Short answer
Near-miss and incident reports are short narratives of what almost went wrong or did go wrong: the task, the hazard, the cause and the fix. They can support hazard recognition, safety reasoning and evaluation of language models that plan around risk. They contain no motion or perception data, and they carry names, injuries and legal sensitivity that make redaction and permission the main hurdles.
What a near-miss report holds
Environment, health and safety (EHS) teams collect near-miss reports because they show a hazard before anyone is hurt. Depending on the form, a record may name the task, where it happened, what the person was doing, what went wrong or nearly did, why, and what was changed afterward. Incident reports follow the same logic but record actual harm, so they add injury, treatment and sometimes investigation findings.
For an embodied-AI buyer the value is in the reasoning. A robot working near people has to anticipate hazards that a task description never mentions: a wet floor, a load above head height, a pinch point. Narratives of near-misses are a compact, real-world list of those situations, with the cause stated in words. Ergonomic injury vocabulary is part of it; OSHA describes disorders such as strains and tendinitis and risk factors such as forceful exertion, repetitive motion and awkward postures (OSHA, Ergonomics overview), and reports may use similar terms; check how consistently a supplier's records do.
What the data can support
- Hazard recognition and classification from text: fall, caught-in, struck-by, chemical, ergonomic, vehicle.
- Safety-aware planning and constraint extraction for language-model components that decide whether to proceed, stop or ask.
- Evaluation sets for hazard reasoning: given a scene or task description, identify what could go wrong.
- Prioritization of which tasks and locations carry risk, useful when scoping what to collect next. See prioritizing maintenance tasks for collection.
- Linking text to other records, such as corrective actions and nonconformance reports.
Reports become stronger when paired with the action taken; see corrective action records. They are close cousins of nonconformance reports, which describe product rather than people-safety events.
Fields that matter
| Field | Why a model or evaluator needs it | Sensitivity |
|---|---|---|
| Task and equipment involved | Grounds the hazard in a real activity | Low to medium |
| Location type (dock, aisle, kitchen) | Lets you split by environment | Low |
| Hazard category | Enables classification and balancing | Low |
| Narrative of the event | The core reasoning content | High (names, details) |
| Root cause | Supports causal reasoning | Medium |
| Corrective action | Closes the loop for planning | Medium |
| Severity or potential severity | For prioritization and evaluation | Medium |
| People involved | Rarely needed by a model | Very high |
Illustrative package: redacted near-miss narratives
# Illustrative only - fields and values are examples, not an available dataset
package: near_miss_narratives_v0
source: EHS system exports from 2 operating businesses
records: text only, English, 40-300 words per narrative
fields:
task: "restocking top shelf from step ladder"
location_type: dry_goods_aisle
hazard_category: fall | struck_by | caught_in | ergonomic | chemical | other
potential_severity: low | medium | high # reporter's estimate
root_cause_text: free text, redacted
corrective_action_text: free text, redacted
report_year: year only
redaction:
removed: names, employee IDs, exact dates, site addresses, witness names, medical details
reviewed_by: two reviewers, spot-check log kept
excluded: photos, video, injury treatment records, legal correspondence
rights: AI training and internal evaluation; no redistribution
Permissions and confidentiality
Incident records can attach to claims, insurance matters, regulatory filings and disputes. The business may treat them as confidential, privileged, or subject to retention duties. Employee names and descriptions of injuries are personal information, and some records may fall under health-related or employment-related rules. Which rules apply depends on the facts and jurisdiction; this is not legal advice.
Questions to put to the supplier include: who owns the report content, whether it was prepared at counsel's direction or for an investigation, what employees or unions were told about its use, and whether a license could conflict with obligations to insurers or regulators. Redaction reduces what a narrative reveals, but redaction alone does not establish permission to license the underlying record. A closely related topic is confidentiality in defect data.
Limitations: what hazard narratives cannot do
A narrative describes an event in words, so it contains no sensor data, motion, force, or geometry. It cannot train physical control, and it does not show what a hazardous situation looks like to a camera. A robot needs images or video of the scene types involved; compare environmental hazard images.
Reporting culture can shape the data. Ask the supplier how reporting is encouraged and whether reports can be anonymous: where people fear blame, near-misses may go unreported, so the mix of categories reflects what employees chose to write down rather than the true distribution of risk. High-severity events may be few, narratives are written by non-experts, and the same event may be described differently by different reporters. Free text often contains abbreviations and site slang that need normalization.
The text also describes the human side of work. It will not tell you how a humanoid's own size, speed and strength change the hazard picture, so treat the content as input to safety reasoning, not as a safety case. Hazard annotations on objects themselves are covered by hazard properties of objects. For wider reasoning about why tasks fail, see failure reasoning data.
Acceptance checks
- Read a random sample of narratives: are they specific enough to name a task, hazard and cause?
- Check redaction on a sample for names, dates, places and indirect identifiers such as a rare job title.
- Test category balance; confirm rare, high-severity categories are present or note their absence.
- Check duplicates and templated text that inflate counts.
- Confirm that a held-out set can be separated by site and time.
Using the text with other inspection records
Narratives are strongest when joined to other records about the same task or location. A near-miss about a pinch point becomes more useful beside the inspection report text for the same equipment, the pre-use checklist that missed it, and the corrective action that followed. If you plan to evaluate a language model on hazard reasoning, hold out whole sites or time periods, so the evaluation does not reward recalling a report the model saw in training. Keep the evaluation narratives out of every training package.
Questions to ask suppliers
- Where were the reports created: an EHS system, paper forms, or a vendor platform?
- Who owns the report text, and was any of it prepared for counsel, an insurer or a regulator?
- What did employees understand about how their reports would be used?
- Which fields identify people, and how are names, IDs and indirect identifiers removed?
- Are injury and medical details excluded?
- Are near-misses separated from actual incidents, and are severities defined?
- How many sites and task types are covered, and over what years?
- Are corrective actions and root causes recorded in structured fields or only free text?
- Is there a policy of anonymous reporting that affects how the data can be shared?
- Can records be withdrawn or deleted if a claim or investigation requires it?
Discuss your requirements
If you need hazard narratives for a defined set of tasks and environments, discuss your humanoid training-data requirements with SourceX. SourceX reviews the requirement and follows up about potential US business sources, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match.
See the humanoid training data guide and the inspection, defect and failure data hub. For related catalog categories, see inspection reports and manufacturing quality records.
Discuss your humanoid training-data requirements with SourceX
Sources
- Ergonomics - Overview — ergonomic disorders and risk factors that incident narratives commonly describe.