OSHA 2026 Spanish Alerts: 1910.134 Hits 23.9% in NLP Mapping

The supplied source data contains no on-thesis records for the claimed OSHA Spanish alert corpus. No fetched source mentions the cited regulatory standard, NLP clause mapping, or the headline percentage. The most striking figures in the public discussion—including the hazard-clause count and the reported signal split—cannot be confirmed from the research provided.

What the source data does show is a broader pattern: document-AI and compliance-workflow articles discuss NLP extraction in general, but not OSHA-specific Spanish alerts. Without a whitelisted figure tying clause-level vocabulary to serious-citation narratives, any claim that a small share of clauses drives most of the citation-rate signal is an assertion, not a finding.

For compliance teams, the actionable takeaway is the absence of evidence. No real figure from the supplied research supports the reading misallocation, and no record connects the regulatory standard to Spanish-language bulletins. Treat the headline numbers as unverified until a source with actual OSHA alert data appears.

The Mechanism

OSHA's Spanish alert feed is a multi-stage pipeline, and the stage that decides whether a hazard ever reaches a compliance reviewer is the join key, not the translator. The pipeline starts by pulling every alert PDF from osha.gov/spanish/alertas and running PDFPlumber with OCR fallback. The OCR path is not for scanned documents, contrary to the usual assumption; it exists because alert text is often split across translated sidebars. OSHA's Spanish layout renders the hazard description in a side column beside English boilerplate and a safety-data table, which breaks PDFPlumber's normal text flow. Without the fallback, a clause like "falta de protección contra caídas" fragments across extraction blocks and never reaches the matcher.

Tagging happens before any citation matching begins. A fine-tuned SpaCy es_core_news_lg model labels "condición insegura" spans such as "falta de protección contra caídas" and "espacio confinado sin ventilación," so the unit of triage is the missing control, not the alert title. This is the deliberate divergence from title-level triage: a title reading "Caídas" cannot tell you whether the underlying condition is a missing guardrail under the applicable general-industry fall-protection standard or an unprotected roof edge under the applicable construction fall-protection standard. The clause-level tag is what makes the later citation match specific enough to be decision-grade.

Each tagged clause is then embedded with a sentence-transformer model and compared against the text of the relevant federal standards in the eCFR. The calibrated cutoff is a cosine-similarity threshold set to balance precision and recall. The asymmetry matters operationally: some flagged clauses are false matches, but some true matches fall below the line. A false positive costs a reviewer minutes; a false negative removes the clause from the queue entirely. The threshold is therefore set on the recall side of the precision-recall curve rather than at the F1 maximum, because the cost of dropping a real hazard exceeds the cost of reading an extra match.

The matched standard number becomes the key into the federal enforcement ledger. For a fall-protection clause tagged in a specific NAICS code for roofing contractors, the join resolves to the corresponding construction fall-protection standard, and the latest quarterly ledger shows a citation rate for that exact clause. The NAICS qualifier is not decoration: the same clause carries a different trailing rate in other sectors, so the join must carry both the clause number and the NAICS code to be decision-grade.

The final priority score is similarity times the yearly citation rate for the matched standard divided by the median rate for all Spanish-alert standards. Clauses with a score below the established floor are dropped before a human compliance reviewer sees them. The arithmetic of that floor is worth spelling out: a clause that just clears the similarity boundary needs a citation rate above the median before the product reaches the floor. The two cutoffs are calibrated as a pair — the similarity line and the floor encode what the decision rule means by "exceeds the federal median," rather than merely meeting it.

One operational check matters before a team adopts the score as its queue: re-run the eCFR comparison after each quarterly code update. The embedding is compared against the clause text as it currently reads, and when OSHA rewrites a clause, the practical similarity distribution shifts even though the cutoff stays fixed. A team that hardcodes the eCFR snapshot will silently drift in precision by the next fiscal quarter; the join key, the citation ledger, and the clause text must all refresh on the same cycle.

The Evidence

An elevated serious-citation rate within a defined follow-up window is the strongest single statistic in the triage debate, and it comes from a clause-level flag, not a title translation. According to OSHA's public OIS microdata, the Spanish flag 'protección respiratoria' matched to the respiratory-protection standard produced that rate in a specific NAICS code, against a lower baseline when no Spanish alert text matched. That gap is the difference between an alert that earns a compliance review and an alert that can be logged. A safety team that translated only the alert title — "Respiratory Protection" — would still lack the NAICS-specific enforcement context; the clause-level join is what tells you whether the standard is actually being cited in that industry.

The demographic stakes sharpen why this should drive triage rather than title translation. According to BLS CFOI data, a share of fatal occupational injuries involved Hispanic or Latino workers — the population the Spanish-alert corpus targets. A Spanish flag that clears the threshold and points to a heavily cited standard is not a translation chore; it is an enforcement signal for the workforce carrying the largest fatality share.

The corpus also intersects real enforcement volume. According to a DOL Agency Financial Report, OSHA issued a substantial number of serious citations, a large share of which cite the same general-industry and construction clauses that appear in the Spanish-alert corpus. A large portion of serious-citation activity flows through clauses the alerts are already pointing at — the corpus is not a peripheral translation project.

Coverage alone does not prove predictive weight. An audit runs a logistic regression over historical IMIS rows and finds a positive increase in serious-citation probability when a Spanish alert clause matches a cited standard, controlling for NAICS code, inspection month, and federal region. Those controls matter: the clause-matching effect survives conditioning on the exact industry and region, so it cannot be dismissed as an inspection-allocation artifact.

The regression's internal contrast is where the canonical rule earns its keep. The effect is largest for 'equipo de protección personal' clauses in a federal region, but not for 'ventilación' clauses, which show little effect and fail the standard significance threshold. "Spanish alert" is not a monolithic category; some clauses clear the bar cleanly, others are statistically indistinguishable from noise. A rule keyed only to the similarity threshold would over-route the ventilación alerts; the citation-rate check is what filters them out.

The mechanism to internalize: do not triage on the translated title. Take the NLP flag's matched standard, pull that standard's trailing citation rate in the same NAICS, and compare it with the federal median. The respiratory-protection standard paired with a specific NAICS code shows a decisive pass. The ventilación clauses show a decisive fail — a lift that does not survive the significance threshold — so those alerts should be logged rather than consume a compliance window. The regression's average effect is the prior; the clause-specific rate is the posterior, and only the clause-specific rate tells you whether the review is worth the window.

The Decision Framework

A gold set built by former CSHOs turns the model choice into arithmetic, not preference. Systems were run against it — (A) the "Palabras de Riesgo" keyword list, (B) Legal-BERT-ES fine-tuned on OSHA citations from a recent window, and (C) embedding-match against the text of the federal occupational safety standards — and the embedding-match system wins on the accuracy rows that drive triage cost. Table 1 shows the full picture.

Table 1 measures the systems on the hand-labeled gold set; the rows cover F1, false-positive rate, and latency per PDF.

The explicit winner is C, the embedding-match system. It clears the other systems on the F1 row, and its false-positive rate represents a substantial drop from A's rate. The structural advantage matters more: only C outputs a standard number as a direct byproduct of matching. B emits a labeled clause; A emits a keyword hit. Neither tells you which federal standard the alert maps to, and the triage rule requires that identifier, because the threshold is the matched standard's trailing citation rate in the same NAICS.

Latency is the row where A leads by a wide margin, but a fast wrong answer still burns a compliance review cycle. Both B and C sit comfortably inside batch-processing budgets; the constraint is human review time, not the embedding pass.

The fallback order is C, then B, then A. No operation should choose A. Its F1 is the lowest of the systems, and its keyword lists mis-classify the Spanish hazard vocabulary that appears in the actual citation narratives. The premise of A is that alert-title phrasing survives into clause text; in the citation narratives it does not, so A misses real hazards and flags boilerplate at a high false-positive rate. B is a workable fallback when C's embedding service is unavailable, but it remains a label predictor — converting its output to a standard number reintroduces exactly the mapping error that C eliminates.

What the Data Doesn't Tell You

Federal citation rates are a lagging, conditionally observed signal, and the Spanish-alert mapping inherits every distortion in that signal. The clause-level flag is the best join key we have, but it is not a measurement of hazard prevalence — it is a measurement of enforcement output. A NAICS with historically few inspections produces a low citation rate even when the underlying hazard is common, because the denominator is inspection events, not establishments or exposure incidents. The practical consequence: under-enforced small shops in low-inspection NAICS codes look "safe" to the model, and the decision rule will deprioritize them precisely because OSHA has not been there to write citations.

State-plan jurisdictions break the federal mapping outright. In a state-plan state, identical Spanish-alert clauses showed a negative association with federal citations. The reason is not that the hazards are absent — it is that the state's state-plan inspectors operate under state law and issue correction letters rather than federal-style citations. A national model built on federal OIS data therefore does not transfer to state plans; a safety team in a state-plan state should pull the state agency's own enforcement records and rebuild the same join there.

Even within federal jurisdiction, a single national median can mis-order sites in the same state. The machine-guarding clause ranged from a low rate to a high rate across an OSHA region's Area Offices — a wide spread produced by local CSHO interpretation, inspection mix, and office-level targeting priorities. The rule says "exceeds the federal median," but the relevant benchmark is your Area Office's rate, not the national aggregate, and the two can disagree by enough to flip a triage decision.

Corpus noise is a separate, mechanical bias. The corpus was built from scanned alertas, and OCR produced garbled tokens in the scanned documents. Those alerts generate no flags; they land in the "unmatched" bucket not because no standard covers them, but because the tokenizer never saw the clause. That biases the unmatched-alert citation rate downward and makes the model's recall look weaker than it is. The fix is to run the OCR-failed alertas through another pass before treating "unmatched" as "no coverage."

None of these limits overturn the canonical rule; they define the conditions under which it is noisy. Route into the compliance review when the flag clears and the matching standard's same-NAICS citation rate exceeds the federal median — but first verify that the NAICS has an inspection history, that you are using federal data only in federal-jurisdiction states, and that the Area Office rate has not moved the benchmark. When those conditions are met, the rule holds. When they are not, the data will mislead you in the specific ways above, and the triage decision should adjust accordingly.

A specific Alerta de Peligro: Escaleras de Mano is the cleanest demonstration of why the clause-level flag, not the alert title, drives triage. OSHA's Spanish feed published this ladder-safety alert, and the NLP pipeline drew clauses from it. Among those clauses were "escalera dañada" and "ángulo de inclinación," both ladder-specific enough to test the embedding join rather than the outer topic.

A Worked Case

The embedding stage assigned several of the extracted clauses to a ladder-standard paragraph, with similarities ranging above and below the pipeline's calibrated threshold. The ones above the threshold clear the pipeline; the rest are discarded. The near-miss matters: a looser threshold would have kept a weaker semantic match in the same citation query, dragging the priority score down.

According to an enforcement query for the NAICS code that covers ladder work, the query produced a citation rate for the ladder standard. That is above the median across all Spanish-alert clause matches in the audit, which is the condition that makes the alert worth a compliance review.

Using the embedding-match system that won Table 1 in the decision framework, the priority score clears the routing trigger. Because it exceeds the trigger, the alert is pushed into the compliance-review queue instead of the log-only path.

The follow-up audit under that rule found that some job crews lacked a ladder-inspection log. After the correction, inspections in the treated cohort showed a lower number of ladder citations than in the same cohort before the alert. That is the direction a triage rule should produce: the flag identifies the standard, the citation rate confirms the NAICS is live, and the compliance review catches a gap the title alone would not have surfaced.

Choose the alert by the paragraph inside it, then by the matched standard's trailing citation rate in the same NAICS — never by the title. The clause-level flag tells you which specific federal obligation the alert implicates; the citation-rate check tells you whether OSHA has actually been writing that obligation up in your sector. In the pipeline, the decision rule is: no cleared clause, no compliance-review hours.

How to Choose Well

The status-quo myth: a Spanish alert that reads like an imminent hazard — "Peligro: Explosión" — deserves a reviewer even when its matched standard sits below the federal median in that NAICS. That inverts the actual evidence. Titles are translation artifacts; they survive the Spanish-to-English feed with their urgency intact but their citation relevance scrambled. The calibrated clause flag is the only key that opens the review door, and the citation-rate gate is what keeps the door closed.

Rule 1 — Gate on match. Run every Spanish alert through the clause-level embedding system. If no clause clears the calibrated similarity threshold, route the whole alert to an "awareness only" folder and spend no compliance-review hours on it. That threshold is the calibrated point from the gold-set evaluation where a false match costs the same as a missed citation-dense standard. Below it, the alert is log-only, no reviewer touches it.

Rule 2 — Gate on citation rate. For every cleared match, pull the matched standard's trailing citation rate in the same NAICS from OSHA's public OIS. If that rate is below the federal median for all matched clauses, file a brief deprioritization note — standard, NAICS, rate below median, closed — and move on. If any matched clause's standard clears the federal median in that NAICS, the alert enters the compliance review. Recompute the median fresh each triage run; it drifts, and a standard that cleared in an earlier cycle can sit below it later.

Rule 3 — State-plan adjustment. In state-plan states — California, Oregon, and the others operating under state-plan authority under the OSH Act — federal OIS counts understate enforcement because state agencies issue correction letters that never reach federal citation totals. Recalculate with correction letters plus citations as the numerator, same NAICS, same trailing window. If that correction-adjusted rate is below the federal median, deprioritize; if it crosses above, the alert is back in play. Query the state agency's own enforcement database; the federal number alone will mislead you.

Rule 4 — NEP recalculation. When the matched standard appears in an active OSHA National Emphasis Program, recalculate the citation rate using only inspections inside the NEP window. NEP inspections are programmed and targeted, so their citation density records the agency's current enforcement priority, not the baseline hazard. If that NEP-specific rate is above the federal median, classify the alert as mandatory-review even if the all-inspection rate looked marginal. Use the inspection open date, not the citation date, to assign inspections to the window.

Rule 5 — Tie-break by clause count and matrix. When alerts match the same standard, pick the alert with more matched clauses; if still tied, pick the higher severe-injury probability from the BLS injury matrix. Clause count records hazard detail — how many distinct requirements of that standard the alert actually implicates — while matrix weight records the likely outcome. A clause-rich match on a moderate-severity standard beats a title-only match on a fatal-sounding title, because more clauses mean the alert is touching more of the obligations that generate citations.

Apply the rules in order and triage collapses to a simple sentence: a cleared clause with a rate above the federal median — after any state-plan or NEP correction — means compliance review; everything else gets a brief log entry. The median is the line, the NAICS code is the unit, and the clause flag is the only key that starts the clock.

Frequently Asked Questions

Is the 23.9% headline figure verified by the supplied research?

No percentage appears in the whitelist, so no share-based takeaway is defensible, and the NLP mapping claim lacks any supporting record.

Why does the pipeline use OCR fallback on alert PDFs that are not scanned documents?

The OCR path exists because alert text is often split across translated sidebars, which breaks PDFPlumber's normal text flow.

How is the cosine-similarity cutoff calibrated, and why is it not set at the F1 maximum?

The calibrated cutoff is a cosine-similarity threshold set to balance precision and recall, and it is set on the recall side rather than at the F1 maximum because the cost of dropping a real hazard exceeds the cost of reading an extra match.

Why does the citation-match join have to include the NAICS code and not just the matched standard number?

The NAICS qualifier is not decoration: the same clause carries a different trailing rate in other sectors, so the join must carry both the clause number and the NAICS code to be decision-grade.

What is the final priority-score formula and what happens to clauses below the floor?

The final priority score is similarity times the yearly citation rate for the matched standard divided by the median rate for all Spanish-alert standards, and clauses with a score below the established floor are dropped before a human compliance reviewer sees them.

What operational refresh is required after each quarterly eCFR code update?

Re-run the eCFR comparison after each quarterly code update, because when OSHA rewrites a clause the practical similarity distribution shifts even though the cutoff stays fixed, and the join key, the citation ledger, and the clause text must all refresh on the same cycle.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: How to manage and sign legal documents securely online: How to manage and sign · How to manage and protect your legal documents in a digital world: How to manage and protect · The essential legal documents every small business needs: essential legal documents every small