Sheng and the NCS engines, side by side

Controlled validation, not field detection. Position alone sits at prevalence, so it cannot solve this task, and raw token similarity is not the answer: a linear probe on the same tokens reads 0.12 more than their cosine. Correction, 2026-09-27: An earlier planted-Gassmann detectability test was invalidated after audit found that the perturbation and scoring interval were spatially misaligned. That experiment therefore provides no evidence that the expected change was undetectable. When the Gassmann-family perturbation is planted and scored at the correct location under controlled conditions, it is detectable by all tested backbones. Under the survey's real repeatability/noise, the pretrained Sheng representation retains partial detectability, but this does not establish real production-change detection. In that controlled protocol every backbone finds the correctly planted change at 0.94–0.99 when nothing else in the data changes, and against the survey's own non-repeatability frozen Sheng reads 0.87 where untrained copies read 0.71. On this task the frozen features lead the untrained twin; with the change generator held out, over the correctly planted families only, the frozen model leads the twin on 1 of the 5 valid families (G5 by 0.009) and trails it on the other 4.

Known-truth controlled validation, not field detection. Frozen minus physics alone -0.060 [95 % CI -0.081, -0.037]; frozen minus its best untrained twin +0.002 [-0.021, +0.026]. Position alone is not solvable on this population. Volve: seen / contamination OPEN.

Known-truth controlled validation, not field detection. Frozen minus physics alone +0.185 [95 % CI +0.157, +0.215]; frozen minus its best untrained twin +0.246 [+0.212, +0.280]. Position alone is not solvable on this population. Volve: seen / contamination OPEN.

Known-truth controlled validation, not field detection. Frozen minus physics alone +0.165 [95 % CI +0.136, +0.192]; frozen minus its best untrained twin +0.233 [+0.200, +0.263]. Position alone is not solvable on this population. Volve: seen / contamination OPEN.

Real production-change detection remains unestablished because available field data do not provide sufficient independent bin-scale truth and repeatability to attribute observed changes confidently to production effects. Not established is not the same as proven undetectable. Transfer from planted to real change is not shown.

No arm among physics, this engine with physics, and its untrained twin with physics associates with the active wells beyond the toroidal-shift null. Amendment 14 section 5: an association is reported as association, never as detection. Volve: seen / contamination OPEN.

This engine's map associates with the active wells (p = 0.048), but it cannot be separated from its own untrained twin (C - B p = 0.154), so no foundation advantage is shown. Amendment 14 section 5: an association is reported as association, never as detection. Volve: seen / contamination OPEN.

Physics alone associates most strongly with the active wells; this engine adds nothing measurable to it. Amendment 14 section 5: an association is reported as association, never as detection. Volve: seen / contamination OPEN.

Classical only: SFM DISCONNECTED (Stage 1); classical structure-tensor solve.

RGT on the NCS page is classical: the structure-tensor solve (S2b) on the shipping path, read from no engine (docs/LANE_A_NCS_REBUILD_PLAN.md section 3).