The proposal we signed in August 2020 and the programme we closed in July 2023 do not describe the same piece of work. On paper, reference DK2018015 dated 27 August 2020 was a geomechanics project: a 10-month effort to estimate in-situ stress and elastic moduli from borehole deformation, with an artificial neural network at its centre. What we actually built, over the following three years, was a detection transformer that reads fractures, bedding planes and vugs off carbonate image logs. Same client, same field, same team. Almost none of the same math. This is the story of how the funded engagement became a different engagement, and why the structure of the contract is what kept the pivot from killing it.
What the August 2020 proposal actually scoped
It helps to be precise about the method we were paid to build, because the distance between it and what shipped is the whole point.
Phase 1, budgeted at roughly four months, was a dataset-and-brittleness effort: a mineralogy-based brittleness index, a slowness-log breakdown of stress, silhouette and k-means clustering to zone the well. Phase 2, roughly five more months, was where the geomechanics lived. We proposed an ANN to recover five unknowns for each depth: Young's modulus, Poisson ratio, the two horizontal stresses, and the borehole size, from two knowns, the long and short caliper diameters, calibrated against cored intervals. The physics of that inversion follows the borehole-deformation relationships in the literature we cited [1]. Solving five unknowns from two measured knowns is not a well-posed problem in the ordinary sense; the whole reason to reach for a learned model was to let core-calibrated priors carry the underdetermined part.
Two caliper diameters in, five geomechanical quantities out. That was the funded product. The planned Gantt ran roughly 1.5 months to mobilise, 5 for the dataset and brittleness work, 5 for the ANN, and 7 for a well-to-well correlation stage, about 19 months across four gates once the later phases were folded in. By the March 2021 revision, proposal V3.0, the total budget stood at USD 533,000. Everything about that document points at rock mechanics.
Where the data and the priorities went instead
Two things moved once the engagement was live. The client's operational question turned out to be less "what is the stress state" and more "where are the open fractures and how do the beds and vugs correlate across the field," which is a detection problem, not an inversion problem. And the data that arrived, high-resolution and compact microresistivity image logs across a set of vertical wells, was far richer for picking geological features off the borehole wall than for feeding a caliper-driven stress inversion. The core-calibrated stress ANN needed cored wells and geomechanical ground truth that were never going to be plentiful. The image logs, by contrast, could be labelled for sinusoids at scale.
So the technical core moved. The clustering and index work of the planned Phase 1 stayed useful as scaffolding, and the geomechanics framing even survived into the Phase-1 planning subtasks as an "exploring geomechanical parameters" line item. But the thing we started building in earnest was a detection transformer, in the DETR family [2], that treats each fracture or bedding plane as an object query and regresses its depth, dip and azimuth directly off the unwrapped image. The shipped programme ran from a December 2021 kickoff to a July 2023 R&D close, with the peer-reviewed papers landing in 2024. Ten planned months for a stress ANN became roughly three years for a detection transformer.
The exhibit above is the arc in one frame. The funded method owns the front of the plan; the shipped method takes over around the Phase-2 gate; and the line down the middle is the only thing that mattered structurally, the point where the technical core swapped while the contract went on unbroken. What is worth sitting with is that this was not a failure of scoping. It was scoping doing its job. A pilot exists to discover which problem is actually worth solving, and ours discovered that the problem worth solving was not the one on the cover of the proposal.
Why the phase gates let the core pivot without the contract dying
A single fixed-deliverable contract for "a geomechanics ANN estimating in-situ stress" would have been in breach the moment we stopped building it. The reason we did not end up there is that the engagement was structured as phase gates rather than one lump of scope, and each gate re-decided what the next phase should contain against what the previous phase had learned. We have written elsewhere about the operating model that runs an oil-and-gas AI programme as staged gates from R&D toward continuous delivery; the mechanics of that model are covered there and we will not re-derive them.[ref: From R&D to Continuous AI] The point specific to this piece is narrower and, we think, sharper.
When you fund a phase gate, you are funding the right to reconsider the method at the boundary. The money attached to Phase 2 was attached to "learn geomechanical structure from these wells," and the honest reading of the data at that gate was that the highest-value thing to learn from these particular image logs was geometry, not stress. Because the gate was a decision point and not a fixed spec, moving from a caliper-to-stress ANN to an image-to-feature transformer was a change of plan, not a change of contract. The USD 533,000 was committed to the problem and the partnership. It was never committed to the specific inversion on page two.
That distinction is the transferable lesson. Fund the problem and the partnership, not the initially-scoped method. If you write the contract around a method, then any evidence that the method is wrong becomes a threat to the engagement, and the incentive is to keep building the wrong thing to stay in compliance. If you write it around gated learning, the same evidence becomes the input the next gate is supposed to consume, and a pivot is just the system working. The geomechanics proposal did not fail. It funded the discovery that the field needed a fracture detector, and the structure let us go build one.
What we would keep, and what we would watch
We would keep the gate spacing. Roughly four to five months per phase was long enough to produce a real result and short enough that a wrong direction did not run for a year before the boundary caught it. We would keep the practice of writing each phase's objective in terms of what the client needs to know, not what algorithm we intend to run, which is precisely what left room for the method to change underneath a stable objective. And we would keep the local capability build that ran alongside it, with young Omani and Arab professionals trained through Muscat University partners, because a pivot is far cheaper to absorb when the people carrying the work are already close to the data.
What we would watch is the temptation to treat every pivot as vindicated. This one was, on the evidence, right: the detection transformer shipped, was published, and became the operational core. But a phase-gate structure that makes pivots cheap also makes them easy, and a programme that pivots at every gate is a programme with no thesis. The discipline is to pivot on evidence about the problem, not on enthusiasm about a method. Our gate moved because the data and the client's question both pointed the same way, not because transformers were interesting.
Limitations
This account is a single engagement in a single carbonate field, and the numbers are drawn from the engagement proposal and the programme record rather than a controlled study. The USD 533,000 figure is the March 2021 V3.0 proposal total and reflects the scope as understood then; the shipped programme's cost and duration evolved with the pivot and are not a like-for-like comparison against that number. The planned Gantt of roughly 1.5, 5, 5 and 7 months is the proposal's own estimate, not an actuals ledger. The claim that phase gates enabled the pivot is a reading of what happened, and it is not a controlled comparison against a fixed-scope engagement that never existed. Read it as one worked example of engagement design under genuine uncertainty, not as a benchmark.
References
- Han, Y. and Yin, S. Determination of in-situ stress and geomechanical properties from borehole deformation. Energies, 2018.
- Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A. and Zagoruyko, S. End-to-End Object Detection with Transformers (DETR). ECCV, 2020.
