Insights

Crustal inversion has an intractable likelihood and a perfectly good simulator, which is the regime simulation-based inference was built for. Train the posterior estimator once and every station in the network costs one forward pass. The part that needs saying out loud is what the calibration check can and cannot see, because simulation-based calibration is computed inside the simulator and passes happily while coverage against real structure collapses.

A consulting firm's full-potential model put $230 billion of AI value on oil, gas and oilfield services. Our own weekly sweep rejected that figure twice, once as commentary naming no operator taking any action and once as out of window, the underlying analysis having been released on 25 August 2026 with only a trade restatement of 2 September inside, and it belongs to no operator. This piece inverts it anyway, because it is the cleanest example of a grammar every value total in our pool shares: a headline that is silently a product of a well count it never states. Divide $230,000,000,000 by the $410,000 average saved per well in an illustrative 94-well drilling scenario we published and the total implies 560,976 well-years. Divide again by the 8-in-25 delivery fraction our own ledger recorded and the population reached is 1,753,049 wells; at the 8-of-80 well-maturity fraction from our interpretation backlog it is 5,609,756. The forecast does not move through any of that. Worse, value per well is not constant in the well count on our own measurement: class error fell from 93.115% at 3 wells to a floor of 0.817% at 11 and then rose to 2.536% at 14 when 3 wells carrying inconsistent picks entered, and carrying that reversal forward at its measured 16.413% per decade puts the turnover at 4,468,141 delivering wells raising $131.7bn, which leaves the headline $98.3bn short at every radius. The largest operator-authored well count anywhere in our pool is more than 1,300 ExxonMobil Permian wells at about 2.2% uplift, and that work is from August 2024 and out of window.

On a synthetic LNG cargo with eight emission rows, measuring the pneumatic devices at 1,050 t instead of a 420 t emission factor moves the methane intensity from 0.134% to 0.192% by mass, and the measured share from 53% to 87%. That one change of method, 0.058 points, is about 19 times the whole-fence intensity that six of seven US export terminals reported for 2023 under EPA's Subpart W (0.0017% to 0.0030%), and the EPA table has no field saying whether any of its rows was measured. The EU Methane Regulation asks importers to show equivalent measurement, reporting and verification behind their cargoes; a Commission proposal of 14 October 2026 would move that first demonstration to 31 May 2029. Either way, the evidence a buyer can check is the list of rows and the method on each.

On 27 August 2026 an SPE Artificial Lift panel heard an Occidental production engineer warn that AI can create false expertise and an ExxonMobil operations supervisor argue for keeping human oversight on safety-critical decisions. On 10 September 2026 a panel at the SPE Subsea Well Intervention Symposium in Galveston said humans must stay in command rather than serve as observers. Both accounts reach us through the Journal of Petroleum Technology rather than from the operators themselves, and neither carries a single number. That absence is the subject here. Oversight only buys assurance when the reviewer's evidence is independent of the model's, and on a borehole image it is not: of the three parameters a fracture sinusoid carries, only the phase has a Fisher information that depends on the amplitude, and the amplitude is exactly what a horizontal borehole destroys. Our own already-published blind run puts azimuth error at 65.4 degrees against roughly 10 on verticals, with depth at 2.11 cm, dip at 11.12 degrees and detection F1 at 60.87%. Calibrated on those two azimuth figures the model implies a 68.6 degree fracture set dip, puts the correlation between the model's error and a reviewer's at 0.929 on a vertical well and 0.998 on a horizontal one, and says that checking all 300 picks in a horizontal well by re-reading the same image catches 2.2% of the out-of-band azimuths against 19.7% on a vertical well. Raising evidence independence to 78% is what buys half of them back.

Four operator-authored papers in one issue use a wellbore or a path measurement to settle a field-scale question. That is sound when the quantity has lateral structure and worthless when it does not, and on our own 13-well carbonate field three of the four curves we produced had none. A curve with no measured correlation informs an area of zero, however many wells carry the tool.

Mozambique LNG worked 22.3 months from its final investment decision on 18 June 2019 to force majeure on 26 April 2021, then spent 54.4 months in force majeure until the consortium lifted it on 7 November 2025. Counted from the lift, its months of work since sanction do not catch up with its months halted until 12 July 2028, a few months before the first LNG expected in 2029. Nothing public describes the state of Area 1's subsurface record, and this piece claims nothing about it. It argues that the first deliverable of any restart this long is an inventory, and that one cell in it, which candidate model was the basis of design, decides which wells need reconciling at all.