diff --git a/docs/claude/phase1-synthesis-2026-11-17.md b/docs/claude/phase1-synthesis-2026-11-17.md index aca72d75d..439232f03 100644 --- a/docs/claude/phase1-synthesis-2026-11-17.md +++ b/docs/claude/phase1-synthesis-2026-11-17.md @@ -15,11 +15,11 @@ Literature is marked as read in full by a tool (not page by page by a person); n - no false connections across gaps, no misclassified roots; - sheet coverage of the true volume is 100 / 80 / 69.23 %, the rest staying 3D. - **0 Greif parts were tried**, because the current certificate only supports axis-aligned regular voxel columns; eligibility of the whole library was not assessed part by part. A general-domain certificate would be new engineering whose cost is unknown. (Wording corrected per MoldAstra: not "mesh-independent", not "0 Greif parts eligible".) -3. **Physics.** Hele-Shaw / 2.5D holds in thin sheets and breaks at junctions, thick regions and the end of filling. Physical structure that can be embedded without new solvers: arrival-time causality along propagation paths, eikonal/level-set structure, and mass conservation on a fixed domain. None of these guarantees front accuracy. +3. **Physics (conditional).** Hele-Shaw / 2.5D can apply where its thin-gap, scale, constitutive/thermal-closure and boundary assumptions hold; junctions, thick regions and the end of filling carry a risk of failure. Neither has been verified region by region on Greif. Embedding conservation structure still requires building the operators and boundaries; it is not free. Physical structure that can be embedded without new solvers: arrival-time causality along propagation paths, eikonal/level-set structure, and mass conservation on a fixed domain. None of these guarantees front accuracy. 4. **Machine learning.** Geometry-aware operators (GINO, Transolver-3 and others) show feasibility at about 400 to 500 geometries, but with random within-family splits and with **no fill-time evidence**. The recommended structure is true local connectivity, coarse global communication, and queries on the original domain. Geometry latents and physics latents should be kept separate. Many "small-data" claims count trajectories, not shapes. 5. **Our own facts.** - The faithful Gre24 GBRT (A) scores 7.08 %; the training-free EIK 9.59 %. - - **EIK finds the last-filled location better than A.** + - On the fixed last-filled-location metric, EIK's point value (0.342 Rg) is better than A's (0.377 Rg). That does not make EIK more accurate about end-of-fill flow in general. A's 7.08 % is the R4bc 552/204 component-mean relative L2, not the same quantity as Gre24's random-split figure. - The learning curve keeps improving over the tested range, but slowly: A's nested pools give 7.703 / 7.308 / 7.083 %. D2 cannot decide whether longer training or more samples would help. - The S2 Transolver recipe scored 10.8 % (a different protocol, and the recipe may not have been adequate). - Greif v2 has 42 parts from 40 source keys not seen in v4 or v5: a candidate fresh hold-out set, not yet vetted. @@ -29,10 +29,10 @@ Literature is marked as read in full by a tool (not page by page by a person); n - Arm (corrected per MoldAstra): EIK's dimensionless prediction *is* the strict volume quantile F, which is already one of A's 19 columns, so adding it as an input adds nothing. What changes is the **target**: F as a fixed skip connection, with the model learning t* - c F from the same 19 features. - Against A, with field L2, last-filled location and Kendall tau all pre-registered as co-primary or secondary. Optionally add a causality penalty along propagation paths. - Question: can one model keep EIK's end-of-fill accuracy and A's field accuracy? - - Cost: hours of CPU. + - Cost: hours of CPU (a planning estimate, not measured). The residual target does not automatically keep EIK's end-of-fill advantage; the joint metrics have to show it. - **H2, geometry-representation route (after H1, if Yanqi chooses) (Yanqi's direction; higher cost):** a geometry-aware neural operator with the structure the ML map recommends, *using the H1 target (correction on EIK) as its physics prior*. - Against A and against H1, with the same metrics. - Question: does a learned geometry representation add information beyond hand-made distance and moment features, at about 400 training parts? - - Cost: GPU days; the recipe needs care (S2's lesson). + - Cost: GPU days (a planning estimate); the recipe needs care (S2's lesson). Using H1's target in H2 is a research choice, not a requirement of geometry operators. Future contracts name these H1/H2 distinctly from the old R4 ablation labels. - **Not proposed now:** the explicit sheet/manifold decomposition. It needs a general-domain certificate first, which is a separate engineering step whose payoff is uncertain. MoldAstra and MeltAstra could scope it on paper if Yanqi wants that route. -- **Protocols:** the 552/204 grouped protocol (reused after exploration, reported as such) plus the v4 random protocol for the paper comparison. If the 42 v2 parts pass vetting (read-only check of LOG/YML settings and field validity), they become a fresh hold-out set, frozen before any result. +- **Protocols:** the 552/204 grouped protocol (reused after exploration, reported as such) plus the v4 random protocol for the paper comparison. The 42 v2 parts / 40 source tokens are file identities only. They could become a fresh hold-out set only after: LOG/YML and field validity pass; same-source and cross-source near-duplicates and component isolation are checked; and their use is frozen before any result. A new source key on its own is not evidence of an independent geometry.