reg-meta-build 0.2.0__tar.gz

reg_meta_build-0.2.0/.gitignore

@@ -0,0 +1,95 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+dist/
+# Root-anchored: only the repo-root packaging dir, NOT the real
+# `reg_monabundle/src/reg_monabundle/build/` source package (the bundle
+# amalgamator). `__pycache__/` above still ignores byte-code anywhere.
+/build/
+
+# Environments
+.venv/
+.env
+.envrc
+
+# Tools
+.ruff_cache/
+.pytest_cache/
+.mypy_cache/
+node_modules/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Project
+.tmp/
+
+# Input data (SCB CSVs, Socialstyrelsen metadata, source PDFs — not committed).
+# Listed per-subdirectory so the maintainer-curated classifications/ folder
+# can stay tracked without an exclude/re-include dance (git doesn't traverse
+# into a parent dir that's been ignored at the directory level).
+reg_meta_build/input_data/*
+!reg_meta_build/input_data/classifications/
+# Under classifications/ only the normalized CSVs (+ manifest.json) are tracked.
+# The raw SOS source workbooks the fetch script downloads are not — they
+# regenerate from scripts/fetch_sos_classifications.py.
+reg_meta_build/input_data/classifications/sos/*.xls
+reg_meta_build/input_data/classifications/sos/*.xlsx
+# landskoder.csv is fetched but deliberately not seeded (see
+# reg_meta_build/CLASSIFICATIONS.md); a bare fetch run regenerates it —
+# keep it out of status noise.
+reg_meta_build/input_data/classifications/sos/landskoder.csv
+
+# mock-data-wizard generated output
+mock_data/
+mock_output/
+mdw_runner.py
+mdw_step1_discovery.json
+mdw_step2_config.json
+mdw_step3_stats.json
+extract_stats*.R
+
+# mock-data-wizard runtime artifacts (transient — fcntl sidecar, run logs)
+.mock_data_config.lock
+mdw_log_*.txt
+
+# Local test workspaces (real user data, not part of the toolkit)
+/covid-education-immigrants-test/
+
+# Personal exploration scripts (not part of the toolkit's curated scripts/)
+scripts/sample_*.py
+
+# MONA probe artefacts -- can contain workspace metadata (paths,
+# hostnames, DSNs); inspect before sharing. The findings that matter
+# live in mock_data_wizard/DESIGN.md.
+mdw_probe_*.log
+mdw_python_probe_*.log
+mdw_py_probe_*.csv
+mdw_upload_probe_*.txt
+
+# reg_meta database (built from SCB exports, not committed)
+*.db
+
+# SCB source PDFs (binary, copyrighted, not committed)
+*.pdf
+
+# Marker raw output (regenerable from PDFs + parser)
+reg_meta_build/docs/_raw/
+
+# Archive (concluded investigations, ad-hoc scripts, internal notes)
+archive/
+*.har
+
+# Claude Code harness state (transient locks, per-user settings, worktree shims).
+# `.claude/skills/` is intentionally tracked for shared skills, so ignore the
+# transient pieces individually rather than blanket-ignoring `.claude/`.
+.claude/scheduled_tasks.lock
+.claude/worktrees/
+.claude/settings.local.json

reg_meta_build-0.2.0/CLASSIFICATIONS.md

@@ -0,0 +1,389 @@
+# Adding canonical code CSVs
+
+Maintainer guide for populating `valid_codes_file` for each classification. Background:
+see `DESIGN.md` § "Classifications" → "Canonical vs observed codes".
+
+## How it works
+
+Each `[[classification]]` in `reg_meta_build/classifications.toml` may declare a
+`valid_codes_file = "<filename>.csv"`. The CSV lives under
+`reg_meta_build/input_data/classifications/` (tracked in git; SCB-sourced CSVs live at
+the top level, SOS CSVs under `sos/`). At build time:
+
+1. Every code in the CSV is ensured to exist in `value_code` (codes that never appeared
+   in any register get inserted as canonical-but-unobserved).
+2. Every `classification_code` row is marked `is_valid = 1` (in CSV) or `is_valid = 0`
+   (observed-only noise).
+3. `classification.valid_code_count` is set to the canonical count.
+
+Without a CSV, every code carries `is_valid = NULL` ("validity unknown").
+
+## CSV format
+
+- Filename: lowercased classification `short_name`, e.g. `sun2000-niva.csv`.
+- Encoding: UTF-8.
+- Header: either `vardekod,vardebenamning` (SCB convention) or `code,label` (universal).
+  Both are accepted; `load_valid_codes` only reads the first two columns — any further
+  columns (`label_en`, `parent_code`, `valid_from`, `valid_to`, …) are silently ignored.
+- One code per row. Whitespace is trimmed on both columns at load time.
+- Duplicate `vardekod`/`code` values → build fails.
+
+Example (SCB convention):
+
+```csv
+vardekod,vardebenamning
+000,Övrig och ospecificerad förskoleutbildning
+001,Förskola
+002,Förskoleklass
+```
+
+Example (universal header, extra columns ignored):
+
+```csv
+code,label,label_en,parent_code
+A01,Tyfoidfeber,Typhoid fever,A00-A09
+```
+
+## Workflow
+
+1. Find SCB's authoritative code list (often a downloadable Excel/CSV on the
+   classification's documentation page).
+2. Save as `reg_meta_build/input_data/classifications/<short_name>.csv` with the header
+   above and commit it.
+3. Add `valid_codes_file = "<short_name>.csv"` to the matching seed entry in
+   `reg_meta_build/classifications.toml`.
+4. Run `reg-meta-build build-db --input-dir reg_meta_build/input_data/`. Build output
+   reports per-classification: canonical / observed-only / canonical-but-unobserved
+   counts.
+5. Spot-check with `reg-meta get classification <SHORT_NAME> --codes --only-valid` and
+   review observed-only and canonical-but-unobserved lists for data-quality issues
+   (mislabeled codes, truncated labels, etc.).
+
+## Status overview
+
+80 classifications (47 per-year LKF entries + 22 SCB-sourced others + 10 SOS code
+systems). All currently declared in `classifications.toml` ship with a
+`valid_codes_file`.
+
+### SCB-sourced classifications
+
+  | short_name           | status | code_count | valid  | notes                                                                       |
+  | -------------------- | ------ | ---------: | -----: | --------------------------------------------------------------------------- |
+  | `SUN2000-NIVA`       | ✓      |        172 |     86 | maintainer-extracted CSV                                                    |
+  | `SUN2000-INRIKTNING` | ✓      |        680 |    647 | maintainer-extracted CSV                                                    |
+  | `SUN2000-GRUPP`      | ✓      |        140 |    139 | from `utbildningsgrupper-sun-2000.xlsx`                                     |
+  | `SUN2020-NIVA`       | ✓      |        106 |     73 | from `sun-2020_niva_inriktning2.xlsx`                                       |
+  | `SUN2020-INRIKTNING` | ✓      |        551 |    531 | from `sun-2020_niva_inriktning2.xlsx`                                       |
+  | `SUN2020-GRUPP`      | ✓      |        231 |    229 | merged main + `sun2020grp_detalj.xlsx`                                      |
+  | `SSYK2012`           | ✓      |      1 125 |    635 | from `ssyk-2012-koder.xlsx`; UTGÅR rows filtered                            |
+  | `SNI2007`            | ✓      |      6 529 |  3 326 | from `sni2007.xlsx`                                                         |
+  | `SNI2002`            | ✓      |      3 151 |  3 088 | from `sni2002.xlsx`                                                         |
+  | `SNI92`              | ✓      |      3 131 |  3 051 | from `sni92.xlsx`                                                           |
+  | `SNI69`              | ✓      |      1 286 |  1 055 | from `sni69.xlsx`                                                           |
+  | `SSYK96`             | ✓      |       ~830 |    501 | scraped from ssyksok.scb.se                                                 |
+  | `NIVA-OLD`           | ✓      |         32 |      7 | hand-written from LISA/UREG docs                                            |
+  | `NIVA-GROV`          | ✓      |         14 |      5 | hand-written from LISA/UREG docs                                            |
+  | `SUN1996`            | ✓      |      4 818 |  4 601 | scraped from metadata.scb.se klassdb (2-, 3-, 5-pos levels)                 |
+  | `LKF{1980..2026}`    | ✓      | 47 entries | varies | per-year split, see "Geography — LKF"                                       |
+  | `ISCED2011`          | ✓      |         53 |     66 | full UNESCO spec via SSB Klass 3426                                         |
+  | `ISCED-F2013`        | ✓      |        164 |    218 | full UNESCO spec via SSB Klass 3428 (+ 2 missing UNESCO codes)              |
+  | `SEKTOR2000`         | ✓      |         53 |     73 | INSEKT 2000 from metadata.scb.se klassdb (Sektor + Undersektor + Delsektor) |
+  | `JURFORM2000`        | ✓      |         45 |     36 | scraped from metadata.scb.se klassdb                                        |
+  | `JURFORM2020`        | ✓      |         39 |     36 | scraped from metadata.scb.se klassdb                                        |
+  | `AGARKAT2000`        | ✓      |         19 |     11 | scraped from metadata.scb.se klassdb (Nivå 1 + Nivå 2)                      |
+  | `AGARKAT2020`        | ✓      |         10 |     11 | scraped from metadata.scb.se klassdb (Nivå 1 + Nivå 2)                      |
+
+`code_count` and `valid` columns reflect the latest build. `valid` may exceed the CSV
+row count when SCB exports carry the same canonical code under multiple labels (each
+label variant becomes its own `value_code` row, all marked `is_valid=1`).
+
+### SOS code systems
+
+These are provider-seeded entries (`provider = "sos"`): they seed canonical codes via
+`valid_codes_file`. PR2 wired SOS→classification linkage via the
+`external_classification` resolver, so SOS variable_states can now carry a
+`classification_id` — the entries are no longer canonical-only. CSVs live under
+`input_data/classifications/sos/`; `manifest.json` there records the per-system source
+URL, sha256, and counts.
+
+  | short_name   | canonical codes | publisher                               | notes                                                    |
+  | ------------ | --------------: | --------------------------------------- | -------------------------------------------------------- |
+  | `ATC`        |          16 264 | Läkemedelsverket (MPA)                  | Swedish ATC register; nightly zip                        |
+  | `ICD-10-SE`  |          38 928 | Socialstyrelsen                         | Swedish ICD-10                                           |
+  | `KVA`        |          13 548 | Socialstyrelsen                         | KVÅ – care measures (KMÅ medical + KKÅ surgical, merged) |
+  | `ICF`        |           1 632 | Socialstyrelsen                         | Functioning, Disability and Health                       |
+  | `KSI`        |           2 074 | Socialstyrelsen                         | Social Services Interventions                            |
+  | `ICD-8-KS68` |           5 600 | Socialstyrelsen                         | Swedish ICD-8, 1969–1986                                 |
+  | `ICD-9-KS87` |           5 679 | Socialstyrelsen                         | Swedish ICD-9, 1987–1996                                 |
+  | `KS87-P`     |             383 | Socialstyrelsen                         | ICD-9 primary care variant, 1987–1996                    |
+  | `DRG`        |           1 983 | Socialstyrelsen / Nordic Casemix Centre | NordDRG diagnosis-related groups                         |
+  | `MDC`        |              29 | Socialstyrelsen / Nordic Casemix Centre | NordDRG major diagnostic categories                      |
+
+## SOS code systems — snapshot fetcher
+
+The SOS CSVs are produced by `scripts/fetch_sos_classifications.py` (committed; the SOS
+analog of `scripts/extract_lkf.py`). This script is run on demand by a maintainer — it
+is NOT executed automatically at build time. The resulting CSVs and `manifest.json` are
+committed under `input_data/classifications/sos/`.
+
+Per-system source URLs and sha256 digests are stored in
+`input_data/classifications/sos/manifest.json`. On each run the script merge-writes that
+manifest, so counts and hashes stay current.
+
+**Refresh caveat:** a bare run fetches all configured source keys, including
+`landskoder` (ISO-3166 country codes) — which we deliberately do NOT seed as a
+classification. To refresh only the 10 seeded keys without touching unrelated entries,
+pass `--only` with the relevant key names:
+
+```bash
+uv run python scripts/fetch_sos_classifications.py \
+    --only atc icd-10-se kva icf ksi \
+             icd-8 icd-9-ks87 ks87-p drg mdc
+```
+
+**DRG / MDC note:** both are drawn from the NordDRG system published by Socialstyrelsen
+in collaboration with the Nordic Casemix Centre.
+
+## Done — extraction details
+
+The conversion was a one-shot script run against XLSX files in `/tmp/scb_xlsx/`. For
+each, this section lists where the source came from and which sheet/columns to use. If
+you need to re-extract, the conversion logic is straightforward enough to redo from
+these notes — there's no committed extraction script because each file's quirks are
+different and the result is what we ship.
+
+### SUN — Svensk utbildningsnomenklatur
+
+Page:
+<https://www.scb.se/dokumentation/klassifikationer-och-standarder/svensk-utbildningsnomenklatur-sun/>
+
+#### `SUN2000-NIVA`, `SUN2000-INRIKTNING`
+
+User-provided. Likely extracted manually from `sun-2000_niva_inriktning.xlsx` on the SCB
+SUN page. CSVs include all hierarchy levels (1-, 2-, 3-digit for NIVA; 1- through
+4-character including suffixes like `010a` for INRIKTNING).
+
+#### `SUN2020-NIVA`, `SUN2020-INRIKTNING`
+
+Source file: `sun-2020_niva_inriktning2.xlsx`
+([download](https://www.scb.se/contentassets/aeeedec0e28c465aa524429407dcd5ba/sun-2020_niva_inriktning2.xlsx))
+
+Layout: each sheet has multiple "Kod \| Benämning \| (gap)" column triples side by side,
+one per hierarchy level (1-siffer, 2-siffer, 3-siffer). Header at row 3.
+
+Extraction rules:
+
+- Use sheets `Nivåer, klartext` and `Inriktning, klartext` (skip the `_alt` variants —
+  those are legacy mappings).
+- For each "Kod" position in the header, read consecutive rows until the first blank in
+  that column. Crucial: the sheet has a legacy-mapping table further down with code `7`
+  "Forskarutbildning" — stopping at the per-column blank avoids picking up these rogue
+  rows.
+
+#### `SUN2000-GRUPP`
+
+Source file: `utbildningsgrupper-sun-2000.xlsx`
+([download](https://www.scb.se/contentassets/aeeedec0e28c465aa524429407dcd5ba/utbildningsgrupper-sun-2000.xlsx))
+
+Layout: code in column 1 (e.g. `01Z`), label in column 2, **multi-line** labels span 2–3
+rows (continuation rows have empty col 1). Section headers like `Allmän utbildning (0)`
+appear in col 2 with no code in col 1.
+
+Extraction: walk rows. When col 1 has a code → start new entry (label = col 2). When col
+1 is empty but col 2 has text → append to current label. Blank row → flush current
+entry.
+
+#### `SUN2020-GRUPP`
+
+**Two source files merged**:
+
+1. `utbildningsgrupper-sun-20202.xlsx`
+   ([download](https://www.scb.se/contentassets/aeeedec0e28c465aa524429407dcd5ba/utbildningsgrupper-sun-20202.xlsx))
+   — main groups, 96 codes (3-char like `01Z`). Header at row 3, code in col 0, label in
+   col 1. Skip row 4 ("0 Allmän utbildning" section header — code is None).
+2. `sun2020grp_detalj_schema-260223_utskrift_260223.xlsx`
+   ([download](https://www.scb.se/contentassets/aeeedec0e28c465aa524429407dcd5ba/sun2020grp_detalj_schema-260223_utskrift_260223.xlsx))
+   — detailed groups, 133 codes (4-char like `01ZA`). Same multi-line layout as SUN 2000
+   GRUPP, but code in col 0 instead of col 1.
+
+Both code spaces map to `value_set_version_label` strings already grouped under
+`SUN2020-GRUPP` (`SUN 2020 - Gruppering` and `SUN 2020 - Gruppering - Detaljerad`), so
+they're merged into one CSV. The detailed file contains a `23XA` 2025 addition that
+hasn't reached our register exports yet — surfaces as canonical-but-unobserved.
+
+### SSYK — Standard för svensk yrkesklassificering
+
+Page:
+<https://www.scb.se/dokumentation/klassifikationer-och-standarder/standard-for-svensk-yrkesklassificering-ssyk/>
+
+#### `SSYK2012`
+
+Source file: `ssyk-2012-koder.xlsx`
+([download](https://www.scb.se/contentassets/0c0089cc085a45d49c1dc83923ad933a/ssyk-2012-koder.xlsx))
+
+Use sheet `Hela strukturen`. Code in col 0, label in col 1. Codes range from 1 (1-digit)
+through 9999 (4-digit) following SSYK's nested hierarchy.
+
+Edge case: 8 placeholder rows have label `UTGÅR[]` ("removed") for codes that were
+retired between revisions (e.g. `215`, `252`, `263`). Filter these out — they're not
+real canonical codes.
+
+#### `SSYK96`
+
+No clean structured XLSX exists. The MIS 1998:3 PDF is scanned (no extractable text),
+and the conversion XLSX (`webb_nyckel_ssyk96_ssyk2012_20160905.xlsx`) has 4-digit codes
+but without canonical category labels (only individual job titles).
+
+Approach: derive the canonical 4-digit codes from the conversion file's `Nyckel_4siffer`
+sheet (355 codes), generate parent levels (1-/2-/3-digit prefixes → 505 codes total),
+then scrape category labels from <https://ssyksok.scb.se/SsykSok/SSYK96/{code}> at
+\~150ms/req. SSYK-Sök's HTML has a stable `<h2><strong>{code}</strong> {label}</h2>`
+pattern. Yields \~501 codes (a handful 404 in SSYK-Sök). The CSV ships as a one-off
+snapshot — it's not regenerated by a script in this repo since the scrape is a
+single-use extraction. If SCB's naming changes, re-extract by running the scrape inline.
+
+### SNI — Svensk näringsgrensindelning
+
+Page:
+<https://www.scb.se/dokumentation/klassifikationer-och-standarder/standard-for-svensk-naringsgrensindelning-sni/>
+
+Source files (one XLSX per version):
+
+- `sni2007.xlsx`
+  ([download](https://www.scb.se/contentassets/d43b798da37140999abf883e206d0545/sni2007.xlsx))
+- `sni2002.xlsx`
+  ([download](https://www.scb.se/contentassets/d43b798da37140999abf883e206d0545/sni2002.xlsx))
+- `sni92.xlsx`
+  ([download](https://www.scb.se/contentassets/d43b798da37140999abf883e206d0545/sni92.xlsx))
+- `sni69.xlsx`
+  ([download](https://www.scb.se/contentassets/d43b798da37140999abf883e206d0545/sni69.xlsx))
+
+Layout: one sheet per hierarchy level (Detaljgrupp/Undergrupp/Grupp/
+Huvudgrupp/Avdelning, plus extras for SNI 69). Column structure varies across sheets:
+
+- Detaljgrupp/Undergrupp/Grupp: `(Officiell kodstruktur, flat code, Benämning)` at cols
+  0,1,2.
+- Huvudgrupp/Avdelning: `(code, Benämning, ...)` at cols 0,1.
+
+Extraction rule: find the column header starting with `Benämning` (some sheets have a
+trailing space) — the **flat code is in the column immediately to the left** regardless
+of sheet. Use that pair across all sheets to build the union.
+
+Notes:
+
+- Use the flat code (no dots), since that's what registers store.
+- For SNI 69, the lowest-level sheet is "Sexsiffer" (6-digit); for the others it's
+  "Femsiffer" (5-digit).
+- SNI 69 has 113 1-/2-/3-digit prefix codes that don't appear in our register exports —
+  surfaces as canonical-but-unobserved.
+- SNI 2007 has the largest observed-only set (\~3 200) because data carries codes in
+  many alt notations not in the canonical list (dotted `01.110`, ranges `102-103`,
+  letters from Avdelning) — most are real references in alt format, not noise.
+
+### SUN 1996, SEKTOR/INSEKT, JURFORM, AGARKAT — scraped from SCB Klassdb
+
+Page: <https://metadata.scb.se/klassdb.aspx>
+
+The SCB Klassifikationsdatabasen UI uses ASP.NET WebForms `__doPostBack` navigation, so
+there are no stable deep-link URLs. The data is reachable through the values pane
+(`tvKlass2`) once a version is selected via
+`__doPostBack('tvKlass', 'sVMF\\<vmf>\\VMK\\<vmk>\\VMV\\<vmv>')`. Each version then
+exposes one or more "Nivå N" levels addressed by appending `\\VMN\\<id1>\\<id2>...`.
+
+These were one-shot scrapes (browser-driven JS in the page DOM, output piped via
+`console.log` or `Blob` download). The CSVs ship as snapshots — re-extract only if SCB
+updates the codes. Reference IDs (in case re-scrape is needed):
+
+  | short_name    | vmk | vmv   | levels                                      |
+  | ------------- | --- | ----- | ------------------------------------------- |
+  | `SUN1996`     | 801 | 781   | 2-pos / 3-pos / 5-pos (4601 codes)          |
+  | `SEKTOR2000`  | 652 | 491   | Sektor / Undersektor / Delsektor (73 codes) |
+  | `JURFORM2000` | 653 | 493   | Nivå 1 (36 codes)                           |
+  | `JURFORM2020` | 653 | 80811 | Nivå 1 (36 codes)                           |
+  | `AGARKAT2000` | 654 | 495   | Nivå 1 + Nivå 2 (11 codes)                  |
+  | `AGARKAT2020` | 654 | 80806 | Nivå 1 + Nivå 2 (11 codes)                  |
+
+Edge cases:
+
+- `SEKTOR2000` toml lists two vardemängdsversionen: the canonical
+  `Standard för institutionell sektorindelning 2000` (53 observed 3-digit codes — match
+  Delsektor) and a register-local `Sektor 2000` (23 codes with a different 1-2 digit
+  code scheme that mostly stays `is_valid=0`). The canonical CSV serves the standard
+  one.
+- `JURFORM 2020` is identical to `JURFORM 2000` except code `84` changed from
+  "Landsting" to "Regioner".
+- `AGARKAT` 2000 and 2020 differ only in code `30/3` Region(kontrollerade) rename.
+
+### Geography — LKF (per-year split)
+
+Page:
+<https://www.scb.se/hitta-statistik/regional-statistik-och-kartor/regionala-indelningar/lan-och-kommuner/>
+
+LKF is split into 47 per-year classifications `LKF1980` … `LKF2026`, each with its own
+canonical CSV. Single-year codes change across the sequence (Skåne 1997, Västra Götaland
+1998, parish mergers ongoing), so each year is its own published snapshot — 665 codes
+had conflicting labels under the previous unified entry. The data already disambiguates:
+`variable_state.value_set_version_label` is year-stamped (`LKF 1990-01-01/...`), so the
+split is purely seed-side — no schema change, no FK rework. `supersedes` chains the
+years sequentially.
+
+CSVs come from `scripts/extract_lkf.py`:
+
+- **1980–2015** (kommun + derived län): from `knkodnyckel.xls`, SCB's kommun-history
+  file with 12 period-snapshots. Script derives 2-digit län codes as kommun prefixes and
+  labels them via a year-aware map (handles 1997 Skåne/Dalarna and 1998 Västra Götaland
+  reforms). \~310 codes/year. **Församlings (6-digit) NOT included** — pre-2016 parishes
+  are a remaining gap.
+- **2016–2026** (full LKF — län + kommun + församling): per-year
+  `lkf{year}.xls`/`.xlsx`. Script handles the three different sheet layouts SCB has used
+  (2018–19 / 2021–22 / 2023+) and the non-standard filenames for 2020 (`_justerad`) and
+  2026 (date-stamped subdir). \~1500–1700 codes/year.
+- **Pre-1980 / 2015 parishes**: not wired. The script can produce 1974–1979 CSVs but the
+  reg_meta data has no instances for those years so no seed entries exist for them. The
+  2015 PDF can be OCR'd to recover that year's parishes.
+
+Re-run extraction:
+
+```bash
+uv run --with openpyxl --with xlrd python scripts/extract_lkf.py \
+    --out reg_meta_build/input_data/classifications/
+
+# Optional: also download PDFs for OCR / cross-checking:
+uv run --with openpyxl --with xlrd python scripts/extract_lkf.py \
+    --download-pdfs
+```
+
+The current toml entries were generated by `--emit-toml` then spliced in. If SCB adds a
+new yearly snapshot, add an `LKF{year}` block manually (or re-run --emit-toml and
+merge).
+
+Then add `LKF{year}` seed entries (`--emit-toml` prints starters; the year-string
+variants need to be reconciled against the existing LKF entry in
+`classifications.toml`). The current single `LKF` entry should then be removed, with
+each year's `value_set_version_label` strings moved to the appropriate `LKF{year}`
+entry.
+
+### Education — ISCED 2011
+
+`ISCED2011` canonical CSV (`isced2011.csv`) was sourced from SSB's Klass register,
+version 3426 (Norwegian translation of UNESCO ISCED 2011 with all three levels: 1-digit,
+2-digit, 3-digit — 66 codes total). All 39 observed Nivå codes match. The 3-digit codes
+I'd previously flagged as "SCB-specific" turn out to be standard UNESCO level-3 detail
+codes.
+
+The "ISCED 2011 AES" vardemängd uses different padded codes (`000`, `200`, `300`, `302`,
+…) which appear to be SCB-specific. Only `100` matches the canonical set; the other 13
+stay as `is_valid=0`.
+
+### UNESCO ISCED-F 2013
+
+`ISCED-F2013` canonical CSV (`isced-f2013.csv`) was sourced from SSB's Klass register,
+version 3428 (UNESCO ISCED-F 2013 — 216 codes spanning broad/narrow/detailed fields).
+124/126 observed `ISCED-F 2013 - Inriktning` codes match. The two SSB drops (`0110`
+Education NFD, `0739` Architecture and construction NEC) were re-added since they're
+standard UNESCO and present in SCB data.
+
+The second vardemängd `ISCED F 2013` (with a trailing space in the source) carries 38
+codes mostly in Swedish (`Pedagogik och lärarutbildning`, etc.) — only 12 match UNESCO.
+This looks like a register-local mapping mislabelled as ISCED-F; left at `is_valid=0`
+for the non-matching codes.