docs: S2 design v0.1 amendments (target in seconds, grouping, R2 stop, slices)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
ClaudeandClaude Opus 5.5 committed 2026-11-14 09:07:16 +00:00
1 parent 3ae8980567
commit fd1db63ee1
1 file changed
+65
+65
View File
@@ -0,0 +1,65 @@
# S2 design v0.1: amendments to v0 (Claude, 2026-11-14)
v0 (`d8aa0f42`, s2-design-v0-2026-11-14.md) still stands except where amended below.
Inputs:
- MeltAstra's target catch (Slack, 17:0x);
- Grok's targeted gaps (17:0x);
- MoldAstra's stage-review note (Drive 10aMWqeBFKuAmKgssaWv-qf-cteNQuVcz).
MeltAstra's execution-level diff still has to be folded in, as v0.2. Nothing runs, and there is no training without Yanqi's go.
## A. Scope statement (Grok)
S2 v0 is a **one-factor ablation of O2 edge-direction readings**: E vs N, with the same backbone and the same g12/g23, plus R2.
It does **not** test the organisation claim we identified as unoccupied. That claim is whether type organisation still helps once a strong whole-part model, a matched generic partition, a control cut and a costed distance baseline are in the same table. S2 results must not be summarised as evidence for or against O1, or for "geometry organisation" in general.
Entering S2 on the edge readings is a *ranking-based* entry. S1 showed only about 1 % coverage structure for them against the bound-preserving control. It is not a confirmation of the original hypothesis.
The first test that can separate O1 from generic local-global organisation (matched generic partition vs type partition vs control cut) is **still unscheduled**. It is the next design item after S2 starts.
## B. Target (replaces v0 §3 "normalised per part by its maximum fill time")
- **Primary target:** Fill_time **in seconds**.
- Training may use log(1 + T/tau), with tau fitted on the training pool only.
- Predictions are mapped back to seconds before scoring.
- No per-part normalisation uses labels.
- In Greif v5 the filling control is Automatic: Moldflow picks injection time from the geometry, so the seconds scale is learnable from geometry inputs. The LOG's injection time/flow rate is **not** an input in S2, because it is ambiguous whether it is a process setting or an outcome. This waits for the in-house stage.
- **Scale-free secondary metrics:** Kendall tau over sampled node pairs; location of the last-filled region.
- **T/T_max with the true T_max** is a labelled oracle diagnostic only. It is never a primary result, and it is never used to report seconds (MoldAstra ②).
- **Coverage:** queries are the S1 mass-PPS draws on d > 0 nodes. Report the covered node mass (positive-distance share x eligible share) per part. Boundary (d = 0) mass is outside the target measure, stated explicitly.
## C. Grouping (MoldAstra ①; replaces "source key union near-duplicate cluster")
1. **Same source key, same side** (hard constraint).
2. **Across keys:** geometry screening with descriptors fixed before any results. For screen hits, registration, then a blind visual review of narrow bridges and holes. Record each confirmed pair as an **evidence edge**.
3. **Folds:** connected components of (same-key ∪ evidence edges) are the grouping units for the 5 folds and the nested training pools.
4. The earlier 0.5 % / 1 % thresholds were exploratory. There is no corpus-wide threshold acceptance and no final group table yet; both are entry conditions.
5. The 249 filename keys are leak-prevention keys, not verified independent CAD families.
## D. Baseline protocol and non-identifiability stop (Grok)
- **R2** follows the Gre24 feature recipe on *our* grouped split:
- geodesic/eikonal distance from the gate;
- local geometric scalars;
- GBRT.
Published Gre24 figures are **not** an external bar, since they use a different corpus version and split.
- **Lee26** reports a *pressure* drop (plate to bracket 11.93 % → 36.06 %, confounded). It is not a fill-time bar.
- **Label-resolution floor:** the Greif Moldflow logs show "Maximum % volume to fill per time step = 1 %". Fill-time labels are therefore resolved to roughly 1 % of each part's filling time.
- **Pre-registered stop:** if R2's source-equal relative L2 error on the 100 % pool is **≤ 2 %** (2x the label resolution), public Fill_time is declared **non-identifiable** for the E vs N question. The S2 comparison is then reported as uninformative, not as a win or a loss. The next step moves to harder targets (aligned pressure) or to in-house data.
## E. Added evaluation slices (Grok, from proposal v2.1)
All slices use geometry-only definitions fixed before the run.
- **Thin-feature slice:** nodes with d/t_part ≤ 0.25 that also lie within 2·t_part of a boundary face whose local opposing wall is closer than t_part. Operational definition to be frozen by MoldAstra; until then only the existing near-wall slice is reported.
- **Connectivity violation ("leakage") on predictions:** the bottleneck test from the 2026-11-13 thread.
- m(x) = min over gate paths of the max predicted T along the path.
- Violation if m(x) > T̂(x) + tol.
- tol is a preset fraction of the predicted part fill time.
- Report count and volume, *relative to the same statistic on the ground-truth labels*.
- Uses true tet adjacency only.
- **Freeze time** stays deferred until its field definition is sourced (G1 is open).
## F. Entry conditions (consolidated)
1. Final grouping table per section C, reviewed.
2. d/t_part on the full corpus, with recorded cost.
3. R2 implemented on the grouped split, with the stop check of section D computable before any neural run.
4. m (10 % candidate), the S2 seed table and slice definitions frozen.
5. Pilot measurement (t_step, memory) on one V100.
6. Yanqi's explicit go for training.