HSE Incident and Near-Miss Reports for AI: Safety Records from Industrial Operators

What public HSE data covers, and where it stops

Public HSE data is good for baselines and weak on near misses and investigation depth. OSHA's recordkeeping rule, 29 CFR Part 1904, defines which work-related injuries and illnesses employers record on the Form 300 log and the Form 301 incident report, and, as of October 2026, larger establishments in designated industries submit case data electronically through OSHA's Injury Tracking Application. That case data is built around recordable injuries and illnesses, not the far larger set of events where nobody was hurt, and its free-text fields are short. MSHA and BSEE publish mining and offshore incident data with similar limits on narrative depth.

Oil and gas indicators are mostly aggregate. API RP 754 sorts process safety events into tiers, with Tier 1 and Tier 2 as the lagging, consequence-based measures, and industry bodies such as API and IOGP mainly publish rates and counts, with only short summaries of selected high-potential events rather than record-level narratives. Tier thresholds have changed across RP 754 editions, so labels drawn from older reports may not match current tier logic.

Research corpora show what labels are possible. HARNESS combines Department of Energy safety and incident reports with event names, dates, locations, summaries and full text [1]. IncidentAI annotates high-pressure gas incident reports for named entities, cause-effect pairs and retrieval [2], and an aviation incident QA set shows how narratives become evaluation questions [3]. None of these supplies multi-site near-miss and observation data from current oil and gas, construction or manufacturing operations, which is the gap a licensed dataset fills.

Which record types matter for SIF-potential and classification models

The most valuable records are the ones that link a low-consequence event to its potential severity. An injury-only dataset teaches a model what already hurt someone; serious injury and fatality (SIF) precursor work needs the near misses and observations that did not. In most EHS platforms these sit in separate modules that share a site, date and asset key.

For upstream operations, linking incidents to activity context improves precursor models; see daily drilling reports with activity codes and NPT narratives and PLC, SCADA and DCS alarm logs. For visual evidence of unsafe acts, the companion guide on workplace safety near-miss video covers footage rather than written reports.

How to harmonize taxonomies across operators

Harmonization means mapping every supplier's codes into one schema before training, and keeping the original codes beside the mapped ones. Operators use different picklists in systems such as Enablon, Intelex, Cority, Sphera or in-house SharePoint forms. A "first aid" in one company is a "minor" in another, and potential severity may be a 1-5 matrix, a 5x5 risk score or a yes/no SIF flag.

Ask for these fields to be mapped explicitly: incident type, energy source (gravity, motion, electrical, pressure, chemical, thermal, and similar), body part and nature of injury, actual severity, potential severity, and the method used to rate potential severity. Record the version of any external framework used, for example the RP 754 edition behind a tier label. Treat the mapping table as a deliverable; ISO/IEC 5259-2 defines a data quality model with measurable characteristics that you can apply to each mapped field [4].

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "record_id": "NM-2025-04417",
  "record_type": "near_miss",
  "site_type": "onshore_well_pad",
  "event_month": "2025-03",
  "task": "pipe handling on catwalk",
  "energy_source": "gravity",
  "narrative": "[WORKER_1] stepped clear as a joint rolled off the catwalk; no contact.",
  "actual_severity": "none",
  "potential_severity": "SIF",
  "potential_severity_method": "operator 5x5 matrix, mapped",
  "barrier_status": "pipe rack stop missing",
  "root_cause_category": "equipment design",
  "corrective_action": "Install rack stops on all catwalks; verify at pre-spud inspection.",
  "action_status": "verified_closed",
  "source_code_original": "NM-LTI-POT"
}

Privacy, privilege and identifiability in safety records

Safety records carry more personal and legal risk than their volume suggests. The OSHA Form 301 incident report includes details about the injured worker, and internal reports add witness names, supervisor names, crew rosters, badge numbers and medical treatment notes. Names, emails, phone numbers and employee IDs should be removed or replaced, and medical details should be generalized or dropped, since workers' compensation and occupational health files can carry their own restrictions.

Indirect identifiers are the harder problem. A fatality at a small site in a named month is often public news, so a narrative can identify a person even with no names; see the guide on indirect identifiers in business text. Generalize dates to month, site to site type, and rare job titles to role families. Ask suppliers to exclude investigation reports prepared at the direction of counsel, along with litigation holds and regulator correspondence, because privilege can be waived by disclosure.

Diligence checklist for buyers

Use this checklist before signing for incident, near-miss and observation data.

• Scope: record types, sites, years, and whether near misses and observations are included or only recordables.
• Label lineage: who set potential severity and root cause, with which method and framework version.
• Taxonomy map: original codes, mapped codes and a list of unmapped values.
• Linkage: keys connecting incidents to investigations and corrective actions, and the share of records that link.
• De-identification: method used, fields treated, and results of a sample review for residual names in narratives.
• Exclusions: privileged investigations, workers' compensation files and records under legal hold.
• Rights: confirmation the supplier owns the records and may license them for model training, with permitted uses and term in the license.
• Documentation: a data card covering sources, collection, annotation and intended use [5].

Keep this distinct from IT outage work: software postmortems are covered by incident postmortem data, and construction project records exclude injury records, as described on the construction project datasets page. Buyers in these sectors can also see construction buyers and manufacturing buyers.

How SourceX sources HSE incident records

SourceX sources operational datasets from US companies on request and manages licensing and ongoing purchases; it does not hold safety data in stock, and a request does not guarantee a match. Buyers describe the records they need, such as near misses with potential severity or investigations linked to corrective actions, and SourceX looks for US operators that hold them. The process runs Find, Assess (data and licensing permissions), Agree (pricing and allowed uses in a license), Transact and Manage, and nothing is contracted until a supplier agrees.

Every dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Names, emails, phone numbers and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Any health records require HIPAA de-identification by Safe Harbor or Expert Determination. Diligence materials on source, rights, preparation and allowed use are prepared per dataset, and delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. You can describe your HSE data requirement to SourceX, and the wider industry operational data guide and AI data hub cover adjacent categories, including report-generation fine-tuning data.

Request HSE incident and near-miss data

SourceX serves AI teams wherever they are based and sources incident, near-miss and investigation records from US companies, with every release approved by the supplying company. Describe the record types, taxonomy fields and intended uses you need, and SourceX will assess data and licensing permissions with potential suppliers. Request HSE incident and near-miss data.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.