Experimental Legal Grounding MCP Server

Vocabulary map of the corpus: obligation clusters (with support status), regulation terms, everyday-phrase bridges, concept bridges (recurring legal phrases the corpus uses but never defines — "public interest", "high risk" — routed to their operative clusters), and corpus boundaries (topics governed by law OUTSIDE this corpus, e.g. cookie consent under ePrivacy). Call this first to translate a user's question into cluster/term selectors, and to learn where honest answers require referral rather than assertion.

Evidence packets for legal topics. Selectors (union): cluster ids and/or term keys from get_catalog, and/or free_text (deterministic lexical+bridge matching only — no semantic search). Each packet returns primary rules PLUS their preconditions and exceptions with byte-exact quotes and stable rule ids. Every served rule includes its enumerated items and conditions with citable ids. Explicit term selectors return their definitions in packets or term_definitions. related_clusters names approved adjacent topics and explains each link. Mandatory packet content is never truncated. Use every part: presenting a duty without its packet's exceptions or preconditions will fail validate_claims. corpus_coverage tells you when the governing law is outside the corpus — qualify or refer, don't guess. If free_text matches nothing exactly, the response carries routing_candidates: clusters whose corpus vocabulary (section headings, cluster names, concept labels) overlaps your text. Inspect matched_vocabulary, choose semantically, and re-request that cluster explicitly — the service names candidates, it never guesses for you. Measured host loop (raises answerable coverage 25%→75%): if no packet returns, ask the user ONE intent question, then re-request with BOTH the chosen cluster AND the clarified free_text — combined selectors surface supplemental rules (e.g. Article 6(4) compatibility) that either selector alone misses. When clusters and free_text are combined, the text refines only within the explicit clusters and never adds another cluster. Before answering, check the substantive rule you need is actually in the response; if the packet is unavailable or incomplete, say so and refer — do not fill the gap from model memory.

Verify a drafted answer's legal citations before presenting it. Each claim: {claim_id, assertion, cluster_id, claim_type, claim_concept, citations: [{rule_id, quote}], propositions: [{duty_rule_ids, exception_rule_ids}]}. ALWAYS supply cluster_id for legal propositions — v0.3.8: WITHOUT it a claim cannot be grounded at all (verdict unclustered_claim), because membership, packet preconditions, required conditions and corpus-boundary warnings are all cluster-scoped. claim_type "exception_scope" is HOST-DECLARED, never inferred: use it only for a claim about an exception itself; it is valid only when every citation is exception-typed (else the claim receives invalid_claim), it exempts the exception-only/unanchored/precondition checks, and it is flagged in the result as host_declared_exception_scope. Any skipped ids are listed in exempted_by_declaration. Consumers MUST inspect that flag and treat a grounded declared claim as resting on the host's own declaration, not on ordinary grounding. claim_concept (optional) requires a concept-labeled cluster. concept_check lists matched, mismatched, and unlabelled citations; use all_citations_match as the strict signal. Verdicts (first match): invalid_claim (this claim violated a claim-scoped declaration rule; sibling claims still get results), unknown_reference (fabricated id), misquoted (canonical quote returned — fix and re-validate), wrong_provision (citation off-topic for the cluster), wrong_concept (you declared claim_concept and no cited labeled rule carries it — recheck your citations or your concept), outside_corpus (abstain/refer to the named external law), qualified_by_exception / qualified_by_precondition (add the returned exception/precondition citations and re-validate, or qualify your prose), exception_only_citations (ALL citations are carve-out rules and claim_type is not exception_scope; they cannot support an affirmative claim), unanchored_exception_citations (an exception's known target is not cited), unclustered_claim (no cluster_id — re-submit naming the claim's topic), mispaired_exception (a declared exception does not target its declared duty), unbound_propositions (strict linkage was requested for an ambiguous citation set without complete proposition declarations), citation_graph_consistent, unsupported. An exception's own conditions must be cited with it. Every cited condition requires its own parent rule. Every cited item of a non-exception rule requires its parent rule; exception items remain independently citable. This includes curated ids standing for items: an Article 6(1) basis requires the Article 6(1) chapeau cited with it. Conditions listed in required_conditions apply only to their named rule. If a submitted id is normalized, citation_checks returns the normalized rule_id and preserves the submitted id in cited_as. Known exception targets apply even when cluster_id is omitted. Every result also carries citation_status (valid/invalid/none) and semantic_status (always not_evaluated). IMPORTANT: citation_graph_consistent certifies that your CITATION SET contains valid relationships — it never certifies that your prose describes those relationships correctly, and it cannot detect prose that cross-pairs a cited exception with the wrong cited duty (round-11 "camouflage" finding: the same citation set can carry a true and a false sentence). Before presenting any claim involving an exception, COMPARE the reconciliation field's exception->limits edges against what your prose actually asserts; if your sentence pairs an exception with a duty that reconciliation does not list for it, your sentence is wrong even though the verdict is citation_graph_consistent. Prefer one claim object per duty-exception proposition over compound claims. Set strict_linkage true when every ambiguous exception-to-duty relationship must be declared and mechanically verified.