Digest · Updated 28 October 2026

Findings digest

The strongest sourced findings across my public EHS research repositories, in one table. Each row gives the finding with its exact number, the repository file it comes from, the limitation that file states, and why it matters for EHS practice.

How to quote it: copy a row with its source link. Numbers are copied from the file on the main branch of each repository, and a test checks that every number on this page appears in the file it cites. Where a source is a simulation, a small sample or a lexical baseline, its limitation says so in the same row. No row is a claim about any specific facility, employer or person.

RMP accident rate by peer group

Accident records per 100 facility-years for NAICS 325 product-type peer groups with 100 or more facilities, with the exact Poisson 95% interval. Highest rate first. The axis starts at zero. Source: ehs-chemical-peers/outputs/accident_by_subsector.csv.

Reporting, not risk: only accidents that meet EPA's five-year-history criteria are on file, and the intervals assume independent accidents, so the true uncertainty is wider than shown. See the first chemical-peers row below.

Findings, sources and limits

One row per finding. The finding and the limitation are quoted or assembled from the linked file on the main branch of each repository.

Finding	Source	Stated limitation	Why it matters for EHS practice
Screening implausible hours raises the pooled 2016-2024 aggregate TRIR from 0.134 to 3.983, a 29.7x correction. The 2.07% of filings flagged hold 96.68% of reported hours.	ehs-osha-analysis/README.mdREADME, Headline results	It does not identify which specific numbers on forms are wrong, nor does it address causes of injury, nor does it represent the broader US workforce.	Injury rates built from unchecked, self-reported hours can look far better than they are. Checking hours before comparing rates keeps a benchmark from rewarding a data-entry error.
Across all 25 pairs of screen bounds, the screened aggregate TRIR runs from 3.963 to 3.993, while the flagged share runs from 1.79% to 4.49%.	ehs-osha-analysis/docs/THRESHOLD_ROBUSTNESS.mdResult 1 and Result 2	The analysis shows the aggregate is insensitive to where the bounds sit. It does not show that every flagged filing is wrong or that the screened rate equals a true national rate; the data are self-reported and not a sample of the US workforce.	A screening rule rests on a judgement call. Showing that the headline holds across reasonable settings, and quoting the share flagged with its range, makes the rule easier to defend.
A site of 1-19 employees that reports zero cases cannot rule out a TRIR of 46.96 (exact 95% upper limit), roughly ten times the band's aggregate rate of 4.94. Detecting a 25% difference at 80% power takes 752.9 site-years per arm in the smallest band.	ehs-osha-analysis/docs/DETECTABILITY.mdResults	Poisson is a best case. The same table shows variance-to-mean ratios from 3.4 (1-19) to 241.4 (1000+) for establishment case counts, far above the Poisson value of 1, so real detectability is worse than reported here.	A year with no recorded injuries at a very small site is weak evidence of a low underlying rate. Comparing small sites takes several years pooled, not one year's number.
RMP accident-record rates differ about 14-fold between the highest and lowest product-type peer groups with 100 or more facilities. Petrochemicals had 8.30 accident records per 100 facility-years (95% interval 6.93 to 9.87; 109 facilities). Ethanol and biofuels had 0.58 (0.32 to 0.95; 199 facilities).	ehs-chemical-peers/docs/insights.mdFinding 1	Only accidents that meet EPA's five-year-history criteria (40 CFR 68.42) are on file. A group with a higher count may report more, hold more, or have more covered processes. Nothing here says why a rate is high or low, and none of it shows that a group is less safe.	Chemical plants are better compared with plants of the same product type than with the whole sector. A single plant's accident count is best read against the closest group's interval.
The 212 facilities in the top 10% by accident count account for 80.18% of all accident records, and the top 1% (22 facilities) for 28.57%. Facilities holding 5 million lb or more of regulated chemicals had 10.48 records per 100 facility-years (46.18% had at least one), against 0.92 (8.36%) for facilities under 100,000 lb.	ehs-chemical-peers/docs/insights.mdFinding 2	Inventory and product type overlap (large holders are concentrated in some product types), so these two findings are not independent.	Because records concentrate in a few facilities and the largest inventories, a list ordered by accident history and inventory puts the facilities behind most records first. It is a prioritisation aid and not a finding about any facility.
Ranking the 13 product-type groups with at least 100 facility-years in each half, the rank correlation between their 2010-2016 rate and their 2017-2024 rate is 0.912 (Spearman). Across the 12 chemical families it is 0.769. Rates were lower in the later half in 11 of 13 product types.	ehs-chemical-peers/docs/insights.mdFinding 3	An accident is on file only if it fell in the five years before a submission. Facilities that stopped filing, or filed more than five years apart, may have unobserved accidents counted as accident-free years. The fall in later-period rates may partly reflect this and also reporting changes; the data cannot separate these.	A peer group's standing relative to other groups holds up across periods, so it is usable context year to year, even though the level of the rate moved.
On 63 factual items, the random floor scores 52.1% [45.9, 58.9] while TF-IDF retrieval scores 27.0% [15.4, 40.0] and BM25 retrieval scores 28.6% [17.2, 40.9].	ehs-ai-grounding-eval/results/baselines_breakdown.mdall_factual row	Caveat: as in the repository's retrieval baselines, the governing passage of every question is in the index (the question text is not removed from it), so retrieval is tested for selecting that passage, not for finding evidence in an unseen corpus. results/assistant_eval.md	Keyword retrieval over regulatory text is not enough on its own to answer EHS compliance questions reliably. Answers need to be checked against the clause they cite.
No retriever solves 45 of the 63 factual items, and the 4 retrievers agree closely (Fleiss kappa 0.980).	ehs-ai-grounding-eval/results/item_analysis.mdDifficulty and Agreement across retrievers	The 4 retrievers are not independent raters. TF-IDF, BM25, BM25+ and RRF (which fuses TF-IDF and BM25) share one tokenizer and one 68-document index, so high agreement is expected by construction and says little about item quality.	Questions that every simple method misses are candidates for review and for human-checked answers. Agreement among similar tools should not be read as accuracy.
The grounded assistant answered 41 of 63 factual items. Accuracy was 22.2% with abstentions counted as not correct, and citation precision on answered items was 78.0%. It declined 5 of 5 category-error items. Against the release gate set before the first run (citation precision on answered at least 0.80) it observed 0.780: FAIL.	ehs-ai-grounding-eval/results/assistant_eval.mdHeadline and Release gate	No language model was used. The answer composer is `MockModel`, which is extractive: it quotes the top retrieved corpus passage verbatim and cites its clause.	A pass or fail gate fixed in advance, and a rule to decline when unsure, make an EHS assistant's limits visible before anyone relies on it.
On 13 held-out CSB investigations (42 coded factor pairs), the curated keyword rules reached precision 0.86 and recall 0.71 (30 true positives, 5 false positives, 12 missed), against precision 0.42 and recall 0.24 for the untuned baseline built from ontology label words.	ehs-human-factors-ontology/incidents/suggester_eval.mdHeld-out cohort	Nothing here is an estimate of performance on other incidents; the held-out cohort is the closest thing to one and it is small, so one pair moves a PIF's precision or recall by many points.	Rule-based suggestions can point an analyst to human-factor conditions worth checking in an investigation report. A person still decides which conditions apply.
The four context dimensions are not new; they are adopted from IDHEAS-G (NUREG-2198, 2021). The repository encodes 20 PIFs with 90 alignments to SPAR-H, CREAM, HFACS and the HSE PIF list.	ehs-human-factors-ontology/README.mdAt a glance and The claim, stated up front	What is contributed is the encoding, the crosswalk and the traceability - and those are worth less than a new taxonomy would be if a new taxonomy were warranted.	Putting the same human-factor conditions from several methods in one traceable table shows where different methods use different words for the same condition.
For a five-latent model with three indicators per latent, the test of close fit reaches 80% power at N = 153. Estimating a single path to a standard error of 0.05 needs N = 369. Telling a path of 0.45 from one of 0.30 needs N = 770, and telling -0.25 from -0.20 needs N = 6,884.	ehs-risk-sem/docs/methods-note.mdAbstract	Every number comes from generated data; no real injury data were analysed.	Before ranking or weighting safety factors from a model, check whether the sample at hand can support the comparison. Passing fit cutoffs does not show that.
Analytic standard errors from the two-step estimator are too narrow at every sample size studied, with 95% coverage between 0.76 and 0.86. A percentile bootstrap restored coverage to 0.925 to 0.975 at N = 500.	ehs-risk-sem/docs/methods-note.mdAbstract	Nothing here analyses real injury data. README.md	An interval that looks precise can mislead. Resampling-based intervals give a more honest range for decisions that depend on an estimate.

How this page is built

The data block above is generated by scripts/build_findings_digest.py, which reads each source file from the main branch of its repository and extracts the numbers with a fixed pattern. Limitations are verbatim sentences from the sources. The “why it matters” lines are the only text written for this page and contain no figures. The same rows are embedded as JSON in the page for reuse.

These are descriptive results from public data and from simulations. They are not legal advice and not an OSHA or EPA publication.