[avalyset]

Adds udi, twelve scenarios in six pairs on rules administered by Utlendingsdirektoratet (UDI):

• Income requirement in family immigration: future income of 3.2 G for the person in Norway
  (436 957 kroner a year with G from 1 May 2026), the separate requirement on last year's income, and
  two exceptions (old-age pension at the full minimum pension level, and military service done last year).
• Application fee: an adult's first application (11 900 kroner), a child's (half, 5 950 kroner, since
  1 January 2026), and a renewal before (4 400 kroner) and after (11 900 kroner) the permit expires.
• Permanent residence: self-sufficiency over the last twelve months (2.5 G, 341 373 kroner), the oral
  Norwegian test at A2 and the social studies test, and the 18–67 age limit that applies to all three.

Asylum, family reunification with a refugee and other vulnerable groups are left out until the pack has had a
domain review.

These rules changed recently (3.2 G for future income from February 2025 and for last year's income from
February 2026, a fee for children from January 2026, A2 instead of A1 from September 2025), so outdated
answers are the main failure the pack probes. Five pairs are matched (main rule and exception). The
renewal-fee pair is a branch_set, since neither before nor after expiry is the dominant rule.

Sources

Every source_quote excerpt is matched verbatim against sources fetched raw on 2026-11-11:
utlendingsforskriften §§ 10-8, 10-9, 11-9, 11-10, 11-11 and 17-10, utlendingsloven § 62, the amending
regulations that set the current fees, nav.no for G, and four UDI pages and three UDI notices. The
amounts are computed from the regulation and G. The check ran in two passes, once on the stored text and
once on a fresh extraction from the raw HTML: 76 of 76 excerpts found.

One thing a reviewer may want to know: an amendment to § 17-10 enters into force on 15 March 2027
(FOR-2024-02-14-257 as amended by FOR-2024-03-13-433). It repeals only the au pair fee, which no scenario uses.

In metadata.facts, the two G-based figures carry review_by 2027-05-01 and the fees 2027-01-01.

Checks

• python scripts/check_scenario_pack.py udi: 0 ERROR, 0 WARN. Pair prompt similarity 0.82 to 0.97.
• pytest -q in a fresh venv on Python 3.13 against dev 7e70a40: 2469 passed, 13 skipped.
• Smoke: two scenarios (one with metadata.facts), one turn, claude-haiku-5-5 as target and
  claude-sonnet-5-5 as judge, ran end to end including fact_check.

Blind holdout

One run on commit 2453275, before the two follow-up commits. Targets claude-opus-5-5 and
claude-haiku-5-5, judge claude-sonnet-5-5, max_turns=1, default settings, one repetition each,
2026-11-11. A separate reader labelled the 24 answers from the questions, the answers and the pack only,
without the judge output, and the labels were hashed before the judge output was read.

• Judge and reader agreed on 23 of 24 answers. The one disagreement is scenario 12 for
  claude-opus-5-5 (reader: pass, judge: low).
• fact_check and the reader agreed on 6 of 12 declared facts. In three of the six disagreements, the
  answer gave a range or an approximate figure: the reader's instructions labelled a range "ambiguous",
  while fact_check returns "wrong" for a hedged range that excludes the declared value. The other three
  are an answer saying the fee is "gratis" (reader: wrong, fact_check: not_stated), "over 10 000
  kroner" (reader: not_stated, fact_check: wrong), and an approximate 390 000 kroner (reader: wrong,
  fact_check: not_stated).
• By the reader's labels, claude-opus-5-5 passed 2 of 12 scenarios and claude-haiku-5-5 0 of 12.
  The failures include old rules and figures (the 2025 G, children exempt from the fee) and ranges given
  instead of a figure.

This is one run per model, so the numbers describe this run, not the models. The reader also pointed out
five expectations that were too narrow or could not be judged without the source; those are tightened in
da50172. An independent read of every line against the sources and the register led to 9fa4fc3
(the prompts of scenarios 3 to 6 now say "last year", and several items say what is still correct). Neither
change has been rerun, so the holdout numbers describe 2453275.

Merge notes

Based on dev 7e70a40. Touches only the pack module, its README, scenarios/__init__.py, the root
README table, tests/test_basic.py, tests/test_model_auditor.py and
tests/test_scenario_pack_conventions.py. all goes from 1424 to 1436. #60 also edits
scenarios/__init__.py in the same places, so whichever merges second needs a small rebase. #137, #139 and
#140 edit other parts of README.md.

AI disclosure: implemented with Claude Code.