Known limitations
Canary AI retrieves passages of federal OSHA standards and interpretation letters and shows them with citations and links to the official text. It is a research aid. It is not legal advice, it makes no compliance or recordability determinations, it has not read any state's standards, its corpus is a dated snapshot, and the optional summary is written by a language model that can be wrong. The text at the link is the authority.
1. Federal OSHA only
Twenty-nine states and territories run their own OSHA-approved plans, often stricter or organized differently. Canary holds none of their standards. When a question names a state, a city in a State Plan state, or a state agency, the answer leads with a jurisdiction note, labels what follows as the federal rule, and links to that state's plan page on osha.gov without characterizing the state standard. Every State Plan state is treated the same way. State standards are often numbered differently from 29 CFR, so the federal citation is where the comparison starts, not where it ends.
The 29 State Plans
Private and public sector
Public sector only
2. No determinations
Canary does not decide whether an injury is recordable, whether a condition is a violation, whether a program is compliant, or what an inspector will do. It shows the criteria and where they come from. Nothing here creates a professional relationship or substitutes for the standard, the state program, or qualified advice.
3. When asked to decide or to write something
| If you ask | What you get |
|---|---|
| "Is this recordable?" | The 29 CFR 1904 decision path as a checklist — work-relatedness, the exceptions, new case, the recording criteria — each step marked by what your question states, linked to eCFR, with the facts still missing. No code path outputs a verdict. |
| "Write me a JHA / incident report" | A note that Canary doesn't write documents, and the passages such a document would cite. |
| "Are we compliant?" "What's the fine?" | What a finding would turn on and where penalties are set (29 CFR 1903.15). No amount, no prediction. |
| "Can they sue?" | A research-aid reminder and the regulatory text that bears on it. |
| Instructions to the tool | No retrieval, no model call. |
The detector is tested in both directions in Part 10 of the accuracy suite: it must not frame an ordinary lookup that merely mentions the 300 log.
4. The corpus is a snapshot
The standards mirror was synchronized to eCFR as of 31 December 2025, with the 2026 amendments to 1910.1200 applied by hand. It does not update itself; check the current text at the link. Known gap: 29 CFR 1926.202 and 1926.203 are absent.
5. Some letters are history
Of the 279 interpretation letters, 25 are ones OSHA has moved to its archive as possibly no longer representing policy. Each carries an archived badge in results and is used by the summary only as historical context.
6. Retrieval can miss; the summary can be wrong
Finding the right passage among 7,220 is a search problem. The engine can rank a neighboring section first, miss a phrasing it hasn't learned, or decline a question it should have answered. The 595-check suite catches the failures seen so far, not all of them. Each answer shows a coverage line (passages, sections, keyword strength). Questions with nothing close in the corpus get no summary. Short questions often match weakly; those still get a summary, on the cheapest model, which says first whether the passages apply. The summary is instructed to use only the retrieved passages and make no determinations. That reduces errors; it does not remove them. The passages are underneath so you can read them yourself.
Before anything is searched, a small model reads the question: what kind it is, which state it names, and how the regulation would phrase it. That reading can be wrong. If it calls an in-scope question off-topic, a fair keyword match overrides it; if it misses that a question asks for a determination, hand-written patterns still catch the clear cases. Basic orientation questions (what OSHA is, what a State Plan is, what a state's own agency is, how 29 CFR is organized, what an interpretation letter is) are answered from a short fact sheet checked against the OSH Act and osha.gov, not from the regulatory text, and say so.
7. The models and their cost
Retrieval is free and always runs; the meaning search uses an open embedding model (bge-base) that runs on the server. Two calls go to Anthropic's API: a small, cheap model reads the question, and the top model tier writes the summary, dropping to a cheaper tier when the day's spending passes 70 percent of its cap and stopping when the cap is reached. Each answer names the tier that wrote it. A question answered once is served from cache.
8. Not in the corpus
Consensus standards incorporated by reference (ANSI, NFPA, ASME, ASTM) are copyrighted and excluded. NIOSH documents, OSHA directives, the Field Operations Manual, letters older than 2016, and all state material are outside the corpus.
9. What the site collects
Nothing that identifies you: no cookies, analytics, accounts or forms. The Testing page publishes the network log showing no third-party host is contacted.
| Thing | What happens |
|---|---|
| Questions you type | Sent to the engine, answered, not stored. Two records are kept without question text or identifier: the model cost per answer, and a cache entry keyed by a one-way hash of the question. Personal identifiers are caught before sending. |
| Who you are | Not read. No IP logging on the public path; a random in-memory token spaces out repeat summary requests. Netlify and Railway keep their own short-lived infrastructure logs. |
| The summary model | Anthropic's API receives the question and the passages; under its commercial terms they are not used for training. |
| Writing to hello@canarysafety.ai is the only way anything about you reaches this project. |