Skip to main content
Reading viewAll insights →
BLOG24 min read

The Corpus Has a Postcode: What a Published Shelf Buys a Subsurface Model, Output by Output

The Corpus Has a Postcode: What a Published Shelf Buys a Subsurface Model, Output by Output
Tarry Singhby Tarry SinghFounder & CEO · 20 Nov 2026
Share

Seven releases sit behind this piece and six of them say nothing about AI at all: Sokkeldirektoratet's Resource Report 2026 and the 2026 map of the Norwegian continental shelf, both on 8 September 2026; SLB and Shearwater starting the Amapa 3D multi-client survey on 8 September, a vendor release naming no operator, where elastic full-waveform inversion is compute-heavy physics and not machine learning; Baker Hughes announcing a bp North Sea stimulation contract on 4 September with no AI, digital, automation or software wording anywhere, and with Baker Hughes speaking rather than bp; TotalEnergies on Acacia-5 on Angola Block 17 on 10 September with no AI content stated; and Santos spudding its first Beetaloo appraisal well on 8 September. The seventh, Baker Hughes and OGDC on 2 September, carries the only AI wording in the pool and places it in an execution phase that starts only after evaluation. The AI question here is ours throughout. Our own record answers it: FORCE 2020 gave us 118 Norwegian Sea wells against a carbonate field of 14 vertical wells, more than eight times the count, and the file's own verdict is that the mechanism is portable while the learned distribution is not. The instrument computes what a published shelf actually covers. Growing the corpus from 14 wells to 118 to 472 takes its volume from 296 cells to 2,497 to 9,988 while the lens over the carbonate field stays at 0 cells the whole way, and the damage lands unevenly: at that field depth still reaches 86.7% and azimuth 50.0%, and on the 5 well horizontal blind run azimuth reaches 0.0% while depth has not moved by a single point. The aggregate slab reads 54.3% there, which is the number a vendor would quote.

On 8 September 2026 Sokkeldirektoratet, the Norwegian Offshore Directorate, published two things on the same page: Resource Report 2026, now available in English, and the 2026 map of the Norwegian continental shelf, which can be downloaded or ordered in print [1]. Neither is a deployment and neither is an intent. They are data products, and they say nothing about artificial intelligence or machine learning, because nothing obliges a national regulator to have an opinion about either. The question in this piece is entirely ours. It is the one that a published national archive raises for anybody training a subsurface model, and it is not the question our nearest published neighbour asked. More Wells Is Not More Signal priced partner share inside one shelf, ran on pick-convention heterogeneity in the labels, took a commercial exploration alliance as its occasion, rested on our well-count ablation and landed on a single scalar crossing point. This piece touches none of that. Its axis is distribution distance out of a published shelf, its occasion is a regulator publishing a report and a paper map, and its result is a disaggregation rather than a scalar: the error is per output, the physics says in advance which output goes, and the aggregate score hides it.

Six releases with no AI in them, and one with a conditional

The pool behind this piece is seven releases and six carry no AI wording at all. That is the frame, not a footnote.

Sokkeldirektoratet's two artefacts of 8 September 2026 are the first two, and they are first-party for the directorate's own data products [1]. We record one practical fact about them because it bears on anyone building a pipeline against that archive: sodir.no returns HTTP 403 to a plain fetcher and reads perfectly well in a browser. The report and the map are recorded here as the occasion and nothing more.

The third is SLB and Shearwater GeoServices, who announced on 8 September 2026, dateline Houston, the start of acquisition on the Amapa 3D multi-client survey on Brazil's Equatorial Margin, about 130 kilometres from the Morpho exploratory well [2]. Two things about it have to be said plainly. It is a vendor speaking, SLB's own account of SLB's own investment, and no operator is named at all, because the survey is multi-client and the customers are prospective. And the only technology claim in it is that Shearwater acquires high density broadband data with Isometrix aboard the Amazon Warrior while SLB processes and images it using elastic full-waveform inversion. Elastic FWI is compute-heavy physics, not machine learning, and the release claims nothing more than that.

The fourth is Baker Hughes announcing on 4 September 2026, dateline Houston and London, a contract from bp for offshore production enhancement and stimulation services across bp's UK North Sea operations, vessel based stimulation using the StimFORCE modular package [3]. This is first-party, but it is the service major's release and not the operator's: Baker Hughes speaking about bp, not bp speaking. There is no contract value, no well count, no duration and no named asset. There is also no mention of artificial intelligence, machine learning, digital tooling, automation, remote operations or software anywhere on it.

The fifth is TotalEnergies, 10 September 2026, announcing the fast-track development of Acacia-5, a new oil discovery on Angola Block 17, where TotalEnergies is operator with 38%, and entry as operator into two new exploration blocks [4]. First-party, and a discovery notice: no named subsurface AI tooling, no seismic imaging vendor, no digital foundation. We read the preview of that release; the full body was not returned to us, so nothing here is claimed about anything beyond the preview.

The sixth is Santos, 8 September 2026, announcing on the ASX that it has spudded its first Beetaloo appraisal well in the Northern Territory following environmental approval in May 2026 [5]. First-party, one contractor named, Exact Contracting for pad construction, and no subsurface technology at all. It is in this piece as supporting geography. The resource figures on that release are promotional undiscovered gas-in-place and have no business in an argument about data or models, so they are not used here.

The seventh is the only one with AI wording on it, and it comes with a conflict we are going to carry rather than resolve. On 2 September 2026 Baker Hughes announced a multi-year contract signed at the headquarters of Oil and Gas Development Company in Islamabad, covering more than 120 wells across OGDC's Tando Alam Oil Complex and the Pirkoh field [6]. Phase one is evaluation. Phase two is execution, and it is phase two that carries AI-enabled chemical injection programmes for flow assurance. Our own sweep rejected this item twice and the re-fetch confirmed the rejection: the phrases "artificial intelligence" and "machine learning" appear zero times on the page, the only AI mention is "AI-enabled chemical injections designed to improve flow assurance", and it sits in a phase that begins only after the evaluation phase finishes. It is first-party for Baker Hughes' account of its own contract; OGDC is the operator and not a co-issuer. So: one AI claim in seven releases, and it describes something that has not started.

Sokkeldirektoratet, SLB, Shearwater, bp, Baker Hughes, TotalEnergies, Santos and OGDC. Eight names across four continents, and between them one conditional sentence about AI. No operator here claimed anything, which means every claim below is ours and is attributed as ours.

Two artefacts from the same shelf, and they are not the same thing

Resource Report 2026 is a resource assessment. It is not a training set. The other Norwegian artefact this piece runs on is FORCE 2020: Zenodo record 4351156, 118 wells from the Norwegian Sea, a machine-learning competition dataset with conventional logs, lithofacies labels and a published train and test split [7][10]. We actually trained on that one. Conflating the two would be exactly the unsupported inference this stream exists to avoid, so we say it once in plain words and then put the same sentence on the face of the instrument, where a reader who never reaches the caption still cannot slide from one to the other.

We used FORCE 2020 for a specific and limited purpose, and our own case study is explicit about it. Building well-to-well lithofacies correlation for carbonate image logs, we refused to debug the machinery on the confidential wells. The target field was, by deep-learning standards, brutally data-poor: 14 vertical wells, two different microresistivity imaging tools logging a fractured carbonate, every log confidential under a strict IP agreement [7]. So we proved the transfer mechanism on 118 public wells first, more than eight times the target field's well count, then carried the mechanism onto the target field and retrained there. And the file states its own boundary in one line: FORCE 2020 is North Sea clastics, the target field is onshore carbonates, and "the mechanism is portable; the learned distribution is not" [7].

That sentence is the whole of what a published shelf buys. Two published insights already say what a public archive is good for: Open Subsurface Data as the Scarcity Antidote on crawling a national well archive, and Rehearsing on Public Wells on building a pipeline against open data before spending an NDA-bound log. This piece is the bound on that same strategy. Those two say what a published shelf is for. This one says the single claim it is never allowed to settle.

Why the damage is not spread evenly

The interesting part of the boundary is not that a distant field is harder. It is that the difficulty does not arrive uniformly across a model's outputs, and the physics says in advance which one goes.

Our fracture detector reads borehole image logs and emits four things about each pick: whether there is a fracture there, and if so its depth, its dip and its azimuth. Read blind on horizontal wells it had never trained near, at a 0.55 probability threshold and the 6.0 cm depth operating point, the detector scored F1 60.87%, with depth MAE 2.11 cm and dip MAE 11.12 deg holding at workable levels, while azimuth MAE reached 65.4 deg against roughly 10 degrees on vertical wells [8]. Our own file puts it in seven words: "Three of four outputs transferred with a haircut; one collapsed." That result rests on a blind evaluation on five horizontal wells, which is a small sample and is itself part of why azimuth degraded so far [8]. That bound travels with the 65.4 wherever it appears, including here.

The reason is geometric and it is not a story told afterwards. A fracture crossing a cylindrical borehole appears on the unwrapped image as a sinusoid. The amplitude of that sinusoid carries the dip and its phase carries the azimuth. A horizontal borehole flattens the sinusoid, so the phase is the quantity most tightly coupled to the exact thing that changed [8]. Depth is coupled to something else entirely: it is anchored in instrument geometry, and it is floored by the pixel. At the native digital log resolution one pixel equals about 3 cm, so a plus or minus 3 cm depth uncertainty is baked in before the model makes a single decision [9]. That floor is the same on a vertical well and on a horizontal one, on North Sea clastics and on Middle East carbonate, because it is a property of the tool and not of the rock.

Generalise it and you have a rule a reader can apply to any published shelf. Outputs anchored in the tool travel. Outputs anchored in the rock and the borehole do not. Norway can publish its entire shelf without moving the second kind by one degree.

There is a matching result from the other side, and it is the sentence no vendor writes. In the same programme, one model revision that pushed the augmented corpus to roughly 92k patches actually regressed, "a reminder that more data volume without more geological diversity does not buy generalisation" [9]. Volume from one distribution is not diversity. The market is buying the first and pricing it as the second.

The arithmetic

Model a corpus as a solid region of a provenance space rather than as a count. The space has three axes, and every one of them is a dimension along which our own proxy corpus and our own carbonate field are on record as differing: rock fabric, clastics to carbonate; imaging instrument, graduated in native cells at 1 pixel to about 3 cm and floored by a hard plate at that resolution; and well geometry, vertical to horizontal. No axis is an error. The box is provenance, not performance.

A corpus body's volume is its published well count, and nothing else:

The published corpus is a solid whose volume is its well count, referenced to the 118 wells of FORCE 2020
VP(N)  =  Vref N118V_P(N) \;=\; V_{\text{ref}}\,\frac{N}{118}

The body grows along its own lengths in fabric and geometry as wells are added, and its height on the imaging axis does not move at all, because every added well from the same shelf is logged with the same conventional suite. The centre never moves, because the wells being added come from the same shelf and the same fabric.

Both bodies are axis-aligned, so the part of the target the corpus actually covers factorises:

An axis-aligned overlap is a product of three per-axis fractions, so one closed axis closes the whole product
cov  =  ∏a∈{u,v,w}∣Pa∩Fa∣∣Fa∣\mathrm{cov} \;=\; \prod_{a \in \{u,v,w\}} \frac{\left|P_a \cap F_a\right|}{\left|F_a\right|}

That product form is the reason a distant field is not merely harder. It is the reason the answer is zero rather than small. One axis with no overlap at all takes the entire coverage to zero regardless of how generous the other two are, and fabric is that axis for a North Sea corpus asked about a Middle East carbonate.

Each output then carries beyond the corpus by as much as it is indifferent to each axis. Write αa\alpha_a for how tightly an output is coupled to axis aa, with 1 meaning it carries no further than the corpus's own character and 0 meaning it carries the whole way across. Depth is fully coupled to the instrument axis and nearly indifferent to the other two, which makes its coverage a ratio of two runs on that one axis:

Depth is anchored in the tool, so its coverage is the run the two bodies share on the imaging axis, and nothing else
covdepth  =  hPhF  =  0.260.30  =  86.7%\mathrm{cov}_{\text{depth}} \;=\; \frac{h_P}{h_F} \;=\; \frac{0.26}{0.30} \;=\; 86.7\%

And detection is the aggregate, which has no couplings of its own. Its coverage is what an aggregate score is:

The single number a vendor quotes is a mean over outputs that behave nothing like each other
covdetection  =  13(covdepth+covdip+covazimuth)\mathrm{cov}_{\text{detection}} \;=\; \tfrac{1}{3}\left(\mathrm{cov}_{\text{depth}} + \mathrm{cov}_{\text{dip}} + \mathrm{cov}_{\text{azimuth}}\right)

The couplings are ours and they are modelled, not measured. What is measured is the four errors they are set against, and those were published before this piece.

The bench

The exhibit below is not a chart. It is five solids sharing one volume, and everything it prints is a volume, an overlap fraction or a count of cells. The aqua body is the published corpus. The steel body is the target the model has to work on. The amber solid between them is the lens: the part of the corpus's learned distribution that actually covers the target. The faint translucent shell enclosing everything is the mechanism hull, the architecture, the 700 data point windowing, the segment-then-correlate protocol and the held-out-well evaluation, and it never changes size, because it is the thing that travels. It measures 343,000 cells at every setting of every control.

Three manipulations of the bodies themselves, and one lighting choice. Publish more shelf grows the corpus from 14 wells to 472. Walk the field carries the target to four provenance detents, keeping its size so that the only thing changing is distance. Augment instead fills the corpus with 92k patches inside a boundary that does not move. And lighting one output at a time draws that output's reach and reads what share of the target it covers.

Published shelf overlap
More published shelf makes a bigger corpus. It does not reach a field one closed axis away.
0.0%of the target covered by 118 published wells, modelled
A surveyed estate: a room whose three axes are rock fabric, left to right from North Sea clastics to onshore Middle East carbonate; the imaging tool, up from a resolution plate at 1 pixel to about 3 cm; and well geometry, front to back from vertical to horizontal. The published corpus stands at the left as a forest of 118 well bars inside its boundary, 2,497 cells. The target, 14 vertical well carbonate field, stands as 14 bars of its own, 296 cells, and does not change size. Each body casts a shadow on the plate, the back wall and the right wall. The lens, the part of the target the corpus covers, is 0 cells, 0.0% of the target. The corpus overlaps the target by 0.0% on fabric, 86.7% on tool and 100.0% on geometry, and coverage is the product of the three. Detection, the aggregate, reads 78.9%, the mean of depth 86.7%, dip 100.0% and azimuth 50.0%, and its slab is drawn spanning the three. A red hatched pane stands wherever an axis has no overlap, and a pale pane stands where a slab face cuts the target. Mechanism hull 343,000 cells, unchanged.
Published corpus
2,497 cells
Target, 14 wells
296 cells
Lens, modelled
0 cells
detection reach, modelled
78.9%
Corpus over target
8.43x
Depth reach
86.7%
Dip reach
100.0%
Azimuth reach
50.0%
Detection, the mean
78.9%
a published well, in the corpusa well of the targetcovered: the lens, and the part a lit slab reachesthe lit slab, modelleda pane where an axis has no overlapa pane where a slab face cuts the targeta wall glows where two shadows meetmechanism hull, 343,000 cells, never changes size
Fabric runs left to right, North Sea clastics to onshore Middle East carbonate. The tool axis runs up from the resolution plate, where 1 pixel is about 3 cm and nothing goes under it. Geometry runs front to back, vertical to horizontal, and is the short axis here.
Measured, ours: 118 Norwegian Sea wells (Zenodo 4351156) against 14 vertical carbonate wells on 2 microresistivity tools, the 700 point window, 1 pixel to about 3 cm, and the 92,000 patch regression. Modelled, ours: every body, the lens, every slab and every coverage fraction. Log image models only, not seismic. The horizontal end rests on 5 wells.

At this setting the corpus overlaps the target by 0.0% on fabric, 86.7% on tool and 100.0% on geometry. Coverage is the product of the three, so the closed axis (fabric) hold the lens at 0.0% whatever the other axes do and however many wells are published.

Two bodies in a provenance room, and the planes that decide how much of one lies inside the other. The three axes are the three dimensions our own published proxy corpus and our own carbonate field are on record as differing along: rock fabric, left North Sea clastics to right onshore Middle East carbonate; imaging instrument, up from a hard plate at 1 pixel to about 3 cm that nothing may go under; and well geometry, vertical to horizontal, front to back, the short axis here. The aqua body is the published corpus, FORCE 2020, 118 Norwegian Sea wells, Zenodo record 4351156, drawn as a forest of well bars inside its boundary. It is not the Resource Report 2026 that Sokkeldirektoratet published on 8 September 2026, which is a resource assessment from the same shelf and a different artefact entirely, and the object says so on its own face. The steel body is the target, 14 vertical wells in a confidential onshore carbonate field logged with two different microresistivity tools, drawn at true relative count so it is 296 cells against the corpus's 2,497. The amber solid is the lens: the part of the published corpus's learned distribution that actually covers the target. The lit slab is the reach of one output, and the detection slab is drawn spanning the other three. A red hatched pane stands wherever an axis has no overlap, which is a separating plane: no growth along the other two axes can make the bodies touch. A pale pane stands where a slab face cuts the target. Each body also casts a shadow on three walls, and where two shadows meet the wall glows amber; two walls can show overlap while the third shows none, and the third one decides. The faint hull enclosing everything is the mechanism, the architecture, the 700 point windowing and the held out well protocol, and it never changes size, because it is the thing that travels. Grow the corpus and watch the lens. Walk the target to the four detents and watch it again. Light one output at a time to see how far each of them carries. Every body, every coverage fraction and the lens are modelled and ours; the well counts, the pixel floor, the 92k patch regression and the four per-output errors are measured, ours and previously published. The claim is about log image models and says nothing about seismic, and the horizontal end rests on five wells. Drag to turn, pinch or hold Ctrl and scroll to zoom, double-click to reset.

What the bench shows

Open it and the two bodies are already at true relative count. The corpus at 118 wells is 2,497 cells. The target, 14 vertical wells in the confidential carbonate field, is 296 cells. That is a ratio of 8.43 to one, which is the well-count ratio and not an artistic decision.

Now drag the corpus down to 14 wells and back up. At 14 published wells the corpus is 296 cells, the same size as the field. At 118 it is 2,497. At 472 it is 9,988. Over that whole range the lens over the carbonate field reads 0 cells. Not a small number of cells. Zero, at every one of the 459 settings of that control, because the fabric axis never closes: the corpus's clastic run stops short of the carbonate field's at every well count the control reaches. Publishing 33.71 times more shelf adds 9,692 cells of corpus and not one cell of overlap with the field the model has to work on.

That is the simultaneity the prose above cannot deliver. A sentence can say a bigger open corpus does not help. It cannot make you watch a body grow by more than 8 times while the lens between it and your own field does not gain a single cubic cell, in one glance rather than in two sentences 40 words apart.

Walk the target inward and the corpus starts paying. At a held-out depth band inside a corpus well, the two bodies are nearly coincident and the lens reads 86.7%, or 257 cells, with the corpus at 118 wells. At 14 published wells that same lens reads 78.7%. So more shelf does buy something: 8.0 points of coverage, on a target that was already inside the shelf. At the second detent, a different well in the same basin and the same fabric, the lens reads 0.0% at 14 wells, 31.9% at 118 and 86.7% at 472. That is the honest version of the case for publishing more: within your own shelf, volume reaches targets it previously did not. The claim this piece refuses is the extrapolation of that gain past the shelf boundary, and the third and fourth detents are where the extrapolation dies.

Then light the outputs, with the target at the 14 well carbonate field. Depth reaches 86.7% of it. Dip reaches 100.0%. Azimuth reaches 50.0%. Detection, the aggregate, reads 78.9%. Four numbers for one model on one target, and a spread of 50.0 points between the best and the worst.

Now walk to the fourth detent, the 5 well horizontal blind run, and read them again. Depth: 86.7%, unchanged, not by 1 point and not by 0.1. Dip: 76.4%, down 23.6 points. Azimuth: 0.0%, down 50.0 points, because the geometry axis has closed and a closed axis takes the product to zero. Detection: 54.3%, down 24.5 points. The aggregate has moved by roughly the same amount as dip, which tells you almost nothing about either of the two outputs that actually decided the answer.

That is the F1 60.87% argument as geometry. One unremarkable middling number, sitting between an output that survived completely and an output that ceased to exist, and concealing both. It is also why the detection slab in the exhibit is drawn spanning the other three rather than standing beside them.

Last, the augmentation toggle. Switch it on and 92k patches of augmented density appear inside the corpus body, about 36.8 patches for every cell the corpus occupies at 118 wells. The boundary does not move by a millimetre. Every label anchor stays exactly where it was, all five of them; the lens stays at 0.0%; and all four output coverages read exactly what they read before. That is our own measured negative control rendered as an object: density rising inside a boundary that does not move is precisely what volume without diversity is, and no sentence makes a reader feel the difference between filling a shape and enlarging one.

Set against all of it is the hull, which never changes, and the numbers it carries. After the mechanism was carried across and retrained on the target wells, the correlation tool's v0.8 product sheet set targets of a 60% lift in interpreter productivity, a 75% improvement in interpretation accuracy, 95% precision on target intervals and 90% stratigraphic-correlation success; those are targets, not measured results [7]. Hull large and unchanging, learned distribution small and stuck, is "the mechanism is portable; the learned distribution is not" as two objects of visibly different size.

Where this model is weakest

The bodies are ours and they are modelled. The couplings that set each output's reach are ours too, and they are a physical argument rather than a fit: depth to the tool, azimuth to the rock and the borehole, dip between them, detection as a mean. A reader who disagrees with those three numbers gets different coverage fractions. What survives any plausible choice is the structure, and the structure is the product form. A single axis with no overlap sends the whole coverage to zero, and no amount of generosity on the other two recovers it.

The measured corners are narrower still. Every one of them is a log-image model doing fracture detection or well-to-well correlation. This exhibit says nothing about seismic, and that sentence belongs here in the body as much as in the caption. The single field behind the target body is one confidential onshore carbonate dataset. The horizontal end of the geometry axis rests on five wells [8]. The 3 cm pixel floor is a property of one class of digital log at one native resolution [9].

One more boundary, and it is the one most likely to be misread. The exhibit prices supervised transfer of a learned distribution. It says nothing about pretraining on a published shelf's images without its labels, which is a different question with a different answer, and nothing about what a national resource assessment is worth to a geologist rather than to a matcher. Resource Report 2026 may be an excellent document. It is simply not a body in this space.

Four questions for anyone who is about to lean on a published archive

Which of your outputs is anchored in the tool and which in the rock? Answer that before you look at a single score, because the answer is available from the physics and does not need an experiment. On our model it is the difference between an output that transfers at 86.7% and one that transfers at 0.0% on the same target.

What does the corpus cover on the axis you are furthest away on? Not how many wells it has. The fraction of your own field it overlaps along the one dimension where you differ most. If that fraction is zero, the product is zero and the well count is decoration.

What is the per-output result, and against what sample? A single aggregate is a mean over outputs that do not behave alike, and it will read middling whatever happens underneath it. Ask for four numbers, and ask how many wells each of them rests on. Ours rest on five for the horizontal case and we publish that alongside the 65.4 [8].

And what changed when the corpus grew, other than the corpus? On our model the honest answer is: taking the shelf from 14 published wells to 472, the lens improved by 8.0 points at the held-out band and by 86.7 points at a different well in the same basin. Off the shelf it improved by nothing, 33.71 times over. If a vendor's answer to that question is a well count, they have answered a different question.

Key takeaways

  1. Six of the seven releases behind this piece carry no AI wording at all: Sokkeldirektoratet's Resource Report 2026 and the 2026 NCS map, both 8 September 2026; SLB and Shearwater's Amapa 3D of 8 September, a vendor release naming no operator, where elastic full-waveform inversion is compute-heavy physics and not machine learning and the release claims nothing more; Baker Hughes on a bp North Sea stimulation contract, 4 September, with no AI, machine learning, digital, automation, remote operations or software wording anywhere, and with Baker Hughes speaking rather than bp; TotalEnergies on Acacia-5 on Block 17, 10 September, no AI content stated and only the preview returned to us; and Santos spudding its first Beetaloo appraisal well, 8 September. The AI question here is ours from the first sentence to the last, and no operator in the pool claimed anything.
  2. The seventh, Baker Hughes and OGDC of 2 September 2026, is the only AI-bearing item and it carries a conflict we do not resolve. It covers more than 120 wells across the Tando Alam Oil Complex and the Pirkoh field, phase one is evaluation and phase two carries AI-enabled chemical injection programmes for flow assurance. Our own sweep rejected it twice on exactly that ground: the phrases artificial intelligence and machine learning appear zero times on the page, and the only AI mention sits in a phase that begins after evaluation finishes.
  3. Resource Report 2026 and FORCE 2020 are two different artefacts from the same shelf. One is a resource assessment published by a regulator on 8 September 2026. The other is Zenodo record 4351156, 118 Norwegian Sea wells, a machine-learning competition dataset we actually trained on before touching confidential logs, against a carbonate field of 14 vertical wells logged with two different microresistivity imaging tools. Our own file's verdict stands: the mechanism is portable, the learned distribution is not.
  4. The damage from distribution distance does not spread evenly across a model's outputs, and the physics says in advance which output goes. A horizontal borehole flattens the fracture sinusoid whose phase carries azimuth, while depth stays anchored in instrument geometry and floored at about 3 cm per pixel. On our blind horizontal run, previously published, depth MAE held at 2.11 cm and dip MAE at 11.12 deg while azimuth MAE reached 65.4 deg against roughly 10 degrees on verticals, on a five-well sample that is itself part of why azimuth degraded so far.
  5. The instrument computes what a published shelf covers, as solids rather than as a curve. Growing the corpus from 14 wells to 118 to 472 takes it from 296 cells to 2,497 to 9,988 while the lens over the 14 well carbonate field stays at 0 cells at every one of those settings, because the fabric axis never closes and an axis-aligned coverage is a product. Within the shelf the same growth is worth something: the lens at a held-out band goes from 78.7% to 86.7%, and at a different well in the same basin from 0.0% to 86.7%.
  6. Output by output at that carbonate field the model reaches 86.7% on depth, 100.0% on dip, 50.0% on azimuth and 78.9% on detection. On the 5 well horizontal blind run depth is unchanged at 86.7%, dip falls 23.6 points to 76.4%, azimuth falls 50.0 points to 0.0% and the aggregate falls 24.5 points to 54.3%. A single F1 of 60.87% is a mean over outputs that behave nothing like each other, and it conceals both the one that survived and the one that did not.
  7. Volume is not diversity. Our own negative control is that a revision which pushed the augmented corpus to roughly 92k patches actually regressed, and the exhibit renders it: switch augmentation on, watch about 36.8 patches per cell fill the corpus, and watch the boundary, all five label anchors, the lens and all four output coverages stay exactly where they were.

References

[1] Sokkeldirektoratet (Norwegian Offshore Directorate). What's new: Resource Report 2026 now available in English, and the 2026 map of the Norwegian continental shelf, available for download or in print. Both 8 September 2026. First-party for the directorate's own data products. No AI or machine-learning content. The site returns HTTP 403 to a plain fetcher and reads normally in a browser. https://www.sodir.no/en/whats-new/news/

[2] SLB. Press release on the start of acquisition of the Amapa 3D multi-client survey on Brazil's Equatorial Margin, about 130 kilometres from the Morpho exploratory well. 8 September 2026, dateline Houston. A partner release: SLB's own account of SLB's own investment, with no operator named. Elastic full-waveform inversion is the only technology claim in it. https://www.slb.com/newsroom/press-release/2026/pr-2026-0908-slb-shearwater

[3] Baker Hughes Company. Baker Hughes Awarded Offshore Production Enhancement and Stimulation Services Contract by bp. 4 September 2026, dateline Houston and London. First-party for Baker Hughes, not for bp: this is Baker Hughes speaking about bp. No AI, machine-learning, digital, automation, remote-operations or software content anywhere on the release. https://investors.bakerhughes.com/news/press-releases/news-details/2026/Baker-Hughes-Awarded-Offshore-Production-Enhancement-and-Stimulation-Services-Contract-by-bp/default.aspx

[4] TotalEnergies. Exploration in Angola: TotalEnergies Announces a New Discovery on Block 17 and Enters Two New Exploration Blocks. 10 September 2026. First-party. No AI content stated. The preview of this page read through; the full body was not returned, and nothing is claimed here beyond the preview. https://totalenergies.com/newsroom/exploration-in-angola-totalenergies-announces-a-new-discovery-on-block-17-and-enters-two-new-exploration-blocks

[5] Santos Limited. Santos Spuds First Beetaloo Appraisal Well. ASX announcement, 8 September 2026. First-party. No AI content; one contractor named, Exact Contracting for pad construction, and no subsurface technology. Cited here for geography only. https://www.santos.com/news/asx-announcements/santos-spuds-first-beetaloo-appraisal-well/

[6] Baker Hughes Company. Press release on a multi-year contract with Oil and Gas Development Company covering more than 120 wells across the Tando Alam Oil Complex and the Pirkoh field. 2 September 2026. First-party for Baker Hughes' account of its own contract; OGDC is the operator and not a co-issuer. The only AI wording in this piece's pool, and it sits in the execution phase that begins after the evaluation phase. https://www.bakerhughes.com/company/news/baker-hughes-collaborates-ogdc-optimize-existing-well-production-supporting-energy

[7] EarthScan. Validating Well-to-Well Transfer on 118 Public Wells Before Touching a Single Client Log. Source file content/case-studies/force-2020-proxy-validation-well-to-well-transfer.mdx. https://earthscan.io/case-studies/force-2020-proxy-validation-well-to-well-transfer

[8] EarthScan. The Generalization Cliff: Turning a Broken Azimuth Metric Into a Deployment Protocol. Source file content/case-studies/the-generalization-cliff-what-horizontal-wells-did-to-our-fracture-model.mdx. https://earthscan.io/case-studies/the-generalization-cliff-what-horizontal-wells-did-to-our-fracture-model

[9] EarthScan. Proving It Generalises: A Blind Zone the Model Had Never Seen. Source file content/case-studies/blind-zone-validation-repeatability-unseen-well-milestone.mdx. https://earthscan.io/case-studies/blind-zone-validation-repeatability-unseen-well-milestone

[10] Bormann, P., Aursand, P., Dilib, F., Manral, S., and Dischington, P. FORCE 2020 Well Well Log and Lithofacies Dataset for Machine Learning Competition. Zenodo record 4351156, 2020. 118 wells, Norwegian Sea. https://doi.org/10.5281/zenodo.4351156

Tarry Singh
Tarry Singh

Founder & CEO

More from EarthScan

Related research

All insights →
More Wells Is Not More Signal: The Partner Share Where Pooled Exploration Data Starts to Pay
Insight

More Wells Is Not More Signal: The Partner Share Where Pooled Exploration Data Starts to Pay

70% Past Peak Is a Rule, Not a Reading: What the Mature-Asset Denominator Is Actually Counting
Insight

70% Past Peak Is a Rule, Not a Reading: What the Mature-Asset Denominator Is Actually Counting

False Expertise Has a Geometry: Why the Output That Breaks Is the One Nobody Can Check
Insight

False Expertise Has a Geometry: Why the Output That Breaks Is the One Nobody Can Check

Stay ahead

EarthScan insights, in your inbox.

Field-tested research on subsurface and energy-transition AI. About twice a month. No noise.

We use your email only for this newsletter. Unsubscribe anytime Privacy.