Accuracy
Every tool of this kind says it is accurate. This page says what Tarn.hr actually does to be accurate, what that has caught so far and what it still does not prove.
Where your data lives and who touches it is a different question, answered on the security page.
15
Situations, each with its own written statement of current law
12
Primary sources behind them, each dated the day it was read
8
Out of date answers the eval watches for
Guidance last checked against primary sources on 17th August 2026. That date is one constant in the codebase. It is what the footer of every page prints, it is what this page prints and it is what the build measures against, so there is no version of it that can be older than it looks.
A general AI tool answers an employment law question out of whatever it absorbed in training, which is a mixture of the law as it stands, the law as it used to stand and confident guesswork. It has no way to tell you which it reached for.
Each of the 15 situations Tarn.hr handles carries its own statement of the law that applies to it, written by hand, checked against the primary sources listed below. That statement goes into the request, and the guidance is written from it. It is the difference between asking somebody to recall the law and handing them the law before they start.
This is also why the product refuses to state a pending change in the present tense. Several of the rules that matter most to a UK employer right now are announced rather than in force, and the two are a year apart.
Every statute, set of regulations and statutory code the briefings lean on, with the day its page was last opened and read against the prose. The links go to the publisher rather than to anybody’s summary of it, so you can check the wording yourself.
Read 28th September 2026
Unfair dismissal, the right to be accompanied's statutory backdrop, statutory sick pay's parent provisions, and redundancy as a fair reason for dismissal. Also section 86, the minimum notice bands behind lib/tools/notice-period.ts: a week from one month, a week per complete year, twelve weeks at the top, and nothing at all below a month.
Read 30th August 2026
The 2026 and 2027 changes the scenarios turn on: day-one statutory sick pay, the doubled protective award, sexual harassment as a qualifying disclosure, and the unfair dismissal qualifying period dropping to six months.
Read 30th August 2026
Protected characteristics, harassment, discrimination arising from disability, and the duty to make reasonable adjustments.
Read 28th September 2026
The statutory right to be accompanied at disciplinary and grievance hearings. Section 10 also carries the only fixed period the Acas Code contains, the five working days within which a postponement for an unavailable companion must fall, and section 13 defines a working day as excluding weekends, Christmas Day, Good Friday and bank holidays.
Read 30th August 2026
Whistleblowing: what makes a disclosure qualifying and protected, and detriment and dismissal for having made one.
Read 30th August 2026
TUPE transfers: what transfers, the duty to inform and consult, and dismissals connected with a transfer.
Read 28th September 2026
Disciplinary and grievance procedure, what the Code does and does not apply to, and the tribunal uplift for unreasonably failing to follow it. lib/tools/disciplinary-timetable.ts leans on it for which steps are required, and for the fact that it sets no timescale for any of them beyond promptness and no unreasonable delay.
These change without the law changing, which makes them the ones most likely to go stale while still reading as correct. The April rates pages are the first to be re-read each year.
Read 30th August 2026
Eligibility, the 28 week limit, fit notes and form SSP1.
Read 30th August 2026
The statutory sick pay rate the absence scenario quotes: the lower of £123.25 a week or 80 per cent of average weekly earnings. This page is superseded every April, so it is the first one to re-read.
Read 27th September 2026
Statutory redundancy pay: who qualifies, how it is calculated, and the caps. Also the two figures lib/tools/redundancy-cost.ts is built on, the £751 cap on a week's pay and the £22,530 maximum, both in force from 6th April 2026. Revised every April, so this is the first page to re-read.
Read 28th September 2026
The second source for the statutory notice bands, agreeing with section 86 exactly: one week from a month to two years, a week per year to twelve, then twelve weeks. Also that a contract may give more notice than the statute and may not give less.
Read 30th August 2026
Which measures are in force and which are only announced. The scenarios are careful never to state a pending change in the present tense, and this is the page that settles what is pending.
None of these is a review process that somebody has to remember to run. They are automated checks, they run on every proposed change and a failure stops the change reaching the site.
Wordings that were wrong, or that were right until the law moved, are held in a list. Any of them reappearing in the briefings fails the build. This exists because a correction was once applied to one of the two files carrying the law and never carried across to the other, and nothing noticed.
If a briefing names an Act or a set of Regulations that the source register does not cover, the build fails. It fails in the other direction too, so an entry nothing cites any more cannot sit there looking like diligence.
The date below is a constant, and a test fails once it is more than 182 days old. That is a deliberate time bomb. The law layer is the one part of this product that can become wrong while nobody touches it, so it is the one thing set to fail on a date rather than on a change.
Every phrase the eval treats as wrong is checked against the product's own legal context on every pull request. The first draft had five patterns that would have failed a correct answer, including one matching a sentence that is simply true. A broken instrument sends somebody hunting a bug that is not there.
Tests of the kind above check the code. They cannot check whether the answer the model actually writes is any good. So there is a second set of 8 fixed situations with known right and wrong answers, run against the live model and scored, which is what turns “the guidance is accurate” from a claim into a number.
The strongest check in it is negative. Each case lists phrases that would be actively wrong in UK employment law today, most of them the position before a recent change, which is exactly what a model reaches for when it is answering from memory. A phrase is either present or it is not, so there is nothing to argue about. Inventing a case name is a hard failure for the same reason: a citation nobody can check is the most damaging thing a model can produce, because a wrong case reads exactly like a right one.
Every run so far is below, including the ones that failed. Across the 4 runs, 25 of 32 cases have passed. Last updated 27th September 2026.
Run 1
5 passed, 3 failed
All three failures were faults in the eval rather than in the guidance. Two cases were written so loosely that a fair answer could not pass them, and one demanded a particular phrase rather than the point the phrase was standing for.
Run 20d7e703
7 passed, 1 failed
The eval again. Three cases asserted the same point with three separate lists of accepted wordings, one list had been widened and the other two had not, so a correct answer failed for a reason already fixed elsewhere in the same file. The lists are one shared constant now.
Run 39005925
5 passed, 3 failed
The first run that found a real defect, and it found two instances of it. The closing disclaimer came back wrapped in Markdown emphasis on the two most serious answers in the run, and nothing in the rendering path parses Markdown, so a reader would have seen the asterisks. Fixed in the prompt and repaired at render time. The third failure was the eval.
Run 4ba60106
8 passed, 0 failed
Every case passed.
Most of those failures turned out to be faults in the eval rather than in the guidance, which is worth saying plainly because it cuts both ways. The guidance held up under every case that was fairly put. The instrument has so far been less reliable than the thing it measures, which is normal for a new instrument and is not a reason to trust it less.
The eval catches drift, a model that starts writing differently and the specific confident wrongness the legal briefings exist to prevent. It does not check whether the law was applied correctly to your facts. Nothing automated can. That still needs a person, and on a serious matter that person should be a solicitor.
Tarn.hr is guidance, not legal advice, and it says so in every answer it writes. It tells you when a situation has gone beyond what it can help with rather than pressing on. The disclaimer sets out the limits in full.
There is no accuracy percentage on this page, and there will not be one. A set of 8 cases is too small to carry a figure like that honestly, and any number produced from it would say more about the cases than about the product. What the set is for is catching the thing that has gone wrong, which it has done twice.
If you find guidance on Tarn.hr that is wrong, please tell us. It is read by the person who can fix it, and a case that catches it goes into the set above so it cannot come back.