Defects — rolling stock faults and restrictions, from the LINX consist feed

What actually breaks

Free text written by depot staff, bucketed by keyword. These are a reading aid, not a taxonomy: a fault can land in more than one bucket ("door heater fault" is both), so the buckets overlap and do not sum to the total. Bars are scaled to the largest bucket.

The feed's own fault codes, decoded from the text

Every defect arrives with a two-character code — 423 of them — and no code book we have any access to. But each code's descriptions are written by people describing the same kind of thing, so the words that are disproportionately common inside a code, measured against how common they are across all 65,000 descriptions, say what that code is about. This is inference, not the operator's definition, and it could be wrong for any given code — but where it can be checked it reads true: W* comes out as toilets, B* as brakes, UW as wheel flats. Codes with fewer than 20 distinct faults are left out rather than characterised from a handful of sentences.

The oddest things anyone has written down

Faults reported at most twice whose wording uses words almost nothing else in the corpus uses. That is a rarity measure over vocabulary, not a judgement about the fault — a unit that has hit a wheelie bin ranks highly because no other report mentions one, which is exactly the intent. Misspellings would otherwise fill this list (a typo is by definition a word seen once), so a rare word within one edit of a word the corpus uses constantly is treated as a typo and counted in the section below instead. That test is imperfect both ways and some garbled reports still get through. Reports naming an individual, and anything involving a death or an injury, are excluded outright — they are not curiosities.

What the free text is actually like

Worth stating plainly on a page built almost entirely out of it, because all three of these cap what any keyword bucketing — including the themes above — can achieve.

Descriptions are hard-truncated at exactly 50 characters, mid-word — a third of every fault we hold is a half-sentence. Some carry a (..R2) marker, which means the sender split the text across two of its own records. The second record is never transmitted. Not as another row, not in another field: DefectDescription is the only element on this feed that ever reaches 50 characters, and it never exceeds it. So the remainder is not something we are failing to match up — it does not arrive at all, and cannot be recovered from here. The marker does not even reliably mean words were lost: some records split after a complete sentence. And a fault written announcementrs will never match a keyword search for announcements.

Longest examples, with what the corpus suggests was meant: internittently → intermittently · investigatiion → investigation · abnnouncements → announcements · investifgation → investigation · announcementrs → announcements · intermitiantly → intermittantly · intermittment → intermittent · extingguisher → extinguisher. Note the suggestion for one of them is itself a misspelling — it appears often enough to count as standard usage here, which is its own small finding.

Carried longest — same fault, most re-reports

One vehicle, one fault, sent again on every subsequent run. This is the closest thing here to a "nobody has fixed this" list. It is not a safety judgement — most of these are comfort or cosmetic items that are legitimately deferred, and the feed gives no severity to sort on.

Most distinct faults, by vehicle

Counted on distinct descriptions, so a unit is not penalised for running often. Bear in mind the feed only covers operators who publish consist data — this is a league table of what we can see, not of the national fleet.

By operator

Read the per-vehicle column, not the total. An operator with 815 vehicles having more faults than one with 45 is arithmetic, not a story. Operator codes here are the feed's own four-digit resource codes, which are not ATOC codes and we have no name mapping for — they are shown as-is rather than guessed at.

Operating restrictions

The other note kind, and the more operationally pointed one — these constrain where or how a set may work rather than describing something broken. They attach to the whole formation, not to one vehicle.

Passing Newton Wamphray (WCML, Lockerbie–Beattock)

Workings booked through here in the last two days whose formation carries at least one recorded fault. Only operators who publish consist data appear — most freight and much of the WCML overnight will never show up here at all.