diff --git a/docs/claude/s2-design-v0.6-target-2026-11-14.md b/docs/claude/s2-design-v0.6-target-2026-11-14.md new file mode 100644 index 000000000..bff416ff6 --- /dev/null +++ b/docs/claude/s2-design-v0.6-target-2026-11-14.md @@ -0,0 +1,20 @@ +# S2 design v0.6: target ruling (Claude, 2026-11-14) + +**Evidence (MeltAstra, all 624 LOGs, no labels read):** +- `Fill time = s` appears inside the "Filling Control:" settings block. +- It sits between `Filling control type = Automatic` and `Stroke volume determination = Automatic`, before "Filling Phase:" and before the first time-series header. +- Examples: 850 at 0.40 s (line 242), 778 at 0.05 s (line 240), 666 at 0.40 s (line 250). +- It is therefore a pre-solve process setting t_set. + +**Ruling (replaces v0.1 B and v0.5 item 1):** +- **Primary target:** t* = T / t_set, with t_set taken from the LOG settings block and given identically to every arm as a process input. +- **Scoring:** seconds = t*·t_set. The per-part relative L2 is the same in seconds and in t*. +- **B0 in seconds:** V*_f · t_set. +- **Scope statement:** + - The prediction is *process-conditioned*: deployment must supply the fill-time setting. This is not claimed as pure-geometry prediction. + - Under Automatic control, Moldflow chose t_set from the geometry. It is a given setting, not a label. It never substitutes for the true T_max, and t_set is never compared with or mixed with actual end-of-fill values. +- **Not used:** + - "Maximum machine injection rate" (a machine limit, the same for all parts); + - the time-series Flow Rate column; + - any V/t-derived V̇. +- The training-pool regression fallback ("predicted scale") from the 18:38 ruling is no longer needed. It stays documented as the rule should t_set be missing for any part; such a part fails closed and is reported.