# Tasks in workplace incidents: task download guide

610 O*NET 27.2 tasks with activity documented in reviewed workplace incident narratives. Membership reflects AI-reviewed evidence of activity, not injury probability or a ranking of danger.

## Start here

Use dangerous_tasks.csv or dangerous_tasks.json for the task list. Join task_id to the O*NET 27.2 task inventory. task_description is the exact official description; evidence_status is ai_confirmed_incident_activity. No posting data are included.

## Files

- [ai_confirmed_onet30_tasks.csv](ai_confirmed_onet30_tasks.csv): 590-task O*NET 30.0 compatibility subset: IDs, descriptions, evidence quotations and AI review decisions.
- [ai_confirmed_onet30_ids.txt](ai_confirmed_onet30_ids.txt): 590 O*NET 30.0-compatible task IDs, one per line, for joining.
- [ai_confirmed_historical_tasks.csv](ai_confirmed_historical_tasks.csv): 610 supported historical tasks, including 20 O*NET 30.0 identity exceptions.
- [inventory.csv](inventory.csv): All 6,060 candidate tasks and their final review decisions, including partial and unconfirmed tasks.
- [confirmed_evidence.csv](confirmed_evidence.csv): Source narratives, quotations and review provenance: 610 incident-linked tasks plus 93 separately labeled context-only tasks. Context-only tasks are excluded from the primary bundle; multiple review records can refer to one task.
- [ai_confirmed_context_only_mentions.csv](ai_confirmed_context_only_mentions.csv): Contextual activities kept outside the primary incident-linked bundle.
- [review_protocol.md](review_protocol.md): AI evidence-review protocol and decision rules.
- [PROVENANCE.md](PROVENANCE.md): Source and review provenance.
- [VALIDATION_REPORT.md](VALIDATION_REPORT.md): Review results, interpretation limits and reproduction instructions.
- [review_summary.json](review_summary.json): Machine-readable review summary.
- [release_integrity_audit.json](release_integrity_audit.json): Source release integrity checks.
- [dangerous_tasks.csv](dangerous_tasks.csv): Compact primary list: 610 rows; task_id, task_description, onet_version, evidence_status.
- [dangerous_tasks.json](dangerous_tasks.json): The same 610 records as a JSON array; IDs are strings.
- [export_downloads.py](export_downloads.py): Code selecting the compact list from the reviewed release and building this download page. The original source-release directory is required to rerun the full exporter.
- [narrative_task_counts.csv](narrative_task_counts.csv): Pre-review task match occurrences and distinct investigation counts for the 610 IDs.
- [occupation_task_panel.csv](occupation_task_panel.csv): Only the occupation-task frequencies displayed in the joint injury explorer.
- [OSHA_AUDIT_AND_LIMITS.txt](OSHA_AUDIT_AND_LIMITS.txt): Dated Codex computational audit, corrections and interpretation limits.
- [osha_source.bib](osha_source.bib): OSHA source citation.
- [injury_sources.bib](injury_sources.bib): BLS CFOI and SOII source citations.
- [osha_run_full_analysis.py](osha_run_full_analysis.py): Original extraction/aggregation source code; historical pct_investigations uses occurrence counts, as disclosed in the audit.
- [osha_build_onet_task_inventory.py](osha_build_onet_task_inventory.py): Deduplicate report IDs and audit task identity.
- [osha_prepare_onet_list_validation.py](osha_prepare_onet_list_validation.py): Construct original task evidence packets.
- [osha_collect_onet_list_validation.py](osha_collect_onet_list_validation.py): Collect frozen AI review decisions; does not perform the reviews.

## Interpretation

The historical_n_investigations and original_rank fields in the full review exports describe original candidate retrieval, not validated incident frequencies or danger rankings. The inventory includes unconfirmed and partial evidence; these are not members of the primary 610-task bundle. Review is AI-based, not human validation. The 610 historical positives include 19 IDs absent from O*NET 30.0 and one description mismatch. All 610 match O*NET 27.2 and remain in the compact list. The 30.0 files are compatibility subsets, not evidence exclusions.

## Provenance and code

Source: TaskMatchAccidents/results/onet_list_validation, source release dated September 27, 2026. Original exports are copied byte for byte. manifest.json records SHA-256 hashes. export_downloads.py builds the compact derivatives; it does not repeat the underlying AI review. See review_protocol.md and PROVENANCE.md for that procedure. Original release reports describe the 30.0 compatibility subset; this guide and the current benchmark methods document the full 610-task 27.2 bundle. This download guide and page description were generated by Codex.


## Full-corpus retrieval

[All narrative-task matches](full_corpus_task_matches.csv.gz): investigation IDs, sentence text, all 6,187 raw task IDs, and flags for review inventory and retained bundle. [All investigation counts](full_corpus_investigation_counts.csv.gz). [Field definitions](FULL_CORPUS_MAPPING_README.txt). These are retrievals, not individually validated matches.