SFM 4D Cloud · Volve 2002 to 2010
Frozen seismic models on one Volve pair
The engines are shown side by side. They are not ranked.
The Sheng planted-change row is the earlier P7-v2 run (F-031); the NCS rows are the preregistered NCS Volve run (F-072). They are different runs, so no number of one is set against a number of the other.
Sheng study → /sheng/ · NCS study → /ncs/ · Comparison → /compare/
Sheng
The frozen Sheng SFM's measured record in this project, each row read from the receipt that measured it.
- Frozen model
- Sheng SFM-Base-512 · sheng-mae · 89,726,208 parameters, 0 trainable
- Training corpus
- F3: possibly seen (Sheng lists 13 SEG open-data surveys, no names)
- Data
- Volve, the 2002 baseline and the 2010 monitor over the Hugin reservoir; F3 for the facies test.
What was measured
F3 facies · F-071 · each model against its own untrained twins
| Model | kNN | Linear probe | Own twins, kNN | Decision |
|---|---|---|---|---|
| Sheng SFM-Base-512 | 0.571 | 0.542 | 0.312-0.314 | Qualifies on F3 |
This is an F3 result; it says nothing about Volve.
Volve planted 4D · PR-AUC on a change of known size planted in real Volve seismic · controlled validation, not field detection
| Model | Run | Frozen + head | Physics alone | Untrained twin | Receipt state |
|---|---|---|---|---|---|
| Sheng SFM-Base-512 | P7-v2 · F-031 (historical) | 0.799 | 0.700 | 0.718 | Known-truth validation |
Each row's arms are read on that row's own test pixels. A cell is compared with the controls in its own row.
Volve real pair · each model against its own null
| Model | Real 2002 to 2010 pair |
|---|---|
| Sheng SFM-Base-512 | No association with active wells beyond the null (P15-E) Not demonstrated |
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.
NCS
The NCS engines' measured record in this project, each row read from a sealed receipt.
- Frozen model
- NCS-v1 3D · ncs-vit3d · 88,209,408 parameters, 0 trainableNCS-v1 2.5D · ncs-vit25d · 85,857,024 parameters, 0 trainableNCS-v1 2D · ncs-vit2d · 85,798,656 parameters, 0 trainable
- Training corpus
- NCS-v1 3D: F3: outside corpus (F3 is Dutch sector; NCS is DISKOS-only)NCS-v1 2.5D: F3: outside corpus (F3 is Dutch sector; NCS is DISKOS-only)NCS-v1 2D: F3: outside corpus (F3 is Dutch sector; NCS is DISKOS-only)Volve: seen / contamination OPEN: the DISKOS training list is unpublished, so Volve cannot be called held out for any NCS model.
- Data
- Volve, the same 2002 baseline and 2010 monitor over the Hugin reservoir; F3 for the facies test.
What was measured
F3 facies · F-071 · each model against its own untrained twins
| Model | kNN | Linear probe | Own twins, kNN | Decision |
|---|---|---|---|---|
| NCS-v1 3D | 0.575 | 0.598 | 0.228-0.232 | Qualifies on F3 |
| NCS-v1 2.5D | 0.587 | 0.598 | 0.216-0.222 | Qualifies on F3 |
| NCS-v1 2D | 0.609 | 0.612 | 0.240-0.244 | Qualifies on F3 |
This is an F3 result; it says nothing about Volve.
Volve planted 4D · PR-AUC on a change of known size planted in real Volve seismic · controlled validation, not field detection
| Model | Run | Frozen + head | Physics alone | Untrained twin | Receipt state |
|---|---|---|---|---|---|
| NCS-v1 3D | NCS Volve 4D · F-072 | 0.641 (layer 3) | 0.700 | 0.638 (best of its twins, seed 1) | No advantage over its own untrained twin NO_DEMONSTRATED_FOUNDATION_ADVANTAGE |
| NCS-v1 2.5D | NCS Volve 4D · F-072 | 0.886 (layer 6) | 0.700 | 0.640 (best of its twins, seed 2) | Adds 4D value over physics and its own twin N_ADDS_4D_VALUE |
| NCS-v1 2D | NCS Volve 4D · F-072 | 0.866 (layer 12) | 0.700 | 0.632 (best of its twins, seed 2) | Adds 4D value over physics and its own twin N_ADDS_4D_VALUE |
Each row's arms are read on that row's own test pixels. A cell is compared with the controls in its own row.
Volve real pair · each model against its own null
| Model | Real 2002 to 2010 pair |
|---|---|
| NCS-v1 3D | No association with active wells beyond the null Class P15-E |
| NCS-v1 2.5D | Associates, but not separable from its own twin Class P15-C |
| NCS-v1 2D | Physics alone associates most Class P15-D |
Real production-change detection remains unestablished (Amendment 14 s5): an association is reported as association, never as detection.
The two frozen models, side by side
The two frozen model families, in the order the receipts hold them
| Sheng SFM-Base-512 | NCS-v1 3D | NCS-v1 2.5D | NCS-v1 2D | |
|---|---|---|---|---|
| Family | sheng-mae | ncs-vit3d | ncs-vit25d | ncs-vit2d |
| Parameters | 89,726,208 | 88,209,408 | 85,857,024 | 85,798,656 |
| Trainable | 0 | 0 | 0 | 0 |
| Training corpus label | F3: possibly seen | Volve: seen / contamination OPEN | Volve: seen / contamination OPEN | Volve: seen / contamination OPEN |
| F3 corpus status | F3: possibly seen (Sheng lists 13 SEG open-data surveys, no names) | F3: outside corpus (F3 is Dutch sector; NCS is DISKOS-only) | F3: outside corpus (F3 is Dutch sector; NCS is DISKOS-only) | F3: outside corpus (F3 is Dutch sector; NCS is DISKOS-only) |
On physically modelled Gassmann-scale changes planted into real Volve seismic, frozen Sheng early-layer representations combined with a lightweight readout outperform physical 4D features alone and an identically structured random backbone.
Known-truth planted-change result under the current padded 512×512 input geometry.
Not licensed: Sheng detects the real Volve production signal.
Sources
Every number above is read from one of these tracked files.
- docs/receipts/f3_backbone_qualification_2026-10-16.json · sha256 9cfb7a3e3795…
- docs/receipts/ncs_volve_4d_2026-10-16.json · sha256 4ba34a40b8e0…
- docs/products/volve/sfm_record.json · sha256 fd9c6048efdb…
- docs/products/volve_ncs/ncs_record.json · sha256 56524295477c…