark structures (samples 1-4) with final models compared quantitatively to SHELXT-derived references, and that about 85 datasets were tested in lottery mode with qualitative validation. The central claim is that the lottery scheme can drive an sR1 calculation toward a correct structure solution.","tokens_in":28281,"tokens_out":4867,"duration_ms":34576,"significance":"If the central claim were fully established, this would be a notable practical advance: it would automate the sR1 method's cycle-to-cycle model selection, which previously required human judgment, and it introduces a stochastic splitting idea (small-child fluctuations) that is original in this context. The paper also provides source code on GitHub, which is a genuine strength for reproducibility, and the quantitative comparison of final models against SHELXT results for four samples is a useful check. However, the current evidence does not yet establish the claimed mechanism, because the acceptance criterion is the same approximate R1 being minimized and there is no independent validation of the intermediate decisions. The significance is therefore conditional on additional evidence linking sR1 minima to structural correctness.","major_comments":[{"comment":"The acceptance gate for retaining a child model is the minimum approximate R1 reached during expansion, and Section 4 states: \"the minimum approximate R1 ever reached is a good measure of how good the resulting model is.\" This is the same quantity that the sR1 search minimizes. Section 2, however, concedes that the sR1 is an ad hoc approximation and is \"not even implicitly related to an electron density of some partial structure.\" The paper provides no evidence that lower sR1 minima track structural correctness; the external SHELXT comparison is applied only to the final models of selected runs, often after manual intervention, not to the intermediate models on which each lottery decision is made. Consequently, the observed drops in sR1 during lottery cycles are equally consistent with greedy descent into a wrong-model minimum, and the data do not demonstrate that the lottery preferentially propagates models with more \"good\" atoms. This is load-bearing for conclusion (1) of Section 6.7 and needs to be addressed, for example by comparing intermediate models at successive lottery cycles against the SHELXT reference.","section":"Section 4 (with Section 2)"},{"comment":"The four quantified examples do not support the abstract's \"care-free\" claim. For sample 3, the text reports that the lottery-cycle result contained ghost atoms that were manually deleted, that atom types were manually corrected, and that missing atoms were found by a separate bond-length-guided sR1 step (Section 5.3, Figure 3). For sample 4, ghost C atoms were deleted after step 2 and the missing C atoms were again found by the bond-length-guided step (Section 5.4, Figure 4). Thus, for the two difficult cases, the lottery mode alone did not produce the final correct model; manual intervention and auxiliary methods were required. The conclusion in Section 6.7 should be restricted to what is demonstrated: the lottery mode contributed to solving these structures within a workflow that still contains user intervention.","section":"Sections 5.3 and 5.4"},{"comment":"The successful single-atom start for sample 3 relies on a post hoc choice of the initial atom position. The default position (0.3, 0.3, 0.3) failed, and (0.325, 0.3042, 0.3) was selected after a 64-point local grid search motivated by that failure. This is a selected trial, not a pre-specified procedure, and it does not provide evidence for a universal care-free starting method. The paper itself says the idea \"currently is only at the initial stage,\" but this qualification is not carried into the general conclusion of Section 6.7, which states that the lottery mode can drive sR1 toward correct solutions without mentioning the dependence on this tuned starting position.","section":"Section 6.4"}],"minor_comments":[{"comment":"The choice of 0.2 Å coordinate precision alongside a 0.4 Å grid step is not justified; the relationship between the grid spacing and the refinement precision could be clarified.","section":"Section 3"},{"comment":"The splitting algorithm uses N both for the parent model size and for the total number of atoms in the unit cell (Section 2), which could confuse readers; consider renaming one of them.","section":"Section 4"},{"comment":"The quantitative comparison with the correct model uses a 0.5 Å criterion, but this is defined only in Section S1; the main text should state the tolerance and the fact that H atoms are excluded.","section":"Section 5.1"},{"comment":"The claim that small child models \"provided most of the improvement\" is based on only four samples and is an observed tendency, not a general result; a more cautious phrasing would be appropriate.","section":"Section 6.2"},{"comment":"For the roughly 85 datasets tested in lottery mode, success is assessed by qualitative visual inspection only, and the table in S9 does not indicate which of these required manual intervention or bond-length-guided steps; adding this information would make the scope of the lottery-mode claim clearer.","section":"Section 6.6 and S9"},{"comment":"The paper repeatedly uses \"care-free\" to describe the method; a more precise term such as \"unsupervised\" or \"automatic\" would better match the actual workflow, which still includes manual steps in the difficult cases.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the author has made the source code available, which is commendable. The central concern is the self-referential acceptance criterion: the search objective and the model-quality gate are the same approximate R1, and the paper does not show that this quantity correlates with structural correctness. The final external validation is applied only after manual interventions for the difficult cases, so the strong claim in the conclusions is not yet supported. I recommend major revision with a request for intermediate-model validation against an external reference, and for a more careful qualification of the workflow's automation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the lottery-mode sR1 is a real and honest attempt to automate the author's phasing-free structure-solution method, and the reported results on four benchmark structures plus about 85 datasets deserve a serious referee. But the paper never establishes the load-bearing assumption: that a lower minimum approximate R1 means a structurally better model.\n\nWhat is actually new: the random split of a parent model into a small (1–20 atom) and a large child, keeping only improvements, is genuinely new relative to the prior sR1 paper and to the Burla et al. random-replacement idea. It makes the method care-free in principle and removes the need for a user to choose fragments or delete ghost atoms between cycles. The paper also credits Kinneging & de Graaff and acknowledges reviewer comments about peaklist optimization and bond-length-guided completion, which is good scholarly hygiene. The code is on GitHub, and the supporting list of 224 datasets with final R-values is unusually transparent for a single-author methods paper.\n\nWhere it is soft: the acceptance gate is the same approximate R1 used as the search objective, and Section 2 concedes that the sR1 is an ad hoc approximation not even implicitly related to a partial electron density. The stress-test concern lands: a wrong-model minimum could propagate confidently, and the paper provides no evidence that sR1 minima correlate with correctness. The only external checks are on final models, often after manual ghost deletion, atom-type correction, and bond-length-guided completion, so they do not validate the intermediate lottery decisions. There is also no baseline against the earlier intelligent mode on the same runs, no error bars, and no commit hash or input data for the code. Sample 3 required a post hoc search over starting positions, which undercuts the claim of a universal startup recipe.\n\nThat said, these are addressable conditions, not contradictions. The reported successes are consistent with what a correct solver would produce, and the author is candid about the role of luck and about the cases where the lottery mode stalled. The central claim is plausible but not proven in its current form. Who this is for: people working on ab initio small-molecule phasing or automated structure solution will want to know this exists. It deserves a serious referee, but the referee should ask for a direct test of the R1-minimum-versus-correctness correlation (for example, on runs where the final model is wrong) and for a cleaner separation of the lottery's contribution from manual intervention.","headline":"A genuine attempt at making the sR1 method push-button, but the lottery's acceptance gate is the same approximate R1 it is minimizing, and the paper never shows that lower sR1 minima track structural correctness.","tokens_in":28912,"tokens_out":1906,"would_cite":false,"duration_ms":13759,"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 random split of the current atomic model into two child models lets the single-atom R1 method solve crystal structures without a human choosing the next partial model.","keywords":["single-atom R1","lottery scheme","crystal structure solution","statistical fluctuation","approximate R1","X-ray crystallography","random splitting","structure solution"],"falsifier":"On a set of datasets with known solutions, take a partially correct model, record the minimum approximate R1 reached after one cycle, then perturb the model by displacing atoms by 0.2, 0.5, 1.0, and 2.0 Å; if the minimum approximate R1 does not increase with displacement, the acceptance criterion can reward models that are farther from the truth.","tokens_in":27829,"feed_emoji":"🎲","tokens_out":6769,"duration_ms":45676,"temperature":0.7,"pith_summary":"This paper claims that a crystal structure can be solved by running the single-atom R1 (sR1) calculation without any human decisions between cycles, if each cycle's finished model is randomly split into two child models and only the child whose expansion lowers the minimum approximate R1 is kept. The sR1 method locates atoms one at a time using an approximate R1 residual; previously a user had to choose recognizable fragments or delete ghost atoms to start the next cycle. The lottery replaces that judgment with random splits, betting that statistical fluctuation will sometimes produce a small child rich in correctly placed atoms. On test structures with 128 to 316 non-hydrogen atoms per cell, the scheme drove calculations to or close to the correct solutions, and about 85 datasets were tested in lottery mode. The paper also reports speed-ups from sharpened intensities and coarser coordinate precision, and a local search over starting positions for difficult cases.","feed_headline":"Random lottery of atomic models solves crystal structures automatically","feed_subtitle":"The single-atom R1 method no longer needs a human to pick fragments or delete ghost atoms between cycles.","key_machinery":"The carrying object is the single-atom R1 (sR1), an approximate version of the traditional crystallographic R1 in which only one undetermined atom's coordinates are optimized while terms containing the other undetermined atoms' coordinates are deleted, though their scattering factors are partly retained. The new mechanism is the lottery split: a parent model is randomly partitioned into a deliberately small child (1 to 20 atoms) and a large child, and the minimum approximate R1 reached during expansion is the objective and acceptance criterion. This criterion is what lets the calculation decide whether a child is an improvement without human inspection. Two implementation changes support the scheme: sharpened intensities replace raw intensities, and single-atom positions are refined to 0.2 Å instead of 0.001 Å.","core_discovery":"The central claim is that the lottery mode can drive an sR1 calculation toward correct structure solutions automatically. After one expansion cycle, atoms whose addition raised the approximate R1 are deleted; the remaining parent model is randomly split into a small child (about 1 to 20 atoms) and a large child. The small child is deliberately small so that random fluctuation can make it nearly all good atoms (producing a rebuild) or nearly all bad atoms (leaving its large sibling slightly richer in good atoms than the parent). The child model whose expansion reaches a lower minimum approximate R1 than the parent becomes the next parent; if neither improves, the unchanged parent continues. The paper reports that in detailed tests the final models matched reference solutions to within 0.5 Å for essentially all atoms, and that repeated runs take different random paths but reach approximately the same final model.","pith_inferences":["The acceptance rule assumes the minimum approximate R1 tracks closeness to the true structure; testing that monotonicity directly on benchmark structures would either support or undermine the lottery's foundation.","The small-child/large-child split resembles an explore-versus-refine balance: the small child can restart from near scratch, while the large child makes incremental corrections; combining this with automated fragment recognition could make the intelligent and lottery modes complementary.","The paper's local search for an optimal starting atom position suggests a universal initialization strategy: run one sR1 cycle from each of 64 nearby grid points and keep the lowest minimum approximate R1, an idea the author states is still at an initial stage.","The paper explicitly acknowledges that sR1 cannot distinguish a solution from its inverted image and that success is judged by chemical recognizability; an unattended pipeline would need to resolve that ambiguity before calling a structure solved."],"forward_implications":["An sR1 calculation can in principle run unattended, removing the need for a crystallographer to recognize fragments or delete ghost atoms between cycles.","Difficult cases that previously required an oriented known fragment can be started from a single atom, at the cost of many more lottery cycles; the paper reports one case needing 243 cycles versus 36 with a fragment start.","Low data resolution makes structures harder but not impossible: a structure that needed one cycle at 0.77 Å resolution was solved after 113 lottery cycles when truncated to 1 Å.","The method is highly parallel because the sR1 map can be evaluated at all grid points simultaneously, so fast completion depends mainly on available computing power.","After the primary structure is found, extra lottery cycles can improve the model, and a bond-length-guided sR1 variant can fill in missing atoms when ghost sites compete."],"supporting_citations":[{"why":"Defines the single-atom R1 method and the cycle expansion that the lottery scheme automates.","marker":"Zhang & Donahue (2024)"},{"why":"Provided the inspiration of replacing poorly located atoms with randomly positioned atoms during ab initio phasing.","marker":"Burla et al. (2018)"},{"why":"Source of the idea that an R1-type calculation can identify poorly located atoms, which the lottery uses before splitting.","marker":"Kinneging & de Graaff (1984)"},{"why":"Supplies the reference solutions against which the paper's lottery-mode results are compared quantitatively.","marker":"Sheldrick (2015a)"},{"why":"Carries out the final refinement of sR1 results, yielding the reported R-values.","marker":"Sheldrick (2015b)"},{"why":"A reviewer comparison noted that the atom-rejection step resembles iterative peaklist optimization maximizing correlation.","marker":"Sheldrick & Gould (1995)"},{"why":"Used to determine the space group and asymmetric unit from the P1 result of the sR1 calculation.","marker":"Spek (2020)"}],"fun_headline_variants":["Lottery mode automates crystal structure solving","Random atomic lottery cracks crystal structures solo","Crystal structure solving gets a lucky break","No human needed: lottery method solves structures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on trusting that every time the approximate R1 number goes down, the model really is closer to the true atomic arrangement.","fun_headline_variants_meta":{"raw":{"variants":["Lottery mode automates crystal structure solving","Random atomic lottery cracks crystal structures solo","Crystal structure solving gets a lucky break","No human needed: lottery method solves structures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000412,"raw_usage":{"total_tokens":2129,"prompt_tokens":940,"completion_tokens":1189,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1135}},"tokens_in":556,"tokens_out":1189,"duration_ms":6964,"temperature":1.0,"reasoning_tokens":1135,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:53:39.546121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a set of datasets with known solutions, take a partially correct model, record the minimum approximate R1 reached after one cycle, then perturb the model by displacing atoms by 0.2, 0.5, 1.0, and 2.0 Å; if the minimum approximate R1 does not increase with displacement, the acceptance criterion can reward models that are farther from the truth.","supporting_citations":[{"cited_title":"& Donahue, J","cited_arxiv_id":null,"evidence_quote":"Defines the single-atom R1 method and the cycle expansion that the lottery scheme automates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Used to determine the space group and asymmetric unit from the P1 result of the sR1 calculation."}],"review_version":1}