feat: verify Global Ask semantic claims with public evidence

AGENTS.md

@@ -177,12 +177,12 @@ Opening a Calendar commitment uses the same focus path (ADR 0094). Do not
 invent a week, a theta, a cutoff body, or a CalDAV event.
 Opening a Customer master related post uses the same focus path (ADR 0095).
 Do not invent a week, a theta, a cutoff body, a CalDAV event, or a customer.
- Opening an Ask Agent cited post uses the same focus path (ADR 0096).
- Do not invent a cited post.
- A linked Event Lineage node opened from that focused popup keeps the
- originating flags (ADR 0097). That open then focuses Keyman as the named
-next read (ADR 0100). Do not invent a week, a theta, a cutoff body,
- a CalDAV event, a customer, or a cited post.
+Opening an Ask Agent cited post uses the same focus path (ADR 0096). Do not
+invent a cited post.
+A linked Event Lineage node opened from that focused popup keeps the
+originating flags (ADR 0097). Do not invent a cited post.
+That open then focuses Keyman as the named next read (ADR 0100). Do not invent
+a week, a theta, a cutoff body, a CalDAV event, a customer, or a cited post.
 
 
 ## Tests

ARCHITECTURE.md

@@ -281,14 +281,13 @@ HTML. `src/api.ts` calls the FastAPI backend directly with the token
 Keycloak issued; `src/App.tsx` renders a git-branch SVG of
 `GET /api/lineage` (click a node to open that post; `post_admin` can
 rebuild), the post list with a named Weekly VOC ISO-8601 week filter
-(ADR 0092; opening that filtered post focuses Event Lineage, ADR 0093).
-Calendar commitments use the same Event Lineage focus path (ADR 0094).
+(ADR 0092; opening that filtered post focuses Event Lineage, ADR 0093),
+Calendar commitments use the same Event Lineage focus path (ADR 0094),
 Customer master related posts use the same Event Lineage focus path
 (ADR 0095). Ask Agent cited posts use the same Event Lineage focus path
-(ADR 0096). A linked Event Lineage node opened from a focused popup
-keeps those flags (ADR 0097) and then focuses Keyman as the named next
-read (ADR 0100).
-The full detail popup includes Korean
+(ADR 0096). A linked Event Lineage node opened from a focused popup keeps
+those flags (ADR 0097) and then focuses Keyman as the named next read
+(ADR 0100). The full detail popup includes Korean
 summary/key-events/R&R, VOC evidence excerpts, an Event Lineage panel
 (direct vs. indirect links; a link opens that post), the Keyman
 affiliate tree (resolved ancestors plus unresolved org roots), Keyman +

CHANGELOG.d/2.20.0-global-ask-public-verification.md

@@ -0,0 +1,7 @@
+# 2.20.0 — Global Ask public semantic verification
+
+- Global Ask may nominate source posts from persisted semantic, ontology, and Knowledge Graph evidence instead of requiring the buyer's term to appear in raw post text.
+- Multilingual contains-search uses one indexable predicate per semantic field and migration 0054 adds matching `pg_trgm` GIN indexes; concatenated expression scans are regression-tested out.
+- An explicit public-verification boundary uses SearXNG retrieval and contextual-orchestrator verification while keeping external URLs separate from internal post citations.
+- Private source evidence, Keyman/person facts, TEPP measurement artifacts, and fast-mlsirm measurement data are ineligible for public-search egress.
+- Public corroboration remains review evidence and never authority-promotes an inferred graph or ontology assertion.

CHANGELOG.d/2.20.0-project-history-contract.md

@@ -0,0 +1,7 @@
+# 2.20.0 Project-history response contract
+
+Project-history storage projections now emit the same strict evidence-bound
+shape that the HTTP response validates. Project identity display names retain
+their authoritative code/key, event source IDs and time basis are explicit,
+and the endpoint no longer fails with a validation error for an authorized
+timeline.

CHANGELOG.md

@@ -34,7 +34,7 @@ All notable changes to this project are documented here. Format follows
   and evaluation as the next read. After an authorized answer, Ask Agent
   names cited posts as current before that open. Home-list opens do not add
   that focus or copy. No TEPP theta is invented. No cited post is invented
-  (ADR 0096 / ADR 0039 / ADR 0016).
+(ADR 0096 / ADR 0039 / ADR 0016).
 
 ## [2.15.0] - 2026-08-19
 

backend/app/global_ask_retrieval.py

@@ -0,0 +1,240 @@
+"""Semantic and Knowledge-Graph candidate nomination for Global Ask.
+
+Candidate nomination is deliberately non-authoritative. This module returns
+post identifiers only. The caller must re-run the normal source-post visibility
+predicate before exposing body text or evidence.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Any
+
+import asyncpg
+
+from lineageweave.claim_verification import ontology_lookup_codes_for_question
+
+_STOP_WORDS = frozenset(
+    {
+        "which",
+        "what",
+        "where",
+        "when",
+        "who",
+        "why",
+        "how",
+        "the",
+        "this",
+        "that",
+        "posts",
+        "post",
+        "글",
+        "게시글",
+        "질문",
+        "관련",
+        "확인되는",
+        "핵심",
+        "사실",
+        "무엇",
+        "무엇인가요",
+        "인가요",
+    }
+)
+_TOKEN = re.compile(r"[^\W_]+(?:-[^\W_]+)*", re.UNICODE)
+_EVIDENCE_POST_IDS = re.compile(r"\[evidence_post_id=([^]]+)\]")
+
+
+def global_ask_query_terms(question: str | None, *, maximum_terms: int = 8) -> tuple[str, ...]:
+    """Return bounded, de-duplicated lexical terms from a Global Ask query."""
+
+    if maximum_terms <= 0:
+        return ()
+    return tuple(
+        dict.fromkeys(
+            token.casefold()
+            for token in _TOKEN.findall(question or "")
+            if len(token) >= 2 and token.casefold() not in _STOP_WORDS
+        )
+    )[:maximum_terms]
+
+
+async def semantic_candidate_post_ids(
+    conn: asyncpg.Connection,
+    question: str | None,
+    *,
+    maximum_candidates: int = 128,
+) -> list[str]:
+    """Nominate posts from persisted semantic and Knowledge-Graph evidence.
+
+    Project mentions, responsibility/affiliation evidence, Keyman names, and
+    organization/team catalogs use one ``ILIKE`` predicate per indexed column.
+    Search-corroborated raw/canonical organization-name pairs additionally
+    nominate direct organization mentions and affiliated people's posts; pending
+    or uncorroborated aliases do not. This preserves multilingual substring lookup
+    without wrapping indexed fields in an expression that forces a sequential
+    scan. Ontology lookup codes are applied only to graph lookup-code columns.
+    The function never returns source text and nomination never grants access.
+    """
+
+    if maximum_candidates <= 0:
+        return []
+    terms = global_ask_query_terms(question)
+    ontology_codes = ontology_lookup_codes_for_question(question or "")
+    if not terms and not ontology_codes:
+        return []
+    rows = await conn.fetch(
+        """
+        with query_terms as (
+            select unnest($1::text[]) as term
+        ), verified_organization as (
+            select distinct entity.corporate_entity_id
+              from organization_name_resolution resolution
+              join corporate_entity entity
+                on entity.entity_name = resolution.resolved_organization_name
+              join query_terms term
+                on resolution.raw_organization_name ilike '%' || term.term || '%'
+                or resolution.resolved_organization_name ilike '%' || term.term || '%'
+             where resolution.verification_status_code = 'verify_corroborated'
+        ), candidate_post as (
+            select mention.post_id, post.created_at
+              from post_project_mention mention
+              join source_post post on post.post_id = mention.post_id
+             where exists (
+                    select 1 from query_terms term
+                     where mention.project_name ilike '%' || term.term || '%'
+                        or mention.evidence_text ilike '%' || term.term || '%'
+                        or mention.ontology_iri ilike '%' || term.term || '%'
+                )
+            union all
+            select role.post_id, post.created_at
+              from post_summary_role role
+              join source_post post on post.post_id = role.post_id
+             where exists (
+                    select 1 from query_terms term
+                     where role.actor_name ilike '%' || term.term || '%'
+                        or role.responsibility ilike '%' || term.term || '%'
+                        or role.affiliated_organization_name ilike '%' || term.term || '%'
+                )
+            union all
+            select mention.post_id, post.created_at
+              from post_person_mention mention
+              join cataloged_person person on person.person_id = mention.person_id
+              join source_post post on post.post_id = mention.post_id
+             where exists (
+                    select 1 from query_terms term
+                     where person.person_name ilike '%' || term.term || '%'
+                        or person.last_known_job_title ilike '%' || term.term || '%'
+                        or mention.mention_context ilike '%' || term.term || '%'
+                )
+            union all
+            select mention.post_id, post.created_at
+              from post_organization_mention mention
+              join corporate_entity entity
+                on entity.corporate_entity_id = mention.corporate_entity_id
+              join source_post post on post.post_id = mention.post_id
+             where exists (
+                    select 1 from query_terms term
+                     where entity.entity_name ilike '%' || term.term || '%'
+                )
+            union all
+            select mention.post_id, post.created_at
+              from post_organization_mention mention
+              join verified_organization organization
+                on organization.corporate_entity_id = mention.corporate_entity_id
+              join source_post post on post.post_id = mention.post_id
+            union all
+            select mention.post_id, post.created_at
+              from post_person_mention mention
+              join person_affiliation affiliation
+                on affiliation.person_id = mention.person_id
+              join verified_organization organization
+                on organization.corporate_entity_id = affiliation.affiliated_corporate_entity_id
+              join source_post post on post.post_id = mention.post_id
+            union all
+            select mention.post_id, post.created_at
+              from post_team_mention mention
+              join cataloged_team team on team.team_id = mention.team_id
+              join source_post post on post.post_id = mention.post_id
+             where exists (
+                    select 1 from query_terms term
+                     where team.team_name ilike '%' || term.term || '%'
+                        or team.affiliated_organization_name ilike '%' || term.term || '%'
+                )
+            union all
+            select evidence.evidence_post_id as post_id, post.created_at
+              from knowledge_graph_edge edge
+              join knowledge_graph_edge_evidence evidence
+                on evidence.knowledge_graph_edge_id = edge.knowledge_graph_edge_id
+              join source_post post on post.post_id = evidence.evidence_post_id
+             where edge.edge_type_code = any($2::text[])
+                or edge.source_node_type_code = any($2::text[])
+                or edge.target_node_type_code = any($2::text[])
+        )
+        select post_id::text as post_id
+          from candidate_post
+         group by post_id
+         order by max(created_at) desc, post_id desc
+         limit $3
+        """,
+        list(terms),
+        list(ontology_codes),
+        maximum_candidates,
+    )
+    return list(dict.fromkeys(str(row["post_id"]) for row in rows))
+
+
+def graph_fact_evidence_post_ids(fact: str) -> frozenset[str]:
+    """Extract all persisted evidence-post identifiers named by one graph fact."""
+
+    match = _EVIDENCE_POST_IDS.search(fact)
+    if match is None:
+        return frozenset()
+    return frozenset(
+        value.strip() for value in match.group(1).split(",") if value.strip()
+    )
+
+
+def public_external_claim_facts(
+    row: Any,
+    semantic_facts: tuple[str, ...],
+    graph_facts: tuple[str, ...],
+    public_post_ids: frozenset[str],
+) -> tuple[str, ...]:
+    """Return externally searchable facts only for a public source post.
+
+    Source hints, people/Keyman facts, TEPP results, and fast-mlsirm reports are
+    absent by construction. A graph fact is eligible only if every persisted
+    evidence post named by that edge is public in the authorized result set.
+    """
+
+    if row.get("visibility_code") != "public":
+        return ()
+    public_graph_facts = tuple(
+        fact
+        for fact in graph_facts
+        if (evidence_ids := graph_fact_evidence_post_ids(fact))
+        and evidence_ids.issubset(public_post_ids)
+        and "node_person" not in fact
+    )
+    public_semantic_facts = tuple(
+        _public_semantic_fact(fact)
+        for fact in semantic_facts
+        if fact.startswith("project:")
+        and "node_person" not in fact
+        and not fact.startswith(("actor:", "Keyman mention:"))
+    )
+    return tuple(dict.fromkeys(public_semantic_facts + public_graph_facts))
+
+
+def _public_semantic_fact(fact: str) -> str:
+    """Keep only the assertion, never the internal project evidence excerpt."""
+
+    return fact.split(" | evidence:", 1)[0].strip()
+
+
+__all__ = [
+    "global_ask_query_terms",
+    "graph_fact_evidence_post_ids",
+    "public_external_claim_facts",
+    "semantic_candidate_post_ids",
+]