Skip to main content

INVESTIGATE: six two-digit region codes claim to be fylker and are not

Status​

Open — 2026-09-21. Found by ops-dev on urb-agents #1354 as failing relationships tests after the ssb-06913 recovery; the cause is measured and the fix is a decision, not a discovery.

classify_region_code maps ^\d{2}$ to fylke. Six codes in ssb-06913 match that pattern and are not fylker. Measured against Klass 104 (35 codes, 1950→now — none of these) and then against table 06913's own Region labels, which is where these names come from:

21  Svalbard              22  Jan Mayen        23  Kontinentalsokkelen
25  Utlandet              26  Havområder       88  Ikke bosatt i Norge

3,648 rows carry them, uniform across the full 1951–2026 history — a dimension gap, not corruption.

🔵 This is the urb-agents #700 rule — represent, do not cover — in the fylke dimension instead of the kommune one. Those rows keep their region_code; what they must lose is the claim to be a fylke.

Why it is not already fixed​

classify_region_code is used by 22 models. Changing it moves region_kind for every one of them, and ref_region_kind is a closed vocabulary with a relationships test behind it, so new members are a seed change plus a gate change. That is a blast radius, not a line.

⚠️ And the classifier already knows three of these six — as four-digit codes:

svalbard           ^21\d{2}$
jan_mayen          ^22\d{2}$
continental_shelf  ^23\d{2}$

So 2100 classifies correctly and 21 does not, for the same place.

The options, with what each costs​

A — extend the classifier. Bare 21/22/23 join their four-digit forms; 25/26/88 become new vocabulary. Most faithful, and it makes the two-digit and four-digit forms agree. ⚠️ Needs three new ref_region_kind rows with agreed labels, and region_kind changes for any source emitting these codes in all 22 models.

B — narrow fylke only. ^\d{2}$ minus the six, everything else falls to unknown. Smallest change and it stops the false claim. ⚠️ It throws away what SSB actually told us — Utlandet and Svalbard both becoming unknown loses a distinction the upstream makes, and 21 would read unknown while 2100 reads svalbard.

C — fix fylke_nr per model, leave the classifier. ⚠️ Rejected on sight: the same one-off-regex-instead-of-the-macro decision produced the ssb-06913 defect that surfaced this. It would be the third instance in one file.

Recommendation: A, because the names are SSB's own and already measured, and because B makes two spellings of Svalbard disagree.

⚠️ There is no classification to widen to — measured 2026-09-22​

The kommune half of this defect was fixed by widening the Klass window: those twelve codes exist in Klass 131 from 1950 and the ingest simply asked for a range starting in 1960. The fylke half has no such fix. imac measured the deployed host after a re-ingest (urb-agents #1369):

raw.ssb_klass_fylker carries   01..20, 30..56, 99
of the six offending codes     none
ssb_06913 fylke_nr violations  3 648 rows / 6 codes, UNCHANGED by the re-ingest

SSB's population table carries region codes that SSB's own fylke classification does not supply, at any date. So no cadence of re-ingest and no widened window closes this — it has to be decided in the model or the classifier, which is what makes A or B the only live options and removes "wait for better upstream data" as an implicit third.

🔵 ⚠️ A commit titled docs: file the six two-digit codes that claim to be fylker (89a774e) sits in a pinned release range and changes nothing. imac measured it, found no movement, and reported that "its stated effect is not observable here and the pin should not assert it" — correct, because it has no stated effect. A filing whose title names a defect will be read as fixing it; that was a naming error, not a missing prerequisite.

The one thing to check before doing it​

Which other sources emit bare two-digit codes in these ranges? Only ssb-06913 is known, and only because it started delivering. ⚠️ A source whose rows never reached the fact cannot have been seen doing this — the same blind spot that hid the whole defect. Query marts.indicators__* for region_code in ('21','22','23','25','26','88') before assuming the change touches one source.

  • urb-agents #1354 — where ops-dev measured the failing tests
  • urb-agents #700 — represent, do not cover
  • SERVING-A-SOURCE trap 6, "not every region code is a kommune"