law-version-drift-check

name	law-version-drift-check
description	Detects version drift between machine-readable law YAML files and the binding geldende wettekst on wetten.overheid.nl at the YAML's claimed valid_from. Catches the silent failure where authored content lags behind ingevoerde wijzigingen — a YAML dated 2025-01-01 still carrying pre-Stb.-X text. Activate as Step 0 of every regelrecht-stelselanalyse micro-cycle, before any YAML edit. Strict by design: no bypass, no normalization beyond whitespace within a single paragraph; the geldende wettekst is leading and is mirrored verbatim, including errata. Use proactively when starting a stelselanalyse cycle, before harvest/migrate/extend, or when user mentions "drift", "versie", "geldende tekst", or worries that a YAML may be stale relative to wetten.overheid.nl.
allowed-tools	Read, Write, Bash, Grep, Glob, WebFetch, WebSearch, AskUserQuestion
user-invocable	true

Law Version Drift Check — Step 0 of every micro-cycle

Detects whether a YAML's text: blocks (and adjacent fields) match the binding geldende wettekst at the version the YAML claims to represent. The failure this skill catches is silent: a file passes schema-validation, hallucination-checks and BDD tests while its content is from an older version of the law. Every downstream review (wetgevings-fouten, modellering-fouten, untranslatables) then mis-classifies bevindingen until the drift is closed.

This skill belongs in the desk-layer (regelrecht-stelselanalyse) and runs before every other review-as.

Core principle — verbatim spiegelregel

De feitelijk geldende wettekst is leidend. Fouten in de wet worden feitelijk overgenomen. Geen "soll"-fietsen.

The YAML's text: block must equal the geldende wettekst character-for-character. There is one allowed normalization, defined below; every other deviation is drift, regardless of who is "right". Errata, redactionele oddities, dode verwijzingen — all preserved verbatim in the YAML. If the wet contains a fout, the wet itself is the carrier; the YAML mirrors it. A wetgevings-fout (4-weg klasse 2) is reported to the wetgever separately; it is never silently "corrected" in the YAML.

The one allowed normalization

|- block scalars in YAML wrap long sentences at ~N characters for editor readability; wetten.overheid.nl wraps based on its own rendering. Within a single paragraph, multiple whitespace and intra-paragraph line breaks collapse to a single space before comparison. Between paragraphs (between leden, between onderdelen, between verbatim-blocks) structure is preserved exactly.

Anything else — diacritics, quote glyphs («» vs "), capitalization, word order, punctuation — is part of the law's substance and any deviation is drift.

Position: Step 0 of regelrecht-stelselanalyse

This skill is a hard prerequisite for every micro-cycle. Without a CLEAN (or scope-restricted) drift-report, no YAML edit may be made.

Rationale: drift silently mis-classifies findings across the 4-weg-classificatie:

• A claim "the YAML says X but the wet says Y" is a wetgevings-fout only if the geldende wet truly says Y. If Y had already been changed by a Stb. that was not ingeharvest, the same finding is in fact a modellering-fout (drift). Confusing the two breaks the 4-weg discipline that the cycle relies on.

Strict by design: no bypass parameter, no "this cycle is just docs" escape. If a cycle is purely docs and not touching YAMLs, it doesn't reach Step 0 — the trigger is "about to edit a regulation YAML".

Verdict per article — two axes

Structural	Textual	Verdict
match	match	CLEAN
match	diff	DRIFT-tekst (with word-level sub-diff)
mismatch	(skipped)	DRIFT-structureel (with the missing/extra item named)
(any)	not fetchable	NIET-VERIFIEERBAAR (after retry + Staatsblad-fallback)

Structural axis (assessed first): count of leden, ordered list of onderdeel- letters per lid, alinea-count per lid. A mismatch here is decisive — skip textual, report the missing/extra element by name. This is the axis that catches the most costly drift type: an entire missing onderdeel or lid (a downstream wijziging that was never harvested).

Textual axis: per-paragraph word-level diff after applying the single allowed whitespace normalization. The output is a unified diff at word granularity for every paragraph that differs.

Procedure

1. Inventory the target file

Extract from the YAML:

• bwb_id, valid_from
• per article: number, text: block (the literal string after |-), url:
• corpus-level: every competent_authority.name referenced anywhere in the file

2. Structural inventory from the geldende text

For each article, fetch the structural skeleton at https://wetten.overheid.nl/<bwb_id>/<valid_from>#Artikel<N>.

Force enumeration in the prompt. Summary-style prompts produce text that looks complete while quietly dropping the last onderdeel. Required form:

"List the structural skeleton of artikel N at this version: number of leden, for each lid the ordered list of onderdeel-letters (a, b, c, …), and the last letter present. Do not paraphrase, do not omit any onderdeel."

Two independent calls per article (different phrasings) and compare; a disagreement is treated as a fetch reliability problem and triggers Staatsblad- fallback for that article.

3. Calibration ijkpunten — verplicht

Before any drift-report is emitted, the skill must independently re-find a set of pre-recorded known drifts for the corpus under review. The ijkpunten file lives in the corpus (docs/drift-ijkpunten.md or equivalent), not in this skill, so the skill stays dossier-agnostic.

If calibration fails to recall ≥ 80% of known drifts: WebFetch is not reliable for this run. Retry with stricter prompts; on a second failure, fall back to Staatsblad-stack for the entire corpus. A drift-report without passing calibration is invalid and must be marked as such.

4. Verbatim text fetch (textual axis)

For articles that pass the structural axis, fetch the verbatim text per article with an enumeration-forcing prompt:

"Give the literal full text of artikel N at this version, lid by lid, onderdeel by onderdeel, verbatim. State the last onderdeel letter explicitly so completeness can be verified. Do not summarize, do not paraphrase, do not omit."

5. Diff per article

• Apply the single allowed whitespace normalization (collapse intra-paragraph whitespace; preserve paragraph boundaries).
• Word-level diff per paragraph. Output unified-diff at word granularity.
• Diacritics differences count as DRIFT-tekst (per the verbatim principle).
• Errata-style additions (e.g. parenthetical reading notes) present in the geldende text must be present in the YAML; absence is drift.

6. Adjacent fields (same file, same run)

• competent_authority.name values across the file: check against the current formal organ name at the geldende version. A historical name is drift.
• url: anchors per article: must resolve to the artikel at the YAML's valid_from. Anchors pointing at a different version (or a non-existent article) are drift.

7. Staatsblad fallback — five cases

wetten.overheid.nl is primary. Fall back to the Staatsblad-stack (zoek.officielebekendmakingen.nl) only when:

1. Render lag — a recent Stb. is published but not yet rendered in the consolidation; the change exists in the publication of record.
2. Calibration fail for a specific article: distrust the render, verify via the Stb. that introduced the most recent change to that article.
3. Errata-style notations ([bedoeld zal zijn ...] or equivalent): confirm the notation is in the formal Stb. text, not a rendering artefact.
4. Version gap — the YAML's valid_from falls between two consolidations (deferred inwerkingtreding, retroactive effect): the Stb. clarifies the inwerkingtredingsdatum.
5. Historical version reconstruction — older valid_from where the current render no longer shows the historical text cleanly.

Local pdftotext is the final fallback for opaque PDFs; mark such articles as verified via local extraction.

In the report, every article carries its verification mode: primair: wetten.overheid.nl/<datum> or secundair: Stb. <jaartal>, <nr> or tertiair: local-pdf-extract.

8. Partial-failure handling — scope-restrictie

After retry + Staatsblad-fallback, articles that remain NIET-VERIFIEERBAAR are frozen for the duration of the micro-cycle: any edit touching their text block, competent_authority value, or url anchor is blocked until the next drift pass. Other articles in the same file remain editable. The report lists frozen articles with the failure reason (TLS, rate-limit, persistent render issue) and a suggested retry timeframe.

No bypass. The skill never permits "skip and continue".

Output

Drift-rapport (per file)

# Drift-rapport — <file path>
bwb_id: ...
valid_from: ...
calibratie: ✅ (X/Y ijkpunten gerecalled) | ❌ (rapport invalide)

| Artikel | Verdict | Bron-modus | Detail |
|---------|---------|------------|--------|
| 1       | CLEAN   | primair    | — |
| 2       | DRIFT-structureel | primair | onderdeel z ontbreekt in YAML |
| 3       | DRIFT-tekst | primair | lid 2: "<woord1>" → "<woord2>" |
| ...     | ...     | ...        | ... |
| N       | NIET-VERIFIEERBAAR | — | TLS-fout na 3 pogingen; bevroren |

## Aggregaat
- CLEAN: a
- DRIFT-tekst: b
- DRIFT-structureel: c
- NIET-VERIFIEERBAAR (bevroren): d

## Diakritiek-aggregaat (planning, geen excuus)
Aantal artikelen waar diakritiek de enige driftbron is: e

## Bevroren artikelen (scope-restrictie)
- art. N — reden — suggested retry

Resolutie-tracker entries

Each DRIFT entry becomes a row in the corpus's resolutie-tracker.md with 4-weg klasse 1 (Modellering-fout) and a link to the geldende wettekst. The fix lands in modellering-fixes-plan.md (corpus-side, not in this skill).

Sub-rapport diakritiek

Aggregate list of articles whose only deviation is diacritics. This is NOT an exemption — diacritics-only drift still counts as DRIFT-tekst in the verdict table. The sub-rapport exists only to make a corpus-wide reconciliation plannable (one PR that restores accents corpus-wide, rather than per-article).

Integration with regelrecht-stelselanalyse

• Inserted as Step 0 of every micro-cycle that may touch regulation YAMLs.
• Findings → resolutie-tracker.md (status: drift-gerapporteerd / drift-opgelost).
• Fixes → modellering-fixes-plan.md as 4-weg klasse 1.
• A micro-cycle may only collect findings on other axes (wetgevings-fouten, untranslatables, coverage, source-refs) on articles that are CLEAN or whose DRIFT-tekst is closed within the same cycle. Otherwise the finding is mis-classified and rejected at synthesis time.

What this skill does NOT do

• It does not edit YAMLs. It only reports.
• It does not "fix" errata in the geldende wet. Errata are preserved verbatim in the YAML; the wetgevings-fout is reported elsewhere.
• It does not decide on diacritics policy. It reports the drift; a separate corpus-wide reconciliation cycle decides whether to align the YAMLs to the wet (verbatim) or to leave them — but the report is unambiguous.
• It does not bypass under load. A flaky network produces scope-restrictie, not a green report.

Files in this skill

• SKILL.md — this document (process).
• reference.md — URL patterns, the single allowed whitespace normalization algorithm, prompt templates, verdict criteria, output format.

Hard rules

• Dossier-agnostisch. No corpus content, no concrete drift examples, no references to specific findings — those belong to the corpus the skill runs against, not to the skill itself.
• Verbatim is wet. No normalization beyond the single intra-paragraph whitespace rule.
• Structure before text. Always.
• Calibratie vóór rapport. A report without passed calibration is invalid.
• Strict, no bypass. Scope-restrictie handles partial failure; nothing handles "skip this cycle's drift check".
• No push without permission. Inherits the stelselanalyse protocol.