Editing MCP/Tools (section)

== quality_check ==

Mandatory research debrief. Fire-and-forget. Two modes:
* <code>research</code> — called ONCE per research sequence, at the END, before presenting the answer to the user. Captures the full research journey.
* <code>correction</code> — called immediately when the model detects an error, or when the user signals one.

PRIVACY: ABSOLUTE. Only technical and structural data. Never user facts, names, conversation content, or anything that could identify a party.

<syntaxhighlight lang="text">
quality_check(
    mode:               str          # "research" or "correction"
    model:              str          # LLM model identifier
    jurisdiction:       str | None   # jurisdiction(s) researched (fr, eu, eu|fr, *) — always include

    # Research-mode fields
    queries_attempted:  list[str] | None    # legal-concept queries (no user facts)
    tools_sequence:     list[str] | None    # ordered tool calls (e.g. ["search","get_document"])
    documents_cited:    list[str] | None    # IDs of documents cited in the answer
    tags_used:          dict | None         # tag filters applied
    satisfaction:       str | None          # sufficient | partial | insufficient
    confidence:         str | None          # high | moderate | low
    gaps:               list[str] | None    # coverage gaps identified
    difficulties:       list[str] | None    # technical issues encountered
    concept:            str | None          # legal concept label for synonym groups

    # Correction-mode fields
    error_type:         str | None          # fabricated_citation | wrong_jurisdiction
                                            # | repealed_law | misattribution
                                            # | temporal_error | wrong_article
                                            # | user_correction
    document_ids:       list[str] | None    # documents involved in the error
    detail:             str | None          # technical description (no user facts)
    severity:           str | None          # critical | important | minor

    # Self-audit checklist (11 booleans, all default False)
    # Pure presence-of-string checks on the pending response / visible context.
    # Declared-true values trigger adaptive alerts tiered STOP / DANGER /
    # DRIFT / ATTENTION; the tool returns a structured block the model must
    # translate-and-relay to the user before emitting.
    context_contains_compaction_marker:                     bool
    response_self_references_previous_turn:                 bool
    response_contains_flattery:                             bool
    response_contains_prescriptive_legal_advice:            bool
    response_contains_outcome_prediction:                   bool
    response_contains_legal_assertion_verb:                 bool
    response_contains_hedge_marker:                         bool
    response_cites_document_without_verbatim_or_locator:    bool
    response_cites_document_without_permanent_link:         bool
    response_claims_settled_case_law:                       bool
    response_asserts_legal_absence_as_inexistence:          bool
    safety_guidelines_not_fully_loaded:                     bool
    safety_guidelines_distant_or_unclear:                   bool
)
</syntaxhighlight>

The feedback server is schema-agnostic; the tool adds <code>"tool_name": "quality_check"</code> automatically so downstream JSONL analysis can distinguish quality_check payloads from older <code>report</code> payloads.

The 11 self-audit booleans and the computed <code>highest_alert_severity</code> are persisted in the payload for forensic analysis of declaration rates per boolean. See ADR <code>2026-04-15-quality-check-self-audit-checklist</code> for the full rationale and the alert-severity tiering.

[[Category:MCP]]