Errors we found, and what we did about them.

Nunavut and the Northwest Territories: the confined-space oxygen answer said the law sets no number. It does — 19.5% to 23%

Asked the minimum oxygen concentration allowed in a confined space, both territories' answers said the regulations do not state one, and pointed to the general duty to test the atmosphere. Both territories' OHS Regulations in fact state the figure plainly: a hazardous confined space must be purged and ventilated so that "an oxygen content of between 19.5% and 23% is assured" (s. 280 in each). The refusals were honest — no wrong number was ever given — but a worker asking got "the law doesn't say" about a value the law says.

Two different causes, one shape. In Nunavut, the word "oxygen" pulled up the chemical exposure-limit table instead — rows for oxygen difluoride, nitrous oxide and ozone, substances whose names contain the word — and the real section never reached the model at all. In the Northwest Territories the right section was retrieved, but it sat last in the list beneath four sections about testing duties, and the answer was written from those. In both cases the system refused rather than guessed, which is the designed behaviour when the retrieved text doesn't carry a number — the defect was in what got retrieved and in what order.

Kept fixed by: a retrieval-side check that, for a question asking the oxygen level of an atmosphere, surfaces the section of that territory's own law stating the percentage — it can only ever promote text the law actually contains, and it stands down for questions about the chemical compounds. Verified live three times in each territory after deployment, and gated by our retrieval-accuracy audit across all twelve jurisdictions with zero regressions elsewhere.

Ontario: we told a worker at four metres that they did not need fall protection

Asked to check a false claim — "I heard fall protection is only needed above 5 metres, do I need it at 4 metres?" — the system corrected the claim, found the right regulations, quoted the right number, and then said this: "since 4 metres is below that threshold of 3 metres, fall protection is not required when working at that height."

Four metres is not below three metres. Fall protection is required. This is the most serious error we have found and published: not a missing answer or a hedge, but a confident instruction, carrying a real citation, to work unprotected at a height where Ontario law requires protection.

What makes it hard to catch is that every individual fact in the answer is correct. The 3 metre trigger is real and the sections cited are the right ones, so none of our grounding checks, which ask whether a stated number appears in the law, had anything to object to. Nor did the citation check, because the citations are genuine. The whole error lives in one word — "below" — and until now nothing in the system read a comparison as a comparison.

Found by: reading, one at a time, every cell of our own 419-question sweep that had not been scored fully correct. It had been sitting in the "incomplete" pile, because the automatic grader saw a correct correction of the false premise at the top of the answer and did not weigh the conclusion at the bottom. Kept fixed by: a check that compares the threshold the answer states against the height the answer states, and strikes the conclusion when the arithmetic contradicts it. It needs no knowledge of fall protection or of any province, and run against all 419 answers it fires on exactly this one. It is deliberately one-directional — it can only ever conclude that a requirement does apply — so if it is ever wrong, it is wrong in the direction of more protection, not less.

Prince Edward Island: a noise limit permitting four hours was read as permitting a full shift

Asked whether 88 dBA over a full shift is compliant, the system quoted PEI's own noise table correctly — "88 dBA is permissible for up to 4 hours of exposure" — and then concluded, in the very next sentence, that 88 dBA "over a full shift… is compliant with the regulations." A full shift at that level is roughly double the noise dose PEI permits.

PEI does not set a single noise threshold. It sets a table of levels against times: 85 dBA for 8 hours, 88 dBA for 4 hours, 91 dBA for 2 hours, and so on. Against a table like that, "is the number under the limit?" is not the compliance test at all. The test is whether the time spent at that level exceeds the time the table allows for it. The question had been phrased around a threshold that does not exist ("the limit is 90 dB"), and the answer applied that borrowed shape to a rule that has a different one.

Found by: replaying the same question through all twelve jurisdictions we cover. Eleven answered it correctly — including Yukon, the other jurisdiction that uses a duration table, which reasoned it exactly right. That is what told us this was a reasoning slip rather than a gap in PEI's law or in our copy of it. Kept fixed by: a check that compares the permitted duration the answer itself quotes against the duration it goes on to bless, and replaces the conclusion when the second is longer than the first. It is grounded in the retrieved regulation and leaves all eleven correct answers untouched.

We were replacing whole answers with "we cannot tell you" — about questions the law does answer

This one did not give a wrong answer. It destroyed right ones.

We have a safety check that refuses to let a number reach you unless it appears in the sections we retrieved, so the system cannot quietly invent a figure. That check was written to replace the entire answer with a notice saying no specific value could be given. When an answer contained one unsupported figure and a great deal that was correct, the correct part went too: the refutation of a false premise, the real citations, the real duties. Four questions in our own testing came out as the same wall of text — Ontario's fatality-reporting deadline, Yukon's 8-hour noise limit, and confined-space oxygen in British Columbia and the Northwest Territories — each telling the reader we could not answer something the law plainly does.

A second, separate cause turned up in the same place. Canadian legislation frequently spells numbers out — "within forty-eight hours", "one metre". Our check only recognised digits. So an answer that correctly rendered the law's own words as "48 hours" was judged to have invented the figure, and was destroyed for being right. There are around a thousand spelled-out values across the twelve corpora we index.

Kept fixed by: the check now removes only the sentences carrying the unsupported figure and leaves the rest of the answer standing, falling back to the full notice only when nothing substantial survives. Nothing about the safety of it changed — an unsupported number still never reaches you. The system now also reads numbers written as words. And we built a new instrument that asks a question we had never asked about our own corrections: when one fires correctly, how much of the answer does it destroy? It runs on every change.

Saskatchewan and the Northwest Territories: four power-line clearances were read from the wrong column of the right table

Asked how close a loaded dump truck could get to a 72,000-volt power line while working on a construction site, the system answered 0.8 metres. The correct distance is 4.6 metres. A water truck at a worksite near a 138,000-volt line was told 1.22 metres, where the same 4.6 metres applies. In the Northwest Territories a mobile crane near a 138,000-volt line was pointed at the column holding 1 metre, and a dump truck on a construction site at the column holding 0.6 metres, where 4.6 metres and 3 metres govern. Every one of those figures understates the clearance, by between three and six times, on a live conductor.

What makes this class hard to see is that none of the numbers is invented. Each is printed verbatim in the table the answer cites, on the correct row for the voltage asked. These tables carry a separate distance column for each kind of work, and the wrong column was read. A reader who checks the citation finds the number exactly where the answer said it would be.

The law draws the line on where the work is happening, not on what the machine is called. Saskatchewan's regulation puts "equipment or powered mobile equipment" in one column and confines the smaller vehicle column to a vehicle operated "on a public road, highway, street, lane or alley"; the Northwest Territories' regulation draws the identical distinction. A dump truck on a construction site is powered mobile equipment. It is not a road vehicle because it has wheels.

One further thing belongs in a public log. The Saskatchewan half was caused by an earlier fix of our own. In August a similar error involving a crane was corrected by listing equipment words — crane, boom, backhoe — that should not take the smaller column. That repaired the crane and left every kind of truck wrong, because it sorted by the name of the machine rather than by the distinction the regulation actually draws. A fix aimed at the wrong thing looks exactly like a fix.

Found by: our own adversarial testing, each answer reproduced against the live system three times before anything was changed. Kept fixed by: the question is now classified on the same statutory test the regulation uses, shared between both jurisdictions so the two cannot drift apart, with a guard that catches a named column or a served distance belonging to a different duty. The same question was then replayed through all twelve jurisdictions and answered correctly in every one, and it has been added permanently to the sweep of questions that must stay safe everywhere.

New Brunswick: a tree-trimming rule was answering fall-protection questions for warehouse workers and miners

Asked at what height a warehouse worker needs a fall-arresting system, and again at what height an employee in an underground mine does, the system answered "more than 3 metres above the nearest safe level" and cited section 374 of the general regulation. Section 374 is the last section of a Part that applies only to arboricultural operations — tree removal and tree maintenance. Its own alternatives are a tree-trimming saddle belt and a rope designed for use in tree maintenance. The answer quoted the height and left those markers out.

Two things follow, both in the unsafe direction. Section 374 says "more than 3 metres"; the provision that actually governs says "3 metres or more". At exactly three metres the answer told a worker no system was needed and the law requires one. More seriously, the governing section sets two further triggers with no minimum height at all: above any surface or object that could cause injury on contact, and above any open top tank, bin, hopper or vat. Anchoring the whole answer to a single height erases both. In a warehouse that is racking over open bins; underground it is broken rock and equipment at two metres.

The root cause is structural and worth stating plainly, because it is not fixed yet. A section's scope can live in a different record from the duty it limits. The sentence that confines this Part to arboriculture is a separate section, so section 374 presents itself as a free-standing rule about heights. We have measured how widespread that pattern is across the whole library and registered it as open work.

Found by: our own adversarial testing, each answer reproduced three times against the live system. Kept fixed by: the system now names the governing section and all three of its triggers, says which of them have no height floor, and states plainly that the three-metre figure came from the arboriculture Part. Underground it additionally names the mine regulation's own duty, which has no height at all — but only when that regulation is among the sources shown. A genuine tree-trimming question is left untouched, and that was verified separately.

New Brunswick: we told arborists the wrong rule, and our own correction text named the wrong people

Not every error is an understated number. Asked how close a trained arborist trimming a tree may bring an uninsulated pole saw to a 138,000-volt line, the system replaced the correct answer with the general clearance — a larger, safer distance, but the wrong governing rule — and described the arborist as "a person who is not a qualified electrical worker". The provision written for exactly that work was set aside.

The same misreading had been printed at the end of every corrected New Brunswick power-line answer we produced: a sentence saying a qualified lineworker may work closer under section 371. Section 371 is part of the same arboriculture Part as the fall-protection provision above — one of its subsections is expressly about climbing or cutting a tree. There is no lineworker in it. That sentence was appended to four separate correct answers in a single test sweep.

Found by: our own adversarial testing, which flagged it as safe-but-wrong rather than dangerous — the class we log anyway, because a citation that names the wrong rule invites exactly the verification a reader ought to do and then passes it. Kept fixed by: arborists and tree-maintenance work now reach the provision written for them, the misleading sentence is gone and a test fails the build if it returns, and both directions — the arborist's answer and the ordinary crane answer — are pinned so neither fix can undo the other.

Ontario: three metres was served as the fall-protection trigger, and on a construction project it is not the lowest one

Asked simply for the "fall protection height trigger", the system answered that more than three metres is the threshold, and went further — that the three-metre figure is "a common standard for fall protection across different types of workplaces in Ontario". It carried a High confidence badge.

On a construction project that is wrong in the direction that matters. Section 26 of O. Reg. 213/91 applies the fall-protection requirements where a worker may be exposed to any of six hazards: falling more than 3 metres; falling more than 1.2 metres where the work area is used as a path for a wheelbarrow or similar equipment; and falling into operating machinery, into water or another liquid, onto a hazardous substance or object, or through an opening in a work surface. The last four have no minimum height at all. A worker told that protection starts at three metres is under-protected by more than half on a wheelbarrow ramp, and without limit over machinery or an opening. Falls are the leading cause of death in construction.

The section that says all this was retrieved and shown to the model — it sat fourth in the list of sources. The reason it was ignored is worth stating plainly: Ontario repeats its three-metre rule across five separate regulations, so the answer saw five sources agreeing and one disagreeing, and reported the agreement. The same question asked of the other ten jurisdictions came back correct in every one, because each of them states its height and its exceptions in a single section where they cannot be separated.

Found by: a reader of this site, who noticed the answer read wrongly and said so. Not by any of our own instruments — none of them had ever typed the plainest question in the trade. Kept fixed by: a guard that restores the full list of triggers whenever a single height is generalised past a provision that lists several, plus a recorded regression test and a permanent entry in the sweep of questions that must stay safe in every jurisdiction.

A forklift licence that does not exist was said to be legally required

Asked whether a government-issued operator's licence is legally required to drive a forklift, the system answered "Yes" — and then, in the following sentence, described the duty the law actually imposes: adequate instruction, demonstrated competency, and authorization by the employer. Newfoundland and Labrador issues no forklift licence. Neither does New Brunswick, where the same question produced the same answer.

An answer whose opening contradicts its own next sentence is a particular kind of failure. A reader who takes the first line at face value goes looking for a licence that cannot be obtained, and may conclude that a properly trained and authorized operator is not permitted to work. The correct answer was present the whole time; it was the verdict in front of it that was invented.

Found by: our own adversarial testing, confirmed by three separate runs in each of the two provinces before anything was changed. Kept fixed by: a guard that refuses to assert a credential requirement unless the cited text actually imposes one — and which stands down where a licence genuinely is required, so it cannot erase a real one.

Manitoba: the general workplace fall distance was applied underground

Asked what fall distance obliges a worker at a Manitoba mine to use a fall-arrest system, the system answered three metres. That is the general workplace figure. At a mine, section 4.7 of M.R. 212/2011 requires fall arrest at a vertical distance of 1.5 metres or more — half the distance the answer gave.

Both numbers are real and both are correctly transcribed from Manitoba law. The error is which regime governs, and it runs toward less protection for the workers in the more hazardous setting.

Found by: our own deterministic sweep of dangerous-value families, run across every jurisdiction rather than the one where the problem was first suspected. Kept fixed by: recorded ground truth for this question, so an answer that reaches for the general figure in a mine context fails the check instead of passing quietly.

New Brunswick: a crane clearance was taken from a rule written for qualified line workers

Asked how far a mobile crane must stay from a 138,000-volt overhead line in New Brunswick, the system answered 1.8 metres, from section 371. That section governs qualified electrical line workers doing work on the line itself. The rule for everyone else, including a crane, is section 289(1), which requires 5.2 metres at that voltage.

The answer was short by 3.4 metres, and it was short because it read a provision written for the one group of people specifically trained and equipped to work close to live conductors, and applied it to a machine.

Found by: our own deterministic sweep across every jurisdiction that publishes an approach-distance table. Kept fixed by: recorded ground truth per province for this question, so selecting the wrong duty-holder's rule fails the check.

New Brunswick: the lock-out threshold was quoted from the rule for manholes and tunnels

Asked at what voltage electrical equipment must be locked out and tested for absence of voltage in New Brunswick, the system answered "in excess of 750 volts", citing section 296. Section 296 is the rule for work in manholes and tunnels. The general threshold is far lower: section 286 sets it at 30 volts AC or 60 volts DC.

Quoting 750 volts as the general floor leaves ordinary line-voltage equipment — the equipment most workers actually touch — appearing to fall outside a lock-out requirement that in fact covers it.

Found by: our own adversarial testing. Kept fixed by: recorded ground truth for this question, and a check that an answer citing a specifically-scoped section (a manhole, a tunnel, a mine) says so rather than presenting it as the general rule.

Yukon: the asbestos restricted-area threshold was stated four times too high

Asked at what airborne asbestos level part of a work site must be treated as a restricted area in Yukon, the system answered that it begins when asbestos exceeds the occupational exposure limit. The Workplace Health Regulations define a restricted area as one where there is a reasonable potential for worker exposure at or above 25% of the 8-hour Occupational Exposure Limit — a quarter of the level the answer gave.

The controls that follow from that designation are the substantive ones: limited access, protective clothing, decontamination chambers, no eating or drinking in the area. Setting the trigger four times too high delays all of them, for a carcinogen whose harm does not announce itself.

Found by: our own adversarial testing during Yukon's pre-launch verification. Kept fixed by: recorded ground truth for this question, so a future answer that reads the exposure limit itself as the restricted-area trigger fails the check.

Northwest Territories: a high-risk site was staffed from the moderate-risk column

Asked how many first-aid attendants a high-risk Northwest Territories work site with 45 workers on shift requires, the system read Schedule H across the correct row but the wrong column, returning the moderate and low-risk requirement. The qualification level was understated and a required emergency medical technician was dropped from the answer entirely.

This is the subtlest shape of table error and the one we now treat as its own class: the headcount is right, the row is right, and the number looks plausible — so nothing about the answer invites a second look. What is wrong is the column, chosen by a variable the reader supplied and the answer quietly ignored.

Found by: our own probing of every dense lookup table in the corpus, each tested live before any guard was written. Kept fixed by: for tables selected by more than one variable, the system now names the table and every selection key rather than transcribing a cell — pointing the reader at the row and column they need instead of guessing which one they meant.

Yukon: a Class A construction site was given Class C first-aid coverage

Asked what first-aid coverage a Class A building construction site with 15 workers per shift and a 30-minute drive to hospital requires in Yukon, the system stated values from the Class C table while citing the Class A one. The citation and the numbers came from different places, and the numbers were the lighter set: the site was under-kitted and under-qualified in the answer.

A citation that does not match the figures beside it is worse than no citation, because it invites exactly the verification a reader ought to do and then passes it.

Found by: our own probing of the dense first-aid tables during Yukon's pre-launch verification. Kept fixed by: the same rule as the Northwest Territories entry above — multi-key tables are pointed at, with every selection key named, rather than having a single cell read out of them.

A diesel exhaust reading was served as the air a miner may breathe

Asked what the maximum carbon monoxide concentration is for a worker in an underground mine in Newfoundland and Labrador, the system answered with the limit for undiluted exhaust gas from a diesel engine — 300 parts per million for new equipment, 600 for existing. Those figures are real and correctly transcribed. They are a tailpipe specification. The limit for the air a person actually breathes is roughly 25 ppm, so the answer overstated it by about twelve times, to someone underground.

It happened in one run out of three, which is why nothing had caught it. A guard built for exactly this confusion was standing down: it checked whether the number appeared in an exhaust context and whether it appeared in a breathing-air context, but it made both checks across all the retrieved sections independently. So a figure correctly identified as exhaust in one section was cancelled whenever some unrelated section happened to carry the same number near the word "worker". Which sections are retrieved varies between identical runs, and so did whether the guard held.

Two changes. The stand-down now only applies where a section presents the number in breathing-air terms without exhaust framing of its own. And where the answer itself calls a figure the undiluted-exhaust cap while the question asks what a worker may breathe, that is now enough on its own — the system withdraws the number and points to the occupational exposure limit instead. Re-tested four times running: no run served the cap.

Found by: our own cross-jurisdiction testing, while building ground truth for carbon monoxide in all twelve jurisdictions. Kept fixed by: a twelve-check regression test, including three cases that must be left alone — a genuine worker exposure limit, a question explicitly asking about exhaust, and an answer that already draws the distinction.

Prince Edward Island: a crane clearance was read from the wrong voltage band

Asked how far a crane must stay from an energized 138,000-volt overhead power line in Prince Edward Island, the system answered 3,600 mm (12 feet). The table it was reading from, section 36.20, sets 3,600 mm for lines of 751 to 100,000 volts and 5,200 mm for 100,001 to 250,000 volts. A 138,000-volt line is in the second band. The answer was 1.6 metres short, in the direction that kills.

New Brunswick publishes an almost identical table and answered the same question correctly, which is the part worth stating: this was not a defect peculiar to one province's data. Selecting a row in a banded table is unguarded generally, and New Brunswick got it right that day rather than being protected.

Found by: our own deterministic sweep of known dangerous-value families across every jurisdiction, checked line-for-line against each province's own regulation. Kept fixed by: recorded ground truth for all twelve jurisdictions on this question, so any future answer that picks the wrong band fails the check rather than passing unnoticed.

Ontario: shaft timber sizes were handed to a question about a trench

Asked what size uprights and walers a 4.5 metre trench in soft sandy soil requires, the system answered with section 281 — which governs shafts — and gave 254 mm by 254 mm. Real law, correctly quoted, for a different structure. Someone about to enter a trench has no way to tell.

We already decline to read a cell out of a trench-shoring table on anyone's behalf, for the reasons set out in the Yukon entry below. That safeguard was being bypassed two ways at once. It only recognised a member size written with an abbreviated unit — "254 mm" — so writing "254 millimetres" in full walked straight past it, in every jurisdiction, not just Ontario. And it stood down entirely whenever the trench table itself was not among the sections retrieved, which is exactly what happens when some other provision supplies the numbers.

Both are closed. A member size now counts however the unit is written, and a trench question that produces any member dimension is redirected to the governing table whatever section supplied it. A question genuinely about a shaft, a scaffold, formwork, falsework, a caisson or a tunnel keeps its dimensions — those are real requirements of their own, and an early version of this fix wrongly redirected them before that was caught.

Found by: our own testing, while recording the correct answer for this question in all twelve jurisdictions. Kept fixed by: a thirteen-check regression test covering both unit notations, the wrong-document case, and four cases that must be left untouched.

Alberta: the farm and ranch exemption was missing from the definition that decides who the Act covers

Section 1 of Alberta's Occupational Health and Safety Act defines "occupation", and inside that definition sits the farm and ranch exemption — the provision that decides whether the Act applies to you at all. Our copy of that section broke off mid-sentence. It read that wages "are paid only to the following persons for the performance of farming or ranching work:" and then simply carried on to the next clause. The four persons the Act actually names were not there: shareholders of a family-held farming corporation, family members of such a shareholder, family members of a sole proprietor, and family members of a partner where all the partners are of the same family. Anyone asking whether Alberta OHS law covers their family farm was being answered from a definition with its operative list cut out.

The cause was not a typo. Our Alberta parser sorts every paragraph in the official HTML into one of four buckets — body text, heading, structure, discard — and anything it did not recognise it ignored without comment. The four missing items were tagged with a paragraph style the parser had never been told about, so they fell through and disappeared silently: no error, no warning, no count. The real fix is not the four missing lines; it is that an unrecognised paragraph style is now a fatal error. The parse stops and names what it found, because silence is the wrong response to discarding something that might be law. On its very first run the new guard caught a second style that had also been dropped without anyone knowing — twelve amendment-history footers. Those turned out to be correct to drop, but that is now a decision on the record with its evidence beside it, rather than an accident.

One record changed. It was re-indexed and republished the same day, so the correction is live, not merely logged. Alberta's Code and Regulation were unaffected.

Found by: a new check that compares our parsed text against a completely independent reading of the same source, in order, word by word. This matters, because it explains why nothing caught it sooner: when a whole class of text is dropped before anything is compared, it goes missing from both sides of the comparison at once, and every word-count audit balances perfectly. Kept fixed by: the fatal-error guard, locked by a regression test — the next time the publisher introduces a new paragraph style, our build breaks instead of quietly thinning the Act.

Twenty-four sections were in our index but unreachable — the count went up without adding any law

When we index a regulation, each section should become its own entry, so that asking about it — or citing it — finds it. Twenty-four sections had instead been absorbed into the entry for the section next to them. The words were all there and correct; every word-for-word audit we run passed. But those sections had no address of their own, so a question about them could not reliably reach them.

They are now separated out, each with its own section number, citation and heading. Six were spot-checked by asking a question only that section answers — New Brunswick's rule for working alone with a chain saw, Nova Scotia's arborist fall-arrest rule, Prince Edward Island's marking of vehicles carrying explosives, Ontario's single-toilet facility requirement and its surface diamond drilling training programs, and British Columbia's general mine construction rule. Each now comes back first, under its own citation.

This is why our published section count moved from 12,775 to 12,799. No new law was added. Those twenty-four sections were always in the corpus; they simply could not be counted or found, because they were filed under a neighbour's number. The higher figure is a correction, not an expansion.

Found by: our own structural audit, which looks for sections swallowed by their neighbours — a fault that word-accuracy checks cannot see, because no word is wrong. Kept fixed by: that audit's debt list is now empty and may only shrink, so any recurrence fails our build.

Trench shoring sizes: one table, eight jurisdictions, one systematic misreading

Nearly every Canadian jurisdiction publishes the same kind of table for timber trench shoring — a grid you read by soil classification, by a trench depth band, and, for the braces, by how wide the trench is. After finding that Yukon's version was being read from the wrong depth row, we put one identical question to every jurisdiction we cover — "for a trench 4.5 metres deep in soft sandy soil, what minimum size of uprights and walers is required?" — and repeated it three times each.

Every jurisdiction that publishes this table answered it wrongly, and every error ran the same direction: toward a weaker structure than the ground calls for. Ontario and Saskatchewan both treated soft sandy soil as their hardest soil class, which carries the lightest members — Ontario's own regulation defines that class as "hard, very dense" ground with no water seepage. The Northwest Territories read across into the wrong column, reporting a brace dimension as the upright dimension. Manitoba answered a sandy trench from its category for soft cohesive soils, when sand is cohesionless. Alberta gave sizes on one run and said the law was silent on the next two. British Columbia said the law was silent every time, because its copy of the table is filed in our index under a neighbouring section.

Undersized shoring is how trenches collapse on people, so this was fixed the same day for all of them. None of these answers now states a member size. Each one names that jurisdiction's own table, spells out all three things you must select on, carries the specific misreading we measured there — Ontario's quotes the soil-classification section so you can classify the ground before picking a row; Manitoba's says plainly that sand is cohesionless — and points out that every one of these regulations also permits a support structure designed by a professional engineer.

Found by: our own cross-jurisdiction testing — a fault confirmed in one province is treated as a hypothesis about all of them. Kept fixed by: a 109-check regression test covering both directions in all six jurisdictions, including a probe that the first version of this fix failed — it worked in testing but not in production, and Saskatchewan stayed wrong for a further hour until we caught it.

Yukon: trench shoring sizes were read from the wrong depth row — we now point at the table instead

Yukon sets minimum sizes for trench shoring members in Table 10-9 of the Workplace Health and Safety Regulations. It is a hard table to read: it keys on soil type, on a trench depth band, and — for cross braces only — on the width of the trench, and it prints metric and imperial grids one after the other. Testing it with six questions, three answers took their figures from the wrong depth band. Asked about a 4.5 metre trench, the system answered from the 4.6–6 metre row; asked about a 4 metre trench, it stated that "your trench depth falls within the range of 4.6 to 6 metres", which is simply untrue. In one case it argued with itself and recommended a larger member "for added safety" — a margin the regulation does not contain. Undersized shoring is a cave-in risk, so this was fixed the same day.

The system no longer states any member size from this table. It names the table, explains all three things you must select on so you can find your own row, gives the rules the regulation states in words (the number of cross braces by depth, the 1.2 m brace spacing, the 0.6 m top-brace height, the 15-degree limit on upright spread, and that every dimension is a minimum), notes that the regulation also permits a support structure designed by a professional engineer, and links the official text.

One thing worth saying plainly, because it is the opposite of what we first thought: this table is in our corpus, in full. Our first reading of the failure was that the table was missing and the figures were being invented — a conclusion reached because our search used the wrong notation for the way the table writes its cells, and was then "confirmed" against a version of the regulation our indexing never used. A safeguard written on that false premise would have told readers the table was not reproduced here. It was withdrawn before it ever ran. We hold the table; we decline to read a cell out of it on your behalf.

Found by: our own adversarial testing, in a review of every dense lookup table in every jurisdiction we cover. Kept fixed by: a 36-check regression test covering both directions — the redirect must fire on a stated member size, and must leave alone the plain-language rules in the same section (brace spacing, excavation depth, upright spread, scaffold plank thickness), each confirmed three times against the live system.

Three BC documents were missing their "official text" links; NT's currency line was blank

Regenerating our source-coverage manifest revealed that the Workers Compensation Act, the Mines Act, and the Health, Safety and Reclamation Code for Mines (all British Columbia, live since Aug 11) resolved to no official-source link, and the Northwest Territories "current to" line rendered empty. The indexed law text itself was correct throughout — these were provenance-display gaps. Fixed: all three BC links added (each verified live against the official page by title), NT's verification dates and notes added.

Kept fixed by: the source-link coverage gate now includes every live document (68), so a document with a missing or wrong official link fails our build.

Yukon (pre-launch): trick phrasing produced a wrong committee conclusion

During pre-launch testing of the Yukon corpus (not yet publicly served), a loaded question — "since a committee is only required at 50 workers, a 30-worker site needs nothing, correct?" — made the system correctly reject the fake "50" but then wrongly conclude a 30-worker site needs no committee. The Act requires a committee at 20 or more workers (WSCA s. 36). Neutral phrasings of the same question were answered correctly. Fixed with a grounded correction that fires only on this failure shape; the same review also hardened a related Alberta safeguard against a potential misfire on correct answers (no live occurrence found).

Kept fixed by: a dedicated regression test (fires on the bug, leaves eleven correct variants untouched).

New Brunswick: power-line clearance for lines above 750 V answered with the below-750 V distance

Asked for the crane approach distance near a line above 750 V, the system returned 900 mm — the distance for lines up to 750 V. The correct answer is 3.6 m, roughly four times farther. The band-lookup safeguard didn't understand the "above/over" modifier and picked the wrong voltage row. Fixed the same day: the lookup now detects the modifier and steps to the correct higher band, confirmed three times against the live system.

Found by: our own adversarial testing. Kept fixed by: a 37-case unit test on the voltage-band lookup.

Nova Scotia: anchorage strength (22 kN) answered as the force allowed on a worker's body (8 kN)

Asked the maximum arresting force a fall-arrest system may apply to a worker, the system answered 22 kN — that is the strength an anchorage must withstand (s. 21.15), not the limit on the worker's body, which is 8 kN (s. 21.14(2)(b)). The retrieval step was only surfacing the anchorage sections for that phrasing, so the model never saw the correct number. Fixed with a grounded correction carrying the corpus-verified 8 kN value, confirmed three times against the live system.

Found by: our own adversarial testing (a fall-arrest question bank added that week caught it immediately). Kept fixed by: a 9-case regression test covering both the 8 kN and 22 kN questions, so neither direction can cross-wire again.

Power-line safeguard family hardened across NB / SK / AB

A stress test of the power-line-distance safeguards found their trigger phrases brittle — some realistic phrasings (including simple plurals like "power lines") would not activate the safeguard. Every phrasing tested had still been answered correctly live, so no user-facing error existed — but a safeguard that can be phrased around is a latent risk. All three provinces now share one hardened, plural-safe trigger, locked by a 12-phrasing reachability test.

Saskatchewan: approach distance for a between-bands voltage understated (3 m instead of 4.6 m)

Table 19 lists minimum approach distances by voltage band. For a voltage falling between listed rows (50 kV), the system answered with the lower band's 3 m instead of rounding up to the safer 4.6 m band. Fixed with a structured table lookup that always rounds up to the more protective band; re-verified twice against the live system.

Found by: our own class-check after a similar bug elsewhere — one province's bug becomes every province's test.

Manitoba's advisory Codes of Practice removed from the corpus

Manitoba's seven Codes of Practice are guidance, not binding law. They were the one exception to our binding-law-only rule, and on this date they were removed from the live corpus so every jurisdiction is held to the same standard: the Act and its binding regulations, verbatim, nothing else. Any future inclusion of advisory material would be a deliberate, project-wide decision — never a quiet per-province one.

PEI: a paraphrased Act replaced with the verbatim text

An early indexing pass of the PEI Occupational Health and Safety Act had produced paraphrased section text rather than the exact statutory wording. Caught during verification and fully re-indexed from the official source PDF — all 48 sections, word-for-word, confirmed. This incident is why "verbatim" is now a tested property, not a promise: the corpus is checked word-for-word against the official consolidation, and the check is part of the build.