Replace the reconstructed test_dse_agent.py with the canonical version and add
the companion unit tests (classification_gate, embedding_recall) covering the
recovered v3 modules. Include the curated DSE criteria backup + changelog
(legal-note rationale per control), the v1 validation writeup, and the
multi-company DSE ground-truth fulltexts (elli/eto/mercedes/safetykon) used
for threshold calibration.
18 DSE tests green offline (DB/embedding/LLM stubbed).
dev-only, no deploy.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The calibrated DSE engine (4-layer: regex-boost / keyword / BGE-M3 embedding
recall @0.65 / semantic-validator) existed ONLY in the running macmini
container (docker cp'd, never committed) — at risk of loss on any container
rebuild. This recovers it into git and wires it into the live check path.
- Recover dse/{agent,v3_engine,_embedding_recall,_classification_gate,
regex_boost,mcs,deep_check}.py. DSEAgent (v3, BaseSpecialistAgent) replaces
the keyword-only stub: delegates MC-loading to the main engine
(rag_document_checker._load_controls), deterministic cached embedding recall
(reachability-gated), semantic-validator LLM layer honoring skip_llm,
third-country -> HIGH on documented transfer.
- Wire "dse" into _agent_outputs._TOPIC_AGENTS -> live check emits a validated
DSE tab (was snapshot/legacy-only).
- Tests rewritten for v3 (DB/embedding/LLM stubbed offline): regex-boost
detection, embedding-recall reachability guard, result->Finding conversion,
third-country HIGH; topic-wiring asserts "dse".
- deep_check.py recovered for preservation (alternate LLM-judge path, unwired).
Runtime data deps for full live behavior (note for prod): doc_check_controls
in DB + /data/mc_classification.db embedding sidecar + embedding-service; all
degrade gracefully (keyword layer carries) if absent.
dev-only, no deploy.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Linter: FORBIDDEN_OUTPUT_TERMS per Wortgrenze → 'Schutzgarantien'/'geeignete
Garantien' (Art. 46) passieren, 'garantiert'-Claims bleiben geblockt.
- DSE: L2-Detail wird übersprungen statt 'na', wenn die L1-Pflichtangabe fehlt
(kein irreführendes 'nicht anwendbar' für z.B. Transfermechanismus).
- DSE: Drittland → HIGH bei dokumentiertem Drittlandtransfer (scan_context via
AgentInput.context) — BMW (Konzern, US-Provider) ist kein weiches MEDIUM.
- DSE: Titel/Maßnahme kurz (treibt den Recommendation-Titel); ausführliche
Begründung als evidence — behebt 120-Zeichen-abgeschnittene Überschriften.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
DSEAgent wrappt die existierende ART13_CHECKLIST (33 kuratierte Pflichtangaben
L1 + Detailchecks L2) → strukturierter AgentOutput, NICHT der 90k-Library-
Firehose (eCall/Gesundheit/Telekom-Lärm). GET /snapshots/{id}/dse-check spiegelt
impressum-check; doc_input_from_snapshot generalisiert. Frontend: generischer
AgentModuleTab (lazy → AgentResultTab) für Impressum + DSE; DSE-Tab in der
Snapshot-Seite. Plus HRB-Pattern \d→\d+ (volle Registernummer als Beleg).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Benjamin Admin