National importance
This page maps each of my public research repositories to a US federal interest that a federal agency has documented in its own words. Every federal statement links to the agency's page. Every result links to the file in the repository that produced it. No number on this page was estimated or rounded beyond its source.
In short: federal law requires employers to report injury data and chemical facilities to report accident histories, and federal agencies use and publish those records. My work tests how reliable those public records are when people compute rates and rankings from them, and how reliable AI tools are when they answer questions about the regulations behind them.
Repository to federal interest
Bracketed numbers point to the Sources list below. Repository figures are quoted from the linked file on each repository's default branch.
| Repository | Problem it addresses | Federal interest | Evidence in the repository |
|---|---|---|---|
| ehs-osha-analysis | Injury rates aggregated from OSHA's public Form 300A filings can be distorted by a few filings with impossible hours-worked values. | Establishments with 250 or more employees, and those with 20-249 in designated industries, must submit Form 300A data to OSHA every year by March 2 [1]. OSHA publishes the data [2] and, in its 2023 tracking rule, states that public access "will ultimately result in the reduction of occupational injuries and illnesses" [3]. | 2,801,064 filings, 2016-2024. 57,857 (2.07%) fail a plausibility screen yet hold 96.68% of all reported hours. Aggregate TRIR is 0.134 unscreened and 3.983 screened, a 29.7x correction.outputs/summary.json |
| ehs-benchmarks | Employers cannot compare their injury rate with similar establishments, because the raw public file is not usable as a benchmark. | The same OSHA electronic submission data [1][2]. BLS reports a 2024 private-industry total recordable case rate of 2.3 per 100 full-time workers [4]. | TRIR and DART percentiles for 3,842 NAICS by size-band peer groups. In the CY2023-2025 files, 2.24% of filings carry 86.9% of reported hours. Superseded by ehs-osha-analysis.README.md |
| ehs-chemical-peers | Chemical plants are ranked on injury rate, which says little about the risk of a hazardous chemical release. | EPA's Risk Management Program (40 CFR Part 68, Clean Air Act 112(r)) covers facilities holding more than a threshold quantity of a regulated substance; plans include a five-year accident history and are resubmitted every five years [5][6]. Employers must report amputations, eye losses and in-patient hospitalizations to OSHA within 24 hours [7]. | 10,597 NAICS 325 plants joined across OSHA and EPA records. Plants with a release in the past five years had one the next year in 12.06% of plant-years, against 1.47% without (8.2x). 18.21% of serious injuries with a usable prior-year log occurred at plants that recorded zero injuries the year before.outputs/summary.json |
| ehs-ai-grounding-eval | AI assistants can answer a safety-regulation question about the wrong but closely related device or rule, and the answer still reads as correct. | NIST AI RMF 1.0 lists "valid and reliable" as a necessary condition of trustworthy AI, and "safe" as not leading to a state where "human life, health, property, or the environment is endangered" [8]. | 68 benchmark items sourced from 29 CFR, 40 CFR and 46 CFR. Non-LLM baselines: TF-IDF retrieval 27.0%, BM25 28.6% accuracy on 63 factual items. No language model has been evaluated yet.results/baselines_table.md |
| ehs-human-factors-ontology | Human-factor frameworks used in incident analysis group their factors differently, so findings do not transfer between them. | The US Nuclear Regulatory Commission's Office of Nuclear Regulatory Research published IDHEAS-G (NUREG-2198, May 2021), a general human reliability analysis methodology [9]. | A machine-readable encoding of IDHEAS-G's 20 performance-influencing factors in four contexts, with 90 cited alignments to SPAR-H, CREAM, HFACS and the HSE factor list.crosswalk/crosswalk.csv |
| ehs-risk-sem | Safety risk scores and predictive alerts are published with more precision than the data can support. | BLS reports a 2024 private-industry rate of 2.3 recordable cases per 100 full-time workers [4]. The simulations use that rate as their base rate. | At that rate, a predictor with 80% sensitivity and 95% specificity raises 680 alerts per true event. Telling a path of -0.25 from -0.20 needs 6,884 observations.results/study02_alert_burden.csvresults/study01_requirements.csvSimulated data |
| ehs-capitals-calculator | Safety investments lose capital contests because costs that land in other budgets are left out of the business case. | OSHA states that injury costs "include direct and indirect costs," with indirect costs such as training replacement employees and accident investigation [10]. | The same loading-station investment gives 147% five-year ROI under traditional accounting and 315% when human-capital costs are counted. Inputs are user-supplied; the tool cannot check them.README.md |
| grounded | The repositories above are separate; readers need one place that states what is and is not established. | Ties the programme to the interests above: OSHA and BLS injury data [1][4] and NIST AI RMF trustworthiness [8]. | Landing site for the programme. Status is draft: all 45 references in the OSHA paper were verified on 2026-10-27 (reference ledger); reference verification for the other paper is incomplete, and no external peer review has been done.README.md |
Why this work matters nationally
Work injuries are common and costly. Private-industry employers reported 2.5 million nonfatal injuries and illnesses in 2024 [4], and 5,070 workers died from work injuries that year, one every 104 minutes [11]. OSHA cites almost $1 billion per week in direct workers' compensation costs alone [10].
GAO has found gaps in the injury data itself. GAO found that employers reported no summary injury data for more than half of the establishments it estimated were required to report [12]. GAO has also reported that workers may not report injuries for fear of job loss and that employers may not record them for fear of higher workers' compensation costs [13], and in 2024 that the extent of underreporting "is unknown on both OSHA recordkeeping forms and BLS surveys" [14].
My repositories address a related, measurable gap: errors inside the records that are submitted. A small share of filings can move a national aggregate rate by a large factor, and a plant's injury rate can look clean while its release history points the other way. The methods, code and outputs are public and tested, so OSHA, EPA, researchers or employers can rerun them on the same public files. The AI benchmark measures one specific failure that bears on the NIST "valid and reliable" characteristic: answers about safety regulations that cite the wrong but closely related entity.
Sources
All accessed 27 October 2026.
- 29 CFR 1904.41, Electronic submission of Employer Identification Number (EIN) and injury and illness records to OSHA. eCFR. ecfr.gov/.../section-1904.41
- OSHA, Establishment Specific Injury and Illness Data (Injury Tracking Application). osha.gov/Establishment-Specific-Injury-and-Illness-Data
- OSHA, Final Rule to Improve Tracking of Workplace Injuries and Illnesses. osha.gov/injuryreporting/final-rule
- BLS, Employer-Reported Workplace Injuries and Illnesses, 2024 (released 22 January 2026). bls.gov/news.release/osh.htm
- EPA, Risk Management Program (RMP) Rule Overview. epa.gov/rmp/risk-management-program-rmp-rule-overview
- 40 CFR 68.1, Scope (Part 68, Chemical Accident Prevention Provisions; authority 42 U.S.C. 7412(r)). eCFR. ecfr.gov/.../section-68.1
- 29 CFR 1904.39, Reporting fatalities, hospitalizations, amputations, and losses of an eye to OSHA. osha.gov/.../1904.39
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023. nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf
- US NRC, The General Methodology of an Integrated Human Event Analysis System (IDHEAS-G), NUREG-2198, May 2021. nrc.gov/.../sr2198
- OSHA, Business Case for Safety and Health: Costs. osha.gov/businesscase/costs
- BLS, Census of Fatal Occupational Injuries Summary, 2024 (released 19 February 2026). bls.gov/news.release/cfoi.nr0.htm
- GAO-21-122, Workplace Safety and Health: Actions Needed to Improve Reporting of Summary Injury and Illness Data (January 2021). gao.gov/products/gao-21-122
- GAO-10-10, Workplace Safety and Health: Enhancing OSHA's Records Audit Process Could Improve the Accuracy of Worker Injury and Illness Data (October 2009). gao.gov/products/gao-10-10
- GAO-24-106413, Workplace Safety and Health: OSHA Should Take Steps to Better Identify and Address Ergonomic Hazards at Warehouses and Delivery Companies (September 2024). gao.gov/assets/gao-24-106413.pdf
Repository figures come from committed outputs and READMEs on each repository's default branch. ehs-risk-sem uses only simulated data; ehs-ai-grounding-eval reports non-LLM baselines only; ehs-benchmarks is superseded by ehs-osha-analysis, which reconciles the two panels.
Not legal advice.