[bdb-dd]

Stacked on #109 (the Parquet source). Review that one first. This PR's own diff is the four commits on top of it.

What

A local-config example that uses the Parquet source on a public dataset, Every AI Policy Ever (CC-BY-4.0). The dataset is a register of about 400 AI laws, regulations, frameworks and strategies from 87 countries and international bodies, plus defined terms quoted verbatim.

• Two datasets with weekly pipelines from Hugging Face:
  • digdir/ai-policy: the records, filterable by jurisdiction, country, region, instrument type, legal force, status, verification level, safety topics and year.
  • digdir/ai-policy-definitions: the definitions, filterable by term, term group and language.
• The record text carries short title, dates published, in force and last updated, the English version and its translator, and the verification note. Answers can then build timelines and say when an instrument binds.
• KI-regelverksguiden, an agent-loop service over both datasets. For each instrument it gives jurisdiction, type, legal force, status, dates, identifier and verification level, warns at level C, says the summaries are AI-made, and ends with the CC-BY attribution.
• Four skills, pinned by the guide: compare jurisdictions, quote definitions (original language plus the English rendering), build a timeline, and check whether something is binding.
• Two step flows with services:
  • ai-policy-brief: binding instruments on a topic, as a table and a short brief.
  • ai-policy-comparison: a side-by-side table across jurisdictions.
• A draft test set, packages/eval/testsets/ai-policy-guide-draft.yaml: 10 cases, including one not-covered case, all needsReview.
• Seeding: local config now seeds skills/ before services/, so a service that pins a seeded skill is no longer skipped with "no published skill". README and development docs mention it.

Tests

• definition-sources.test.ts:
  • seeds skills before services;
  • every pinned skill is seeded at the pinned version;
  • every skill is valid, named after its file and has a must_reject case.
• testsets.test.ts: the draft stays a draft, has sources in both datasets, and has a not-covered case.
• Pipeline seed tests validate both pipelines.
• Checked end to end locally: ingest, then answers through the API and the portal playground. Citations were valid against the register in every round checked: 19 of 19 for the guide, and 37 of 37 and 11 of 11 for the flows.

Notes

• Definitions have their own dataset, because two pipelines on one dataset currently replace each other's documents (A pipeline run replaces the documents of the dataset's other pipelines #107, fix in Keep the other pipelines' documents when a pipeline runs #108).
• Services use foundry-openai/gpt-4o-2m-tps, like the other local services.

Draft, for review.

Follow-up commit (b624e7d)

• Eval grounding. Every judged case states that the register is a dated snapshot whose 2025–26 instruments are real, and when the answer passes. The criteria judge only sees the reference text, so it had graded against its own knowledge. Two eval runs scored 0.73 and 0.82, with no "made up" verdicts.
• New train case: deepfake labelling, where the EU AI Act was missed.
  • The binding-check skill now filters on both the narrow and the broad topic.
  • The guide reranks in the same round as it reads, because generate_response without a rerank sees only the first ten chunks in insertion order.
  • It passes chunk ids to read_chunks.
  • EU-03 was cited in 0 of 3 runs before and 3 of 4 after.
• Definitions take status, dates and verification level from the law with the same Record id, instead of "ikke oppgitt". Colorado was right in 3 of 3 runs, against 0 of 5 before.
• scripts/ai-policy-jurisdictions.ts regenerates the comparison flow's jurisdiction list from records.parquet, or checks it with --check, which reports 141 names matching. Tests keep the bokmål mapping inside the list. The flow says a name outside its list is unknown to it, not missing from the register.

Independent verification

A separate verifier found three issues, all fixed:

• A comparison where every jurisdiction was unknown ran an unfiltered search. It now answers without searching.
• The context limit could silently drop the sixth jurisdiction. Comparisons are now capped at 4 jurisdictions, 3 chunks each, with a context sized for the worst case. A fifth requested jurisdiction is reported, not dropped.
• The pipelines read main weekly while the answers claimed a fixed snapshot. Both pipelines are now pinned to Release v1.18 (2026-10-19), with no schedule, and the texts state that version.