{
  "model": "J1J2",
  "lattice": "square",
  "n_sites": 324,
  "boundary": "P",
  "params": {
    "J2": 0.46
  },
  "instance_id": "J1J2/square_324_P_0.46",
  "rows": [
    {
      "energy": -656.2296,
      "sigma": null,
      "energy_variance": null,
      "dof": 324,
      "einf": 0,
      "v_score": null,
      "method": "CNN",
      "method_detail": "",
      "method_as_published": "CNN model with 106529 variational parameters",
      "family": "CNN / ResNet",
      "bound_type": "variational",
      "bound_type_reason": "variational ansatz at a stated size; energy is an upper bound (assigned during source reading)",
      "reference": "Li, Chen, Xiao, Wang, Jiang, Zhao, Lin, An, Liang & He, Bridging the Gap between Deep Learning and Frustrated Quantum Spin System for Extreme-Scale Simulations on New Generation of Sunway Supercomputer, IEEE Trans. Parallel Distrib. Syst. 33, 2846 (2022), doi:10.1109/TPDS.2022.3145163, arXiv:2108.13830",
      "peer_reviewed": true,
      "source": "sweep-allresults-2026-10-07",
      "provenance": "primary",
      "verified": {
        "checked_on": "2026-10-07",
        "method": "source text of arXiv:2108.13830 read locally (arXiv HTML or pypdf layout text); value copied from the harvested cell and the text, no LLM transcription of numbers; single reading",
        "reported_as": "-0.50635",
        "note": "Table 4 ('Energies for varied J2/J1 && system scales'), row 18 x 18, column 0.46; no error bar printed. Li et al. (IEEE TPDS 33, 2846, Sec. 3.2.1) discard samples whose local energy is far from the standard result and restart those chains, so the energy is a mean over a window, not a plain average; kept variational (qmbl-verify 2026-10-07, ruling 4, Tristan) | Energy per site in S.S units, periodic boundary conditions (text: 'periodic boundary condition'); the 10x10 and 16x16 cells at J2 = 0.50 equal the rows QMBL already carries from this table.",
        "secondary_of": null
      },
      "corrections": [
        {
          "field": "method",
          "to": "CNN model with 106529 variational parameters",
          "reported_as": "label: 'The CNN model with 106529 variational parameters is well optimized by millions of samples with the SR method and multiple stages of transfer learning' (arXiv v1-v4 and journal, contributions list, p. 2); value: '-0.50635'",
          "location": "label: arXiv:2108.13830v4 p. 2; value: Table 4 ('Energies for varied J2/J1 && system scales'), row '18 × 18', column 0.46 (arXiv v4 p. 10; journal p. 2856; Table 3 in v1/v2)",
          "version_read": "arXiv:2108.13830v1, v2, v3, v4 (PDFs downloaded by this reader, md5 equal to the shared-cache copies; pypdf plain + layout text) and the published IEEE TPDS 33(11), 2846-2859 (2022) PDF (IEEE Xplore copy of 2026-10-03 from qmbl-verify-2026-09-29/sources/journal, re-extracted here with pypdf); all five print the same cell",
          "conversion": "-0.50635 x 4 x 324 = -656.2296 (stored -656.2296: agrees; energy not changed)",
          "checked_on": "2026-10-07",
          "reason": "The stored published string is not printed anywhere in the source. Its parenthetical names the source, which RULES 3 keeps out of method_as_published and method_detail. Ruling B4 (Tristan, 2026-10-02): a corrected published string quotes a label the source prints, word for word, and what explains it goes to method_detail via scripts/method_names.mjs. The paper calls its method 'CNN'. The bare key 'CNN' already maps to ['CNN', ''] for Choo et al.'s different network, which shares square_100_P_0.55 and _0.6 with these rows. So the proposal quotes the longer printed label, which gives this network its own key and a detail: large model, Marshall sign (Eq. 9), C4 rotation sum (Eq. 10), fine-tuned by transfer learning from the 24 x 24 state (Sec. 4.3.3, Fig. 5). The 106529 parameters go to compute.parameters through the mapping's extra. Minimal alternative: 'CNN' alone, which leaves the detail empty, as it is now. The sibling row J1J2/square_256_P_0.5 (-508.17024, from the 2026-09-15 pass, outside this group) carries the same string and needs the same change. Keep the old method_names key until that row moves too.",
          "source_entry": "RP1-arxiv-2108-13830#46",
          "from": "CNN (Li et al. 2022, sunway supercomputer)"
        },
        {
          "field": "verified.note",
          "edit": [
            "; no error bar printed |",
            "; no error bar printed. Li et al. (IEEE TPDS 33, 2846, Sec. 3.2.1) discard samples whose local energy is far from the standard result and restart those chains, so the energy is a mean over a window, not a plain average; kept variational (qmbl-verify 2026-10-07, ruling 4, Tristan) |"
          ],
          "checked_on": "2026-10-07",
          "reason": "Ruling 4 of qmbl-verify 2026-10-07: the outlier window stated on the row.",
          "source_entry": "RP1-arxiv-2108-13830 (ambiguous, bound_type); ruling 4, Tristan 2026-10-07",
          "from": "Table 4 ('Energies for varied J2/J1 && system scales'), row 18 x 18, column 0.46; no error bar printed | Energy per site in S.S units, periodic boundary conditions (text: 'periodic boundary condition'); the 10x10 and 16x16 cells at J2 = 0.50 equal the rows QMBL already carries from this table.",
          "to": "Table 4 ('Energies for varied J2/J1 && system scales'), row 18 x 18, column 0.46; no error bar printed. Li et al. (IEEE TPDS 33, 2846, Sec. 3.2.1) discard samples whose local energy is far from the standard result and restart those chains, so the energy is a mean over a window, not a plain average; kept variational (qmbl-verify 2026-10-07, ruling 4, Tristan) | Energy per site in S.S units, periodic boundary conditions (text: 'periodic boundary condition'); the 10x10 and 16x16 cells at J2 = 0.50 equal the rows QMBL already carries from this table."
        }
      ],
      "compute": {
        "parameters": 106529,
        "gpu_hours": null,
        "device": null,
        "n_devices": null,
        "samples": null,
        "wall_clock": null,
        "cpu_core_hours": null,
        "bond_dimension": null,
        "iterations": null,
        "reported_as": "method string: \"CNN model with 106529 variational parameters\"",
        "source": "the row's published method string (scripts/method_names.mjs, 2026-09-17)",
        "scope": "row",
        "confidence": "high",
        "note": "Only the numbers the method string states; nothing else about this run's cost is recorded."
      }
    },
    {
      "energy": -656.43696,
      "sigma": null,
      "energy_variance": null,
      "dof": 324,
      "einf": 0,
      "v_score": null,
      "method": "CNN",
      "method_detail": "CNN1, 1 Lanczos step",
      "method_as_published": "CNN1",
      "family": "CNN / ResNet",
      "bound_type": "variational",
      "bound_type_reason": "variational ansatz at a stated size; energy is an upper bound (assigned during source reading)",
      "reference": "X. Liang, M. Li, Q. Xiao, J. Chen, C. Yang, H. An and L. He, Deep learning representations for quantum many-body systems on heterogeneous hardware, Mach. Learn.: Sci. Technol. 4, 015035 (2023) (doi:10.1088/2632-2153/acc56a), Table 1 (journal version only; arXiv:2204.07816v1 treats J2 = 0.5 alone)",
      "peer_reviewed": true,
      "source": "sweep-allresults-2026-10-07",
      "provenance": "primary",
      "verified": {
        "checked_on": "2026-10-07",
        "method": "source text of arXiv:2204.07816 read locally (arXiv HTML or pypdf layout text); value copied from the harvested cell and the text, no LLM transcription of numbers; single reading",
        "reported_as": "−0.506 51",
        "note": "Table 1 of the published article ('achieved by CNN1 with 106 529 parameters. The values are obtained by the Lanczos step with p = 1'), row 18 x 18, column 0.46; no error bar printed | Energy per site in S.S units on periodic lattices (abstract: 'rectangular lattices within periodic boundary conditions'); its 10x10 cells at J2 = 0.50 and 0.60 equal the rows QMBL already carries.",
        "secondary_of": null,
        "second_read": {
          "checked_on": "2026-10-07",
          "scope": "value",
          "method": "Mach. Learn.: Sci. Technol. 4, 015035 (2023), journal version: IOPscience full-text HTML, two independent browser snapshots (run sources/doi-10.1088-2632-2153-acc56a.iop.html, 2026-10-07; qmbl-verify-2026-09-29/sources/liang2023-mlst-acc56a.html), Table 1 parsed locally with node (grids identical); arXiv:2204.07816v1 PDF (only arXiv version, md5 aadd407a...) via pypdf layout and plain text for the version comparison",
          "location": "Table 1, row 18×18, column 0.46",
          "reported_as": "−0.506 51 (energy per site, S.S units, J1 = 1; no error bar printed)",
          "conversion": "-0.50651 x 4 x 324 = -656.43696",
          "note": "S.S per site x 4N = Pauli total (journal Eq. (1) spin operators, J1 = 1; the authors' released MCMC.py divides the Pauli local energy by 4 x site_number); L x L torus ('rectangular lattices with PBC'); J2 from the Table 1 column header; p = 1 Lanczos energy, variational (the 10x10 and 12x12 J2 = 0.50 cells equal the text's p = 1 values, not its extrapolations); no error bar or variance printed in the journal, arXiv v1 or the code release, so sigma and energy_variance stay null and the row is eligible for nothing (RULES 6); not a duplicate: no other row on the instance carries this number from this or a quoting paper (checked data/<instance>.json read-only; the same group's row on every instance is Li et al., IEEE TPDS 33, 2846 (arXiv:2108.13830), a different paper with a higher energy). Cosmetic: the stored verified.method says 'source text of arXiv:2204.07816', but Table 1 is in the journal version only (arXiv v1 fixes J2 = 0.5 and has no such table); the loader template in scripts/add_allresults_rows.mjs fills arXiv:${r.arxiv} although make_rows.mjs read sources/doi-10.1088-2632-2153-acc56a.txt.",
          "source_entry": "RP5-arxiv-2204-07816#18"
        }
      }
    }
  ],
  "url": "https://qmbl.org/i/J1J2/square_324_P_0.46/",
  "per_site_divisor": 1296,
  "per_site_label": "E/N (S.S)",
  "record": null,
  "no_record_reason": "sampled rows carry no error bar"
}