Insights

On an applied-research engagement with a major operator in Oman, the annotation work our models needed was slow, and the obvious move was to escalate it as a cost. In a leadership meeting on 24 May 2022 we made the opposite recommendation: hire young Omani interns to do the collection and labelling, and treat that spend as the intake of a talent pipeline. Interns become junior interpreters; junior interpreters become AI-literate staff; and the knowledge those people carry stays in the building as senior interpreters retire. This is how a vendor turns its own data-supply blocker into a client capability-building play, with the funnel and the retention offset made explicit.

Screening a candidate injection site has traditionally meant weeks of specialist review, which caps how many sites a team can consider and quietly biases portfolios toward the ones already believed to be good. Amortizing the physics into a trained surrogate moves the cost from per-site to once, so screening becomes something you run on the whole portfolio rather than something you spend on your favourites. All figures here are from a synthetic demonstration.

The companion architecture note describes what an on-prem integration looks like once it works. This is the month-by-month checklist for getting there: a three-month runway to move a research sinusoid model into an operator's in-house geoscience platform in Oman. January gathers requirements - VPN, SDK docs, API templates, a client DEV greenfield copy, a locked v1.0.0 feature list, a requirements questionnaire. February integrates the Sinusoids API through DEV then STAGING, where load testing and security live. March demos one well section for stakeholder sign-off. The plan reads like an engineering schedule, but the one beat that actually set the pace was not code. The DEV greenfield copy the client had to create was recorded Pending, and every February and March beat waited behind it.

A technical AI programme dies in the management chain when it is legible to one reader and opaque to the next. Our Phase-2 board decks solved that by layering a single status update for three audiences at once: domain experts got metrics and figures, the CIO and CDO got architecture, latency and cost, and the CEO and board got business impact and risk. This is the concrete slide-layering discipline behind that, drawn from the 18 October and 10 December 2022 board meetings: one update, split into three disjoint payloads, with named section owners on each track and a three-tier compute story (1080Ti to DGX A100 to SuperPod) that lives in exactly one layer. The point is not eloquence. It is that a programme stays fundable only while every reader in the chain can find the one layer written for them without wading through the two written for someone else.

Enterprise AI programmes are usually diagnosed at the model. The dataset was too small, the architecture was wrong, the metric was soft. On the engagement behind this whitepaper, a three-year subsurface-AI programme with a major operator in Oman, the model was never the thing that stalled. Data collection was. Midway through, with a downstream phase gated on a minimum of ten to fifteen wells and wells arriving at a pace that put the timeline at risk, we stopped building and ran a data-management intervention, captured in a thirty-nine-slide workshop deck. What that deck argues, and what this whitepaper generalises, is that failed data initiatives share a pathology with a shape. It is a Situation-Complication-Resolution loop: a tactical fix for one team, a subject-matter-expert hire whose remit confirms the first fix, a data-strategy exercise that produces a strategy rather than access, a fall back to another tactical fix, and data debt that compounds every turn the loop runs. Underneath the loop sit three questions an executive can pose in any language and rarely gets answered: I cannot access the data, I do not trust the data in the report, and it takes too long to get to the data. We argue three things with three instruments. First, the fizzle is a loop, not a cliff, and its cost is the compounding debt, not the failed fix. Second, AI-readiness is not one switch but eight independent dimensions from the McKinsey MLOps frame, and a fizzle is what happens when a few are lifted and the rest are left at their before-state. Third, the exit is a practice-and-platform operating model delivered as three ordered four-day workshops, define requirements, design practices, architect the platform, sequenced across a three-horizon capability climb on a five-fiscal-year roadmap. The deliverable is a way for a board to recognise the pathology early and buy the fix that ends the loop, which is an operating model, not another model.

Most receiver-function pipelines compute the transverse component, glance at it, and discard it, on the textbook reasoning that a flat isotropic crust puts all the converted energy in the radial-vertical plane. The reasoning fails the moment the crust stops being flat and isotropic, which is almost everywhere interesting. That energy is the dip and anisotropy signal, and its backazimuth periodicity tells you which one you are looking at.