diff --git a/docs/claude/s2-design-v0.1a-2026-11-14.md b/docs/claude/s2-design-v0.1a-2026-11-14.md new file mode 100644 index 000000000..4e1e3350c --- /dev/null +++ b/docs/claude/s2-design-v0.1a-2026-11-14.md @@ -0,0 +1,21 @@ +# S2 design v0.1a: replaces v0.1 sections D (stop rule) and E (thin slice) (Claude, 2026-11-14) + +Source: MoldAstra's D/E review (Drive 18X07uNtM1j697MCztXryzmITCcud3_T4). Accepted. + +## D, replaced +- Withdrawn: "1 % volume per step → about 1 % label time resolution → 2 % non-identifiability stop". + - A volume step limit does not bound time resolution: a 1 % volume step can span a large share of total time. + - Even under constant flow, there is no positive lower bound on node write-out or interpolation error, nor on relative L2. +- No mandatory stop. In its place: + - Report R2's absolute source-equal relative L2 next to every E vs N result. + - The six-region decision table alone decides E vs N. + - If a later independent engineering basis justifies a budget stop line, it must be pre-registered before the pilot. When it triggers, the result reads "E−N not tested". + +## E, replaced +- **Primary slice:** commonly eligible nodes with 0 < d/t_part ≤ 0.25 (the near-wall layer). +- **Thin-feature slice (MoldAstra definition):** + - For each boundary face, march along the inward normal from the face centroid to the first exit from the cavity. + - Keep faces whose ray hits an opposing wall with chord < t_part. + - Select near-wall nodes within a true tet-edge path distance ≤ 2·t_part of those faces. + - Until this mask is validated, report only the near-wall slice, with "unknown" counted separately. +- Connectivity-violation (bottleneck) metric: unchanged from v0.1.