Part One. The answer is not a verdict. Chapter two.
When an Exception Changes the Answer
Contents of Canons
…have the right to compensation by the operating air carrier in accordance with Article 7, unless: Regulation (EC) No 261/2004, Article 5(1)(c)
A flight inside the European Union, twelve hundred kilometres, is cancelled. The passenger had a confirmed booking and learned of the cancellation the day before departure. Airport notices often summarise passengers’ compensation rights: €250 for shorter flights, with higher amounts for longer distances.
Is the passenger owed two hundred and fifty euros? Take a moment. Nearly everyone says yes, and nearly everyone is right.
The machine agrees. The model contains executable rules linked to every article of Regulation (EC) No 261/2004, drawn from the official English text. This does not mean that every possible interpretation or consequence has been formalised. In particular, the model does not incorporate the subsequent case law. Asked about this passenger, the machine says: established, two hundred and fifty euros. Alongside the answer, unasked, it lists the duties the cancellation created for the carrier at that same moment: offer a refund or re-routing, offer meals and phone calls, hand over a written notice of the passenger’s rights, explain the alternative transport.
That is the easy case. The chapter is about the word that follows the right in the regulation’s own sentence. Unless.
The carrier objects
The carrier writes back: we informed you two weeks in advance. The regulation does say that a passenger informed at least two weeks before departure gets no compensation.
Change one thing in the case. The carrier states that the passenger was informed on the twelfth of August at ten in the morning, for a departure on the twenty-sixth at ten: fourteen days to the hour. It has not yet produced proof. Is that statement enough to take the money away?
Most people say no. The machine says: established, two hundred and fifty euros.
This is not a slip, and it is the first place where you should look for the person who wrote the model. The regulation adds, in a separate paragraph, that the burden of proof concerning whether and when the passenger was informed rests with the carrier. That sentence was written into the model as a presumption: until the carrier has proved the notification, the passenger counts as not informed, and an exception that begins with “they are informed” cannot start. A date in the file is a claim. The carrier’s proof is a separate fact, and it is not there.
Add it. The carrier proved the notification. Ask again.
Refuted: under the rules counted, the right does not exist. And the amount: not established, because with no right there is nothing to price. Two answers to two questions. The compensation that was owed a minute ago is, on this reading, not owed.
What “refuted” is made of
Look inside that answer: it is a different kind of thing from “not established” in the first chapter. The machine did not fail to find a right. It found two rules that both fit the facts. One is the general rule: in case of cancellation the passenger has the right to compensation. The other is the exception: not if informed at least two weeks before. Both fired, and their conclusions are opposite, and the answer carries a record of what happened between them. The general rule was defeated by the exception, and the record names the ground: a rule of precedence, declared by name, which says that the exception is the more specific provision and prevails.
In the first chapter two arguments of equal standing collided, nothing in Gellius ranked them, and the machine handed back the collision. Here the text ranks them: “unless” subordinates the right to its exceptions. So one rule wins and the machine returns the winner. It does not erase the loser. The general rule is still in the answer, marked as defeated and by what, so that a reader who wants to argue that the exception should not apply knows which sentence to argue with.
Here is the second place to look for the author. That “unless” means “the exception beats the right” is a reading. It is a very natural one, and few lawyers would read it otherwise, but it was declared by a person, who wrote down which rule has precedence, and it is visible in the answer rather than buried in the order in which the rules happened to be written. Had the author declared nothing, the machine would have returned “established and refuted”, exactly as in the first chapter, and the passenger would have been told that the text itself does not settle the question. The ranking is the author’s claim that the text does settle it, and you may dispute the claim.
One hour
Move the notification by one hour. The twelfth of August at eleven, not ten; departure still on the twenty-sixth at ten; the carrier’s proof still in the file. Two weeks less one hour. Who gets the money now?
Established. Two hundred and fifty euros, back on the table.
The regulation says “at least two weeks before the scheduled time of departure”. The model counts hours, because the text gives a scheduled time, not a date: fourteen days are three hundred and thirty-six hours, and the carrier managed three hundred and thirty-five. A person could count calendar days instead and call the twelfth “two weeks before” the twenty-sixth. That is the third choice, and the least visible of the three: it sits in the rule that compares two moments, and nothing in the answer flags it as a choice. You have to know to look.
| The case | Right to compensation |
|---|---|
| Cancelled; passenger told the day before | established, 250 EUR |
| Told two weeks before; carrier has not proved it | established, 250 EUR |
| Told two weeks before; carrier proved it | refuted; no amount |
| Told two weeks less one hour before; proved | established, 250 EUR |
Four rows, one passenger, one flight. The money appears, disappears and reappears. The output identifies the operative rules and the provisions supporting them. It does not yet expose every interpretive choice embedded in those rules, including the decision to measure this deadline in hours. Both implementations agreed on all four cases.
Where the model stops
Two of the carrier’s other defences are not decided here. If the carrier pleads extraordinary circumstances, the machine does not decide whether a storm qualifies; it says that a court must, and keeps the right standing until the court answers. And if the flight was delayed rather than cancelled, the machine says compensation is not established, although European courts have awarded it for long delays since 2009: the model was built from the text alone, and the court decisions are not in it. That is a true statement about the model and a misleading one about the law as applied, and Part Two is about how such gaps are measured and declared.
What is left to argue about
A good lawyer keeps a list of the unlesses that could undo a claim. The machine keeps the same list, written down, ranked, and attached to every answer. That is not a small thing, but it is not a verdict either. Here is what it leaves you.
If you were the passenger in the third row, you lost, and you can see the sentence you lost to. You can also see the two choices that carried it there: that the carrier’s proof, not the date, is what counts, and that the exception outranks the right. Both are readings of the text, and both are named in the answer.
If you were the carrier in the second row, you lost with a true date in your hands, and you can see why: the model treats your notification as a claim until you prove it. That is the regulation’s own allocation of the burden, read by a person into the machine. If you think the reading is wrong, you know what to say.
If you were either of them in the fourth row, you are one hour from a different answer, and the choice that put the line there is the one the answer does not flag. A calendar would have moved it. Nobody has proved that hours are the right way to read “two weeks”; the model takes it as an assumption.
A third, independent program re-checked the rule steps of both machines and, for every defeat recorded in these answers, that the declared ranking really does make the winner the stronger rule. What it does and does not vouch for is in the experimental notes; the positive answers here are proved more strongly than the defeats, and the defeats more strongly than the silences.
None of this decides the case. What it does is turn “unless” from a footnote into a rule that beats another rule, with the ranking declared and the defeat recorded, so that the passenger and the carrier argue about the same sentence and know whose reading they are arguing with.