From 439c02803c2cb3c66af78dafa5bc4025ed992255 Mon Sep 17 00:00:00 2001 From: Codex Date: Thu, 19 Nov 2026 21:49:47 +0800 Subject: [PATCH] Archive received SR-20261119-2137 contributions --- docs/stage-review-SR-20261119-2137/grok.md | 25 +++++++++++++++++++ docs/stage-review-SR-20261119-2137/meltwin.md | 5 ++++ docs/stage-review-SR-20261119-2137/tron.md | 5 ++++ .../undermind.md | 5 ++++ 4 files changed, 40 insertions(+) create mode 100644 docs/stage-review-SR-20261119-2137/grok.md create mode 100644 docs/stage-review-SR-20261119-2137/meltwin.md create mode 100644 docs/stage-review-SR-20261119-2137/tron.md create mode 100644 docs/stage-review-SR-20261119-2137/undermind.md diff --git a/docs/stage-review-SR-20261119-2137/grok.md b/docs/stage-review-SR-20261119-2137/grok.md new file mode 100644 index 000000000..68d4bb091 --- /dev/null +++ b/docs/stage-review-SR-20261119-2137/grok.md @@ -0,0 +1,25 @@ +# Grok · 独立文献意见 + +原稿发表于正式配对结果之前。原稿当时的点估计状态按原文保留。 [Slack 原稿](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1795095723719089?thread_ts=1794219487.255049&cid=C0C78UFKGRM) + +SR-20261119-2137, Grok, questions 3 and 4. Independent, public sources only; no new compute; I have not read other members' reports. Brief at `e0c8a27` [read]. H2-P is *points only, formal verdict pending*. Internal Greif/H2 figures are taken from that brief and are not re-audited here. + +*3. What the matched ShapeNet result says — and does not say — about SOTA* + +The only *verified same-protocol* published number is GINO's own Table 2: GINO (decoder) 7.12% *de-normalised* test L2 on 611 watertight ShapeNet-car shapes, 500/111. [read: Li et al., NeurIPS 2023, §4.2 and Table 2, ]. The paper does not print the split IDs, so later “500/111” tables are not automatically the same split unless they use the same package IDs. §4.2 also calls the 111 a *validation* set, while Table 2 says *test error*. [read] + +The project's three-seed reproduction on that package (brief: 7.43%, seed spread 0.8 pp) is consistent with 7.12% and supports a matched protocol. It does not move any public leaderboard. [brief; not a public paper] + +OTNO(Plan) 6.70% ([JMLR 26:25-1380](https://www.jmlr.org/papers/volume26/25-1380/25-1380.pdf), Table 1) and PCNO 7.04% ([arXiv:2501.14475](https://arxiv.org/pdf/2501.14475), §5.4) remain *undetermined*: split IDs, de-normalisation, and seed count are not stated. OTNO's own GINO baseline is 7.21%, which does not match Table 2. [read]. Transolver and the sota2 board use other splits, fields, or metrics and stay out of this table. [read] + +H2-P points in the brief (mean 10.01% vs reproduced GINO 7.43%) sit behind GINO on this task, but that is not a closed comparison until the paired verdict. Even after a verdict, public literature only licences “better or worse than GINO under GINO's protocol”, not “SOTA”: there is no verified post-GINO table on the same IDs, metric, and seed rule. [inferred from the sources above] + +The result also does not speak to molding, or to generality. ShapeNet-car is one object class, one RANS setup (k–ε, 20 m/s, Re ≈ 5×10^6), and one surface-pressure field. [read, GINO §4.2] + +*4. Is GINO-Greif the right next test? What would probe generality?* + +GINO-Greif is the right *next architecture-transfer test*, not a generality test. The only verified 3D operator on the ShapeNet protocol is GINO; the natural check is whether that operator, under Greif's grouped fill-time protocol, matches or beats H2. If it does, a GINO-type default is warranted on both public tasks we have. If it does not, H2 has a molding-specific advantage worth isolating. Suggested stop: a pre-registered paired comparison against H2 on the frozen Greif protocol; stop when GINO is clearly ahead, clearly behind, or tied within the pre-registered interval — no informal retuning. [inferred] + +It does *not* probe generality. Every public Greif–Meyer fill-time release I found (v1–v5 and the 2025 warpage variant) uses one solver, one material, and one process rule (a random surface gate). Packing is zeroed only in the warpage set; that should not change the fill phase [inferred]. There is no official independent-family split. [read: ; ; warpage ] + +A genuine generality benchmark would need all of: (a) multiple independent part families with an official held-out-family split; (b) explicit and varying gates; (c) material and process-window variation, not a single setpoint; (d) per-node fill time or the actual decision field; (e) a public licence. I have not found such a public set. [read; item 1 in this thread] diff --git a/docs/stage-review-SR-20261119-2137/meltwin.md b/docs/stage-review-SR-20261119-2137/meltwin.md new file mode 100644 index 000000000..abda379fd --- /dev/null +++ b/docs/stage-review-SR-20261119-2137/meltwin.md @@ -0,0 +1,5 @@ +# MeltWin · Adviser 导入与导出实查 + +由 TRON 逐字转达。 [Slack 原稿](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1795095768451039?thread_ts=1794219487.255049&cid=C0C78UFKGRM) + +已收到 SR-20261119-2137-MeltWin。在 ARON 上,Greif 778 原始 BDF 的完整性和网格计数已核实,但 Adviser 2027 的导入列表没有 BDF,通过 All Files 选择也无法读取,因此未能建立保留原四面体网格的 study,也未验证逐节点 Fill Time 数值导出。求解 0 次,云积分 0。 若未来另行授权替代路线,须先确认许可适用,并在不求解的情况下验证原网格、单位和浇口位置,同时确认数值结果能通过 ID 或坐标关联回同一网格。停工状态继续有效。 diff --git a/docs/stage-review-SR-20261119-2137/tron.md b/docs/stage-review-SR-20261119-2137/tron.md new file mode 100644 index 000000000..881370486 --- /dev/null +++ b/docs/stage-review-SR-20261119-2137/tron.md @@ -0,0 +1,5 @@ +# TRON · 流程意见 + +[Slack 原稿](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1795095656167739?thread_ts=1794219487.255049&cid=C0C78UFKGRM) + +TRON process note. Keep three questions separate: retrainable architecture, unseen-family generalization within the same physical task, and weight transfer across tasks. Publish outcomes and failure boundaries together, including the last-fill tradeoff, the car pressure result, and unknown protocol equivalence of outside rankings. Use the same tasks, IDs, metrics, inputs and budget accounting in comparisons; keep test targets sealed before the scoring protocol is frozen. Disclose results and decide next tests without treating this meeting as launch approval. diff --git a/docs/stage-review-SR-20261119-2137/undermind.md b/docs/stage-review-SR-20261119-2137/undermind.md new file mode 100644 index 000000000..3f45ee88b --- /dev/null +++ b/docs/stage-review-SR-20261119-2137/undermind.md @@ -0,0 +1,5 @@ +# Undermind · 工具辅助文献意见 + +由 TRON 报告在现有 Undermind 项目中的只读结果。 [Slack 原稿](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1795095656167739?thread_ts=1794219487.255049&cid=C0C78UFKGRM) + +Undermind literature note. H2-P changed inputs and outputs and trained from scratch on the car data. This demonstrates an adaptable architecture, not transfer of molding-trained weights. [GINO §4.2/Table 2](https://doi.org/10.48550/arXiv.2309.00583) reports surface pressure, 500 training/111 held-out cars and de-normalized test error, but no per-case IDs in the paper. The project matched the supplement IDs; H2-P 10.01% vs reproduced GINO 7.43% supports the narrower baseline comparison, not a SOTA ranking. [Multi-Physics Pretraining §§3–4](https://doi.org/10.48550/arXiv.2511.10829) explicitly compares adapting pretrained models with training from scratch; using target-task labels is not zero-shot. After a matched Greif comparison, test unseen part families with verified gates/process and separate changed-process cases. To claim cross-physics weight transfer, compare pretrained vs random initialization under fixed target data and compute; reserve zero-shot for no target-label training or tuning. Both PDFs were checked in the existing Undermind corpus.