diff --git a/docs/claude/s2-design-v0.3-2026-11-14.md b/docs/claude/s2-design-v0.3-2026-11-14.md new file mode 100644 index 000000000..688e794a0 --- /dev/null +++ b/docs/claude/s2-design-v0.3-2026-11-14.md @@ -0,0 +1,54 @@ +# S2 design v0.3: rulings on remaining gaps G2–G5 (Claude, 2026-11-14) + +Integrates MeltAstra's diff at 33b18e08, §2. G1 (grouping calibration) stays with MoldAstra and is frozen by me later. G6 is a status table kept by MeltAstra. + +No execution of any kind (preprocessing measurement, pilot, fits or training) without Yanqi's explicit approval for that scope. The entry list is an order, not an authorisation. + +## Clarifications (MoldAstra, accepted) +- **Bootstrap:** each replicate resamples components only. The three matched seed runs of each component travel with it, and seeds are not resampled. Both arms are recomputed on the same indices, giving Δ* = 1 − Ē_E*/Ē_N*. The interval is conditional on the fixed folds and the trained runs; it does not cover retraining variance. +- **Reused S1 draws:** S2 eligibility is re-applied from the raw IDs and counts, never from S1's already-filtered sets. + +## G2: sampling and inputs +- **Draw:** the frozen implementation of commit 94011fd. prefix = integer numeric prefix. Candidate order = ascending original tet-node ID. +- **Eligibility:** + - valid spectrum = O2-rep v1.1a status valid; + - edge block complete = every incident true-edge end evaluated, at least one channel-valid finite end, and finite summaries. This is the definition used by driver 5e4b406. +- **Target transform:** enabled. + - τ = draw-count-weighted median of T over the eligible draws of the learning pool (per fold and pool). + - y = log(1 + T/τ); predictions are inverted before scoring. +- **Graph distance:** ℓ_g is Dijkstra on tet edges from the node nearest the YML gate; ties go to the smaller ID. + - All 624 tet subdomains are single components (audit 2026-11-13), so "unreachable" should not occur. + - If it does, or if any encoder input is non-finite, the part fails closed with a reason. No imputation. + +## G3: code and training +- **Code:** + - primary: thuml/Transolver @ 75e0f67643806a81cd1d3f6adc88dd8c02416fe7 (MIT, LICENSE sha256 2c919cd0…); + - backup: neuraloperator @ a6648eeeb60fe0837f2d7d021517a945a159a0c8 (MIT). + - Any modifications are recorded as a diff and never called a reproduction. +- **Configuration:** + - 4 layers, 8 heads, 32 slices, width ∈ {128, 256}; + - query decoder MLP with 2 hidden layers of the same width; + - N decoder input = [embedding, g12, g23]; E adds the 4 edge q values. Parameter counts are reported. +- **Optimisation:** + - AdamW, lr 1e-3, cosine schedule, weight decay 1e-5, batch = 1 part; + - loss = draw-count-weighted MSE in y, normalised so that parts within a component and components overall are equally weighted; + - at most 200 epochs, early stopping on inner validation with patience 20. +- **Capacity choice:** lowest inner-validation component-equal error; a tie goes to the smaller width. The chosen width is applied to both N and E. + +## G4: R2 and auxiliary statistics +- **R2:** sklearn HistGradientBoostingRegressor, version pinned at implementation, random_state=0, early_stopping=False (its internal split is not group-aware). + - Grid (12 configurations): learning_rate {0.05, 0.1} x max_leaf_nodes {31, 63} x max_iter {200, 500, 1000}; min_samples_leaf = 20. + - Selected on inner validation. +- **Neural seeds:** {0, 1, 2}, each applied to the torch, numpy and data-order RNGs. +- **Kendall tau-b:** 5,000 random pairs of distinct eligible IDs per part, seed SeedSequence([2026111410, prefix]). Pairs tied in the truth are excluded. +- **Last-filled location:** among eligible query nodes, the distance between argmax T̂ and argmax T, divided by R_g. Labelled as being within query coverage only. +- **NA handling:** a paired contrast uses components where both arms are valid. The excluded count is reported. + +## G5: resources (upper limits; still require approval) +- **10-part measurement:** tertiles by node count, 3 parts per tertile drawn with seed 2026111411, plus the corpus's largest part (00000666). + - Per part, the whole preprocessing pipeline: 30 min wall time, 1.25 GiB RSS across the whole process tree. + - Run sequentially; total ≤ 3 h; disk ≤ 5 GiB. + - A failure is recorded and the part is not swapped. +- **Pilot runs:** per run, wall time ≤ 30 min, VRAM ≤ 15 GiB, host RAM ≤ 16 GiB; 0 retries. + - The overall cap stays at 4 runs and 2 GPU-h. + - Test fold sealed.