2.8 KiB
2.8 KiB
R4(a) contract: propagation algorithm with a shared element speed (Claude, 2026-11-16)
Basis: Mold's two specifications (references/2026-11-16/gre24/mold-r4-shared-speed-spec.md, Drive 16do4F5l…, SHA 8e8e1744…) and my rulings 1794819308 and 1794822106. Descriptive and exploratory: no gate, no change to any frozen S2/v5/v4 result.
Question
On the S2 grouped protocol, how much does the distance propagation algorithm change the naive volume-map predictor, when the element speed, gate, volume measure, clock and scoring are identical?
Arms (both with the same positive element speed a_K = mean of the four tetFIM vertex D_w; all-wall tets use the centroid wall distance, exactly as frozen in gre24_features_v2)
- EIK: the tetFIM D_f with metric a_K^2 I. Reuse the frozen v5 preparation output where it exists.
- GRAPH: Dijkstra on tet edges from the same gate node set. Edge weight w_e = L_e / max over elements K containing e of a_K. Edges whose adjacent elements all have a_K = 0 do not exist (should not occur given the centroid rule; count them).
- B0_S2 (reference row, separate attribution): the historical S2 B0 predictions, unchanged. Its difference from the others is attributed only to "the whole distance construction".
- Prediction in each arm: T_hat = t_set x c_fold,pool x F_<(D_f), with the strict full-tet lumped-volume CDF and the 15 frozen learning clocks from the naive-ordering batch. Ties are kept and the [F_<, F_<=] band is reported.
Data and roster
- The 617-part / 209-component S2 roster, 5 folds and 3 pools, with the S2 query IDs and multiplicity.
- EIK fields: reuse the frozen v5 D_f for roster parts that passed v5 preparation (611 candidates per MeltAstra's check); verify each by hash.
- For the 6 remaining roster parts, compute only D_w, D_i and D_f (no spatial moment) with the frozen code path, and label them.
- Failures are kept; no substitution.
Scoring (frozen S2 functions, no new code paths)
- Component-mean relative L2 for EIK, GRAPH and B0_S2 at each pool. Paired contrasts:
- Delta = 1 - E_GRAPH / E_EIK (propagation algorithm);
- 1 - E_B0_S2 / E_EIK (whole construction), descriptive.
- Component bootstrap, 2000 replicates, seed 2026111408.
- Also Kendall tau-b and last-filled location, as in S2. No six-region verdict: descriptive only.
Resources and order
- At most 10 worker CPU-h, at most 4 workers on cores not used by the v4 run, 1800 s / 4 GiB per part, no retries.
- Order: implement, then synthetic tests (a tet chain, where EIK and GRAPH should agree along an edge path; a planar front at constant speed, where GRAPH overestimates by a known factor), then compute the GRAPH arm and the 6 missing EIK fields, then freeze the predictions, then score.
- One receipt with all hashes. May run in parallel with v4.