How a read is made, and what it will not do
The failure mode of a confidence tool is being confidently wrong. Everything below is the machinery for not being — and it is stated so you can check it rather than take it, because a page that just promised carefulness would be the least trustworthy page here.
How a read is made
Five questions in, three categories out, by a rules engine over structured records. No language model is involved at any point.
A ship record holds what an operator knows about a hull — where to put people, what sits above and below the cabins, what the balconies actually look at, what the client will be surprised by. The engine reads that record and the five answers about the booking, and emits a call, the flags that qualify it, and the reasoning behind it. Nothing is generated: every ship-specific sentence in a read is a sentence somebody wrote into a record and signed.
The consequence is that the same booking returns the same answer every time. That is testable, so it is tested: a snapshot of all 93 covered hulls against all 32 combinations of the four client questions — 2,976 reads — is hashed and checked on every build. A change to the reasoning that nobody intended fails the build; a change that was intended has to be regenerated and committed, so it shows up in review as a diff rather than passing unnoticed.
There are exactly three categories — cabin, money, traps — and there is deliberately no fourth. A category nobody has worked up for a ship renders as uncharted rather than disappearing, because a read showing two cards and a read showing three look identical if the missing one is simply not drawn, and they mean opposite things.
What verified means, and what it does not
Three states, and the distance between the second and the third is the one that matters.
- Verified
An operator has checked this against the line's own deck plans and policy pages and signed it.
- Researched
Researched from sources but not signed off by an operator. Check it against what you know before you act on it.
- Uncharted
Not enough reliable evidence to recommend from it. This is not a clean bill of health — nobody has checked it.
Those three words are the same ones the badges use, drawn from the same constant — this page does not get its own definition of them.
Verification is per block, not per ship. Cabin advice comes from somebody who has walked the hull; money and trap policy is researched off the line. Those become true at different moments, so they are markable separately — otherwise nothing could ever be signed without over-claiming the rest. Of the 93 charted hulls, 93 are signed end to end.
Sourcing is recorded per ship rather than per block, which is a real imprecision and is stated on every ship page rather than hidden: the badge says when the ship was last checked, not when that particular block was.
The rules it runs
General operator judgment, held once and applied everywhere, rather than restated on every hull.
The sandwich test
A deck is a quiet candidate when it has cabins directly above it and directly below it and nothing public sharing it. Plenty of cabins on the deck itself proves nothing — the deck above is what wakes people up. It is arithmetic, so it runs rather than being asserted, and it narrows the field without picking the cabin: the operator's own call is allowed to be narrower than the table and often is.
The quiet-cabin guide →What sits above and below, ranked
12 sources of noise, ordered worst first, with what a traveller actually hears from each. The ranking and the wording are operator judgment and hold on any hull, so a ship record says only which of them it has and where — it never restates the judgment.
The full ranking →“Obstructed” is at least 11 things
A steel bulkhead takes the horizon; a lifeboat roof takes the view straight down and leaves the horizon; a public walkway takes privacy and no view at all. Those lead to opposite advice and arrive as the same word on a booking screen, so the taxonomy separates the cause from the effect. It is populated only where the mechanism is actually established — 60 hulls have one on record, and 21 are published as obstructed with no mechanism stated, which renders as exactly that rather than as a guess.
The obstruction guide →Where class rules come from
Computed by comparing the ships, never written down.
A hand-written class rule is a second source of truth that agrees with the ship records until somebody edits one of them. So there is no class content: a field byte-identical across every hull in a class is reported as inherited, a field that varies is reported as a per-ship exception, and both fall out of the records with nothing restated. Add an exception to a ship and the class page reports it on the next build. The page cannot claim a rule is class-wide when the records say otherwise, because it is reading the records.
That also surfaces something no written page would: how much a class actually behaves as one thing. Only 4 of the 26 charted multi-ship classes are uniform. The rest diverge somewhere, which is the case for reading a class page before quoting a sister ship from memory.
What it refuses to do
Each of these costs coverage or convenience. They are the product.
No score, ever
No rating, no percentage, no confidence number anywhere. A score is one more thing to weigh and it implies a precision nobody has — and a number is harder to argue with than a sentence, which is the wrong property for advice that might be wrong.
No inferred cabin facts between sisters
Cabin numbers, obstruction lists and deck bands are not carried from one hull to another even inside a class. Several records say explicitly that a documented pair on one ship does not reproduce on its sister — that is a finding, and inheritance would have erased it.
Uncharted is never a clean bill of health
An empty block, an empty section and an empty availability list all say so in words. They are drawn rather than hidden, because a page with one fewer section reads as a ship with one fewer problem.
Absence is not evidence
"Nobody checked" and "there is nothing there" are different values in the data model, not the same empty field. Reading the first as the second is the error that produced several of the corrections in this project's history.
Coverage, and how it grows
Breadth in the picker must never become breadth in the calls.
The catalog carries 195 ships across 14 lines so an advisor can find their ship. 93 of them have a read, on 3 lines. The other 102 say plainly that nobody has worked them up — they have no page at all, because a page that exists reads as a page that has been checked.
Coverage grows by somebody working a hull up and signing it, one class at a time, and the order is driven by what advisors actually ask for. The directory says which lines are which.