Precision Is Not a Property of the Model: What Cheaper Prospect Screening Does to the Funnel

On 18 August 2026 Saudi Aramco and Maaden announced a shareholders' agreement for a joint venture in mineral exploration and hard-rock mining, and put the method in one sentence: "Leveraging advanced computational algorithms, AI, and high-performance computing, the JV intends to target areas most likely to contain copper and valuable minerals, accelerating the path from regional screening to target definition and discovery" [1]. In ExxonMobil's 2Q26 Prepared Remarks, dated 31 July 2026, the company says it has "created sophisticated exploration models powered by both AI and the world's largest seismic database", that for one model trained on Guyana discovery data "validation testing showed it could identify nearly 90% of known discoveries", and that "to date, this model identified four new Guyana prospects" [2]. On 24 August 2026 Equinor, Aker BP and Vår Energi said they plan to mature and test a portfolio of approximately 20 to 25 exploration opportunities over four to five years, "with an ambition to drill around five high impact exploration wells annually" [4].

Read them together and the shape of the industry's year is visible. Two operators are making the front of the funnel cheaper and wider. The third has published what the end of the funnel can absorb. Nobody has published the number that connects them.

That number is the precision of the screen on the population it is actually pointed at, and it is the number a vendor cannot give you, because precision is not a property of a model. A screen has a recall and a false positive rate; the precision it shows is what those two produce against a base rate. Widen the funnel and the base rate falls, because the candidates you add are the ones you ranked below the ones you already had. So the same screen, unchanged, gets worse at the only thing the target-definition team cares about, and it does so faster than the extra screening adds prospects.

Three releases, read exactly

What each says and does not say is the whole of the evidence, so it is worth reading to the letter. All three are the companies speaking on their own sites and in their own filings. None of the three is a service company describing a customer.

Aramco and Maaden is the one that names the funnel. The release is an intent, and it says so: the JV "intends to target" the areas, "the effectiveness of the shareholders' agreement and the incorporation of the JV is conditional upon the fulfillment of certain condition precedents", and the ownership split of 51% Maaden to 49% Aramco is "expected" [1]. The scale figures belong to the exploration ground, not to any system: "spanning approximately 182,000 square kilometers, nearly 10% of Saudi Arabia's total land area", and "over 90 years, Aramco has accumulated and analyzed the largest amount of geological and geophysical data ever acquired in a single basin for the Kingdom" [1]. The commodities are copper first, then "other energy transition minerals including zinc, lead, and rare earth elements" [1]. This is a minerals venture built on a petroleum operator's subsurface record, and no number in it counts a prospect, a target or a hole. It is quoted here for one thing only: an operator writing down, in its own words, that the point of the compute is to compress regional screening into target definition.

ExxonMobil is the one that gives a model number, and the number is a recall. "Validation testing showed it could identify nearly 90% of known discoveries" is a statement about how many of the discoveries the company already knew about the model found again [2]. It is not a precision. It says nothing about how many things the model flags that turn out to be nothing, which is the quantity that lands on a maturation team. The four new Guyana prospects in the next sentence are "ones previously not identified through traditional methods" [2], and they are prospects, not wells and not discoveries. The 13 August trade-press item that carried this into the news cycle is not the company speaking; it quotes ExxonMobil's vice president of exploration on expanding the use of these technologies and carries no validation figure at all [3]. When this piece needs a recall it uses ExxonMobil's, dated 31 July, and calls it a recall.

Equinor, Aker BP and Vår Energi is the one that publishes the end of the funnel, and it is the leg with no artificial intelligence in it. Read the release and the words AI, artificial intelligence, machine learning and algorithm do not appear anywhere on the page [4]. What does appear is a rate: 20 to 25 opportunities matured over four to five years, and an ambition of around five high impact wells a year [4]. That is roughly five opportunities matured a year against five wells a year, which is a funnel whose two ends are already the same width. The AI reading of that release is ours, not theirs.

What widening the funnel does to the base rate

Rank the candidate inventory by prior, best first, which is what a team screening at specialist cost is forced to do. Take the prior of the candidate at rank $x$ to decay exponentially, $p(x) = p_0 e^{-x/k_0}$. That shape is an assumption and the plate says so, but its consequence is not negotiable: the play holds $p_0 k_0$ real prospects however deep you dig, and the mean prior over the first $n$ candidates is

$$\bar{p}(n) \;=\; \frac{1}{n}\int_0^{n} p_0 e^{-x/k_0}\,dx \;=\; p_0\,\frac{k_0}{n}\left(1 - e^{-n/k_0}\right)$$

FormulaLaTeX

Copy LaTeX

which falls like $1/n$ once $n$ is past $k_0$. A screen with recall $R$ and false positive rate $f$ then raises

$$T(n) = n\,\bar{p}(n)\,R \qquad\text{and}\qquad F(n) = n\left(1-\bar{p}(n)\right) f$$

FormulaLaTeX

Copy LaTeX

and the two behave completely differently. $T(n)$ saturates on $p_0 k_0 R$, because there are only so many real prospects in the ground. $F(n)$ grows without limit, at slope $f$, because there is no limit to how many things are not prospects.

Nobody publishes $f$, so the bench asks for something a vendor will actually say: the precision $P_0$ the screen shows on a curated set whose prevalence is $p_0$. Bayes gives the rest.

$$f \;=\; \frac{p_0\,R\,(1-P_0)}{P_0\,(1-p_0)}$$

FormulaLaTeX

Copy LaTeX

Downstream, a target-definition team can mature $M$ prospects a year, and it cannot tell a true flag from a false one before maturing it, which is what maturing is for. Working the flagged list, it gets through $\min(T+F,\,M)$ of it, and the true ones among those are

$$D(n) \;=\; T(n)\,\min\!\left(1,\; \frac{M}{T(n)+F(n)}\right) \;=\; \begin{cases} T(n) & T+F \le M \\[10pt] M \cdot \dfrac{T(n)}{T(n)+F(n)} & T+F > M \end{cases}$$

FormulaLaTeX

Copy LaTeX

The second branch is the whole argument. Past the width at which the flagged list fills the year, the prospects that reach a decision are the team's capacity multiplied by the realised precision of the screen, and realised precision falls with every candidate added. Not levels off. Falls. So beyond that width, screening more takes prospects off the drilling schedule.

One assumption inside that second branch is worth naming, because it is the first thing a practitioner reaches for. The team works the flagged list without a further ranking of it. If the screen returned a calibrated score rather than a flag, the team could mature the strongest flags first and the dilution would bite later than it does here. That is a real answer and it is also a different product: none of the three releases describes a screen that ranks, and a ranking that decides maturation order has to hold across the whole widened population rather than across the favourites it was demonstrated on. The bench holds the harder case, which is the one a screen procured on a precision figure actually delivers.

The bench

The exhibit below computes that funnel rather than drawing it. Left to right is the cost of screening one candidate, from 200 specialist hours at the left to 3.6 seconds at the right, on a log scale. Back to front is the funnel width, 10 to 10,000 candidates screened a year, also log. The solid surface is true prospects reaching a well in that year; the translucent sheet is the other output of the same funnel, the target-definition slots spent on flags that were not prospects. Height is prospects a year, log scale. The amber sheet is the drilling ceiling. The white curtain is your funnel width.

Four numbers open the bench, and they come from four different places. The maturation capacity of 632.5 prospects a year is derived on the plate from an illustrative scenario we published: eleven geoscientists and four days per prospect both come from that composite interpretation-loop case study, not from a client engagement [5], and the 46 week year of five day weeks is a stated assumption. The recall of 90.0% is ExxonMobil's validation figure of 31 July 2026 [2], used as a recall. The drilling ceiling of five wells a year is the NCS alliance's stated ambition [4], used as the fixed end of the funnel because it is the only published drilling rate in this pool; the turnover does not depend on it, because the expression for $D(n)$ does not contain it. The precision of 50% is yours: no release named here gives one.

The inventory shape and the screening budget are assumptions too. The budget of 20,240 specialist hours a year is the same eleven people at eight hours a day, and it is what makes the cost axis do something: a funnel of 1,000 candidates a year needs the cost per candidate down to 20.24 hours before that budget will pay for it, and 10,000 needs 2.02 hours.

candidates screened a year, 10 to 10,000: 1,000screen recall, 50.0% to 100.0%: 90.0%precision at the reference base rate, 5.0% to 95.0%: 50.0%target definition capacity, 1.00 to 2,000 a year: 632.5

Readouts are taken at the cheapest cost on the axis, where the budget never binds. A funnel of 1,000 candidates a year needs the cost per candidate down to 20.24 h to be paid for out of that budget.

What the bench shows

At the opening settings the hero reads 5.00, the drilling ceiling, and the false-flag readout reads 477.8 a year. Realised precision reads 2.57% against the 50.0% quoted. That is the first thing worth sitting with: a screen a vendor could honestly describe as half right on your favourites shows 2.57% precision on a funnel of 1,000 candidates a year, and the whole of that gap lands on the people who mature prospects.

Now drag the width control from 1,000 to 10,000. The hero falls from 5.00 to 1.64. The false-flag readout goes from 477.8 to 630.9. Realised precision goes from 2.57% to 0.260%. Ten times as much screening, and a third as many prospects reach a well. The bench names the two widths that bracket it: the flagged list fills the year at 1,293 candidates, and 3,277 is the widest funnel that still puts five true prospects in front of the drill.

That is the finding, and the shape of it is what matters rather than the numbers. Below 1,293 candidates the maturation team keeps up, every true flag is matured, and screening wider genuinely adds prospects. Above it the team is the constraint and the arithmetic inverts: the true flags have already saturated at the play's endowment, the false flags have not, and the true ones are progressively crowded out of a queue that is worked in the order it arrives. At 10,000 candidates the team spends 630.9 of its 632.5 slots on things that were never prospects.

Leave the funnel wide open at 10,000 and there are three other controls to reach for. Only two of them put the hero back on the ceiling. Raise the target-definition capacity to 2,000 a year and the hero goes back to 5.00, with the widest-funnel readout moving to 10,000; the load readout then says 1,995, which is the honest price. Raise the quoted precision to 95% and the hero also returns to 5.00, load 254.7. Now drag the recall control across its whole range and watch the hero at 10,000: it does not move. It reads 1.64 at 50.0% recall and 1.64 at 100.0%, because the false positive rate the quoted precision implies is proportional to recall, so raising recall raises the true flags and the false flags in the same proportion and leaves realised precision exactly where it was. Recall moves the capacity width, from 2,337 candidates at 50.0% to 1,163 at 100.0%.

That is uncomfortable in a specific way, because recall is the one figure any operator in this pool has published.

The other end, set to what an operator published

Set the target-definition capacity to five a year, which is what 20 to 25 opportunities over four to five years comes to [4], and the bench stops being about screening at all. The widest-funnel readout says "never reaches 5" and the hero reads 2.27 at 10 candidates a year, 1.07 at 100, and 0.13 at 1,000. With a maturation capacity of five, no screening cost and no funnel width anywhere on the block puts five true prospects in front of the drill, because the flagged list overflows the year at the very first width on the axis.

That is not a criticism of the NCS alliance, whose release is about pooling exploration acreage and expertise and makes no claim about screening throughput at all. It is the point of the exercise. Their two published rates, five opportunities matured a year and around five wells a year, describe a funnel that is already matched end to end. Pour a cheaper screen into that funnel without touching the maturation end and the arithmetic above says what happens, and it is not more wells.

Where the maturation default comes from, and what it is not

The one throughput on the plate comes from an illustrative scenario we published, a composite case study rather than a client record. In it, an agentic interpretation loop runs over a diverse exploration portfolio: a foundation model proposes horizons and faults, a review agent scores every pick for geologic plausibility and flags what it is unsure of, and a geoscientist reviews only the flagged zones and commits corrections that feed the next prospect. Interpretation cycle time goes from six weeks per prospect to four days, geoscientist hours per prospect fall 72%, and cross-team pick consistency rises 38 percentage points [5]. The team is eleven people carrying a queue of twenty-plus prospects, and the case study says in its own words where the constraint went: "the bottleneck is shifting from how fast we can interpret to how fast we can decide what to drill" [5].

Two things that figure is not. It is not a target-definition rate: interpretation is one step of maturing a prospect and the bench uses it as a stand-in because it is the only interpretation throughput we have published, and that figure is illustrative. And the 46 week year that turns four days per prospect into 632.5 prospects a year is an assumption we made, printed on the plate, not something we measured. Put your own capacity on the slider; the shape does not move.

The cost collapse on the other axis is also ours, and it is not a deployment. We published a synthetic demonstration in August in which a screen that traditionally takes about three weeks of specialist review ran in under a minute, and said in that piece that every number in it is illustrative and comes from the demonstration rather than a deployment [7]. It belongs here as the mechanism, not as evidence of a rate: amortised inference moves the expensive computation from per-site to once, so the marginal candidate becomes nearly free and the rational move becomes to screen everything. This piece is the other half of that argument. Screening everything is right, and it is only right if somebody downstream can absorb what comes back.

We have also paid the false-positive bill directly. On a Middle East carbonate operator's image logs, the binding problem was not finding pores; it was that every dark blob became a candidate, and the fix was not a bigger network but an engineered gate, a circularity band and an area window backed by contrastive and vicinity filtering, that exploits the physical signature of a real dissolution pore [6]. That gate is what the maturation team is, in the exploration funnel, except that it is made of people and its throughput is fixed.

Three questions for a screening programme

What is the precision on the population you will actually point it at, not on the curated set it was demonstrated on? A recall figure, even an operator's own, does not answer it, and the bench shows why: recall does not move realised precision at all once the quoted precision is fixed.

How many prospects a year can target definition absorb, and is that number moving? Until it moves, the widest funnel that still fills the drilling schedule is a fixed property of your organisation, and screening past it is a transfer of work from a cheap step to an expensive one.

What happens to a flag that turns out to be nothing? A screen that costs seconds and a maturation step that costs weeks are not two halves of one system unless something between them is allowed to reject cheaply. Two of the three releases at the top of this piece describe the front of that funnel and one publishes the rate at its end. Not one of them carries a number about the middle.

Key takeaways

1. 1Saudi Aramco and Maaden state the intent plainly: advanced computational algorithms, AI and high-performance computing to accelerate the path from regional screening to target definition, over roughly 182,000 square kilometres, for copper and other energy transition minerals, conditional on precedents and with a 51/49 split described as expected.
2. 2ExxonMobil's 2Q26 Prepared Remarks of 31 July 2026 give the only model figure in this pool, and it is a recall: validation testing showed one model could identify nearly 90% of known discoveries. The four new Guyana prospects in the same paragraph are prospects, not wells, and no precision is published. The 13 August trade-press item is not the company speaking.
3. 3The Equinor, Aker BP and Vår Energi release publishes the end of the funnel, 20 to 25 opportunities matured over four to five years and an ambition of around five high impact wells a year, and contains no AI wording at all. The AI reading of it is ours.
4. 4Precision is not a property of a model. Fix the recall and the false positive rate and the precision you see is set by the base rate, and widening the funnel lowers the base rate, because the candidates you add are the ones you ranked lower.
5. 5Past the width at which the flagged list fills the maturation year, true prospects reaching a decision equal maturation capacity times realised precision, and realised precision falls with every extra candidate. At an illustrative throughput of 632.5 a year and a screen quoted at 50% precision on the favourites, going from 1,000 to 10,000 candidates screened takes the hero from 5.00 true prospects a year to 1.64 and the false flags from 477.8 to 630.9.
6. 6Two controls hold the hero on the ceiling at 10,000 candidates: target-definition capacity and the quoted precision. Recall does not move it at any setting, because the false positive rate a quoted precision implies scales with recall. Recall is the one figure an operator in this pool has published.

References

[1] Saudi Arabian Oil Company (Aramco). Aramco and Maaden sign shareholders' agreement. 18 August 2026. https://www.aramco.com/en/news-media/news/2026/aramco-and-maaden-sign-shareholders-agreement

[2] Exxon Mobil Corporation. 2Q26 Prepared Remarks. 31 July 2026. https://d1io3yog0oux5.cloudfront.net/_ffcbd7a566640bfe3df0b1e14b50bc22/exxonmobil/db/2404/22710/pdf/2Q26+Prepared+Remarks.pdf The document's own creation date is 31 July 2026; it is cited here rather than the later news coverage because it is the company's own text.

[3] World Oil. ExxonMobil identifies four Guyana exploration opportunities using AI. 13 August 2026. https://www.worldoil.com/news/2026/8/13/exxonmobil-identifies-four-guyana-exploration-opportunities-using-ai/ Trade press, cited only as the in-window pointer to [2]. It carries no validation figure.

[4] Equinor ASA. Equinor, Aker BP and Vår Energi join forces. 24 August 2026. https://www.equinor.com/news/20260824-equinor-aker-bp-and-var-energi-join-forces

[5] EarthScan. Illustrative scenario: From weeks to days: an NOC's agentic seismic-interpretation loop. A composite case study, not a client engagement. https://earthscan.io/case-studies/noc-agentic-seismic-interpretation-loop

[6] EarthScan. Killing Vug False Positives: The Circularity-and-Area Gate That Separates Real Pores From Image-Log Artefacts. https://earthscan.io/case-studies/vug-false-positive-filtering-circularity-area-resistivity-gate

[7] EarthScan. Amortized Induced-Seismicity Screening for Injection Sites. 14 August 2026. https://earthscan.io/insights/amortized-induced-seismicity-screening-for-injection-sites Every figure in that piece is from a synthetic demonstration, not a deployment, and it says so.