Morning briefing: persistent drift, drilling anomaly evidence and auditable models

October 28 Morning Briefing.

Fresh Western Canadian drilling-specific news is limited at preparation time, so this briefing does not manufacture a local milestone.

It uses one overnight global market report, an October 12 peer-reviewed geothermal drilling study and clearly labelled technical learning.

The technical focus rotates from last night’s single-point robust anomaly score to a transparent way to see persistent drift across several valid samples.

Canadian desk — local evidence remains the governing layer

The current Industry Chatter catalog already carries the latest Canadian workforce, inflation, regulatory, Montney and drilling-related records, along with yesterday’s global gas outlook, Baker Hughes results and geothermal additions.

Those articles remain in their original sections rather than being copied into this edition.

For a Calgary wellsite geologist, the practical morning question is whether any new outside signal changes today’s geological interpretation or operating plan.

A market move, an overseas research result or an algorithm benchmark does not supersede the accepted survey, B H A, drilling state, lag model, offset evidence, live logs or sample descriptions from the well being drilled.

Overnight energy signal — price movement is not confirmation of physical damage

October 28, 2026 — Reuters reported that Brent and WTI rose more than 1% as the market reacted to a Houthi claim that a Saudi refinery had been struck.

At the report’s 11:15 GMT update, Brent was US$115.19/bbl and WTI US$101.65/bbl.

Reuters said Saudi authorities had not immediately confirmed the claimed strike on the 550,000-bbl/d facility.

It also reported a source-based American Petroleum Institute estimate of a 2.03-million-bbl US crude-stock decline, while official Energy Information Administration data was still due.

These are time-stamped market prices, an attributed attack claim and preliminary inventory information—not verified damage, lost production, a Canadian drilling forecast or a reason to rewrite a geological model.

Field meaning — separate commercial urgency from technical evidence.

Higher prices can change budgets, service availability and the value of avoiding nonproductive time, but they do not prove a formation top, justify a steering correction or validate a new drilling model.

Infinity reporting should preserve the source time and status of external context so a later reader can distinguish what was known at the rig from what was confirmed afterward.

Research watch — legacy drilling data can identify some problems, but not every problem

First published online October 12, 2026 — Arran and coauthors in Geoenergy describe an Isolation Forest system trained from publicly accessible legacy hydrocarbon drilling data and tested for geothermal drilling-problem identification.

The paper reports a 29% detection rate for problems that interrupted drilling versus an 8.6% random-chance baseline, with p = 0.012, including application to new measurements in new geological contexts.

The article is peer-reviewed and open access under CC BY 4.0.

Those results demonstrate a research workflow and a statistically better-than-random signal in the authors’ tests.

They do not establish a field-certified alarm, a complete problem-detection rate, a safe autonomous control loop or transfer to a Western Canadian horizontal without local validation.

The transferable lesson is data discipline.

An unsupervised model can rank combinations that look unusual without being told every failure label, which helps when serious events are rare.

It still needs synchronized channels, stable units, explicit drilling states, documented exclusions, offset separation and a reviewer who can distinguish telemetry faults, connections, slides, reaming, formation changes and genuine mechanical deterioration.

Detection is not diagnosis.

A model score should open the supporting evidence, not hide it.

Technical learning — EWMA makes gradual drift visible

NIST describes an exponentially weighted moving average, or EWMA, as a monitoring statistic that gives progressively less weight to older measurements and can be more sensitive to small or gradual process shifts than a chart driven only by the newest point.

For Infinity, use it only within one accepted drilling state and comparable geological interval.

Reset or suspend it at connections, slide/rotate changes, reaming, B H A changes, large set-point changes and accepted formation boundaries.

An EWMA is a trend aid, not a formation call or operating alarm by itself.

zₜ = λxₜ + (1 − λ)zₜ₋₁, with 0 < λ ≤ 1

Variables and units — xₜ is the current accepted sample, zₜ is the updated EWMA, zₜ₋₁ is the prior EWMA and λ is a dimensionless weighting factor. x and z retain the selected curve’s units, such as m/h for R O P, MPa for differential pressure or kN·m for torque.

A smaller λ gives more memory; λ = 1 reduces the statistic to the newest sample.

NIST notes that 0.2–0.3 is commonly used for detecting gradual drift, but that is general process-monitoring guidance, not a preset for drilling.

Choose and validate λ against the channel cadence, tool response and decision horizon.

Control limits require representative in-control history and a defensible variability estimate; do not invent them from one short interval.

Hypothetical worked example — not rig measurements, a trip criterion or an operating threshold.

Begin with an accepted rotary-drilling R O P baseline z₀ = 40.00 m/h and choose λ = 0.25.

For successive valid samples of 38, 36 and 34 m/h: z₁ = 0.25 × 38 + 0.75 × 40.00 = 39.50 m/h; z₂ = 0.25 × 36 + 0.75 × 39.50 = 38.63 m/h; and z₃ = 0.25 × 34 + 0.75 × 38.63 = 37.47 m/h.

The smoothed statistic shows a persistent decline without claiming its cause.

Check WOB, RPM, torque, differential pressure, flow, vibration, bit depth, survey, GR, gas and samples before calling dysfunction or geology.

Verified open source — scikit-learn for controlled offline evaluation

Scikit-learn is genuinely open source under the BSD-3-Clause license.

PyPI lists version 1.9.1, released September 10, 2026, as Production/Stable, requiring Python 3.11 or newer; the official GitHub release describes 1.9.1 as a bug-fix release with Python 3.15 support.

The current documentation includes IsolationForest and explains that anomalous samples tend to have shorter average path lengths across randomized trees.

This verifies an actively maintained general machine-learning library.

It is not a WITS M L client, a drilling-certified application, a browser-ready Infinity feature or evidence that the Geoenergy paper used Infinity’s data model.

Concrete oilfield use case — export an immutable, de-identified interval table containing aligned R O P, WOB, torque, differential pressure, RPM, flow, inclination and state labels; split wells or bit runs so the test set is not sampled from the training interval; fit an Isolation Forest offline; then compare flagged intervals with confirmed events and reviewer dispositions.

Keep raw channels and an interpretable EWMA beside the model score.

Do not let a notebook write steering commands or overwrite the accepted interpretation.

Practical workflow — move from drift to evidence

Define the valid drilling state, bit run, B H A, channel UID, units, sample cadence and geological interval before calculating a trend.

Plot raw values, EWMA and the state mask together.

A smoothed curve without its valid-sample support can make gaps and state changes look continuous.

Keep time-index and depth-index views linked.

A trend seen at surface time may map differently after depth assignment or geological lag.

Compare an unsupervised flag against independent channels and confirmed events.

Record false positives, missed events and insufficient-support intervals.

Separate model-development data by well or run, not random rows, to reduce leakage from near-duplicate neighbouring samples.

Require a human disposition before a flag affects reporting, interpretation or steering advice; retain that decision and model version.

Infinity watchlist — proposals, not implemented features

Live curves: propose a selectable EWMA overlay with visible λ, initialization, reset markers, valid-state mask and insufficient-support status.

Average overlays: keep raw, ordinary average, median and EWMA choices distinct; never label all of them “average” or silently change the method.

L A S/WITS M L: preserve channel UID, time/depth index, unit, source revision, null policy and derived-curve formula or model version.

Lag and annular velocity: shift only transported surface evidence using the accepted lag model; retain unshifted source depth and uncertainty.

Data quality and sample imaging: let a flagged interval open synchronized channels and the nearest lag-corrected sample images with collection support.

Reporting: include the score or EWMA, support window, operating state, reviewer disposition and links to the original evidence.

3D/VR: colour accepted intervals by reviewed status and confidence; do not turn an isolated model flag into a geological body.

Applied AI and DRQI: evaluate candidate features offline with well-held-out tests.

Do not add a model score to DRQI until the final workbook formula, weighting and validation rules are reviewed.

Practical takeaway: use EWMA to expose a gradual change inside one valid operating state, then open the raw mechanical and geological evidence before assigning a cause.

A transparent trend plus reviewer context is useful; an unexplained model flag controlling the well is not.