6.7 KiB
S2 design v0.2: rulings on MeltAstra's execution diff (Claude, 2026-11-14)
Integrates MeltAstra's docs/s2-execution-diff-v0-2026-11-14.md (codex/s2-execution-diff-20261114 @ c35b02d9). The diff is adopted except where a ruling below says otherwise.
Prior versions (v0 d8aa0f42, v0.1 fd1db63e, v0.1a 4e435a03) stand unless amended here.
No computation or training starts without Yanqi's explicit go.
R1. Target scale (D1)
The log scale τ is fitted separately for each (outer fold, budget pool), on that learning pool only. A larger pool never supplies the scale for a smaller one.
R2. Splits (D2)
- Components: source key ∪ confirmed near-duplicate evidence edges, following MoldAstra note14.
- The rule must be frozen before the candidate list is regenerated.
- Unresolved pairs that could change the grouping put S2 on HOLD.
- Outer folds: 5, over components, with a fixed label-free ordering and seed 2026111409.
- Inner validation: a fixed 20 % of each outer fold's non-test components (adopted). The 25/50/100 % pools are nested prefixes of the remaining learning pool. Checkpoints and HPO see inner validation only.
- development_exposed: the S1 panel and the previously diagnosed parts are flagged. The main table keeps the pre-registered 5 folds. Exposed and unexposed subsets are reported as a secondary breakdown. There is no separate untouched set.
R3. Target domain and eligibility (D3)
-
Primary domain: commonly eligible nodes in the d > 0 PPS draw. Statements are restricted to this domain and never called whole-part results.
-
S2 common eligibility:
- d > 0;
- O2 spectrum valid;
- N_eff ≥ 64;
- edge block complete and finite (needed by arm E, and therefore applied to every arm).
The S1 gate-direction requirement is dropped: gate q is not an S2 input. This is recorded as a difference from S1.
-
Draws: the frozen S1 draws are reused for the 24 panel parts. New parts get the same rule, with seed [2026111400, prefix]. The 2,000 queries are fixed per part and not refreshed per epoch.
-
Per-part error (adopted): e_p = sqrt(Σ c E (T̂−T)² / Σ c E T²). No second mass factor. A zero denominator is NA with a reason.
- Aggregation: parts are equally weighted within a component, components equally weighted overall. Source-key weighting is descriptive only.
- Coverage table per part: h, draws, distinct IDs, K, K/2000, h·K/2000, the near-wall share, 1−h, and rejection reasons.
-
"Last-filled region" and Kendall τ are defined only over the query coverage, and are labelled that way.
R4. Connectivity bottleneck metric (D8): deferred
It needs predictions on the full tet graph, plus full-node O2 for arm E, and neither is in S2's budget. No substitutes are allowed: no ground-truth infill, no query-induced graphs, no zero-filled O2. It moves to a later full-field evaluation stage. The thin slice follows v0.1a.
R5. Baselines and backbone (D4/D5)
- R2 = "R2-G, Gre24-style approximation".
- Features, as listed in the diff §5.1:
- (x−x̄)/R_g;
- d/t_part;
- ‖x−gate‖/R_g;
- ℓ_g/R_g, where ℓ_g is the tet-edge Dijkstra distance from the gate's nearest node;
- R_g, V, t_part.
- This is not called a Gre24 reproduction.
- Grok gets one targeted task: Gre24's exact feature list and its eikonal setup. If a feature can be matched cheaply, it is added before the freeze; R2-E (eikonal) is otherwise deferred.
- GBRT: sklearn HistGradientBoostingRegressor with sample_weight. Weights make components equal, parts equal within a component, and use draw counts within a part.
- A fixed HPO grid of at most 12 configurations, selected on inner validation.
- With a fixed random_state it is deterministic, so it is counted as 15 independent fits (5 folds x 3 pools), not 45.
- Features, as listed in the diff §5.1:
- N/E backbone: primary candidate is Transolver (thuml, MIT) as the whole-part encoder on all tet nodes with the non-O2 base inputs. Backup is GINO (neuraloperator, MIT).
- The query decoder is an MLP over [node embedding, g12, g23] for N and [..., 4 edge q] for E.
- Query nodes are mesh nodes, so no interpolation is needed.
- The licence and the exact commit are checked before the pilot.
- Two capacity tiers are pre-registered (width 128 and width 256, 4 layers), chosen on inner validation. The same tier is used for N and E.
R6. Statistics (D6)
- Primary contrast (single): Δ = 1 − Ē_E / Ē_N.
- Measured at the 100 % pool, whole primary domain.
- Ē is the component-equal error averaged over 3 matched seeds.
- m = 10 %, now frozen.
- The six-region table applies.
- Interval: bootstrap over components, with matched seed blocks kept intact.
- 2,000 replicates, seed 2026111408, two-sided 95 % percentile.
- Both arms are recomputed on the same indices before the ratio is taken.
- Stated limitation: overlapping training sets and fixed folds are not resampled.
- Secondary (reported, no gate, no multiplicity claim): the 25 % and 50 % pools, the near-wall slice, the exposed/unexposed split, Kendall τ.
- R2 usefulness: report Δ_N/R2 and Δ_E/R2 with the same machinery. The neural route is called useful only if N's lower bound against R2 is > 0.
- The query-level Monte Carlo error is not merged into the model-generalisation interval.
R7. Resources (D7)
- Preprocessing: before any corpus-wide run, MeltAstra measures a 10-part stratified sample. It covers decode, d/t_part, ℓ_g, the draw, the O2 count and the O2 kernel, with CPU time, RSS and disk recorded for each part. The full-corpus table then lists cases above the limits. Limits stay at 30 min / 1.25 GiB per part; no part is swapped.
- Pilot:
- one outer fold, the 100 % pool, one seed, N and E, both capacity tiers;
- at most 4 runs, at most 2 GPU-h in total, one V100 16 GB, FP16 AMP with FP32 loss and normalisation;
- test fold sealed.
- Exceeding a limit means stop and report; nothing is shrunk silently.
- Full matrix accounting: 90 neural runs + 15 R2 fits + HPO + pilot + reruns, all counted. GPU-h is computed only from pilot measurements.
R8. Extensions (not in S2 v0.2)
- N0 (an arm without O2): conditional. It is added only if Yanqi wants the value of the whole O2 block tested; that costs 45 more neural runs.
- O1: stays open, next design item. Order: type definition, then calibration on 890 plus one uncached part, region tables, capacity LP, then acceptance.
- In-house CAE pair: prepare the data-request and control table (constant Δp and fixed flow) only; no solving yet.
Entry sequence (unchanged)
- Grouping rule and components.
- Domain and coverage.
- Features and backbone with licences.
- Statistics frozen (this document).
- 10-part preprocessing measurement.
- Pilot receipts.
- Full-matrix budget.
- Stage review, then Yanqi's go.