Der harte relevant=true-Filter versteckte ~25% des Korpus (40.926 Atome),
~70% davon echte Pflichten (500er-Validierung). relevant wird zur Stufe:
- Service: tier-Param (core=Default schuetzt Agent/CRA; all=alles inkl. review),
ORDER BY relevant DESC; pro Control relevant/tier/source_type
(own_library bei license_rule=3, sonst derived) + source_regulation/article;
core_count/review_count. Pure Helper tier_label + source_type (+ Tests).
- Route: optionaler tier-Query (default core) — contract-safe (additiv).
- Frontend: Coverage-Drill-down /sdk/coverage/[useCase] — Kern-Pflichten vs.
"zur fachlichen Pruefung", je mit Herkunfts-Badge; Uebersicht zeigt Delta.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reads compliance.atom_classification (Haiku pass: relevant + sub_topic +
canonical_obligation) when present -> precise, sub-topic-organized controls per
topic; master-grain seed stays as fallback for unprocessed topics. New optional
sub_topic filter + subtopic_counts facet + granularity flag in the response.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Read-only layer (service + thin route + tests) that returns the controls
mapped to a use-case/topic, ranked by a deterministic precision proxy
(is_primary + mapping confidence + registry keyword relevance) over the
existing mc_use_case_mappings seed. No schema change.
Shared handoff point: the document specialist agents AND the CRA
finding-mapper draw from this one controls index instead of separate
retrievals.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Benjamin Admin