Skip to main content

Plan: Ingest Bufdir's Barnevern kommunemonitor

Ingests Bufdir's child-welfare (barnevern) kommune monitor as a new Atlas source, reusing the bufdir-barnefattigdom ZIP-bulk-export pattern — adapted, not copied, because the two monitors' workbooks are shaped differently (verified below).

IMPLEMENTATION RULES: Before implementing this plan, read and follow:

Status: Completed​

Completed: 2026-10-01

Goal: Add bufdir-barnevern as a served Atlas source, plugging the barnevern axis that Report #2 (Child Welfare / Vulnerability Composite) is missing today.

Last Updated: 2026-10-01

Investigation: INVESTIGATE-new-norwegian-public-sources.md §Tier 1 #1 ([Q1]–[Q3])

Prerequisites: None. The "Phase 0 schema prep" this investigation originally pointed to was a phantom (no provider enum, no dim_period/dim_indicator tables exist) — see the investigation's 2026-10-01 correction. bufdir is already a valid publishers.yaml provider (used by bufdir-barnefattigdom), so no new publisher metadata is needed either.

Pattern reference: atlas-data/ingest/src/sources/bufdir-barnefattigdom/ (index.ts, parse.ts, fetch_retry.ts). Adapt the shape of this module, not the literal code — Phase 1 below records concrete, verified differences between the two monitors' ZIPs that make a byte-for-byte copy wrong.

Problem Summary​

Atlas ingests Bufdir's Barnefattigdom (child-poverty) kommunemonitor but not its sibling Barnevern (child-welfare) kommunemonitor, even though both are published the same way, from the same CMS, and the investigation named this as the #1 gap-fill candidate five months ago. Today Report #2 (Child Welfare / Vulnerability Composite) has child low-income, persistent low-income, single-parent share, bullying, and overcrowded housing — but no barnevern axis at all: no measure of how many children are actually in contact with, or removed into, the child-welfare system. Bufdir's Barnevern monitor is the canonical kommune-level measurement for that axis.

Phase 1: Confirm the real shape — DONE (verified 2026-10-01, recorded here so it isn't re-derived)​

The investigation's own description of this source ("machine-readable JSON... open data API") was wrong — bufdir-barnefattigdom's own history (INVESTIGATE-bufdir-barnefattigdom-zip-ingest.md, INVESTIGATE-bufdir-upstream-restructure.md) already corrected that to "ZIP bulk export of Indikator_*.xlsx workbooks, discovered by scraping the monitor landing page." This phase checked whether Barnevern follows the same mechanism and the same internal shape. It follows the mechanism; it does not follow the same internal shape.

Tasks​

  • 1.1 Fetch the live monitor page (https://www.bufdir.no/statistikk-og-analyse/monitor/barnevern/) and confirm it publishes a bulk ZIP the same way Barnefattigdom does. Confirmed — <title>Barnevern kommunemonitor | Bufdir</title>, same Strapi/Azure Container Apps CMS, canonical download link: https://ca-statistikk-strapi-prod.whitesea-89be7839.norwayeast.azurecontainerapps.io/uploads/Kommunemonitor_barnevern_2026_09_04_cedc894149.zip (Last-Modified: Fri, 04 Sep 2026, content-type: application/zip, 1,295,973 bytes). ⚠️ Filename shape differs from Barnefattigdom's. Barnefattigdom: YYYY_MM_DD_barnefattigdom_monitor_<hash>.zip. Barnevern: Kommunemonitor_barnevern_YYYY_MM_DD_<hash>.zip — different word order, different leading token. discoverZipUrl's four regex tiers in bufdir-barnefattigdom/parse.ts all require the literal barnefattigdom substring (confirmed by reading the regexes directly); none would match a barnevern URL. This needs its own discovery function with its own four tiers matching barnevern/kommunemonitor_barnevern, not a parameterized reuse of the existing one — don't force a shared abstraction across two literal strings seen once each; copy the shape (progressive tiers, sole-upload last resort, loud failure on ambiguity) into a new, separate function.

  • 1.2 Download the live ZIP and inspect its contents. Confirmed — 24 workbooks, 1,435,207 bytes uncompressed:

    1A_andel_ant_barn_tiltak_ilaaret_ialt_0_17.xlsx   1F_fosterhjem_i_familie_nære_nettverk.xlsx
    1B_barn_med_undersøkelse_ift_innbyggere_0_17.xlsx 2A_utgifter_per_barn_i_befolkningen.xlsx
    1C_barn_utenfor_hjemmet_ift_innbyggere_0_17.xlsx  2B..2F (utgifter-familien)
    1D_barn_med_hjelpetiltak...xlsx                   3A, 3C, 3I, 3L, 3M (frister/bemanning/akutt)
    1E_barn_med_undersøkelse_el_tiltak_per_årsverk.xlsx 4A..4F (befolkning/sosioøkonomi)
    Turnover_kommunalt barnevern_2016-2024.xlsx        (one non-conforming filename, has a space, no code prefix)

    ⚠️ Filenames use an alphanumeric indicator code (1A, 1E, 3M, 4F, …), never the purely numeric Indikator_<n> scheme Barnefattigdom uses. One of 24 files doesn't even carry a code (Turnover_kommunalt barnevern_2016-2024.xlsx) — it's the same indicator as 3M_turnover_...xlsx with a different (older?) year range; needs a decision in Phase 2 (dedupe against 3M, or keep both and let a dbt test catch any disagreement).

  • 1.3 Inspect the internal workbook structure (sheet name, header row, columns) and compare to Barnefattigdom's Data sheet / Region, Regionnavn, Enhet, Tallformat, <years> shape. Confirmed different, not a variant — genuinely different layout:

    • Sheet name is Sheet1, not Data (checked 1A_...xlsx). The one non-conforming file (Turnover_...xlsx) has two sheets, Info and Turnover saksbehandlere — a third shape again.
    • The header row is row 3, not row 1 — row 1 holds a single cell equal to the filename stem (e.g. 1A_andel_ant_barn_tiltak_ilaaret_ialt_0_17), not a separate human-readable title sentence the way Barnefattigdom's row-1 title apparently is. There is no richer title string in the sheet to parse — indicator_title for Barnevern likely has to be a prettified version of the filename slug, at least for 1A. Re-check 2–3 more files in Phase 2 before finalizing; don't assume this holds for all 24 on the strength of one file.
    • Columns: Region | Regionnavn | Tallformat | <year columns 2015..2025> — no Enhet column at all. Barnefattigdom's category_unit dimension (barn/husholdning) doesn't exist here, because Barnevern's indicators are already child-population rates, not paired child/household tuples. raw.bufdir_barnevern must drop category_unit from its primary key and columns entirely — it is not "always null", it is not a column.

  • 1.4 Before writing the parser, sample more workbooks across the numbered families to confirm the Sheet1 / row-3-header / no-Enhet shape holds across the whole bundle, not just 1A. Confirmed on 2C_brutto_utgifter_...xlsx, 3A_fristbrudd_...xlsx, 4F_barn_med_bekymringsmelding_...xlsx: all four (including 1A) are Sheet1, row 1 = filename-stem title, row 3 = header, no Enhet column. 2C's shared-strings pool doesn't surface Regionnavn/Tallformat near the front (it's dominated by a long run of region codes before those strings first appear), which looked like a possible wrinkle — checking 2C's actual <row r="3"> cells directly (not the string pool order) resolved it: same Region | Regionnavn | Tallformat | <years> header as the other three. The shape holds across all four samples spanning 1*–4*.

Validation​

✅ Confirmed 2026-10-01 — four sampled workbooks across 1*–4* all share the same shape (Sheet1, row-3 header, Region | Regionnavn | Tallformat | <years>, no Enhet). Phase 1 is DONE.

Phase 2: Ingest module + raw table — DONE​

⚠️ Caught during this phase's own live validation, not by the plan's Phase 1 research: every standard workbook's Tallformat column uses the literal value Andel for decimal-formatted figures, not the sibling Barnefattigdom source's Prosent. A first draft copied the sibling's prosent check without re-verifying it against Barnevern's own data; running the real ingest against the live ZIP (not just the unit tests) showed 9 of 23 workbooks returning zero rows and the rest returning roughly half their true row count — 70,899 rows instead of the correct 191,673. Fixed by checking for andel instead of prosent throughout parse.ts. Also: andel is not limited to 0–100 percentages — some andel-formatted indicators are kroner-per-child figures, so it means "Norwegian-decimal formatted", not "is a percentage". See parse.ts's parseCell doc and the manifest's category_format dimension note.

Also added during this phase, not originally listed in the task breakdown below: Dagster asset registration (atlas-data/dagster/atlas_data/assets/raw_other.py and schedules.py's _ANNUAL_SOURCE_IDS). Required by the standing "every ingested source appears in a scheduling grouping or cadence.UNSCHEDULED_SOURCES" invariant (schedules.py's own comment) — without it, bufdir-barnevern would need a human to trigger it by hand, exactly the failure mode PLAN-zero-touch-ingest-automation (#1793) fixed for every other source this session.

⚠️ CI caught three more generated-artifact drift gates on first push (PR #489), none anticipated by the task list — recorded so the next new-source PLAN budgets for them:

  • tags.topic: child-welfare isn't a registered category — atlas-data/ingest/src/sources/topics.yaml is a short, curated, editorially-maintained list (check-manifests.sh + the website's generate-sources-registry.mjs both enforce it), not free text. Changed to topic: social ("the bridge between need and the services that respond to it" — the existing category description already fits barnevern without adding a new one for a single source).
  • atlas-data/dbt/seeds/sources/_sources_manifest.csv / _sources_dimensions.csv are generated from every manifest.yml by scripts/build_sources_seed.py and must be regenerated and committed in the same PR as any manifest change — check-sources-seed-is-current.sh diffs the committed seed against a fresh regeneration. Ran it; committed the regenerated seed.
  • atlas-data/template-info.yaml's first_data.takes / first_data.first_load state the raw-table count in prose ("~4.1M rows across 52 raw BASE TABLEs...") and render-template-info.sh checks that number against grep -c 'create table raw\.' across migrations/*.sql. Bumped 52 → 53 in both occurrences — the row-count/timing narrative around it is a real prior measurement (urb-agents #1027) and was left untouched; only the table-count fact changed.

Tasks​

  • 2.1 Create atlas-data/ingest/src/sources/bufdir-barnevern/:

    • fetch_retry.ts — copy verbatim from bufdir-barnefattigdom (generic retry wrapper, no barnefattigdom-specific logic; the file's own header already notes this pattern is duplicated per-source rather than shared, so a fourth copy matches existing convention).
    • parse.ts — new file, same shape as the sibling (discoverZipUrl-equivalent, surrogateIndicatorApiId-equivalent, sheet parser), different literals:
      • Discovery tiers match barnevern / kommunemonitor_barnevern, not barnefattigdom.
      • Surrogate id: /^(\d+[a-z])_/i captures the 1a/3m-style code → bv_zip_ind_<code> (e.g. bv_zip_ind_1a). Hash fallback (bv_zip_<sha256-24>) for the one non-conforming Turnover_...xlsx filename, same defensive shape as the sibling's hash-fallback tier — do not hardcode a special case for that one file. Use a bv_ prefix, not bf_, so the two sources' surrogate ids are visibly distinguishable in raw/marts even though a collision is already unlikely (Barnevern's codes are inherently alphanumeric, Barnefattigdom's are purely numeric).
      • Sheet reader targets Sheet1 at header row 3 (confirmed in Phase 1), no Enhet/category_unit column to unpivot.
    • index.ts — adapted from the sibling's orchestration (fetch monitor page → discoverZipUrl → download → extract → parse each workbook → upsert raw.bufdir_barnevern), same matchTier !== "canonical" warn-log pattern.
    • manifest.yml — hand-authored, matching the sibling's style: 
      source_id: bufdir-barnevern
      upstream_id: barnevern-monitor
      upstream_url: https://www.bufdir.no/statistikk-og-analyse/monitor/barnevern/
      upstream_landing_page: https://www.bufdir.no/statistikk-og-analyse/monitor/barnevern/
      upstream_title: "Barnevern kommunemonitor"
      publisher: Barne-, ungdoms- og familiedirektoratet
      license: NLOD
      license_url: https://data.norge.no/nlod/no/2.0
      periodicity: P1Y
      eu_theme: SOCI
      tags:
        provider: bufdir
        topic: social
        geo: kommune
        cadence: annual
      suggested_joins:
        - bufdir-barnefattigdom
      dimensions: block uses the final confirmed column set: indicator_api_id, region_code, category_format (Tallformat — values antall/andel, no category_unit), year, values_json.
    • README.md — written, matching sibling's structure, documenting the andel vocabulary finding.
    • __tests__/ — 33 tests: discovery tiers (barnevern URL shapes, ambiguous-ZIP refusal), isStandardIndicatorWorkbook filter, parseCell (including the andel-is-not-always-a- percentage case), surrogate id, and an end-to-end parse of the real 1A fixture plus a test that the non-conforming Turnover fixture throws rather than silently mis-parsing.
    • Also registered in atlas-data/ingest/package.json (ingest:bufdir-barnevern script) and in the Dagster asset groupings (see Phase 2 header note above) — not originally itemized here.

  • 2.2 atlas-data/migrations/056_raw_bufdir_barnevern.sql — create table raw.bufdir_barnevern with primary key (indicator_api_id, region_code, category_format, year) (no category_unit — see Phase 1.3), plus comment on table/column written fresh for the actual Barnevern shape (including the andel vocabulary note) rather than copied from the sibling's slightly-stale comments.

Validation​

cd atlas-data/ingest && npm test -- bufdir-barnevern   # 33 passed
cd atlas-data/ingest && npm test                        # 181 passed, whole package — no regressions
cd atlas-data/ingest && npm run typecheck                # clean
env -u DATABASE_URL node_modules/.bin/tsx src/sources/bufdir-barnevern/index.ts  # live run, NDJSON-only

✅ Done 2026-10-01. Live run against the real upstream ZIP: 23 workbooks, 0 workbook.no_rows warnings, 191,673 rows, 489 distinct region codes. (First live run, before the andel fix, produced 70,899 rows with 9 workbooks at zero — exactly the "zero rows silently dropped" failure mode this validation step exists to catch; not caught by the unit tests alone, only by running the real ingest.) Needs Node ≥22 (engines in package.json) — Node 20.11 fails at vitest startup (styleText not exported from node:util).

Phase 3: dbt staging and marts — DONE​

⚠️ A second real defect caught by building against real ingested data, not an empty schema — the same class of bug as Phase 2's andel/prosent miss, one layer deeper: the bare region_code ~ '^[0-9]{4}$' / '^[0-9]{2}$' pattern (copied from the sibling model, matching its current code) calls Svalbard a kommune/fylke. Barnevern's own ZIP carries a Svalbard placeholder row at BOTH levels — region_code = '2111' (4-digit, 394 rows) and a bare '21' (2-digit, 394 rows), every value null across all 23 indicators. This is exactly the defect macros/classify_region_code.sql was built to fix (urb-agents #700) — confirmed live, not inferred, via a local Postgres loaded with the real 191,673-row ingest. Fixed kommune_nr via the existing region_code_to_kommune_nr macro and added a region_kind column (matching the established indicators__ssb_07459.sql convention). fylke_nr needed a narrow, local exclusion of region_code = '21' instead — classify_region_code's own fylke branch is a bare ^\d{2}$ catch-all with the identical latent gap, but fixing that macro site-wide is a bigger blast radius than this PLAN's scope; noted in the model's own comment for whoever investigates the macro next.

Tasks​

  • 3.1 Added raw.bufdir_barnevern to atlas-data/dbt/models/indicators/sources.yml.
  • 3.2 atlas-data/dbt/models/indicators/indicators__bufdir_barnevern.sql — same shape as indicators__bufdir_barnefattigdom.sql, minus category_unit, using 'bv_' || indicator_slug || '__' || category_format for contents_code. kommune_nr goes through region_code_to_kommune_nr, not a bare regex (see defect note above); region_kind added.
  • 3.3 Documented the model's columns in atlas-data/dbt/models/indicators/schema.yml — ran check-osmosis.sh against a local throwaway Postgres (initdb, TCP, short /tmp socket path — the scratchpad directory's own path was too long for a unix socket) with raw.bufdir_barnevern actually loaded via the real ingest. All columns documented, 566 total in that file.
  • 3.4 atlas-data/dbt/models/marts/api/mart_indicators__bufdir_barnevern.sql — api passthrough, mirroring the sibling. Documented in atlas-data/dbt/models/marts/api/schema.yml.
  • 3.5 dbt build against the real loaded data: 1 table model, 21 data tests, 1 view model, all green (23/23) after the region_kind/fylke_nr fix.

Also regenerated, not originally itemized: api_v1_generated.sql / api_v1_state.json (regenerate-api-v1.sh), lineage.csv / lineage_direct.csv (extract_lineage.py), api_v1_relations.csv, the website's sources-registry.json and new per-view dataset page, and template-info.yaml's raw/marts table-count prose (53 raw / 83 marts tables / 69 marts views — each bumped and re-verified via render-template-info.sh and generate-holdings.py --check).

Validation​

cd atlas-data/dbt && dbt build --select indicators__bufdir_barnevern mart_indicators__bufdir_barnevern
# against a local Postgres loaded via: npm run migrate && tsx src/sources/bufdir-barnevern/index.ts

✅ Done 2026-10-01. 23/23 pass against 191,673 real rows (not an empty schema — an empty build passes checks that prove nothing, per [[ci-builds-from-empty-so-state-defects-are-invisible]]). check-osmosis.sh, check-api-v1.sh, check-lineage-is-current.sh, check-every-source-is-served.sh, check-sources-seed-is-current.sh, check-manifests.sh, render-template-info.sh, and generate-holdings.py --check all green.

Phase 4: Catalogue, deploy, and verify arrival — DONE​

Tasks​

  • 4.1 Website catalogue already regenerated as part of Phase 2/3's artifact fixes (44 sources, 9 publishers — bufdir-barnevern scoped under the existing bufdir publisher, no new publisher work needed).
  • 4.2 atlas-data/dbt/check-every-source-is-served.sh passes locally (44 sources, 0 exempt, 0 deferred) — confirmed as part of Phase 3's validation, re-confirmed after merge.
  • 4.3 Filed the deploy request — not a ~/home file (this project's actual convention, per the bus history: cluster-deploy/verification requests go to imac, who operates the live cluster, not a for-ops-*.md drop or ops-dev). Sent via urb send --to imac --from atlas --context atlas, filed as urb-agents#1796. Named the exact relations (raw.bufdir_barnevern, api_v1.indicators__bufdir_barnevern/ mart_indicators__bufdir_barnevern), the code-location image digest (ghcr.io/terchris/atlas-data:v20261001-1b16c96, sha256:c8fb9ce64920ebb1e58b5564019f238bf972690cd09faab48292b60c44caf9de, labelled as the code-location digest, not a UIS install-artifact digest — don't have one of those from here), the derived LANDS WITH (annual_sources_refresh then transform_and_publish — atlas-data/uis/lands-with.sh 8cbcb84..1b16c96), and an expected row count (191,673, stated explicitly as a prediction to be measured, not a certainty — Bufdir could republish the bundle between my local fetch and imac's run, as the sibling source already has once). Also flagged: weekly_polled()'s AutomationCondition.on_cron won't self-materialize a brand-new asset before the next Sunday 02:00 tick, so the refresh needs triggering deliberately if this should land sooner.
  • 4.4 Verified arrival, 2026-10-01 — not just imac's green run report, independently re-checked against the live public API myself before closing the task: 
    GET /meta_sources?source_id=eq.bufdir-barnevern&select=served_as
    -> [{"served_as":["indicators__bufdir_barnevern"]}]

    GET /indicators__bufdir_barnevern?limit=1
    -> 200, a real row (bv_zip_ind_1a, region_kind "nasjon", category_format "andel",
       full 2015-2025 values_json spine)

    HEAD with Prefer: count=exact
    -> content-range: 0-0/191673, cf-cache-status: BYPASS (not a stale cached read)
    191,673 rows — an exact match to both this plan's own local prediction and imac's independent cluster-side measurement. No upstream drift between the local validation run and the live deploy. imac also confirmed no regression elsewhere (bufdir-barnefattigdom 81,568 rows unchanged; FHI sources unchanged) and caught one more naming gotcha worth keeping: the public API serves indicators__bufdir_barnevern, not mart_indicators__bufdir_barnevern — the mart_ prefix is a marts-schema internal naming convention, not part of the published api_v1 view name (the same mistake was made once before on brreg_enhet). Full exchange: urb-agents#1796, closed completed.

Validation​

✅ Done 2026-10-01, verified independently, not taken on trust. Live curl against the public API returns real Barnevern rows through indicators__bufdir_barnevern (191,673, confirmed via Content-Range, not a cached response), and meta_sources.served_as for bufdir-barnevern is non-empty.

Outcome​

Shipped end to end, 2026-10-01: ingest (23 workbooks, 191,673 rows, zero dropped) → dbt staging and api_v1 publication → live cluster deploy → independently verified arrival. Two real data defects were caught by validating against real ingested data at each phase rather than an empty schema or a green CI run — the andel/prosent vocabulary mismatch (Phase 2) and the Svalbard-as-kommune/fylke misclassification (Phase 3) — both documented in place rather than only in this plan, so a future reader of the code finds them without needing this file. bufdir-barnevern plugs the barnevern axis Report #2 (Child Welfare / Vulnerability Composite) was missing.

Acceptance Criteria​

  • bufdir-barnevern ingests cleanly from the live ZIP with zero rows silently dropped — 23 workbooks, 191,673 rows, verified by running the real ingest, not just unit tests.
  • raw.bufdir_barnevern correctly has no category_unit column (verified against Phase 1's finding, not assumed).
  • indicators__bufdir_barnevern and mart_indicators__bufdir_barnevern build and test clean (23/23 data tests, against real loaded data).
  • bufdir-barnevern appears in meta_sources.served_as after a real deploy, with rows confirmed via a live curl, not inferred from CI. Deployed by imac (urb-agents#1796), verified independently by this agent against the live public API: 191,673 rows, non-empty served_as.
  • Golden-file tests cover both the 1A-style shape and the Turnover_... outlier shape (33 tests total, including an end-to-end parse of the real 1A fixture).
  • The investigation (INVESTIGATE-new-norwegian-public-sources.md) and 1PRIORITY.md are updated to mark this candidate shipped, same as the other three corrections made 2026-10-01.

Implementation Notes​

  • Don't force a shared discovery/surrogate-id module across the two Bufdir sources. Each has exactly one literal string to match; a shared parameterized function buys nothing two separate, readable functions don't, and couples two sources' deploy risk together for no reason.
  • The Turnover_kommunalt barnevern_2016-2024.xlsx vs 3M_turnover_saksbehandlere.xlsx overlap is the one real open modelling question this plan doesn't resolve ahead of time — Phase 2's completeness check (mirroring what PR #67 did for the sibling) is where to decide whether they're duplicates, a methodology revision, or genuinely different year ranges worth keeping both.
  • Do not assume Phase 1's 1A findings hold for all 24 workbooks. Task 1.4 is unchecked for a reason — this plan's own confidence in the raw-table shape is bounded by two sampled files out of twenty-four, which is exactly the "scattered sampling misses clustered failures" risk this project has been burned by before. Finish 1.4 before writing the migration DDL.

Files to Modify​

  • atlas-data/ingest/src/sources/bufdir-barnevern/manifest.yml (new)
  • atlas-data/ingest/src/sources/bufdir-barnevern/index.ts (new)
  • atlas-data/ingest/src/sources/bufdir-barnevern/parse.ts (new)
  • atlas-data/ingest/src/sources/bufdir-barnevern/fetch_retry.ts (new, copied)
  • atlas-data/ingest/src/sources/bufdir-barnevern/README.md (new)
  • atlas-data/ingest/src/sources/bufdir-barnevern/__tests__/ (new)
  • atlas-data/migrations/056_raw_bufdir_barnevern.sql (new)
  • atlas-data/ingest/package.json (ingest:bufdir-barnevern script)
  • atlas-data/dagster/atlas_data/assets/raw_other.py (asset registration, weekly-polled cadence)
  • atlas-data/dagster/atlas_data/schedules.py (_ANNUAL_SOURCE_IDS)
  • atlas-data/dbt/models/indicators/sources.yml
  • atlas-data/dbt/models/indicators/indicators__bufdir_barnevern.sql (new)
  • atlas-data/dbt/models/indicators/schema.yml
  • atlas-data/dbt/models/marts/api/mart_indicators__bufdir_barnevern.sql (new)
  • atlas-data/dbt/models/marts/api/schema.yml
  • atlas-data/dbt/api_v1_generated.sql, atlas-data/dbt/api_v1_state.json (regenerated)
  • atlas-data/dbt/seeds/sources/lineage.csv, lineage_direct.csv, api_v1_relations.csv (regenerated)
  • atlas-data/template-info.yaml (raw/marts table counts), website/docs/developers/index.md (regenerated)
  • website/src/data/sources-registry.json, website/docs/datasets/indicators__bufdir_barnevern.mdx (regenerated)
  • website/docs/ai-developer/plans/backlog/INVESTIGATE-new-norwegian-public-sources.md (mark shipped)
  • website/docs/ai-developer/plans/backlog/1PRIORITY.md (mark shipped)