10-Session Programme · 25 Minutes Per Day · Free · For leaders in Nigeria’s oil and gas sector
WEEK 1 OF 2 · LEADERSHIP, FOUNDATIONS AND INTELLIGENCE IN THE OILFIELD · SESSIONS 1–5 · WORK AT YOUR OWN PACE
Orientation · The ThesisSESSION 1
Leadership That Closes the Loop: The Six Tools
Why This Matters
The book asks you to map one decision's loop before choosing a platform, so this programme begins with a decision. Over ten sessions you will write a signed First-Year AI Operating Programme for your own company, one section per session, around one current decision you choose today. You will also say which side of the operator–contractor relationship your company sits on. Today you learn the six tools that every later session uses, and you map the full loop of your decision. If nobody can say who owns the evidence and who may act on it, that is your first finding.
25-Minute Module
0:00 – 5:00
The finish line, the house rules and one disclaimer
- The book opens with a personal note, and I give it to you as written: AI can improve an operator's decisions. Leadership must turn those decisions into safe operations, reliable production and lasting value. That sentence is the brief for the next ten sessions.
- Week 1, Leadership, Foundations and Intelligence in the Oilfield. Week 2, The Service Company, Trust and Delivery, ending in the capstone.
- The finish line is a signed First-Year AI Operating Programme for your own company: nine sections written one per session, assembled and signed in Session 10.
- House rules: write about your own company; bring a real decision; no benefit claim without a counterfactual; no recommendation without an owner and a permitted action; report a blocker the same day. The pass mark is 70.
- Said once, in the book's own words: it presents a leadership doctrine and proposed operating methods. It does not replace approved engineering practices, statutory obligations, professional judgment or site-specific safety procedures. Its numerical examples are illustrative, not reported results from Nigerian assets.
Bring a decision you will still care about in Session 10.
5:00 – 10:00
Why operator and service company sit in one programme
- A local decision can move value elsewhere: increasing a well's rate may overload water handling. A contractor can deliver a technically successful job and still lose money because the scope changed without commercial authorisation. The operator holds reservoir history and production constraints; the service company holds tool behaviour, execution experience and maintenance records. Shared intelligence improves decisions only when the parties agree on definitions, rights, evidence and authority.
- THE DOCTRINE
A model must answer a defined question, use information available at the decision time, beat an appropriate alternative and remain useful within its stated limits.
- The same standard continues: a recommendation must identify the evidence, uncertainty, owner and permitted action; a benefit claim must distinguish observed performance from an estimated counterfactual.
- Nigeria is the operating context. Conditions differ between an onshore location, a swamp development and an offshore installation, and must be measured during discovery rather than assumed from national averages. The book avoids a fictional typical Nigerian field and does not promise universal savings.
- The governance foundation draws on the voluntary NIST AI Risk Management Framework, alongside Nigerian law and upstream regulatory sources [1-6].
Voluntary means voluntary. Do not let anyone describe a framework as a legal requirement when it is not.
10:00 – 17:00
The six tools, in order of use
Tool What the book says The question it forces The Operating Intelligence Loop Sense, decide, act and learn. A dashboard covers only part of the loop. Where does our loop stop before the action and the learning? The Data Foundation Precondition Readiness is specific to a decision and includes ownership and access rights. Which records are essential, and which failures invalidate the recommendation? The Autonomy Ladder Five permissions, each with its own assurance. Which permission are we actually granting? The Value versus Risk Map Assess value and consequence separately; keep safety constraints out of the monetary score. Is it valuable but too consequential to deploy early? The Sequencing Discipline Define, evidence, compare alternatives, evaluate on withheld conditions, shadow, then controlled use; expand only after outcomes are understood. Which step are we tempted to skip? The Operations Excellence Lens Measure the result where it occurs. A better prediction that arrives after the mobilisation deadline is inferior to a timely one. Which availability, production, margin, invoice or cash figure moves? - The five permissions of the Autonomy Ladder: retrieve evidence; draft an interpretation; recommend an action; execute a bounded administrative action; execute an operational action under separately engineered control.
- THE DOCTRINE
An agent allowed to create a draft work order is not automatically authorised to release it. A model that identifies an operating opportunity does not inherit permission to change a setpoint.
- Sequencing moves quickly for an administrative application and takes longer for a decision connected to well integrity. The difference reflects consequence, not resistance to innovation.
Learn the six until you ask their questions without thinking about it.
17:00 – 22:00
Choose your side, choose the decision, map the loop
- Say which your company is, as the introduction frames it: an asset owner (operator), a service company or contractor, or both. If both, choose the side on which your decision sits.
- Choose one current decision, described as a decision, not a technology.
- Record its loop as Chapter 0 asks: the source of the signal, the recipient of the recommendation, the permitted response and the evidence that closes the action.
- THE OPERATOR · THE WORKED CASE
- Operator. A fictional operator has a model that predicts a pump problem. No reviewable maintenance decision is created, so the signal reaches a screen and stops. In the book's words, it leaves value unrealised.
THE SERVICE COMPANY · THE PARALLEL- Service company. A fictional contractor's job loop crosses organisational boundaries: operator approval, contractor execution, client acceptance and feedback into future planning. If one of those handoffs is undefined, the organisation cannot establish whether its AI improved delivery.
- THE DOCTRINE
If the handoffs across the contract boundary are undefined, the organisation cannot establish whether its AI improved delivery.
Map the loop as it runs today, not as the procedure says it should.
22:00 – 25:00
Write it down, and Section 1
- Before you close the laptop, write three lines: your company and its side; the decision; the point where its loop stops.
- Section 1 of your programme is Our company, our side of the operator–contractor relationship, and the one decision whose loop we will map.
- The leadership charge of Chapter 0: choose one current decision and map its entire loop. Specify who owns the evidence, who can act and how you will recognise a useful outcome. Do this before choosing a platform.
- The book's own advice: read Chapter 0 before approving an AI investment.
If you cannot name who can act, write that down. It is a finding, not a failure.
Key Terms
Operating Intelligence Loop
Sense, decide, act and learn: observations with quality and provenance, an action selected within authority, execution through a named owner, and review of the outcome.
Autonomy Ladder
Five distinct permissions, from retrieving evidence to executing an operational action under separately engineered control. Each requires its own assurance.
Data Foundation Precondition
Readiness judged for a specific decision: which records are essential, which can be missing, which failures invalidate the recommendation, and who owns and may access them.
Counterfactual
The outcome that would have occurred under an alternative action or without the intervention. It is generally not directly observed for the same event.
Self-Check — Before You Submit
1. A fictional operator's model flags pump problems accurately on a screen. No maintenance decision is ever raised. What does Chapter 0 say?
- A. The model needs better accuracy
- B. The loop stops after sensing; without a reviewable maintenance decision, value is unrealised
- C. The application is complete because the screen exists
- D. The model should be allowed to release work orders
B. The loop stops after sensing; without a reviewable maintenance decision, value is unrealised — The book says a model that predicts a pump problem without creating a reviewable maintenance decision leaves value unrealised. Granting release permission is tempting but skips the Autonomy Ladder: creating a draft is not authority to release.
2. Which of these is not one of the Autonomy Ladder's five permissions?
- A. Retrieve evidence
- B. Recommend an action
- C. Execute a bounded administrative action
- D. Approve an operational action without separate engineering control
D. Approve an operational action without separate engineering control — The fifth permission is to execute an operational action under separately engineered control. The option removes exactly that control, so it is not a rung on the ladder; the other three options are among the book's five permissions.
3. A fictional contractor can have a sophisticated forecast two days after the mobilisation deadline, or a cruder one in time. Which tool decides, and how?
- A. The Operations Excellence Lens: the timely, less precise forecast is operationally superior
- B. The Data Foundation Precondition: the sophisticated forecast has better data
- C. The Value versus Risk Map: the crude forecast is lower risk
- D. The Sequencing Discipline: always wait for the better model
A. The Operations Excellence Lens: the timely, less precise forecast is operationally superior — The book says a technically improved prediction that arrives after the mobilisation deadline is operationally inferior. Waiting for the better model looks rigorous but makes accuracy the objective.
Today’s Assignment
Name your company and its side of the operator–contractor relationship, choose one current decision, and map its entire loop.
- Your company and its side. What your company does; it need not be named. Asset owner, service company or contractor, or both, and why.
- The decision. One current decision, stated as a decision with its timing, and why it matters to your business.
- The loop. Sense, decide, act and learn as they run today: the source of the signal, the recipient of the recommendation, the permitted response and the evidence that closes the action. Name who owns the evidence and who can act, by role, and how your company would recognise a useful outcome.
- The handoffs. Where the loop crosses the contract boundary, and whether each handoff is defined.
- The honest part. What your company currently gets wrong in this loop. If your company has no AI in use yet, that is normal: work from the decision in front of you. Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 1 — Our company, our side of the operator–contractor relationship, and the one decision whose loop we will map.
How this is marked — the rubric
| States the company's side of the operator–contractor relationship with a reason | 15 |
| Names one current decision as a decision, with its timing, not a technology | 20 |
| Maps all four loop stages, naming the evidence owner and the person who can act by role | 35 |
| States how a useful outcome will be recognised, in operating terms rather than model accuracy | 15 |
| The honest paragraph names a specific failing in the student's own company | 15 |
| Total · pass mark 70 | 100 |
Submitting session 1
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 1, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
The book's intellectual standardThe six reader-toolkit toolsLoop mapping across the contract boundaryReading the Autonomy LadderProgramme authorship
Part One · FoundationsSESSION 2
The Combined Operating System and the Field Data Contract
Why This Matters
An operator and a service company can each improve their own measure while the combined result gets worse. The operator may prioritise maximum rate; the contractor may prioritise billable utilisation. Neither tells you whether the right work happened at the right time with acceptable risk and economics. This session gives you two disciplines: locating the real constraint, and agreeing what a field observation means before anyone models it. Both expose ownership gaps. For your company, today's work will show which decision is worth pursuing, who owns it, and which five variables you could not yet defend in front of an auditor.
25-Minute Module
0:00 – 6:00
Begin with the constraint
- Map two chains: from reservoir opportunity to accepted production, and from service opportunity to collected revenue. Then locate the current constraint using evidence.
- The book's candidates: well deliverability, facility capacity, access, equipment availability, materials, approvals or financing. A production opportunity without evacuation capacity is not immediately realisable value.
- THE DOCTRINE
AI applied upstream of an immovable constraint can produce recommendations the organisation cannot use.
- Create a joint review for decisions that cross the contract boundary. The operator supplies the authorised scope and operating constraints; the contractor supplies resource availability, execution assumptions and the evidence needed for job acceptance. Record disagreements rather than allowing the model to resolve them implicitly.
- Each party may keep detailed cost data while sharing the operational quantities needed for coordination.
Ask where the constraint is before asking where the data is.
6:00 – 11:00
A decision portfolio, and visible accountability
- Describe candidate applications as decisions: which well to investigate, which pump to inspect, which job to mobilise, which bid to pursue. A portfolio of decisions lets you compare value, evidence availability, decision frequency and consequence of error. A list of technologies does not.
Role What it holds Executive sponsor Executive sponsorship of the decision (one per decision) Domain owner Technical acceptability Data owner Record integrity and access Model owner Maintaining the application Action owner Control of execution Finance Verifying the benefit method - In a small company these can be one person, but the responsibilities must still be distinguishable.
- Separate operating results from incentives. An operator's production gain is not automatically a contractor's gain. A contractor should not be held to a reservoir response its work cannot determine. Conversely, a contractor should not count time saved by an operator’s unrelated shutdown change as its AI benefit. Define the interface measure before deployment: mobilisation readiness at the agreed cut-off, first-pass acceptance of deliverables, failure of equipment supplied to the job, elapsed time from approved completion to invoice acceptance.
- At the leadership meeting, a red item needs a consequence, owner and next decision. A green item needs evidence.
If the benefit depends on an action nobody owns, there is no benefit yet.
11:00 – 18:00
The field data contract
- A rate may be an instantaneous estimate, a test result or a daily allocation. Pressure may be gauge or absolute. Downtime may include planned work in one report and exclude it in another. No algorithm repairs these distinctions by being more sophisticated.
- A field data contract records each variable's definition, unit, physical location, sampling interval, timestamp convention, quality flag, responsible steward and permitted uses. Keep the original observation, every transformation and the reason for any correction.
- Separate event time from ingestion time. In the book's example, a field report recorded at 18:00 may arrive the next morning, so a model for a 20:00 decision cannot use it, even though its event time falls before the decision.
- Missing data can mean communications loss, planned shutdown, maintenance or an observation never taken. Interpolating across a shutdown can create fictional operating behaviour; some decisions should stop when required data become stale. A pump moved between sites needs a stable identity plus dated installation relationships.
- THE OPERATOR · THE WORKED CASE
- Operator. A fictional operator supplies approved well information, constraints and history for an intervention, and must decide what the contractor may retain afterwards.
THE SERVICE COMPANY · THE PARALLEL- Service company. The fictional contractor wants to train a model across clients. Permission to perform a job is not permission to reuse client records for a model trained across clients. Ownership, confidentiality, training use, retention, return and deletion must be agreed.
- The legal context, as the book states it: the Petroleum Industry Act provides the upstream legal context [2]. The regulator lists National Data Repository material on its draft-regulations page, which is not evidence of a final operative instrument [13]. Personal records also require assessment under the Nigeria Data Protection Act [5].
- Test the contract continuously: impossible unit combinations, duplicated identities, backward timestamps, stale measurements, unexplained schema changes. A sudden improvement after a source change may be leakage rather than progress.
A model is only as honest as the definition of its inputs.
18:00 – 22:00
The decision record and approval by consequence
- Appendix A uses one class record per decision class and a linked event record for each material use: what was known, what the system proposed, what the owner decided and what occurred. The ten fields: Decision, Owner, Boundary, Evidence, Alternatives, Model, Uncertainty, Constraints, Approval, Outcome. If a later review changes the interpretation, append the correction rather than rewriting history.
Use Approval evidence the book proposes Retrieval and drafting Source accuracy, access control, handling of unsupported output Operational advice Add domain validation, uncertainty, response capacity and evidence that users understand limits Administrative execution Add tool restrictions, transaction integrity and rollback Operational actuation A separate engineered assurance process with the applicable technical standards and approval structure - These categories are a proposed governance approach, not a statutory approval route.
- THE DOCTRINE
Ask a reviewer who did not build the application to attempt a plausible failure. A superficial sign-off by someone unable to inspect the method adds little assurance.
Write the record so that a stranger could reconstruct the decision a year from now.
22:00 – 25:00
Write it down, and Section 2
- Before you close the laptop, write your binding constraint in one line and the five variables that drive your decision's recommendation.
- Section 2 of your programme is Our binding constraint, decision portfolio and field data contract.
- The leadership charges: Chapter 1, require every AI proposal to identify the business constraint it addresses and the person who can change that constraint. Chapter 2, make the data contract an operating responsibility, and ask for the definition and provenance of the five variables that drive your most important recommendation.
If you cannot define the five variables, you have found the first piece of work.
Key Terms
Interface measure
A measure agreed between operator and contractor before deployment, such as first-pass acceptance of deliverables, that neither party needs to disclose its whole business to report.
Field data contract
A record of each variable's definition, unit, location, sampling interval, timestamp convention, quality flag, steward and permitted uses.
Event time and ingestion time
When something happened versus when the record became available. Historical evaluation must respect availability, or it measures information the real application would not have had.
Data leakage
Use of information in training or evaluation that would not legitimately be available to the deployed prediction, or that compromises the independence of evaluation.
Self-Check — Before You Submit
1. A fictional operator's model recommends higher rates on six wells. There is no spare evacuation capacity. What follows?
- A. Approve them; the model has found value
- B. Retrain the model with facility data
- C. The opportunity is not immediately realisable value; the constraint lies elsewhere
- D. Pass them to the contractor to resolve
C. The opportunity is not immediately realisable value; the constraint lies elsewhere — AI applied upstream of an immovable constraint produces recommendations the organisation cannot use. Retraining is tempting, but the problem is not the model; it is the location of the constraint.
2. A field report with an event time of 17:30 reaches the database at 07:00 the next day. Can a model supporting a 20:00 decision use it in historical evaluation?
- A. No: it was not available at 20:00
- B. Yes, because the event happened before the decision
- C. Yes, provided the report is accurate
- D. Only if the model is retrained daily
A. No: it was not available at 20:00 — The book separates event time from ingestion time. The tempting answer is that the event preceded the decision, but availability, not event time, governs what the model may use.
3. A fictional contractor holds data from interventions for several clients and proposes one model trained across all of them. What must come first?
- A. Removing client names from the records
- B. An agreement covering ownership, confidentiality, training use, retention, return and deletion
- C. Confirmation that the National Data Repository material permits it
- D. A test of the model's accuracy
B. An agreement covering ownership, confidentiality, training use, retention, return and deletion — Permission to perform a job is not permission to reuse client records. Citing the National Data Repository material is the trap: the book says it sits on the draft-regulations page and is not proof of a final instrument.
Today’s Assignment
State your company's binding constraint, build a short decision portfolio, and draft the field data contract for your chosen decision using the Appendix A decision record.
- The constraint. Trace your company's chain to accepted production or collected revenue and name the current constraint, with the evidence for it, and the person who can change it.
- The portfolio. Three to five candidate decisions, each with its value, evidence availability, frequency and consequence of error, and the decision from Session 1 among them.
- The roles. For your chosen decision: sponsor, domain owner, data owner, model owner, action owner and the finance role, by role not name.
- The data contract. For the five variables that drive the decision: definition, unit, timestamp convention, quality flag, steward, how missingness is treated, and what crosses the operator–contractor boundary with which rights.
- The decision record. Complete the ten Appendix A fields for the decision class, marking any field you cannot fill.
- The honest part. Which definition your business currently gets wrong or cannot reconstruct. Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 2 — Our binding constraint, decision portfolio and field data contract.
How this is marked — the rubric
| Names the binding constraint with evidence and the role that can change it | 25 |
| Builds a portfolio described as decisions, with the chosen decision's roles distinguishable | 20 |
| Drafts the data contract for five variables, covering definition, timing, missingness and exchange rights | 35 |
| Completes the Appendix A fields and marks gaps honestly rather than filling them | 20 |
| Total · pass mark 70 | 100 |
Submitting session 2
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 2, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
Locating the binding constraintBuilding a decision portfolioDefining interface measuresField data contractsDecision records
Part One · FoundationsSESSION 3
Models That Fit the Decision, and Honest Evaluation
Why This Matters
Chapters 3 and 4 are the technical heart of the book's standard, and they are easy to delegate. You do not need to build a model. You do need to know what to ask the people who do: what it must beat, what exactly it predicts, when it will refuse to answer, and whether its test resembles the way it will be used. A single impressive accuracy figure can describe a system that never detects the event you care about. By the end of this session you should be able to read an evaluation for your company's decision and say whether it supports deployment, and to ask the one question that exposes a weak one.
25-Minute Module
0:00 – 6:00
Start with a credible alternative and state the task
- A model deserves deployment when it improves a relevant alternative under operating conditions. That alternative may be a rule, a persistence forecast, a decline curve, a reliability policy or an experienced engineer's documented process. Define the comparison before fitting the candidate.
- THE DOCTRINE
A weak baseline creates the illusion of AI value.
- Then insist the task is stated. A rate forecast states its horizon, whether it predicts oil or total liquid, and the operating state assumed. A failure classifier states the failure mode and time horizon. An optimisation states the objective, decision variables and constraints.
- Minimising prediction error is not the same as optimising a decision. A maintenance decision depends on the costs and consequences of false alarms, missed failures and delayed action. Ask for the prediction to feed an explicit decision policy, not a policy hidden inside a score.
- On methods, the book is measured: XGBoost is an established boosting method, not a guarantee of performance on field data [7]; retrieval-augmented language systems can help locate and synthesise documents [8]. Match each method to its information and failure modes.
Ask to see the baseline before you ask to see the model.
6:00 – 11:00
Complexity is a cost; refusal is a feature
- Complex models add obligations: inference time, compute, training stability, monitoring, explainability, specialist availability and recovery procedures. A marginal statistical gain may not justify them; equally, simplicity is not enough if the decision needs relationships a simple model cannot represent.
- An abstention policy lets the application say the evidence is insufficient. The book distinguishes four conditions: low predictive confidence, out-of-domain conditions, missing required observations and conflicting source documents. Each needs a response: manual review, a simpler approved baseline or suspension of the recommendation.
- A generic warning attached to every output has little operational value.
A system that never says no has not been tested on the cases that matter.
11:00 – 17:00
Evaluate the way it will be used
- Randomly splitting rows from the same well across training and testing can reward memorisation of that well's operating state. Use chronological evaluation for future behaviour; withhold entire wells or equipment identities to test transfer; where both matter, design separate tests.
Illustrative case (Appendix C, Model 2) Detected False alerts Recall Precision Accuracy Predicts no event for all 10,000 observations (100 events) 0 0 Zero Not defined: it raises no alerts 9,900/10,000 = 99 percent Candidate model 80 of 100 400 of 9,900 80/100 = 80 percent 80/(80 + 400) = 16.67 percent (80 + 9,500)/10,000 = 95.8 percent - The less accurate classifier detects events, but its review burden may still be unacceptable. For rare failures, report precision, recall, false alerts per equipment-month and lead time under a stated event-matching rule, and evaluate workload: an engineer cannot investigate every minute-level alert.
- THE OPERATOR · THE WORKED CASE
- Operator. A fictional operator's warning model is tested on random rows from the same wells. The honest test withholds future periods and whole wells, and counts false alerts per equipment-month against the engineers' review capacity.
THE SERVICE COMPANY · THE PARALLEL- Service company. A fictional contractor's duration forecast feeds mobilisation. Its credible alternative is a simple planning estimate, and the question is whether its interval is narrow enough to commit equipment to the next job.
- For probabilities, examine calibration: among events assigned a stated risk range, how often did the event occur? A large model's raw score is not automatically a calibrated probability [9].
Always ask for the denominator.
17:00 – 21:00
Intervals, causes and the selection matrix
- Standard split conformal prediction can provide marginal coverage under exchangeability and other stated conditions. Time dependence and distribution change need further analysis; a nominal 90 percent interval is not a promise for every individual well or operating state [10]. Test empirical coverage on the populations that drive decisions and record interval width.
- Appendix C, Model 8, illustrative: a service duration forecast has a central estimate of six days and an interval of four to ten days. A plan needing the equipment again on day seven is not assured by the central estimate. If the interval is too broad to select an action, the system has not solved the decision.
- THE DOCTRINE
Require separate evidence for future prediction, transfer to new assets and improvement caused by action. A single impressive accuracy figure cannot establish all three.
- Association is not intervention. High-performing wells may use a setting because engineers chose it for favourable wells; it does not follow that the setting will improve another well. Statistical adjustment cannot remove an unmeasured cause merely because the model is flexible.
Decision family Candidate approach Principal assurance question Structured forecast Regression or tree ensemble Does it beat the baseline on future periods and relevant assets Temporal warning State-aware statistical or sequence model Is the warning early and workload manageable Time to failure Survival model Are censoring, exposure and competing events handled Resource assignment Constrained optimisation Do constraints represent actual readiness and authority Document support Retrieval plus generation Are sources current, accessible and sufficient Bounded administration Agent with restricted tools Can content or errors trigger unauthorised action
Six of the book's eight rows are shown; often the right result of discovery is an engineering rule or a better workflow.
21:00 – 25:00
Write it down, Section 3, and Pop Quiz 1
- Before you close the laptop, write the credible alternative your decision must beat.
- Section 3 of your programme is The credible alternative, the model task, the abstention policy and the evaluation design.
- The leadership charge of Chapter 3: ask the model team to show the best alternative, the intended decision and the reason for added complexity, and require a demonstration of when the system refuses to recommend an action.
- Pop Quiz 1 opens now: five questions, closed book, on Sessions 1 to 3. Sit it before Session 4.
If you cannot name what it must beat, you cannot know whether it won.
Key Terms
Credible alternative
The rule, persistence forecast, decline curve, reliability policy or documented engineering process a model must outperform under operating conditions.
Abstention
The system's decision to withhold an answer or recommendation when the evidence, domain or authority is insufficient.
Calibration
The relationship between predicted probabilities and observed event frequencies under the stated evaluation conditions.
Exchangeability
The condition that the joint distribution of observations is unchanged under permutation. Standard conformal guarantees rely on it, and it must not be assumed casually for field time series.
Self-Check — Before You Submit
1. In the illustrative Appendix C example, 10,000 observations contain 100 events. A system that predicts no event every time scores 99 percent accuracy. What is its recall?
- A. 99 percent
- B. 1 percent
- C. Zero
- D. It cannot be calculated
C. Zero — It detects none of the 100 events, so recall is zero and it cannot provide an early warning. The 99 percent figure is the trap: accuracy is dominated by the 9,900 non-events.
2. A fictional vendor presents a failure-warning model tested on rows randomly drawn from the same wells it was trained on, and reports excellent results. What should you ask for?
- A. A larger random sample
- B. A comparison against a more complex model
- C. Accuracy reported to more decimal places
- D. Chronological evaluation and withheld whole wells or equipment, with false alerts per equipment-month and lead time
D. Chronological evaluation and withheld whole wells or equipment, with false alerts per equipment-month and lead time — Random splits within a well can reward memorisation of its operating state. A larger random sample repeats the same flaw at greater cost.
3. Under the abstention policy, which response does the book treat as having little operational value?
- A. A generic warning attached to every output
- B. Manual review when required observations are missing
- C. Falling back to a simpler approved baseline out of domain
- D. Suspending the recommendation when source documents conflict
A. A generic warning attached to every output — The book distinguishes the conditions and gives each a response. A blanket caveat looks cautious but changes no action, so it has little value.
Today’s Assignment
For your chosen decision, name the credible alternative it must beat, the mathematical task, the abstention policy and how the evaluation will match deployment.
- The credible alternative. What your company does today for this decision, described well enough to be measured against.
- The task. Target, prediction time and horizon, or objective, variables and constraints, or permitted questions and authoritative sources. Say how the output feeds an explicit decision policy.
- Complexity. What added obligations a model would bring to your business, and what gain would justify them.
- Abstention. For each of the four conditions, what the application does and who responds.
- Evaluation design. Chronological, withheld assets or both; the measures you will require, including workload; what you will ask about intervals; and how you would separate association from intervention.
- The honest part. Where your company would currently accept a weak baseline or an impressive single figure. Four hundred to nine hundred words; if no AI is in use, work from the decision in front of you.
YOUR 90-DAY PLAN
Section 3 — The credible alternative, the model task, the abstention policy and the evaluation design.
How this is marked — the rubric
| Describes the credible alternative concretely enough to be measured | 25 |
| States the task with its timing, and links the output to an explicit decision policy | 20 |
| Gives a distinct response and responder for each abstention condition | 25 |
| Designs evaluation that matches deployment, with workload and separate evidence for prediction, transfer and causal effect | 30 |
| Total · pass mark 70 | 100 |
Submitting session 3
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 3, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.ASSESSMENT NOW OPEN
Pop Quiz 1 — The Loop, the Data Contract and the Model opens now. Sit it before you start the next session — it is how you find out whether the doctrine landed, and the practice papers may be resat as often as you like.
Sit Pop Quiz 1 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Demanding the credible alternativeReading an evaluationAbstention policyInterval literacyQuestioning a model team
Part Two · Intelligence in the OilfieldSESSION 4
Subsurface, Reservoir and Drilling Decisions
Why This Matters
Subsurface interpretation, in the book's words, is an inference problem under incomplete observation. The danger with AI here is that a detailed output can look like a fully observed reservoir. A single coloured volume or a confident optimum can make uncertainty disappear from the room just when the investment decision needs it. If your company is an operator, this session is about which development, reservoir or drilling decision would actually change. If your company is a contractor, it is about the client decisions that create or cancel your work, and about the job evidence you return that becomes the asset's memory. Either way, the test is the same: does more interpretation change the action?
25-Minute Module
0:00 – 6:00
Preserve the alternatives
- Subsurface interpretation is an inference problem under incomplete observation. AI can propose seismic features, classify log intervals or help reconcile records, but the reservoir does not become fully observed because the output is detailed.
- Ask what a seismic model predicts: a geological feature, an interpreter's annotation or a measured property. These targets differ. For petrophysics, a model trained against another computed estimate may reproduce that estimate's assumptions.
- Link the model to the development choice. Propagate geological scenarios through the downstream evaluation: uncertainty in connectivity, contacts or permeability can change well ranking even when point estimates look stable.
- THE DOCTRINE
Do not allow predicted reserves to inherit authority from an AI label. A screening model can identify information gaps; it cannot independently certify a volume or commercial classification.
If only one interpretation is on the screen, ask to see the competing ones and the evidence that distinguishes them.
6:00 – 12:00
Buy information only when it changes action
- Value of information compares the best expected value you can achieve now with the expected value you could achieve after learning more. The difference is the expected value of sample information before its acquisition cost. The book's Appendix C example is illustrative, in millions of generic currency units.
Illustrative step (Appendix C, Model 4) Calculation Result Undertake now, equal priors 0.5 × 10 − 0.5 × 6 2 million: undertake After favourable signal (80 percent correct) 0.8 × 10 − 0.2 × 6 6.8 million: undertake After unfavourable signal 0.2 × 10 − 0.8 × 6 −2.8 million: decline, value zero Value with study, before cost 0.5 × 6.8 + 0.5 × 0 3.4 million Expected value of sample information 3.4 − 2 1.4 million Net of a 0.4 million study 1.4 − 0.4 1.0 million Expected value of perfect information (0.5 × 10 + 0.5 × 0) − 2 5 − 2 = 3 million - The study's accuracy, priors and project values are assumptions. Real information value must include whether the study arrives in time, whether the organisation can act on it and whether the test's performance applies to the geological setting.
- THE DOCTRINE
A fictional precision in the prior distribution will produce a fictional precision in the information value. If all credible scenarios select the same action, more interpretation may have little immediate decision value.
- Appendix C, Model 7, illustrative: a Beta(1, 9) prior has mean 0.10. After two events and eighteen non-events it becomes Beta(3, 27), mean 3/30 = 0.10. Same estimate, different strength of evidence. Two assets with the same estimated risk may need different review.
Ask what you would do differently if the study came back the other way.
12:00 – 17:00
Physics in the model, and its limits
- Machine learning can approximate a simulator, learn a residual error around a physics model, or support parameter estimation. These are different functions. A surrogate approximates another model's outputs; matching the simulator does not establish that the simulator matches the reservoir.
- Physics-informed neural networks add equation residuals and boundary or initial conditions to the training objective. Adding a physics penalty does not guarantee accurate or physically admissible predictions; the book cites the foundational research and later failure-mode analysis [11, 12].
- The book's single-phase conservation illustration omits multiphase behaviour, relative permeability, capillarity and other effects. A model enforcing an incomplete equation can fit the wrong abstraction very well. Ask which physics are represented, approximated and absent, and whether dimensional consistency was checked.
- Validate across the decision domain, withholding regions of the operating space. Outside the validated domain, return to the full engineering workflow and keep a route to the reference simulator.
An equation in the loss function is a design choice, not a validation.
17:00 – 22:00
Drilling and completion: state, sequence, envelope, evidence
- Drilling data contain changes in formation, operating mode, equipment and crew practice. Without a validated state classification, a model may treat a normal connection or planned transition as an anomaly.
- Evaluate warnings as complete sequences on unseen wells; a persistence rule or engineering threshold may be a strong baseline. Do not use model output to suppress independent alarms.
- Setpoints stay inside an approved envelope defined by competent engineers; the predicted optimum is a candidate for review. A reinforcement-learning agent's reward does not establish safe behaviour. The authorised drilling programme, equipment limits and well-control arrangements remain the governing framework.
- THE OPERATOR · THE WORKED CASE
- Operator. A fictional operator's completion assistant must retrieve the approved version, identify the applicable well and show the source of each extracted requirement. An absent record remains absent; the assistant must not manufacture a plausible value to complete the pack.
THE SERVICE COMPANY · THE PARALLEL- Service company. The fictional contractor's job close-out is, in the book's words, an input to future technical decisions, not merely a billing attachment. Accepted execution evidence feeds later production and intervention analysis as part of the asset's operating memory.
Treat the close-out pack as the asset's operating memory, not a billing attachment.
22:00 – 25:00
Write it down, and Section 4
- Before you close the laptop, write one decision in your business that would change if a subsurface, reservoir or drilling interpretation changed. If your company does no subsurface or drilling work, name the client decision that drives your work.
- Section 4 of your programme is Where subsurface, reservoir or drilling intelligence changes an action in our business, and what information is worth buying.
- The leadership charges: Chapter 5, ask which development decision changes when the interpretation changes, and require at least one credible alternative and the observation that would distinguish it. Chapter 7, demand proof that a warning arrives before the decision it is intended to improve.
If no action changes, the information is interesting, not valuable.
Key Terms
Expected value of information
The expected improvement in the selected decision's value from additional information, stated explicitly as before or after its cost.
Surrogate
An approximation of another model or simulation. Its approximation quality and the reference model's real-world validity are separate questions.
Physics-informed learning
A learning method or objective that incorporates physical relationships. It does not automatically guarantee a physically correct solution.
Approved envelope
The range of operating settings defined by competent engineers within which any model-proposed setpoint must remain, as a candidate for review.
Self-Check — Before You Submit
1. In the illustrative Appendix C, Model 4, the expected value of sample information is 1.4 million and the study costs 0.4 million. What is the net illustrative information value?
- A. 3.4 million
- B. 1.0 million
- C. 1.4 million
- D. 3 million
B. 1.0 million — 1.4 minus 0.4 is 1.0 million. 3.4 million is the expected value with the study before cost, not the value of the information; 3 million is the value of perfect information.
2. A fictional team presents a physics-informed reservoir model whose loss includes a conservation equation, and asks for approval. What does Chapter 6 require?
- A. Approval, because the physics penalty guarantees admissible predictions
- B. A larger neural network
- C. Evidence that the equation, numerical training and validation domain match the decision being made
- D. Replacement of the reference simulator with the faster model
C. Evidence that the equation, numerical training and validation domain match the decision being made — The book says adding a physics penalty does not guarantee accurate or admissible predictions. Retiring the reference simulator is the dangerous option: the book requires a route back to it.
3. A fictional drilling optimiser proposes a setting slightly outside the engineers' approved envelope, predicting faster progress. What should leadership do?
- A. Trial it on one well to gather evidence
- B. Accept it if the model's historical accuracy is high
- C. Let the reinforcement-learning agent decide
- D. Treat it as a candidate for engineering review; setpoints stay inside the approved envelope
D. Treat it as a candidate for engineering review; setpoints stay inside the approved envelope — Setpoints must remain inside an approved envelope defined by competent engineers; the model's predicted optimum is only a candidate for review. A one-well trial looks like learning, but it still moves the setting outside the envelope the engineers approved.
Today’s Assignment
Identify where subsurface, reservoir or drilling intelligence would change an action in your business, and what information is worth buying.
- The decision. One development, reservoir, drilling or completion decision your company makes, or, for a contractor, a client decision that drives your work.
- The alternatives. At least one credible alternative interpretation and the observation that would distinguish it.
- Information worth buying. Using the logic of value of information, without inventing figures, say whether additional data would change the action, whether it would arrive in time and whether your company could act on it.
- Limits. Where a model's authority must stop: reserves certification, the approved envelope, well-control authority or the validated domain.
- The evidence you return or receive. What close-out or completion evidence crosses the contract boundary, and whether it is usable as operating memory.
- The honest part. Where your company currently lets a single interpretation stand without its alternatives. Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 4 — Where subsurface, reservoir or drilling intelligence changes an action in our business, and what information is worth buying.
How this is marked — the rubric
| Names a decision that changes with the interpretation, or the client decision driving the contractor's work | 25 |
| Gives a credible alternative and the distinguishing observation | 25 |
| Applies value-of-information reasoning, including timing and ability to act, without invented figures | 30 |
| States where model authority stops and what job evidence must return | 20 |
| Total · pass mark 70 | 100 |
Submitting session 4
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 4, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
Keeping geological alternatives visibleValue-of-information reasoningQuestioning physics-informed modelsDrilling within the approved envelopeClose-out as operating memory
Part Two · Intelligence in the OilfieldSESSION 5
Production, Reliability, Integrity and Field Continuity
Why This Matters
These four chapters name three ways a benefit claim goes wrong: the same gain counted twice, a failure that did not happen credited to a model without evidence, and a clean summary covering an incomplete inspection. This session gives you the questions that keep those claims honest, and the rules that keep protection independent of any analytics system. Whether your company produces the barrels or services the equipment, you will leave with one field application defined by its connected-system constraint, its alert policy, its integrity evidence and what happens when the connection drops. Week 1 closes here.
25-Minute Module
0:00 – 6:00
Optimise the connected system
- A well's apparent optimum can be infeasible for the facility. A well-by-well ranking is a useful screen; it cannot independently select a feasible field schedule.
Illustrative (Appendix C, Model 3) Well A (20 percent water) Well B Facility check (liquid 1,500; water 500 bbl/d) Oil, bbl/d Plan at assumed 60 percent water cut on B 1,000 500 Liquid 1,500; water 200 + 300 = 500: both bind 800 + 200 = 1,000 Same plan if B is really 70 percent water 1,000 500 Water 200 + 350 = 550: limit violated Not feasible Corrected plan at 70 percent 1,000 300/0.7 = 428.57 Water 200 + 300 = 500 800 + 0.3 × 428.57 = 928.57 - Illustrative only: a real problem must include approved well limits, pressure behaviour, artificial-lift performance, shared gas or power, stability, measurement uncertainty and applicable safety constraints.
- Keep measurement separate from optimisation. A virtual flow estimate is not automatically acceptable for custody transfer, allocation or regulatory reporting; those uses need their own approvals and measurement assurance.
- THE DOCTRINE
Do not count the same gain once as improved well performance and again as reduced facility downtime. Make production improvement a field-level claim.
Ask where the extra liquid goes before you ask how much oil it brings.
6:00 – 12:00
Reliability: define failure, then convert risk into policy
- Equipment histories mix repairs, planned work, interruptions and failures. Define the failure mode and horizon first, and measure exposure: an idle unit does not accrue the same exposure as a heavily used one.
Illustrative (Appendix C, Model 1) Calculation Expected cost No preventive work, p = 0.12 0.12 × 500,000 60,000 Planned work, p_after = 0.02 40,000 + 0.02 × 500,000 50,000 Difference 60,000 − 50,000 10,000 in favour of planned work Break-even p − p_after > 40,000/500,000 = 0.08 p = 0.10 when p_after = 0.02 - If the plausible range for p is 0.06 to 0.18, the decision is sensitive to uncertainty, and inspection could be valuable if it changes the estimate enough to select an action. The model omits risk aversion, catastrophic consequences, interacting failures and statutory or mandatory maintenance. Never read the comparison as permission to delay required work.
- A remaining-life prediction of twenty days can be misread as a permitted operating life. Run alerts through consolidation, a review queue, acknowledgement, diagnosis, action and confirmed outcome. If a warning needs a spare unavailable for six months, the value may lie in inventory planning.
- Successful maintenance may prevent the failure that would have been the positive label. The absence of a failure after a warning does not make it false, and counting actions after alerts does not establish avoided failures.
An alert nobody has capacity to answer is a cost, not a control.
12:00 – 18:00
Integrity evidence and the service company
- Well integrity is a technical assurance discipline. AI may connect pressure trends, inspection records, barrier documentation and maintenance history, and rank records for review. It does not establish barrier integrity from an absence of alarms or a favourable score.
- Respect correlated uncertainty: a longer operation may increase vessel cost and production deferment together. Competent engineers establish feasibility; the economic model compares options inside that boundary.
- Assess the entire inspection chain: access, coverage, image quality, model detection and technician disposition. AI-generated visual reconstructions must never become inspection evidence. An inadequate image means further inspection.
- THE OPERATOR · THE WORKED CASE
- Operator. A fictional operator issues the authorised scope and current technical information, and must refuse any integrity summary that hides an incomplete inspection or an unverified barrier record.
THE SERVICE COMPANY · THE PARALLEL- Service company. The fictional contractor returns structured as-found and as-left records with tool configuration, calibration, deviations, approved changes and acceptance. That record supports the next intervention and protects the commercial record when additional work is requested.
- An anonymised learning library needs data rights; removing a client name alone may not prevent identification of a distinctive well.
A clean summary is only as good as the evidence chain beneath it.
18:00 – 22:00
Twins, edge deployment and independent protection
- A digital twin should specify the physical system, its state, update mechanism, fidelity and supported decisions. Mark scenario outputs as scenarios, or users may read a simulated rate as an observation.
- Choose cloud or edge from the actual location's latency, bandwidth, confidentiality, hardware, power and maintenance needs. An edge application needs versioned packages, secure updates, local audit records and rollback.
- Design degraded operation: administrative drafts can wait; engineering recommendations may revert to an approved conventional process; safety systems require their own engineered arrangements. Do not replay queued actions blindly when a connection returns.
- THE DOCTRINE
Keep analytics access and operational control permissions separately engineered. A digital twin does not replace an independent protective function merely because it models the same equipment.
- The book cites the NIST OT security guide, Special Publication 800-82 Revision 3, for the performance, reliability and safety characteristics of operational technology [6]. Its source list notes that an initial public draft of Revision 4 exists, and that a draft must not be treated as a final standard.
Ask to see the system with its best connection unplugged.
22:00 – 25:00
Write it down, Section 5, and Practice Exam 1
- Before you close the laptop, write the application you will use, its connected-system constraint and what happens when its data source disappears.
- Section 5 of your programme is Our field application: the system it optimises, the alert policy, the integrity evidence and degraded operation.
- The leadership charges: approve an alerting application only with a response process, a defined failure taxonomy and a method for assessing workload and outcomes (Chapter 9); require the integrity application to show the evidence chain and unresolved uncertainty (Chapter 10); ask the team to demonstrate degraded operation and recovery (Chapter 11).
- Week 1 closes here. Practice Exam 1 opens now: fifteen questions, closed book, covering Sessions 1 to 5. Sit it before Week 2.
Protection that depends on the analytics being available is not protection.
Key Terms
Virtual flow estimation
Inferring flow where direct measurement is intermittent. It is not automatically acceptable for custody transfer, allocation or regulatory reporting.
Right censoring
An event not yet observed by the end of the observation period. A unit removed because damage was suspected is not randomly censored.
Digital twin
A defined representation of a physical system with a stated relationship to observations and supported decisions. The term alone says nothing about accuracy or assurance.
Degraded operation
The planned behaviour of an application when a sensor fails, communications stop, the model service is unavailable or power is lost.
Self-Check — Before You Submit
1. In the illustrative Appendix C, Model 3, Well B's water cut turns out to be 70 percent rather than 60 percent. With Well A held at 1,000 barrels per day, what is the most Well B can produce within the 500 barrels per day water limit?
- A. 500 barrels per day
- B. 428.57 barrels per day
- C. 300 barrels per day
- D. 714.29 barrels per day
B. 428.57 barrels per day — Well A contributes 200 barrels of water, leaving 300; 300/0.7 = 428.57. Keeping 500 is the trap: it produces 550 barrels of water and violates the limit.
2. In the illustrative maintenance example, expected cost favours delaying a piece of statutory maintenance. What does the book say?
- A. Delay it; the arithmetic is clear
- B. Delay it only if the probability range is narrow
- C. Safety requirements remain constraints; never interpret the cost comparison as permission to delay required work
- D. Recalculate with a higher failure cost until planned work wins
C. Safety requirements remain constraints; never interpret the cost comparison as permission to delay required work — The book states the model omits statutory or mandatory maintenance and that safety requirements remain constraints. Adjusting the figures to reach the right answer treats a constraint as a monetary score.
3. A fictional operator's field link is restored after a day offline. Its analytics system holds a queue of actions recommended during the outage. What should happen?
- A. Replay the queue in order to catch up
- B. Delete the queue and restart the model
- C. Execute only the highest-value actions
- D. Do not replay blindly; verify timestamps and current context and authority before any action
D. Do not replay blindly; verify timestamps and current context and authority before any action — The book warns that the operating context or authority may have changed during the outage. Replaying in order looks orderly but acts on stale conditions.
Today’s Assignment
Choose one production, reliability, integrity or twin application in your business and define its connected-system constraint, alert policy, integrity evidence and degraded-operation and independent-protection rules.
- The application and the connected system. What it optimises or warns about in your company, and the facility, equipment or contractual constraint that limits it.
- Failure definition or alert policy. The failure mode, horizon and exposure, and the review queue: who acknowledges, diagnoses, acts and confirms, and with what capacity.
- Integrity and job evidence. What evidence chain supports the decision, and what as-found and as-left records cross the operator–contractor boundary.
- Degraded operation and independent protection. What happens when the sensor, connection, model service or power fails, and how protection stays independent of analytics.
- The benefit claim. How your business would test value after action against a counterfactual, without double counting.
- The honest part. Where your company currently counts a gain twice or relies on a summary it has not verified. Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 5 — Our field application: the system it optimises, the alert policy, the integrity evidence and degraded operation.
How this is marked — the rubric
| States the connected-system constraint the application must respect | 20 |
| Defines the failure mode or alert policy with a named response process and its capacity | 25 |
| Describes the integrity or job evidence chain across the contract boundary | 20 |
| Specifies degraded operation and keeps protection independent of analytics | 20 |
| States a counterfactual test of benefit without double counting | 15 |
| Total · pass mark 70 | 100 |
Submitting session 5
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 5, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.ASSESSMENT NOW OPEN
Practice Exam 1 — Foundations and the Oilfield opens now. Sit it before you start the next session — it is how you find out whether the doctrine landed, and the practice papers may be resat as often as you like.
Sit Practice Exam 1 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Connected-system optimisationAlert and maintenance policyIntegrity evidence chainsDegraded operation designIndependent protection