Method
Governed methodology from raw history to engineer-reviewed action.
Decline-curve analysis extrapolates a regime; it never identifies a mechanism. Everything below exists to keep that boundary visible — from the first data extract to the engineer's recorded verdict.
Fifteen stages
Every stage can refuse to pass a well on.
Each stage is a declared, versioned check. A check that cannot run is reported as not executed, never as passed. Thresholds and policy values are ActionWell's own provisional product decisions, exposed on every artefact and never presented as industry standards.
Inputs and identity
Source ingestion and lineage
Reported production history changes after the fact: regulators accept amendments and volumes are restated months or years later. Each extract is captured as an immutable, dated data vintage, and every forecast, deviation and dossier is pinned to the vintage it was computed from. Every derived value records its source system, source field, the conversion applied and the vintage.A result without a vintage is not reproducible and is not published. Lineage is not quality: a perfectly traced number can still be wrong at source.
Well and completion identity
Volumes are reported against a completion or well event, not the wellbore, and one well may have several. Identity is resolved as a mapping valid over a date interval, since wells are re-linked and change operator. Lease and facility totals are aggregates, not wells: a decline fitted to summed streams takes a shape no component well has.A fitting window that spans an unresolved identity change is refused, not fitted. A proposed identifier link is never used for a published forecast until a person confirms it.
Production-history reconciliation
Before any analysis, the monthly record is reconciled against what was actually filed: superseded amendments resolved, duplicates removed on a declared key, gas reconciled between gross and sales with flare, vent and fuel carried separately. Units, gas base and reference point are declared per source, never assumed. The raw value is never overwritten; every conversion is logged with its reversal.The conditioned series is a fitting input, never a corrected measurement of what the well produced.
Time, quality and regimes
Active-day and downtime handling
A volume divided by calendar days and the same volume divided by hours on production are different rates, and neither is the instantaneous rate a decline equation is written in. Where operating time is reported, both bases are computed and the basis used is printed on every parameter; otherwise the calendar-day basis is stated, not hidden. Hours are never imputed.Whether volume not produced during downtime was deferred or lost is an engineering judgement, not arithmetic. An episode that cannot be classified is reported as unclassified, never defaulted.
Data-quality gates
Every well receives a data-quality profile as a product output: coverage and gaps, zero months by cause, quarantined negative volumes, frozen values, restated months, allocation provenance. A well missing a required field is marked not eligible with the missing field named; nothing is imputed, gaps are preserved, zeros are kept with a recorded cause.A zero month is not a shut-in unless the status record says so. Not eligible describes the method's applicability at one vintage, not the well.
Regime segmentation
A decline curve describes one operating regime. When lift, choke, injection support or the fluid system changes, the analysis is restarted at the change rather than fitted through it. Boundaries are anchored on what the record can evidence — entity or measurement-basis changes and recorded operating events; automated change-point suggestions only ever propose a boundary, never accept one.In the one head-to-head comparison in the published literature — not an ActionWell measurement — an engineer's visual segmentation outperformed automated detection on most wells. Segmentation therefore produces candidates for engineer adjudication; the only claim made for automation is reproducibility, never accuracy.
Models and validation
Candidate decline models
Conventional decline models are empirical: they extrapolate a trend and carry no reservoir model, so every physical constraint is imposed from outside. They are defensible only on a stabilised interval under stable operating conditions. Several admissible models are fitted side by side; where the history cannot separate them, the result is recorded as insufficient evidence rather than a choice.Candidates are never averaged into an ensemble; the spread across them is a disclosure, not an uncertainty interval. Rejected candidates stay in the record with the reason.
Parameter and plausibility checks
An optimiser returns a number whether or not the data can determine it. Parameters are held within physically bounded ranges declared before fitting; a solution that lands on a bound is a finding about the data, not a fit. Residuals are examined for structure, no fit statistic accepts a model on its own, and implied well life and recovery are checked for credibility.Fits are re-run across fitting windows and starting points; the spread is reported, not averaged away. A decline rate without its convention is not a number.
Historical validation
Every surviving candidate is backtested in a design that preserves the order of time: fit on data up to a past origin, forecast forward, score against what happened, repeat at several origins. No later data, event record or restatement enters the fitting window. Out-of-sample evidence, not history-match quality, decides between candidates.A good in-sample fit is not an engineering-valid forecast. Calibration is a property of one asset's data at one vintage and is measured there, never inherited from published studies. No calibration has been measured yet.
Baseline, uncertainty and economics
Do-nothing forecast
The central output is the expected baseline — the do-nothing baseline on the product screens: what the well would produce if operated as it has been, from an accepted model on a qualified segment. That condition is stated on the face of the output. It is not a prediction of what the well will do, and not a reserves quantity. Once published it is immutable; a correction is a new version.A well whose operating conditions have changed is a different object; its forecast is restarted, not extended.
Uncertainty
Uncertainty is reported as separate components — parameter, fitting window, model choice and policy — and never folded into one unexplained band. A deterministic low, best and high case is never labelled with probabilities it does not have. Interval coverage is measured against outcomes on each operator's own population; until it has been, the interval says so.Confidence in data, regime, model, forecast, economics and causal evidence reaches the engineer unaveraged. A single combined score would tell the engineer nothing about which component to check.
Economic limit
An economic limit is the point of maximum cumulative undiscounted net cash flow under one declared set of prices, incremental cash operating costs, taxes and ownership, supplied and owned by the operator. Only costs that would actually stop if the well stopped are counted. Without a complete condition set, no economic figure is produced and the output is a technical forecast only.A rate cutoff used as a proxy is labelled as a proxy. Remaining production is stated to a named truncation; it is not EUR and it is not reserves.
Action and review
Action evidence
A deviation is the signed difference between actual production and the published baseline over a like-for-like window. It is a measurement, not a diagnosis: not deferred, lost, at-risk or recoverable production, which are four different objects a rate series cannot distinguish. Where a material deviation coincides with a recorded operating event, the output is a hypothesis at a stated evidence level, with the alternatives the method cannot exclude listed beside it.Temporal association is not causality. A hypothesis is recorded with the test that would settle it.
Engineer validation
Nothing leaves the system without a named engineer's verdict on each item of the acceptance evidence. The verdict set is closed: accept, accept with recorded amendment, reject the fit, reject the well, or return for a re-run. An amendment is always a re-run, never an in-place edit; dissent is recorded. A causal hypothesis is owned by a named engineer, not by the system.The system may compute, assemble and present; it may not conclude. A gate that is never exercised is not a control, so the share of items accepted without amendment is itself monitored.
Outcome monitoring
Realised production is compared with the published baseline on the same rate basis and stated vintages; the baseline is never re-based to suit the actuals. A breach triggers engineer re-evaluation, never automatic model rejection. Before any job, the do-nothing baseline is frozen, so realised uplift is measured against something not chosen with hindsight.Expected uplift and realised uplift are kept in separate columns and never summed. A realised uplift shows the well produced above its frozen baseline, not that the intervention is why.
Responsibility split
Three actors, three kinds of accountability.
No step delegates a parameter, a threshold, a model choice, a cause or a verdict to a machine.
Proposes, extracts, retrieves, drafts
This split is deliberate: by design the AI lane may only propose, extract, retrieve and draft. No accuracy claim is made for any AI output, and nothing it produces changes a published result before an engineer has reviewed it.
- Document extractionReads operator documents and free-text remarks into dated events and typed fields, each bound to the page and position it came from, or marked “not found”.
- Evidence assemblyAssembles candidate evidence links for a material deviation, each with a resolvable locator; nothing is presented as evidence unless it resolves to a source or a calculation.
- Anomaly and change-point suggestionsProposes candidate regime boundaries and flags atypical points for review; it never selects a segmentation, sets a threshold or deletes a point.
- Comparable-case retrievalRetrieves analogous wells and source passages with their locators as navigation; the reviewer reads the source, not the summary, for anything load-bearing.
- Explanation draftingDrafts what was computed and what to look at first — clearly marked, separable, introducing no number and no causal verb.
Computes, enforces, records
- CalculationsRate-basis conversion, conditioning, deviation arithmetic and backtest scoring are declared, replayable computations whose inputs and reversal are logged.
- Model fittingCandidate models are fitted under a declared objective and declared bounds, with out-of-sample validation that preserves the order of time.
- Parameter constraintsPhysical bounds are policy, not inference; a solution sitting on a bound is routed to review as a finding.
- EconomicsEconomic limit and remaining production come only from the operator's single declared condition set, and are suppressed when it is incomplete.
- VersioningEvery published object is pinned to its data vintage, code version and policy version, and superseded rather than edited.
- RecordsDossier, transform log, suppression register, review record and lineage are kept complete, append-only and versioned, without supplying the judgement.
Judges, owns, decides
- AssumptionsOwns and versions every choice the literature does not settle — rate-basis default, any late-life decline assumption, materiality, the economic condition set.
- Model suitabilityGives the verdict that an interval is eligible for extrapolation, adjudicates segmentation candidates and judges physical plausibility.
- Causal judgementOwns and names every causal hypothesis and makes the deferred-or-lost call. Attribution is a professional act performed by a person, with their name on it.
- Action validationApproves or rejects an intervention with cost, risk, HSE and scheduling inputs the system does not hold. A candidate surfaced by the system is not an approved action.
- Final decision supportSigns publication as the named approver. Investment approval, reservoir-management decisions and any reserves conclusion sit outside the system entirely.
No result is also a valid result.
Insufficient evidence is a designed outcome, not an error path. A well with no qualified stable segment receives no forecast; a well whose history cannot separate the candidate models receives no model choice; a well missing a required field is not eligible, with the field named.
In each case the well still receives a full dossier, with every withheld component present as an explicit suppression stating which evidence is missing, what would resolve it, and when to re-evaluate. A suppressed dossier can be the most actionable object the method produces, because it identifies a data or operating problem no forecast would have revealed.
Method boundaries
What the method does not do.
Every statement here is a cannot, not a cannot yet.
- The methodology is not fit for reserves booking, and no claim is made that a later version will be. Its outputs are estimates under stated conditions — not reserves, not EUR, not a booking, and not a reserves evaluator's opinion.
- Decline-curve analysis does not prove causality. It extrapolates a regime; it never identifies a mechanism. A rate change that coincides with an event is an association, and the system never names a cause.
- A deviation from baseline is not recoverable loss. Deferred, lost, at-risk and recoverable production are four different objects, none of which can be derived from a rate series alone.
- Identified value is not realised production or realised value. It is a queue-level opportunity figure under one assumed economic condition set — conditional and unrealised.
- Segmentation yields candidates for engineer adjudication. The segmentation of record is a human verdict.
- A rank is an allocation of attention, not a conclusion. A ranked queue is not a capital plan; a high rank says where an engineer should look, never what they will find.
- A candidate surfaced by the system is not an approved action. The system recommends nothing, approves no intervention, funds no programme and makes no reservoir-management decision.
- A deterministic low, best and high case carries no probability and is never labelled as one.
- No accuracy, calibration, uplift or performance figure is claimed. Accuracy is measured per asset, on that asset's own history; the claim is that it is measured, not that it is better.
- Every threshold, bound and policy value is ActionWell's own provisional product decision — versioned, exposed on the artefact and sensitivity-tested — and is never presented as an industry standard.
Next step
Test the method on your own history.
A validation is designed to return a defensible no as readily as a yes. Bring one asset and one decision question.