{"id":"9a0f2ef3-52d8-4acb-870f-32d06d694633","arxiv_id":"2607.06604","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"GRAMSCI v2 replaces binary-search N-point enumeration with O(m) merge-walks, adds parity-decomposed and connected 4pCF, and ports the query engine to OpenACC GPUs with out-of-core tiling.","lead":"GRAMSCI v2 speeds up counting of galaxy triplets and tetrahedra with a merge-walk algorithm and a GPU port, plus parity-odd and connected 4-point modes. That makes higher-order BAO and parity tests practical on DESI-scale catalogues on a single consumer GPU.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper’s strongest claim is fully supported by the evidence presented: multi-generation timing curves (Fig. 2), thread scaling (Fig. 3), density and out-of-core overheads (Figs. 4–5), Table 2 DESI timings, and explicit CPU–GPU count agreement at production scale (§4.8). The reader correctly flags the empirical GPU layout under clustered tails as the softest premise, but that premise is already checked by the same production-scale DESI measurement and by the analytic/Quijote validation suite. Remaining limitations (diagonal-only χ², no multi-GPU, no pinned commit hash) are minor and do not undermine the engineering claim. Therefore the ACCEPT / HIGH-confidence verdict stands; no adjustment is warranted.","tokens_in":19616,"tokens_out":587,"duration_ms":5369,"concrete_test":"Re-run the repository regression suite (isolated-pair 2pCF, regular-tetrahedra connected 4pCF, chiral-tetrahedra parity 4pCF) plus the DESI-scale chunked 3pCF comparison on a second GPU architecture (e.g. A100 or consumer Ada); if any binned count differs from the CPU reference by more than floating-point accumulation noise (~10^{-6} relative), the concurrency design would need re-examination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an engineering one: OpenACC GPU kernels for 3pCF/4pCF/parity-4pCF deliver measured 2.6–9× speedups over a 64-thread CPU node on identical estimators, with out-of-core tiling that processes a 9×10^9-edge DESI graph on a 24 GB card at ~20% overhead while matching CPU binned counts. The reader’s weakest assumption (that the gang-per-hub / shared-list / gang-slot / atomic-partial design remains race-free under FKP high-multiplicity tails) is already stress-tested by the paper itself: §4.8 reports exact count agreement for all four kernels in single-pass and chunked modes, including the production DESI 3pCF over 9×10^9 edges (max relative difference 7×10^{-7}, attributed to accumulation order); §4.9 adds analytic chiral-tetrahedron injection and a Quijote parity-null (χ²/dof ≈ 1.10). §5.2 further notes that survey 4pCF cost is dominated by ⟨m³⟩ ≫ ⟨m⟩³ yet still reports connected-4pCF consistency with EZmocks. No internal inconsistency or untested regime that would overturn the measured speedups or correctness claims is evident.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"This paper presents GRAMSCI Version 2, a substantial update to the 2019 graph-database algorithm for configuration-space N-point correlation functions. The main contributions are: (i) a merge-walk sorted-set intersection that reduces the 3pCF/4pCF inner loops from O(m log m) to O(m); (ii) a parity-decomposed 4pCF with deterministic handling of zero-volume tetrahedra; (iii) internal estimation of the disconnected 4pCF on the same graph to return the connected 4pCF; (iv) a NumPy Python interface; and (v) an OpenACC GPU port of the full query engine with out-of-core window tiling. Measured speedups on identical estimators are 2.6× (3pCF) to 9× (4pCF) versus a 64-thread CPU node, with a 9×10^9-edge BAO-scale DESI 3pCF run on a 24 GB card at ~20% overhead. Correctness is checked by exact CPU–GPU count agreement (max relative difference 7×10^{-7}), analytic chiral-tetrahedron injection (0.3% recovery), a Quijote parity-null (χ²/dof ≈ 1.10), and DESI DR1 LRG 3pCF/connected-4pCF consistency with the EZmock ensemble.","tokens_in":19936,"tokens_out":1221,"duration_ms":13041,"significance":"If the reported performance and correctness hold, the work is a practically useful methods contribution for large-scale structure. Configuration-space connected and parity-odd 4pCFs at BAO scales have been expensive; the combination of merge-walk, internal disconnected subtraction, and a single-consumer-GPU out-of-core engine makes DESI-scale covariance estimation and parity cross-checks more routine. Strengths that should be credited explicitly include: public code, exact binned-count agreement between CPU and GPU (including chunked production-scale DESI), analytic injection tests for the parity-odd channel, a field-level Quijote null, and end-to-end application to DESI DR1 LRG versus EZmocks. The paper deliberately avoids cross-code timing claims and reports wall-clock measurements on fixed hardware and identical estimators, which is the right standard for an engineering result.","major_comments":[{"comment":"§5.2 notes that survey 4pCF cost is dominated by the high-multiplicity tail (⟨m³⟩ ≫ ⟨m⟩³) and that NGC at r_max=80 h^{-1} Mpc takes 2.6 h, an order of magnitude beyond the uniform-catalogue scaling of §4.3. Table 2 and Figure 2 report GPU speedups primarily on the quasi-uniform test catalogue (and on DESI 3pCF). For the central 4pCF claim, a short production-scale GPU timing (or at least a clear statement of whether the 9× figure still holds under FKP-weighted clustered tails) would make the headline speedup more representative of the science use case the paper itself emphasizes.","section":null},{"comment":"§3.3 and Eq. (4)–(5): the disconnected subtraction uses an internally computed ξ on the same graph and bins, which is the right design. The manuscript should state more explicitly whether the same Landy–Szalay-style signed-weight estimator and RR normalization are used for every pair partition, and whether any edge-effect or bin-center approximation enters when evaluating ξ(r_ab) for the six edges of a binned tetrahedron. A one-sentence clarification would remove residual ambiguity about self-consistency of ζ_conn.","section":null}],"minor_comments":[{"comment":"Figure 2 caption and §4.3: the fixed ~0.7 s GPU offset is important for interpreting the crossover scales; consider marking the transfer/launch floor explicitly on the plot or in the legend.","section":null},{"comment":"§3.2: the |V|<10^{-9} threshold for degenerate tetrahedra is free but well-motivated; a brief note on sensitivity (or that the even channel is unaffected) would help readers who re-implement the estimator.","section":null},{"comment":"§3.5.3 / Table 2: the out-of-core DESI 3pCF (W=5, ~20% overhead) is a strong result; stating the window-size selection rule or environment-variable override more prominently would aid reproducibility.","section":null},{"comment":"Listing 1 and §3.4: the Python interface is welcome; a one-line note on whether temporary-directory I/O is the only path or whether an in-memory/ctypes path is planned would set expectations.","section":null},{"comment":"§4.7: expected A100/H100 scaling is framed as bandwidth-based expectation, not measurement — keep that distinction clear in the abstract/conclusions if those numbers are quoted elsewhere.","section":null},{"comment":"Minor typography: title and running headers show “F ast” (space in Fast); arXiv abstract has “Sabiu et.al” (missing space/period consistency).","section":null}],"recommendation":"minor_revision","confidential_remarks":"This is a solid methods paper for astro-ph.IM. The reader’s and skeptic’s assessments that the race-free GPU design is already stress-tested by production-scale count agreement are correct; I do not see a load-bearing correctness failure. The two major comments are scoped as clarifications that strengthen the 4pCF performance claim and the connected-estimator description, not as blockers. Fit for a methods-oriented journal is good; novelty is incremental over Sabiu et al. (2019) but the GPU+parity+connected package is useful enough to publish after minor revision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean methods update that actually ships. The headline is the GPU query engine: OpenACC kernels for 3pCF, 4pCF, and parity-decomposed 4pCF, with out-of-core window tiling that runs a 9e9-edge BAO-scale DESI graph on a 24 GB card at ~20% overhead and matches the CPU counts to floating-point noise. Merge-walk replaces the old binary-search inner loop (O(m) instead of O(m log m)), and they add internal disconnected subtraction for the connected 4pCF plus a direction-pixel parity split with a deterministic zero-volume rule.\n\nWhat is new is the combination and the production path, not the individual algorithms. Sorted-set intersection, CSR, and tiling are standard; parity-odd and connected 4pCF are already science targets in the harmonic literature (Encore, Hou/Philcox/Slepian). The paper is honest about being GRAMSCI v2. Credit where due: the validation is careful. CPU–GPU binned counts agree (max relative difference 7e-7 on the full DESI 3pCF), chiral-tetrahedron injection recovers the injected asymmetry to 0.3%, Quijote FoF halos give a parity-odd null with χ²/dof ≈ 1.10, and DESI 3pCF/connected 4pCF track the EZmock ensemble. Code is public, Python wrapper is there, and they flag that survey 4pCF cost is dominated by the high-multiplicity tail (⟨m³⟩ ≫ ⟨m⟩³), so uniform-catalogue benchmarks are lower bounds.\n\nSoft spots are minor and proportionate. Novelty is moderate. χ² values are diagonal-only. Free parameters (volume threshold, window size, pixelization) are documented but not exhaustively ablated. The gang-per-hub / shared-list / atomic-partial design is empirical; the paper already stress-tests it on the production DESI graph and the clustered 4pCF case, so the reader’s weakest-assumption worry does not overturn the claims. Citation pattern is appropriate.\n\nThis is for people who need configuration-space higher-order stats on DESI-scale catalogues and mock ensembles, especially parity or connected 4pCF without committing to a harmonic basis. It deserves a serious referee. I would engage with the code and cite the speed/validation numbers when discussing practical 3/4pCF pipelines.","headline":"Solid engineering v2 of GRAMSCI: merge-walk + OpenACC GPU with out-of-core tiling delivers measured 2.6–9× speedups and usable connected/parity 4pCF on DESI-scale graphs, with careful validation.","tokens_in":20584,"tokens_out":679,"would_cite":true,"duration_ms":7887,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A GPU graph engine makes BAO-scale 3- and 4-point galaxy clustering, including parity-odd 4-point signals, practical on a single consumer card.","keywords":["N-point correlation functions","GPU computing","large-scale structure","parity violation","BAO","graph algorithms","DESI","connected four-point function"],"falsifier":"Run the GPU and CPU 3pCF/4pCF/parity kernels on the same production DESI-scale FKP-weighted graph (or a controlled high-multiplicity synthetic with known analytic counts) and check whether every binned configuration still agrees to floating-point tolerance after out-of-core tiling; any systematic discrepancy falsifies the claim of exact, race-free acceleration.","tokens_in":20466,"feed_emoji":"⭐","tokens_out":1168,"duration_ms":11051,"temperature":0.7,"pith_summary":"Higher-order clustering statistics of galaxies carry non-Gaussian and parity information that the two-point function cannot reach, but counting every triangle or tetrahedron at survey scale is expensive. This paper updates the GRAMSCI graph method so that a single adjacency graph, built once from a kd-tree, supports all subsequent N-point queries. A merge-walk replaces binary search and cuts the inner loop from O(m log m) to O(m). The same graph yields a parity-even/odd 4-point split, an internal disconnected subtraction that isolates the connected 4-point function, and a full OpenACC GPU query engine. On identical estimators a consumer GPU runs 2.6\times to 9\times faster than a 64-thread CPU node, and an out-of-core tiling scheme measures a 9-billion-edge BAO-scale 3-point function on a 24 GB card at about 20 percent overhead. The code is validated against its CPU reference, analytic injections, and DESI DR1 LRG data versus mocks, opening routine BAO-scale 3-point, connected 4-point, and parity tests on modern catalogues.","feed_headline":"GPU graph engine speeds galaxy 4-point stats up to 9×","feed_subtitle":"BAO-scale 3- and 4-point, plus parity-odd, run on one consumer card with out-of-core tiling","key_machinery":"The sorted adjacency graph plus merge-walk (and three-pointer walk for tetrahedra): hub and neighbor lists are already ordered by node index, so their intersection is a single linear two- or three-pointer pass of cost O(m) rather than repeated binary searches; the same CSR layout is mapped to OpenACC gangs (one hub per block) with vector lanes walking a shared list, enabling the GPU kernels and out-of-core tiling.","core_discovery":"The complete GRAMSCI query engine for the 3-point, 4-point, and parity-decomposed 4-point correlation functions can be ported to a single consumer GPU via OpenACC, delivering measured speedups of 2.6\times (3pCF) to 9\times (4pCF) over a 64-thread CPU node on identical binned estimators, while an automatic window-tiling scheme processes graphs larger than device memory (demonstrated on a 9\times10^9-edge DESI BAO graph on a 24 GB card at ~20% overhead) and reproduces CPU counts to floating-point precision.","pith_inferences":["Because the tile decomposition partitions work by hub window with no inter-tile communication, a multi-GPU extension that assigns one hub window per device is a natural next step for even larger surveys.","The same CSR-plus-merge-walk structure is geometry-agnostic and could be reused for higher-order clustering of any discrete point set (halos, voids, radio sources) once the graph is built.","If the connected 4-point residual remains consistent with mocks while the parity-odd channel is non-zero, the code supplies a clean separation between late-time non-Gaussianity and a true parity-violating signal."],"forward_implications":["Connected 4-point measurements across the full BAO range become feasible for DESI-scale catalogues, including the thousands of mocks needed for covariance estimation.","A configuration-space parity-odd 4-point estimator with deterministic treatment of zero-volume tetrahedra is available as an independent cross-check of harmonic-basis parity searches.","Joint 2+3+4-point BAO analyses that bin anisotropic multipoles can extract extra signal-to-noise from the same galaxy sample.","Graphs exceeding device memory (billions of edges) can be measured on a single consumer card with only modest PCIe overhead, removing the need for multi-node CPU clusters for many survey analyses."],"fun_headline_variants":["GPU graph engine hits 9× on galaxy 4-point stats","Consumer GPU ports full 3pCF and parity 4pCF kernels","OpenACC GRAMSCI runs BAO-scale 4pCF 9× faster","Out-of-core tiling enables billion-edge GPU 3pCF","Parity-odd 4pCF joins 9× GPU speedup over CPU node"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That the GPU mapping rules (one block per hub, shared-list vector lanes, gang-slot scratch, and atomic slot-strided accumulation) stay free of race or ordering bugs on the heavy high-multiplicity tails of real clustered, FKP-weighted survey graphs where four-point cost is dominated by rare dense hubs.","fun_headline_variants_meta":{"raw":{"variants":["GPU graph engine hits 9× on galaxy 4-point stats","Consumer GPU ports full 3pCF and parity 4pCF kernels","OpenACC GRAMSCI runs BAO-scale 4pCF 9× faster","Out-of-core tiling enables billion-edge GPU 3pCF","Parity-odd 4pCF joins 9× GPU speedup over CPU node"]},"model":"grok-4.5","effort":"low","cost_usd":0.004162,"raw_usage":{"total_tokens":1391,"prompt_tokens":998,"num_sources_used":0,"completion_tokens":88,"cost_in_usd_ticks":41620000,"prompt_tokens_details":{"text_tokens":998,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":305,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":998,"tokens_out":88,"duration_ms":3805,"temperature":1.0,"reasoning_tokens":305,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T02:09:26.445542+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the GPU and CPU 3pCF/4pCF/parity kernels on the same production DESI-scale FKP-weighted graph (or a controlled high-multiplicity synthetic with known analytic counts) and check whether every binned configuration still agrees to floating-point tolerance after out-of-core tiling; any systematic discrepancy falsifies the claim of exact, race-free acceleration.","supporting_citations":[],"review_version":1}