# October 14 gap review

Three Luna researchers ran a bounded pass across operator disclosures, private-company/vendor cases, and failures or follow-through. The editor re-opened the four strongest original sources and checked the canonical runtime register, including research packets. This is not an exhaustive earnings-call census. Integrated into the October 14 site update: AI-288 (CNX), AI-289 (Aker BP), a separate anonymous failure note and clearer Expand details. Sources, stages and limits remain distinct.

## Findings worth acting on

### CNX: add the safety workflow, with an undated-source qualification

CNX says it uses AI-enabled video analytics to detect unsafe conditions and inform coaching and training. It also describes HSE++ predictive analytics over historical incidents and safety observations. No CNX initiative currently appears in the main adoption register. Earlier October 9 research already flagged separate CNX methane ML; this safety workflow adds another specific use case.

[Original CNX source](https://www.cnx.com/sustainability-radical-transparency/people/workplace-health-and-safety/), Artificial Intelligence section. Accessed October 14, 2026. The page is undated: do not infer deployment start from its 2025 safety milestones. No AI-attributed incident reduction, scale, vendor, false-positive rate or ROI is disclosed. Vehicle telematics is not identified as AI. Recommended department: HSE; describe present use as company-reported, with deployment date unknown.

### Aker BP: capture a specific announced risk/alarm workflow

The May 21, 2026 Cognite/ABB release names Aker BP as the first customer signed on to implement an agentic offering for risk assessment and alarm rationalization using ABB Ability SafetyInsight and AlarmInsight. This exact source was not found in the canonical runtime data.

[Original Cognite release](https://www.cognite.com/en/company/newsroom/cognite-and-abb-collaborate-to-integrate-agentic-ai-into-industrial-applications-to-deliver). Relate to AI-049 industrial agents; evaluate a distinct announced workflow. AI-131 is Microsoft Copilot and is the wrong match. Stage: announcement/implementation commitment, not demonstrated scaled operation. Hundreds of agents and faster workflows are ambitions; no achieved outcome is established for this offering. Prior root-cause-analysis benefits must not transfer to it.

### EarthScan: add a concrete failure and the resulting narrower workflow

A September 10, 2026 vendor account describes an anonymous Oman operator's July 2023 blind test on five horizontal wells. Reported azimuth error reached 65.4 degrees. The vendor says the failed azimuth output was removed from automation; interpreters retained that decision while validated fracture-location/depth outputs remained an aid.

[Original EarthScan account](https://www.earthscan.io/v2/case-studies/the-generalization-cliff-what-horizontal-wells-did-to-our-fracture-model). Put in the failure/methodology appendix, not the named-company adoption count. Source contains ambiguous training descriptions; vertical reference metrics are not a matched controlled comparison. Anonymous customer, small sample, confidential data and vendor-authored results limit verification. No review cost or net return is disclosed. This is a failed output under changed well geometry, not evidence that the entire program was abandoned.

Practical implication: acceptance criteria should be set for each decision-critical output and operating condition. An aggregate score can hide the output that matters most to the engineer.

### Expand Energy: already researched; make the detail easier to find

The Snowflake case covers well forecasting, drilling ML, SCADA-linked site data, Cortex Analyst and M&A data integration. Its estimated 1,000-plus annual labor hours in one business unit and 40% improvement on M&A data work are already in AI-195's `oct6-three-rounds-ai-195` research packet. The source is already registered twice under source aliases. Do not add another project or call this a new discovery.

[Original Snowflake case](https://www.snowflake.com/en/customers/all-customers/case-study/expand-energy/), undated. Existing dossier correctly distinguishes current workflows from the drilling dashboard under development and MLOps roadmap. The headline scopes 40% to time on data-related M&A tasks, while the customer quote also mentions cost; methodology is unspecified. These are not demonstrated AI-only cash savings. Potential improvement: surface the existing details and limitations in the company profile.

## Coverage limits and editorial conclusion

The operator lane returned duplicates, including ADNOC RTOC/AiPSO and PETRONAS programs. PETRONAS's full report could not be opened; indexed passages supported duplicate checks only. Names screened without a qualifying new finding are not companies proved to have no AI work. This pass does not close all earnings calls, private deployments or discontinued-pilot coverage.

Prioritize integrating supported workflows and improving access to evidence already collected. Keep announcements, operating systems, future plans and measured results separate. Another unbounded search would not fix the presentation backlog or produce undisclosed project economics.

## Integration and rebuild

Run `python3 tools/build_data.py`, `python3 tools/integrate_oct10_review.py`, `python3 tools/integrate_flywheel.py`, `python3 tools/integrate_oct14.py`, `python3 tools/build_xlsx.py`, `python3 tools/build_pages.py`, `python3 tools/build_publication_editions.py`, `node tools/build_home_entry.mjs`, and `python3 tools/build_release_manifest.py --write-assets`. `tools/check_oct14.py` checks stage/date boundaries, counting, original-source closure, idempotence and export/history parity.