# Second October 14 research pass

Three Luna researchers screened smaller/private operators, international operators, and business workflows/failures. The editor checked original sources and existing dossiers, and reviewed two additional industry reports. This is a bounded search, not an exhaustive transcript census. Integrated into the second October 14 site update: AI-290 Xinjiang, AI-291 CNX methane, AI-292 HG Oil, expanded AI-139 Santos and two KPMG library entries. Benefits remain source-reported; Santos program totals are excluded from counted results pending scope reconciliation.

## The strongest omission: Xinjiang well diagnostics

[CNPC’s April 15, 2026 article](https://news.cnpc.com.cn/system/2026/04/15/030190061.shtml) describes a one-billion-parameter Transformer model diagnosing rod-pumped-well conditions. The company reports use on more than 3,800 Xinjiang wells, diagnosis accuracy above 90%, and abnormal-well discovery falling from days to minutes. Its 78,000-well data set spans four oilfields; that is not the live deployment population. Staff receive diagnoses, evidence and recommendations. Costs, validation definitions and measurement periods are undisclosed. The architecture does not establish a text LLM or autonomous control.

The specific workflow and source are absent from the register. AI-081’s Kunlun program and AI-236’s Dagang forecasting are adjacent cases. The original article was retrieved and decoded after the web reader failed on its encoding.

## Santos needs a much more specific profile

[Original investor deck, slide 72 / PDF page 74](https://www.santos.com/wp-content/uploads/2026/05/Santos-2026-Investor-Briefing-Day_ASX.pdf#page=74). Its four columns separately describe production-optimization AI ($28m reported 2025 benefits; over $64m potential annual benefit by 2029), an ML pressure model replacing physical gauges (up to about $70m potential capex reduction), rig-connected drilling automation (20% penetration-rate gain), and stimulation logistics (80% productivity gain; not identified as AI). The two potential dollar figures are gross. No complete cost/return bridge is given.

The editor fetched and rendered the original slide, correcting the researcher’s initial column attribution. AI-139 already cites this deck but only describes a generic program. Its research-packet timeline says May 20; the release cover confirms May 26, 2026. Production-optimization work may overlap AI-083/Xecta; resolve that before counting another result. Do not attach its benefit to the pressure model.

## CNX methane: good evidence waiting to be integrated

[CNX’s methane page](https://www.cnx.com/sustainability-radical-transparency/environment/commitment-to-methane-reduction/) describes ML on well/equipment signals to identify release precursors, OGI/aerial analysis to prioritize sources, and operational-data reconciliation with the emissions inventory. It reports qualitative Q1 2026 progress. No AI-specific population, reduction, accuracy or cost is provided. Earlier company emissions reductions must not be assigned to this workflow. Remote shutdown capabilities do not prove autonomous AI control.

This was already an October 9 candidate, so it is an integration backlog item with renewed original-source verification. AI-288 covers separate safety cameras.

## HG Oil: a qualified invoice-processing example

[Preisser’s named customer case](https://preissersolutions.com/case-studies/hg-oil-ai-invoice-processing) describes extraction, coding, anomaly detection, approval routing and human exception review. The vendor reports 75% less manual handling and more than 40 hours/week redirected, with an avoided hire. The page is undated; invoice volume, common measurement period, cost and independently verified savings are absent. The figures cannot reconstruct a net return.

[HG’s own About page](https://www.hgoilholdings.com/about) confirms it is a Kansas oil operator. Its company-wide well count is not an AI deployment population. Human payment approval is explicit on a separate [vendor product page](https://preissersolutions.com/use-cases/ai-invoice-processing-small-business), not the named case alone. Keep that distinction.

## Two reports missing from the library

- [KPMG Global tech report 2026: Energy, Natural Resources and Chemicals](https://assets.kpmg.com/content/dam/kpmgsites/ie/pdf/insights/energy/ie-kpmg-global-tech-report-2026-energy-resources-chemicals.pdf.coredownload.inline.pdf): methodology on printed page 16 identifies 258 technology leaders, including 58 oil/gas respondents, from organizations above $1bn revenue. Printed pages 6–7 report investment priorities. These are self-reported intentions and maturity, not measured project returns. It excludes smaller-company perspectives; future expectations are not achieved scale.
- [KPMG US: How energy can scale advanced tech safely](https://kpmg.com/kpmg-us/content/dam/kpmg/pdf/2026/how-energy-can-scale-advanced-tech-safely.pdf): includes 100 US energy-business executives. The AI/automation and data/analytics maturity comparisons cover broader energy, not upstream alone. Keep its population separate from the global sector report. Both are 2026 editions without an exact publication day established in this review.

A generic KPMG PDF URL returned a financial-services report despite energy passages in search indexing; it was rejected. S&P’s mixed-return survey, Deloitte’s sector AI report and Rystad’s opportunity model were already in the library.

## What we did not establish

No additional named abandoned deployment was substantiated. Oil India’s leak-detection presentation remains held because operating scope is unclear. BKV VADER and QatarEnergy LNG pilot expansion are existing evidence. No search result is treated as proof that an operator has no AI work.

The practical additions concern different sources of value: better field diagnosis, replacing physical measurement hardware, and reducing office processing. They should have separate evidence and economics; they do not support a common ROI ranking.

The accompanying JSON files retain raw findings, editor dispositions and access limits. `editor-review.json` is the consolidated decision record.