fix(classify): Ganymede review fixes — alias cleanup + counter accuracy + handle alignment

1. WARNING — orphan contributor_aliases after publisher/garbage delete:
   Added alias cleanup to the transaction (gated on --delete-events, same
   audit rationale as events). Both garbage and publisher deletion loops
   now DELETE matching contributor_aliases rows. Dry-run adds an orphan
   count diagnostic so the --delete-events decision is informed.

2. NIT — inserted_publishers counter over-reports on replay:
   INSERT OR IGNORE silently skips name collisions, but the counter
   incremented unconditionally. Now uses cur.rowcount so a second apply
   reports 0 inserts instead of falsely claiming 100. moved_to_publisher
   set remains unconditional — publisher rows already present still need
   the matching contributors row deleted.

3. NIT — handle-length gate diverged from writer path:
   Widened from {0,19} (20 chars) to {0,38} (39 chars) to match GitHub's
   handle limit and contributor.py::_HANDLE_RE. Prevents future long-handle
   real contributors from falling through to review_needed and blocking
   --apply. Current data has 0 review_needed either way.

Bonus (Q5): Added audit_log entry inside the transaction. One row in
audit_log.stage='schema_v26', event='classify_contributors' with counter
detail JSON on every --apply run. Cheap audit trail for the destructive op.

Verified end-to-end on VPS DB snapshot:
- First apply: 100/9/9/100/0 (matches pre-fix)
- Second apply: 0/9/0/0/0 (counter fix working)
- With injected aliases + --delete-events: 2 aliases deleted, 1 pre-existing
  orphan correctly left alone (outside script scope), audit_log entry
  written with accurate counters.

Ganymede msg-3. Protocol closed.

diagnostics/app.py

@@ -42,7 +42,7 @@ API_KEY_FILE = Path(os.environ.get("ARGUS_API_KEY_FILE", "/opt/teleo-eval/secret
 
 # Endpoints that skip auth (dashboard is public for now, can lock later)
 _PUBLIC_PATHS = frozenset({"/", "/prs", "/ops", "/health", "/agents", "/epistemic", "/legacy", "/audit", "/api/metrics", "/api/snapshots", "/api/vital-signs",
-                           "/api/contributors", "/api/domains", "/api/audit", "/api/yield", "/api/cost-per-claim", "/api/fix-rates", "/api/compute-profile", "/api/review-queue", "/api/daily-digest"})
+                           "/api/contributors", "/api/domains", "/api/audit", "/api/yield", "/api/cost-per-claim", "/api/fix-rates", "/api/compute-profile", "/api/review-queue", "/api/daily-digest", "/api/search"})
 
 
 def _get_db() -> sqlite3.Connection:
@@ -663,38 +663,115 @@ async def handle_api_domains(request):
     return web.json_response({"domains": breakdown})
 
 
-async def handle_api_search(request):
-    """GET /api/search — semantic search over claims via Qdrant + graph expansion.
+def _qdrant_hits_to_results(hits, include_expanded=False):
+    """Shape raw Qdrant hits into Ship's chat-API contract."""
+    results = []
+    for h in hits:
+        payload = h.get("payload", {}) or {}
+        path = payload.get("claim_path", "") or ""
+        slug = path.rsplit("/", 1)[-1]
+        if slug.endswith(".md"):
+            slug = slug[:-3]
+        results.append({
+            "slug": slug,
+            "path": path,
+            "title": payload.get("claim_title", ""),
+            "domain": payload.get("domain"),
+            "confidence": payload.get("confidence"),
+            "score": round(float(h.get("score", 0.0) or 0.0), 4),
+            "body_excerpt": payload.get("snippet", "") or "",
+        })
+    return results
 
-    Query params:
-      q:          search query (required)
-      domain:     filter by domain (optional)
-      confidence: filter by confidence level (optional)
-      limit:      max results, default 10 (optional)
-      exclude:    comma-separated claim paths to exclude (optional)
-      expand:     enable graph expansion, default true (optional)
+
+async def handle_api_search(request):
+    """Semantic search over claims via Qdrant.
+
+    POST contract (Ship's chat API):
+      body: {"query": str, "limit": int, "min_score": float?, "domain": str?, "confidence": str?, "exclude": [str]?}
+      response: {"query": str, "results": [{"slug","path","title","domain","confidence","score","body_excerpt"}], "total": int}
+
+    GET (legacy + hackathon debug):
+      q: search query (required)
+      limit, domain, confidence, exclude, expand
+      min_score: if set, bypasses two-pass lib threshold (default lib behavior otherwise)
     """
+    if request.method == "POST":
+        try:
+            body = await request.json()
+        except Exception:
+            return web.json_response({"error": "invalid JSON body"}, status=400)
+
+        query = (body.get("query") or "").strip()
+        if not query:
+            return web.json_response({"error": "query required"}, status=400)
+
+        try:
+            limit = min(int(body.get("limit") or 5), 50)
+        except (TypeError, ValueError):
+            return web.json_response({"error": "limit must be int"}, status=400)
+        try:
+            min_score = float(body.get("min_score") if body.get("min_score") is not None else 0.25)
+        except (TypeError, ValueError):
+            return web.json_response({"error": "min_score must be float"}, status=400)
+
+        domain = body.get("domain")
+        confidence = body.get("confidence")
+        exclude = body.get("exclude") or None
+
+        vector = embed_query(query)
+        if vector is None:
+            return web.json_response({"error": "embedding failed"}, status=502)
+
+        hits = search_qdrant(vector, limit=limit, domain=domain,
+                             confidence=confidence, exclude=exclude,
+                             score_threshold=min_score)
+        results = _qdrant_hits_to_results(hits)
+        return web.json_response({"query": query, "results": results, "total": len(results)})
+
+    # GET path
     query = request.query.get("q", "").strip()
     if not query:
         return web.json_response({"error": "q parameter required"}, status=400)
 
     domain = request.query.get("domain")
     confidence = request.query.get("confidence")
-    limit = min(int(request.query.get("limit", "10")), 50)
+    try:
+        limit = min(int(request.query.get("limit", "10")), 50)
+    except ValueError:
+        return web.json_response({"error": "limit must be int"}, status=400)
     exclude_raw = request.query.get("exclude", "")
     exclude = [p.strip() for p in exclude_raw.split(",") if p.strip()] if exclude_raw else None
     expand = request.query.get("expand", "true").lower() != "false"
+    min_score_raw = request.query.get("min_score")
 
-    # Use shared search library (Layer 1 + Layer 2)
+    if min_score_raw is not None:
+        try:
+            min_score = float(min_score_raw)
+        except ValueError:
+            return web.json_response({"error": "min_score must be float"}, status=400)
+        vector = embed_query(query)
+        if vector is None:
+            return web.json_response({"error": "embedding failed"}, status=502)
+        hits = search_qdrant(vector, limit=limit, domain=domain,
+                             confidence=confidence, exclude=exclude,
+                             score_threshold=min_score)
+        direct = _qdrant_hits_to_results(hits)
+        return web.json_response({
+            "query": query,
+            "direct_results": direct,
+            "expanded_results": [],
+            "total": len(direct),
+        })
+
+    # Default GET: Layer 1 + Layer 2 via lib
     result = kb_search(query, expand=expand,
                        domain=domain, confidence=confidence, exclude=exclude)
-
     if "error" in result:
         error = result["error"]
         if error == "embedding_failed":
             return web.json_response({"error": "embedding failed"}, status=502)
         return web.json_response({"error": error}, status=500)
-
     return web.json_response(result)
 
 
@@ -2268,6 +2345,7 @@ def create_app() -> web.Application:
     app.router.add_get("/api/contributors", handle_api_contributors)
     app.router.add_get("/api/domains", handle_api_domains)
     app.router.add_get("/api/search", handle_api_search)
+    app.router.add_post("/api/search", handle_api_search)
     app.router.add_get("/api/audit", handle_api_audit)
     app.router.add_get("/audit", handle_audit_page)
     app.router.add_post("/api/usage", handle_api_usage)

lib/attribution.py

@@ -21,6 +21,14 @@ logger = logging.getLogger("pipeline.attribution")
 
 VALID_ROLES = frozenset({"sourcer", "extractor", "challenger", "synthesizer", "reviewer"})
 
+# Agent-owned branch prefixes — PRs from these branches get Pentagon-Agent trailer
+# credit for challenger/synthesizer roles. Pipeline-infra branches (extract/ reweave/
+# fix/ ingestion/) are deliberately excluded: they're automation, not contribution.
+# Single source of truth; imported by contributor.py and backfill-events.py.
+AGENT_BRANCH_PREFIXES = (
+    "rio/", "theseus/", "leo/", "vida/", "astra/", "clay/", "oberon/",
+)
+
 # Handle sanity: lowercase alphanumerics, hyphens, underscores. 1-39 chars (matches
 # GitHub's handle rules). Rejects garbage like "governance---meritocratic-voting-+-futarchy"
 # or "sec-interpretive-release-s7-2026-09-(march-17" that upstream frontmatter hygiene

lib/contributor.py

@@ -14,7 +14,7 @@ import logging
 import re
 
 from . import config, db
-from .attribution import classify_kind, normalize_handle
+from .attribution import AGENT_BRANCH_PREFIXES, classify_kind, normalize_handle
 from .forgejo import get_pr_diff
 
 logger = logging.getLogger("pipeline.contributor")
@@ -186,7 +186,7 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
     # Refine commit_type from diff content (branch prefix may be too broad)
     row = conn.execute(
         "SELECT commit_type, submitted_by, domain, source_channel, leo_verdict, "
-        "domain_verdict, domain_agent FROM prs WHERE number = ?",
+        "domain_verdict, domain_agent, merged_at FROM prs WHERE number = ?",
         (pr_number,),
     ).fetchone()
     branch_type = row["commit_type"] if row and row["commit_type"] else "extract"
@@ -199,6 +199,10 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
     pr_domain = row["domain"] if row else None
     pr_channel = row["source_channel"] if row else None
     pr_submitted_by = row["submitted_by"] if row else None
+    # Use the PR's merged_at timestamp so event time matches the actual merge.
+    # If a merge retries after a crash, this keeps forward-emitted and backfilled
+    # events on the same timeline. Falls back to datetime('now') in the writer.
+    pr_merged_at = row["merged_at"] if row and row["merged_at"] else None
 
     # ── AUTHOR event (schema v24, double-write) ──
     # Humans-are-always-author rule: the human in the loop gets author credit.
@@ -211,27 +215,33 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
     if pr_submitted_by:
         author_candidate = pr_submitted_by
     else:
-        # External GitHub PRs: git author is the real submitter.
-        rc_author_head, author_head = await git_fn(
+        # External GitHub PRs: git author of the FIRST commit on the branch is
+        # the real submitter. `git log -1` would return the latest commit, which
+        # mis-credits multi-commit PRs where a reviewer rebased or force-pushed.
+        # Take the last line of the unreversed log (= oldest commit, since git
+        # log defaults to reverse-chronological). Ganymede review, Apr 24.
+        rc_author_log, author_log = await git_fn(
             "log", f"origin/main..origin/{branch}", "--no-merges",
-            "--format=%an", "-1", timeout=5,
+            "--format=%an", timeout=5,
         )
-        if rc_author_head == 0 and author_head.strip():
-            candidate = author_head.strip().lower()
-            if candidate and candidate not in {"teleo", "teleo-bot", "pipeline",
-                                               "github-actions[bot]", "forgejo-actions"}:
-                author_candidate = candidate
+        if rc_author_log == 0 and author_log.strip():
+            lines = [line for line in author_log.strip().split("\n") if line.strip()]
+            if lines:
+                candidate = lines[-1].strip().lower()
+                if candidate and candidate not in {"teleo", "teleo-bot", "pipeline",
+                                                   "github-actions[bot]", "forgejo-actions"}:
+                    author_candidate = candidate
         # Agent-owned branches with no submitted_by: theseus/research-*, leo/*, etc.
-        if not author_candidate and "/" in branch:
-            prefix = branch.split("/", 1)[0]
-            # Exclude pipeline-infrastructure prefixes — those are not authors.
-            if prefix in ("rio", "theseus", "leo", "vida", "clay", "astra", "oberon"):
-                author_candidate = prefix
+        if not author_candidate and branch.startswith(AGENT_BRANCH_PREFIXES):
+            # Autonomous agent PR (theseus/research-*, leo/entity-*, etc.) —
+            # credit goes to the agent as author per Cory's directive.
+            author_candidate = branch.split("/", 1)[0]
 
     if author_candidate:
         insert_contribution_event(
             conn, author_candidate, "author", pr_number,
             claim_path=None, domain=pr_domain, channel=pr_channel,
+            timestamp=pr_merged_at,
         )
 
     # ── EVALUATOR events (schema v24) ──
@@ -243,6 +253,7 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
             insert_contribution_event(
                 conn, "leo", "evaluator", pr_number,
                 claim_path=None, domain=pr_domain, channel=pr_channel,
+                timestamp=pr_merged_at,
             )
         if row["domain_verdict"] == "approve" and row["domain_agent"]:
             dagent = row["domain_agent"].strip().lower()
@@ -250,6 +261,7 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
                 insert_contribution_event(
                     conn, dagent, "evaluator", pr_number,
                     claim_path=None, domain=pr_domain, channel=pr_channel,
+                    timestamp=pr_merged_at,
                 )
 
     # Parse Pentagon-Agent trailer from branch commit messages
@@ -257,9 +269,7 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
     # Agent-owned branches (theseus/*, rio/*, etc.) give the trailer-named agent
     # challenger/synthesizer credit based on refined commit_type. Pipeline-owned
     # branches (extract/*, reweave/*, etc.) don't — those are infra, not work.
-    _AGENT_BRANCH_PREFIXES = ("rio/", "theseus/", "leo/", "vida/", "clay/",
-                              "astra/", "oberon/")
-    is_agent_branch = branch.startswith(_AGENT_BRANCH_PREFIXES)
+    is_agent_branch = branch.startswith(AGENT_BRANCH_PREFIXES)
     _TRAILER_EVENT_ROLE = {
         "challenge": "challenger",
         "enrich": "synthesizer",
@@ -285,6 +295,7 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
                 insert_contribution_event(
                     conn, agent_name, event_role, pr_number,
                     claim_path=None, domain=pr_domain, channel=pr_channel,
+                    timestamp=pr_merged_at,
                 )
             agents_found.add(agent_name)
 
@@ -346,6 +357,7 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
                                     conn, handle, "originator", pr_number,
                                     claim_path=rel_path,
                                     domain=pr_domain, channel=pr_channel,
+                                    timestamp=pr_merged_at,
                                 )
 
     # Fallback: if no Pentagon-Agent trailer found, try git commit authors
@@ -364,13 +376,35 @@ async def record_contributor_attribution(conn, pr_number: int, branch: str, git_
                 if author_name and author_name not in _BOT_AUTHORS:
                     role = commit_type_to_role(refined_type)
                     upsert_contributor(conn, author_name, None, role, today)
+                    # Event-model parity: emit challenger/synthesizer event when
+                    # the fallback credits a human/agent for that kind of work.
+                    # Without this, external-contributor challenge/enrich PRs
+                    # accumulate legacy counts but disappear from event-sourced
+                    # leaderboards when Phase B cuts over. (Ganymede review.)
+                    event_role_fb = _TRAILER_EVENT_ROLE.get(refined_type)
+                    if event_role_fb:
+                        insert_contribution_event(
+                            conn, author_name, event_role_fb, pr_number,
+                            claim_path=None, domain=pr_domain, channel=pr_channel,
+                            timestamp=pr_merged_at,
+                        )
                     agents_found.add(author_name)
 
         if not agents_found:
-            row = conn.execute("SELECT agent FROM prs WHERE number = ?", (pr_number,)).fetchone()
-            if row and row["agent"] and row["agent"] != "external":
+            fb_row = conn.execute(
+                "SELECT agent FROM prs WHERE number = ?", (pr_number,)
+            ).fetchone()
+            if fb_row and fb_row["agent"] and fb_row["agent"] != "external":
+                pr_agent = fb_row["agent"].lower()
                 role = commit_type_to_role(refined_type)
-                upsert_contributor(conn, row["agent"].lower(), None, role, today)
+                upsert_contributor(conn, pr_agent, None, role, today)
+                event_role_fb = _TRAILER_EVENT_ROLE.get(refined_type)
+                if event_role_fb:
+                    insert_contribution_event(
+                        conn, pr_agent, event_role_fb, pr_number,
+                        claim_path=None, domain=pr_domain, channel=pr_channel,
+                        timestamp=pr_merged_at,
+                    )
 
 
 def upsert_contributor(

lib/db.py

@@ -9,7 +9,7 @@ from . import config
 
 logger = logging.getLogger("pipeline.db")
 
-SCHEMA_VERSION = 24
+SCHEMA_VERSION = 26
 
 SCHEMA_SQL = """
 CREATE TABLE IF NOT EXISTS schema_version (
@@ -35,6 +35,15 @@ CREATE TABLE IF NOT EXISTS sources (
     feedback TEXT,
     -- eval feedback for re-extraction (JSON)
     cost_usd REAL DEFAULT 0,
+    -- v26: provenance — publisher (news org / venue) + content author.
+    -- publisher_id references publishers(id) when source is from a known org.
+    -- original_author_handle references contributors(handle) when author is in our system.
+    -- original_author is free-text fallback ("Kim et al.", "Robin Hanson") — not credit-bearing.
+    publisher_id INTEGER REFERENCES publishers(id),
+    content_type TEXT,
+    -- article | paper | tweet | conversation | self_authored | webpage | podcast
+    original_author TEXT,
+    original_author_handle TEXT REFERENCES contributors(handle),
     created_at TEXT DEFAULT (datetime('now')),
     updated_at TEXT DEFAULT (datetime('now'))
 );
@@ -207,6 +216,33 @@ CREATE TABLE IF NOT EXISTS contributor_aliases (
     created_at TEXT DEFAULT (datetime('now'))
 );
 CREATE INDEX IF NOT EXISTS idx_aliases_canonical ON contributor_aliases(canonical);
+
+-- Publishers: news orgs, academic venues, social platforms. NOT contributors — these
+-- provide metadata/provenance for sources, never earn leaderboard credit. Separating
+-- these from contributors prevents CNBC/SpaceNews from dominating the leaderboard.
+-- (Apr 24 Cory directive: "only credit the original source if its on X or tg")
+CREATE TABLE IF NOT EXISTS publishers (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    name TEXT NOT NULL UNIQUE,
+    kind TEXT CHECK(kind IN ('news', 'academic', 'social_platform', 'podcast', 'self', 'internal', 'legal', 'government', 'research_org', 'commercial', 'other')),
+    url_pattern TEXT,
+    created_at TEXT DEFAULT (datetime('now'))
+);
+CREATE INDEX IF NOT EXISTS idx_publishers_name ON publishers(name);
+CREATE INDEX IF NOT EXISTS idx_publishers_kind ON publishers(kind);
+
+-- Multi-platform identity: one contributor, many handles. Enables the leaderboard to
+-- unify @thesensatore (X) + thesensatore (TG) + thesensatore@github into one person.
+-- Writers check this table after resolving aliases to find canonical contributor handle.
+CREATE TABLE IF NOT EXISTS contributor_identities (
+    contributor_handle TEXT NOT NULL,
+    platform TEXT NOT NULL CHECK(platform IN ('x', 'telegram', 'github', 'email', 'web', 'internal')),
+    platform_handle TEXT NOT NULL,
+    verified INTEGER DEFAULT 0,
+    created_at TEXT DEFAULT (datetime('now')),
+    PRIMARY KEY (platform, platform_handle)
+);
+CREATE INDEX IF NOT EXISTS idx_identities_contributor ON contributor_identities(contributor_handle);
 """
 
 
@@ -737,9 +773,13 @@ def migrate(conn: sqlite3.Connection):
             ],
         )
         # Seed kind='agent' for known Pentagon agents so the events writer picks it up.
+        # Must stay in sync with lib/attribution.PENTAGON_AGENTS — drift causes
+        # contributors.kind to disagree with classify_kind() output for future
+        # inserts. (Ganymede review: "pipeline" was missing until Apr 24.)
         pentagon_agents = [
             "rio", "leo", "theseus", "vida", "clay", "astra",
             "oberon", "argus", "rhea", "ganymede", "epimetheus", "hermes", "ship",
+            "pipeline",
         ]
         for agent in pentagon_agents:
             conn.execute(
@@ -749,6 +789,62 @@ def migrate(conn: sqlite3.Connection):
         conn.commit()
         logger.info("Migration v24: added contribution_events + contributor_aliases tables, kind column")
 
+    if current < 25:
+        # v24 seeded 13 Pentagon agents but missed "pipeline" — classify_kind()
+        # treats it as agent so contributors.kind drifted from event-insert output.
+        # Idempotent corrective UPDATE: fresh installs have no "pipeline" row
+        # (no-op), upgraded envs flip it if it exists. (Ganymede review Apr 24.)
+        conn.execute(
+            "UPDATE contributors SET kind = 'agent' WHERE handle = 'pipeline'"
+        )
+        conn.commit()
+        logger.info("Migration v25: patched kind='agent' for pipeline handle")
+
+    if current < 26:
+        # Add publishers + contributor_identities. Non-breaking — new tables only.
+        # No existing data moved. Classification into publishers happens via a
+        # separate script (scripts/reclassify-contributors.py) with Cory-reviewed
+        # seed list. CHECK constraint on contributors.kind deferred to v27 after
+        # classification completes. (Apr 24 Cory directive: "fix schema, don't
+        # filter output" — separate contributors from publishers at the data layer.)
+        conn.executescript("""
+            CREATE TABLE IF NOT EXISTS publishers (
+                id INTEGER PRIMARY KEY AUTOINCREMENT,
+                name TEXT NOT NULL UNIQUE,
+                kind TEXT CHECK(kind IN ('news', 'academic', 'social_platform', 'podcast', 'self', 'internal', 'legal', 'government', 'research_org', 'commercial', 'other')),
+                url_pattern TEXT,
+                created_at TEXT DEFAULT (datetime('now'))
+            );
+            CREATE INDEX IF NOT EXISTS idx_publishers_name ON publishers(name);
+            CREATE INDEX IF NOT EXISTS idx_publishers_kind ON publishers(kind);
+
+            CREATE TABLE IF NOT EXISTS contributor_identities (
+                contributor_handle TEXT NOT NULL,
+                platform TEXT NOT NULL CHECK(platform IN ('x', 'telegram', 'github', 'email', 'web', 'internal')),
+                platform_handle TEXT NOT NULL,
+                verified INTEGER DEFAULT 0,
+                created_at TEXT DEFAULT (datetime('now')),
+                PRIMARY KEY (platform, platform_handle)
+            );
+            CREATE INDEX IF NOT EXISTS idx_identities_contributor ON contributor_identities(contributor_handle);
+        """)
+        # Extend sources with provenance columns. ALTER TABLE ADD COLUMN is
+        # idempotent-safe via try/except because SQLite doesn't support IF NOT EXISTS
+        # on column adds.
+        for col_sql in (
+            "ALTER TABLE sources ADD COLUMN publisher_id INTEGER REFERENCES publishers(id)",
+            "ALTER TABLE sources ADD COLUMN content_type TEXT",
+            "ALTER TABLE sources ADD COLUMN original_author TEXT",
+            "ALTER TABLE sources ADD COLUMN original_author_handle TEXT REFERENCES contributors(handle)",
+        ):
+            try:
+                conn.execute(col_sql)
+            except sqlite3.OperationalError as e:
+                if "duplicate column" not in str(e).lower():
+                    raise
+        conn.commit()
+        logger.info("Migration v26: added publishers + contributor_identities tables + sources provenance columns")
+
     if current < SCHEMA_VERSION:
         conn.execute(
             "INSERT OR REPLACE INTO schema_version (version) VALUES (?)",

scripts/backfill-events.py

@@ -46,8 +46,11 @@ PENTAGON_AGENTS = frozenset({
     "pipeline",
 })
 
-AGENT_BRANCH_PREFIXES = ("rio/", "theseus/", "leo/", "vida/", "clay/",
-                         "astra/", "oberon/")
+# Keep in sync with lib/attribution.AGENT_BRANCH_PREFIXES.
+# Duplicated here because this script runs standalone (no pipeline package import).
+AGENT_BRANCH_PREFIXES = (
+    "rio/", "theseus/", "leo/", "vida/", "astra/", "clay/", "oberon/",
+)
 
 TRAILER_EVENT_ROLE = {
     "challenge": "challenger",
@@ -196,6 +199,175 @@ def derive_author(conn: sqlite3.Connection, pr: dict) -> str | None:
     return None
 
 
+def find_pr_for_claim(
+    conn: sqlite3.Connection,
+    repo: Path,
+    md: Path,
+) -> tuple[int | None, str]:
+    """Recover the Forgejo PR number that introduced a claim file.
+
+    Returns (pr_number, strategy) — strategy is one of:
+      'sourced_from' — frontmatter sourced_from matched prs.source_path
+      'git_subject'  — git log first-add commit message matched a branch pattern
+      'title_desc'   — filename stem matched a title in prs.description
+      'github_pr'    — recovery commit mentioned GitHub PR # → prs.github_pr
+      'none'         — no strategy found a match
+
+    Order is chosen by reliability:
+      1. sourced_from (explicit provenance, most reliable when present)
+      2. git_subject  (covers Leo research, Cameron challenges, Theseus contrib)
+      3. title_desc   (current fallback — brittle when description is NULL)
+      4. github_pr    (recovery commits referencing erased GitHub PRs)
+    """
+    rel = str(md.relative_to(repo))
+
+    # Strategy 1: sourced_from frontmatter → prs.source_path
+    try:
+        content = md.read_text(encoding="utf-8")
+    except (FileNotFoundError, PermissionError, UnicodeDecodeError):
+        content = ""
+    fm = parse_frontmatter(content) if content else None
+    if fm:
+        sourced = fm.get("sourced_from")
+        candidate_paths: list[str] = []
+        if isinstance(sourced, str) and sourced:
+            candidate_paths.append(sourced)
+        elif isinstance(sourced, list):
+            candidate_paths.extend(s for s in sourced if isinstance(s, str))
+        for sp in candidate_paths:
+            stem = Path(sp).stem
+            if not stem:
+                continue
+            row = conn.execute(
+                """SELECT number FROM prs
+                   WHERE source_path LIKE ? AND status='merged'
+                   ORDER BY merged_at ASC LIMIT 1""",
+                (f"%{stem}.md",),
+            ).fetchone()
+            if row:
+                return row["number"], "sourced_from"
+
+    # Strategy 2: git log first-add commit → subject pattern → prs.branch
+    # Default log order is reverse-chronological; take the last line (oldest)
+    # to get the original addition, not later rewrites.
+    log_out = git(
+        "log", "--diff-filter=A", "--follow",
+        "--format=%H|||%s|||%b", "--", rel,
+    )
+    if log_out.strip():
+        # Split on the delimiter we chose. Each commit produces 3 fields but
+        # %b can contain blank lines — group by lines that look like a SHA.
+        blocks: list[tuple[str, str, str]] = []
+        current: list[str] = []
+        for line in log_out.splitlines():
+            if re.match(r"^[a-f0-9]{40}\|\|\|", line):
+                if current:
+                    parts = "\n".join(current).split("|||", 2)
+                    if len(parts) == 3:
+                        blocks.append((parts[0], parts[1], parts[2]))
+                current = [line]
+            else:
+                current.append(line)
+        if current:
+            parts = "\n".join(current).split("|||", 2)
+            if len(parts) == 3:
+                blocks.append((parts[0], parts[1], parts[2]))
+        if blocks:
+            # Oldest addition — git log defaults to reverse-chronological
+            _oldest_sha, subject, body = blocks[-1]
+
+            # Pattern: "<agent>: extract claims from <slug>"
+            m = re.match(r"^(\w+):\s*extract\s+claims\s+from\s+(\S+)", subject)
+            if m:
+                slug = m.group(2).rstrip(".md").rstrip(".")
+                row = conn.execute(
+                    """SELECT number FROM prs
+                       WHERE branch LIKE ? AND status='merged'
+                       ORDER BY merged_at ASC LIMIT 1""",
+                    (f"extract/{slug}%",),
+                ).fetchone()
+                if row:
+                    return row["number"], "git_subject"
+
+            # Pattern: "<agent>: research session <date>"
+            m = re.match(r"^(\w+):\s*research\s+session\s+(\d{4}-\d{2}-\d{2})", subject)
+            if m:
+                agent = m.group(1).lower()
+                date = m.group(2)
+                row = conn.execute(
+                    """SELECT number FROM prs
+                       WHERE branch LIKE ? AND status='merged'
+                       ORDER BY merged_at ASC LIMIT 1""",
+                    (f"{agent}/research-{date}%",),
+                ).fetchone()
+                if row:
+                    return row["number"], "git_subject"
+
+            # Pattern: "<agent>: challenge" / contrib challenges / entity batches
+            m = re.match(r"^(\w+):\s*(?:challenge|contrib|entity|synthesize)", subject)
+            if m:
+                agent = m.group(1).lower()
+                row = conn.execute(
+                    """SELECT number FROM prs
+                       WHERE branch LIKE ? AND status='merged'
+                       ORDER BY merged_at ASC LIMIT 1""",
+                    (f"{agent}/%",),
+                ).fetchone()
+                if row:
+                    return row["number"], "git_subject"
+
+            # Recovery commits referencing erased GitHub PRs (Alex/Cameron).
+            # Subject: "Recover <who> contribution from GitHub PR #NN (...)".
+            # Match only when a corresponding prs row exists with github_pr=NN —
+            # otherwise the claims were direct-to-main without a Forgejo PR
+            # record, which requires a synthetic PR row (follow-up, not in
+            # this script's scope).
+            gh_match = re.search(r"GitHub\s+PR\s+#(\d+)", subject + "\n" + body)
+            if gh_match:
+                gh_pr = int(gh_match.group(1))
+                row = conn.execute(
+                    "SELECT number FROM prs WHERE github_pr = ? AND status='merged' LIMIT 1",
+                    (gh_pr,),
+                ).fetchone()
+                if row:
+                    return row["number"], "github_pr"
+
+            # Pattern: bare "Extract N claims from <source-fragment>" (no
+            # agent prefix). Used in early research PRs like Shaga's claims
+            # at PR #2025. Fall back to time-proximity: find the earliest
+            # agent-branch PR merged within 24h AFTER this commit's date.
+            m = re.match(r"^Extract\s+\d+\s+claims\s+from\b", subject)
+            if m:
+                # Get commit author date
+                date_out = git(
+                    "log", "-1", "--format=%aI", _oldest_sha, timeout=10,
+                )
+                commit_date = date_out.strip() if date_out.strip() else None
+                if commit_date:
+                    # git %aI returns ISO 8601 with T-separator; prs.merged_at
+                    # uses SQLite's space-separator. Lexicographic comparison
+                    # fails across formats (space<T), so normalize commit_date
+                    # via datetime() before comparing. Without this, PRs merged
+                    # within the same calendar day but earlier than the commit
+                    # hour are silently excluded (caught by Ganymede review —
+                    # Shaga's #2025 was dropped in favor of later #2032).
+                    row = conn.execute(
+                        """SELECT number FROM prs
+                           WHERE status='merged'
+                             AND merged_at >= datetime(?)
+                             AND merged_at <= datetime(datetime(?), '+24 hours')
+                             AND (branch LIKE 'leo/%' OR branch LIKE 'theseus/%'
+                                  OR branch LIKE 'rio/%' OR branch LIKE 'astra/%'
+                                  OR branch LIKE 'vida/%' OR branch LIKE 'clay/%')
+                           ORDER BY merged_at ASC LIMIT 1""",
+                        (commit_date, commit_date),
+                    ).fetchone()
+                    if row:
+                        return row["number"], "git_time_proximity"
+
+    return None, "none"
+
+
 def emit(conn, counts, dry_run, handle, role, pr_number, claim_path, domain, channel, timestamp):
     canonical = normalize_handle(conn, handle)
     if not valid_handle(canonical):
@@ -332,8 +504,10 @@ def main():
     if not args.dry_run:
         conn.commit()
 
-    print("\n=== Summary ===")
-    for role in ("author", "originator", "challenger", "synthesizer", "evaluator"):
+    # Originator is emitted in the claim-level pass below, not the PR-level pass.
+    # Previous summary listed it here with attempted=0 which confused operators.
+    print("\n=== PR-level events (author, evaluator, challenger, synthesizer) ===")
+    for role in ("author", "challenger", "synthesizer", "evaluator"):
         att = counts[(role, "attempt")]
         if args.dry_run:
             wi = counts[(role, "would_insert")]
@@ -343,18 +517,16 @@ def main():
             skip = counts[(role, "skipped_dup")]
             print(f"  {role:12s} attempted={att:5d} inserted={ins:5d} skipped_dup={skip:5d}")
 
-    if not args.dry_run:
-        total = conn.execute("SELECT COUNT(*) FROM contribution_events").fetchone()[0]
-        print(f"\nTotal contribution_events rows: {total}")
-
     # ── Per-claim originator pass ──
-    # Separate pass: walk the current knowledge tree, parse sourcer frontmatter,
-    # and attach each claim to the merging PR via a claim_path → pr_number map
-    # built from prs.description (pipe-separated claim titles). Imperfect — some
-    # PRs have NULL description or mismatched titles — but recovers the bulk of
-    # historical originator credit.
+    # Walk the knowledge tree, parse sourcer attribution, and attach each claim
+    # to its merging PR via find_pr_for_claim's multi-strategy recovery.
+    # Apr 24 rewrite (Ganymede-approved): replaces the single-strategy
+    # title→description match with four strategies in reliability order.
+    # Previous script missed PRs with NULL description (Cameron #3377) and
+    # cross-context claims (Shaga's Leo research). Fallback title-match is
+    # preserved to recover anything the git-log path misses.
     print("\n=== Claim-level originator pass ===")
-    # Build title → pr_number map from prs.description
+    # Build title → pr_number map from prs.description (strategy 3 fallback)
     title_to_pr: dict[str, int] = {}
     for r in conn.execute(
         "SELECT number, description FROM prs WHERE status='merged' AND description IS NOT NULL AND description != ''"
@@ -367,6 +539,7 @@ def main():
                 title_to_pr[title.lower()] = r["number"]
 
     claim_counts = Counter()
+    strategy_counts = Counter()
     claim_count = 0
     originator_count = 0
     for md in sorted(repo.glob("domains/**/*.md")) + \
@@ -374,13 +547,19 @@ def main():
               sorted(repo.glob("foundations/**/*.md")) + \
               sorted(repo.glob("decisions/**/*.md")):
         rel = str(md.relative_to(repo))
-        # Match via filename stem (with spaces and hyphens) against description titles
         stem = md.stem
-        # Multiple matching strategies
-        pr_number = title_to_pr.get(stem.lower())
+
+        # Strategies 1, 2, 4 via the helper (sourced_from, git_subject, github_pr).
+        pr_number, strategy = find_pr_for_claim(conn, repo, md)
+
+        # Strategy 3 (fallback): title-match against prs.description.
         if not pr_number:
-            # Hyphenated slug → space variant
-            pr_number = title_to_pr.get(stem.replace("-", " ").lower())
+            pr_number = title_to_pr.get(stem.lower())
+            if not pr_number:
+                pr_number = title_to_pr.get(stem.replace("-", " ").lower())
+            if pr_number:
+                strategy = "title_desc"
+
         if not pr_number:
             claim_counts["no_pr_match"] += 1
             continue
@@ -391,6 +570,7 @@ def main():
             continue
 
         claim_count += 1
+        strategy_counts[strategy] += 1
         # Look up author for this PR to skip self-credit
         pr_row = conn.execute(
             "SELECT submitted_by, branch, domain, source_channel, merged_at FROM prs WHERE number = ?",
@@ -419,12 +599,19 @@ def main():
     print(f"  Claims processed: {claim_count}")
     print(f"  Originator events emitted: {originator_count}")
     print(f"  Breakdown: {dict(claim_counts)}")
-
-    final_origin_attempted = counts[("originator", "attempt")]
+    print(f"  Strategy hits: {dict(strategy_counts)}")
+    att = counts[("originator", "attempt")]
     if args.dry_run:
-        print(f"  (dry-run) originator would_insert={counts[('originator', 'would_insert')]}")
+        wi = counts[("originator", "would_insert")]
+        print(f"  {'originator':12s} attempted={att:5d} would_insert={wi:5d}")
     else:
-        print(f"  originator inserted={counts[('originator', 'inserted')]}  skipped_dup={counts[('originator', 'skipped_dup')]}")
+        ins = counts[("originator", "inserted")]
+        skip = counts[("originator", "skipped_dup")]
+        print(f"  {'originator':12s} attempted={att:5d} inserted={ins:5d} skipped_dup={skip:5d}")
+
+    if not args.dry_run:
+        total = conn.execute("SELECT COUNT(*) FROM contribution_events").fetchone()[0]
+        print(f"\nTotal contribution_events rows: {total}")
 
 
 if __name__ == "__main__":

scripts/backfill-synthetic-recovery-prs.py

@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+"""Reconstruct synthetic `prs` rows for historical GitHub PRs lost pre-mirror-wiring.
+
+Two PRs merged on GitHub before our sync-mirror.sh tracked `github_pr`:
+  - GitHub PR #68: alexastrum — 6 claims, merged 2026-03-09 via GitHub squash,
+    recovered to Forgejo via commit dba00a79 (Apr 16, after mirror erased files)
+  - GitHub PR #88: Cameron-S1 — 1 claim, recovered via commit da64f805
+
+The recovery commits wrote the files directly to main, so our `prs` table has
+no row to attach originator events to — the backfill-events.py strategies all
+return NULL. We reconstruct one synthetic `prs` row per historical GitHub PR so
+the events pipeline (and `github_pr` strategy in backfill-events) can credit
+Alex and Cameron properly.
+
+Numbers 900000+ are clearly synthetic and won't collide with real Forgejo PRs.
+
+Idempotent via INSERT OR IGNORE.
+
+Usage:
+  python3 scripts/backfill-synthetic-recovery-prs.py --dry-run
+  python3 scripts/backfill-synthetic-recovery-prs.py
+"""
+import argparse
+import os
+import sqlite3
+import sys
+from pathlib import Path
+
+DB_PATH = os.environ.get("PIPELINE_DB", "/opt/teleo-eval/pipeline/pipeline.db")
+
+# Historical GitHub PRs recovered via direct-to-main commits.
+# Original GitHub merge dates come from the recovery commit messages.
+RECOVERY_PRS = [
+    {
+        "number": 900068,
+        "github_pr": 68,
+        "branch": "gh-pr-68",
+        "status": "merged",
+        "domain": "ai-alignment",
+        "commit_type": "knowledge",
+        "tier": "STANDARD",
+        "leo_verdict": "approve",
+        "domain_verdict": "approve",
+        "submitted_by": "alexastrum",
+        "source_channel": "github",
+        # origin='human' matches lib/merge.py convention for external contributors
+        # (default is 'pipeline' which misclassifies us as machine-authored).
+        "origin": "human",
+        "priority": "high",
+        "description": "Multi-agent git workflows production maturity | Cryptographic agent trust ratings | Defense in depth for AI agent oversight | Deterministic policy engines below LLM layer | Knowledge validation four-layer architecture | Structurally separating proposer and reviewer agents",
+        "merged_at": "2026-03-09 00:00:00",
+        "created_at": "2026-03-08 00:00:00",
+        "last_error": "synthetic_recovery: GitHub PR #68 pre-mirror-wiring reconstruction (commit dba00a79)",
+    },
+    {
+        "number": 900088,
+        "github_pr": 88,
+        "branch": "gh-pr-88",
+        "status": "merged",
+        "domain": "ai-alignment",
+        "commit_type": "knowledge",
+        "tier": "STANDARD",
+        "leo_verdict": "approve",
+        "domain_verdict": "approve",
+        "submitted_by": "cameron-s1",
+        "source_channel": "github",
+        "origin": "human",
+        "priority": "high",
+        "description": "Orthogonality is an artefact of specification architectures not a property of intelligence itself",
+        "merged_at": "2026-04-01 00:00:00",
+        "created_at": "2026-04-01 00:00:00",
+        "last_error": "synthetic_recovery: GitHub PR #88 pre-mirror-wiring reconstruction (commit da64f805)",
+    },
+]
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dry-run", action="store_true")
+    args = parser.parse_args()
+
+    if not Path(DB_PATH).exists():
+        print(f"ERROR: DB not found at {DB_PATH}", file=sys.stderr)
+        sys.exit(1)
+
+    conn = sqlite3.connect(DB_PATH, timeout=30)
+    conn.row_factory = sqlite3.Row
+
+    # Guard against synthetic-range colonization (Ganymede review): check for
+    # any row in the synthetic range that isn't one of ours. INSERT OR IGNORE on
+    # the specific numbers is the real collision defense; this is belt-and-suspenders.
+    max_real = conn.execute(
+        "SELECT MAX(number) FROM prs WHERE number < 900000"
+    ).fetchone()[0] or 0
+    print(f"Max real Forgejo PR number: {max_real}")
+    synth_conflict = conn.execute(
+        "SELECT number FROM prs WHERE number >= 900000 AND number NOT IN (900068, 900088) LIMIT 1"
+    ).fetchone()
+    if synth_conflict:
+        print(f"ERROR: PR #{synth_conflict[0]} already exists in synthetic range. "
+              f"Pick a new range before running.", file=sys.stderr)
+        sys.exit(2)
+
+    inserted = 0
+    skipped = 0
+    for row in RECOVERY_PRS:
+        existing = conn.execute(
+            "SELECT number FROM prs WHERE number = ? OR github_pr = ?",
+            (row["number"], row["github_pr"]),
+        ).fetchone()
+        if existing:
+            print(f"  PR #{row['number']} (github_pr={row['github_pr']}): already exists — skip")
+            skipped += 1
+            continue
+        print(f"  {'(dry-run) ' if args.dry_run else ''}INSERT synthetic PR #{row['number']} "
+              f"(github_pr={row['github_pr']}, submitted_by={row['submitted_by']}, "
+              f"merged_at={row['merged_at']})")
+        if not args.dry_run:
+            conn.execute(
+                """INSERT INTO prs (
+                    number, github_pr, branch, status, domain, commit_type, tier,
+                    leo_verdict, domain_verdict, submitted_by, source_channel,
+                    origin, priority,
+                    description, merged_at, created_at, last_error
+                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+                (
+                    row["number"], row["github_pr"], row["branch"], row["status"],
+                    row["domain"], row["commit_type"], row["tier"],
+                    row["leo_verdict"], row["domain_verdict"],
+                    row["submitted_by"], row["source_channel"],
+                    row["origin"], row["priority"],
+                    row["description"], row["merged_at"], row["created_at"],
+                    row["last_error"],
+                ),
+            )
+            inserted += 1
+
+    if not args.dry_run:
+        conn.commit()
+
+    print(f"\nInserted {inserted}, skipped {skipped}")
+    if not args.dry_run and inserted:
+        print("\nNext step: re-run backfill-events.py to attach originator events")
+        print("  python3 ops/backfill-events.py")
+
+
+if __name__ == "__main__":
+    main()

scripts/classify-contributors.py

@@ -0,0 +1,426 @@
+#!/usr/bin/env python3
+"""Classify `contributors` rows into {keep_person, keep_agent, move_to_publisher, delete_garbage}.
+
+Reads current contributors table, proposes reclassification per v26 schema design:
+  - Real humans + Pentagon agents stay in contributors (kind='person'|'agent')
+  - News orgs, publications, venues move to publishers table (new v26)
+  - Multi-word hyphenated garbage (parsing artifacts) gets deleted
+  - Their contribution_events are handled per category:
+      * Publishers: DELETE events (orgs shouldn't have credit)
+      * Garbage: DELETE events (bogus data)
+      * Persons/agents: keep events untouched
+
+Classification is heuristic — uses explicit allowlists + regex patterns + length gates.
+Ambiguous cases default to 'review_needed' (human decision).
+
+Usage:
+  python3 scripts/classify-contributors.py              # dry-run analysis + report
+  python3 scripts/classify-contributors.py --apply      # write changes
+  python3 scripts/classify-contributors.py --show <handle>  # inspect a single row
+
+Writes to pipeline.db only. Does NOT modify claim files.
+"""
+import argparse
+import json
+import os
+import re
+import sqlite3
+import sys
+from collections import Counter
+from pathlib import Path
+
+DB_PATH = os.environ.get("PIPELINE_DB", "/opt/teleo-eval/pipeline/pipeline.db")
+
+# Pentagon agents: kind='agent'. Authoritative list.
+PENTAGON_AGENTS = frozenset({
+    "rio", "leo", "theseus", "vida", "clay", "astra",
+    "oberon", "argus", "rhea", "ganymede", "epimetheus", "hermes", "ship",
+    "pipeline",
+})
+
+# Publisher/news-org handles seen in current contributors table.
+# Grouped by kind for the publishers row. Classified by inspection.
+# NOTE: This list is hand-curated — add to it as new orgs appear.
+PUBLISHERS_NEWS = {
+    # News outlets / brands
+    "cnbc", "al-jazeera", "axios", "bloomberg", "reuters", "bettorsinsider",
+    "fortune", "techcrunch", "coindesk", "coindesk-staff", "coindesk-research",
+    "coindesk research", "coindesk staff",
+    "defense-one", "thedefensepost", "theregister", "the-intercept",
+    "the-meridiem", "variety", "variety-staff", "variety staff", "spacenews",
+    "nasaspaceflight", "thedonkey", "insidedefense", "techpolicypress",
+    "morganlewis", "casinoorg", "deadline", "animationmagazine",
+    "defensepost", "casino-org", "casino.org",
+    "air & space forces magazine", "ieee spectrum", "techcrunch-staff",
+    "blockworks", "blockworks-staff", "decrypt", "ainvest", "banking-dive", "banking dive",
+    "cset-georgetown", "cset georgetown",
+    "kff", "kff-health-news", "kff health news", "kff-health-news---cbo",
+    "kff-health-news-/-cbo", "kff health news / cbo", "kffhealthnews",
+    "bloomberg-law",
+    "norton-rose-fulbright", "norton rose fulbright",
+    "defence-post", "the-defensepost",
+    "wilmerhale", "mofo", "sciencedirect",
+    "yogonet", "csr", "aisi-uk", "aisi", "aisi_gov", "rand",
+    "armscontrol", "eclinmed", "solana-compass", "solana compass",
+    "pmc11919318", "pmc11780016",
+    "healthverity", "natrium", "form-energy",
+    "courtlistener", "curtis-schiff", "curtis-schiff-prediction-markets",
+    "prophetx", "techpolicypress-staff",
+    "npr", "venturebeat", "geekwire", "payloadspace", "the-ankler",
+    "theankler", "tubefilter", "emarketer", "dagster",
+    "numerai",  # fund/project brand, not person
+    "psl", "multistate",
+}
+PUBLISHERS_ACADEMIC = {
+    # Academic orgs, labs, papers, journals, institutions
+    "arxiv", "metr", "metr_evals", "apollo-research", "apollo research", "apolloresearch",
+    "jacc-study-authors", "jacc-data-report-authors",
+    "anthropic-fellows-program", "anthropic-fellows",
+    "anthropic-fellows-/-alignment-science-team", "anthropic-research",
+    "jmir-2024", "jmir 2024",
+    "oettl-et-al.,-journal-of-experimental-orthopaedics",
+    "oettl et al., journal of experimental orthopaedics",
+    "jacc", "nct06548490", "pmc",
+    "conitzer-et-al.-(2024)", "aquino-michaels-2026", "pan-et-al.",
+    "pan-et-al.-'natural-language-agent-harnesses'",
+    "stanford", "stanford-meta-harness",
+    "hendershot", "annals-im",
+    "nellie-liang,-brookings-institution", "nellie liang, brookings institution",
+    "penn-state", "american-heart-association", "american heart association",
+    "molt_cornelius", "molt-cornelius",
+    # Companies / labs / brand-orgs (not specific humans)
+    "anthropic", "anthropicai", "openai", "nasa", "icrc", "ecri",
+    "epochairesearch", "metadao", "iapam", "icer",
+    "who", "ama", "uspstf", "unknown",
+    "futard.io",  # protocol/platform
+    "oxford-martin-ai-governance-initiative",
+    "oxford-martin-ai-governance",
+    "u.s.-food-and-drug-administration",
+    "jitse-goutbeek,-european-policy-centre",  # cited person+org string → publisher
+    "adepoju-et-al.",  # paper citation
+    # Formal-citation names (Firstname-Lastname or Lastname-et-al) — classified
+    # as academic citations, not reachable contributors. They'd need an @ handle
+    # to get CI credit per Cory's growth-loop design.
+    "senator-elissa-slotkin",
+    "bostrom", "hanson", "kaufmann", "noah-smith", "doug-shapiro",
+    "shayon-sengupta", "shayon sengupta",
+    "robin-hanson", "robin hanson", "eliezer-yudkowsky",
+    "leopold-aschenbrenner", "aschenbrenner",
+    "ramstead", "larsson", "heavey",
+    "dan-slimmon", "van-leeuwaarden", "ward-whitt", "adams",
+    "tamim-ansary", "spizzirri",
+    "dario-amodei",  # formal-citation form (real @ is @darioamodei)
+    "corless", "oxranga", "vlahakis",
+    # Brand/project/DAO tokens — not individuals
+    "areal-dao", "areal", "theiaresearch", "futard-io", "dhrumil",
+    # Classic formal-citation names — famous academics/economists cited by surname.
+    # Reachable via @ handle if/when they join (e.g. Ostrom has no X, Hayek deceased,
+    # Friston has an institutional affiliation not an @ handle we'd track).
+    "clayton-christensen", "hidalgo", "coase", "wiener", "juarrero",
+    "ostrom", "centola", "hayek", "marshall-mcluhan", "blackmore",
+    "knuth", "friston", "aquino-michaels", "conitzer", "bak",
+}
+# NOTE: pseudonymous X handles that MAY be real contributors stay in keep_person:
+#   karpathy, simonw, swyx, metaproph3t, metanallok, mmdhrumil, sjdedic,
+#   ceterispar1bus — these are real X accounts and match Cory's growth loop.
+# They appear without @ prefix because extraction frontmatter didn't normalize.
+# Auto-creating them as contributors tier='cited' is correct (A-path from earlier).
+PUBLISHERS_SOCIAL = {
+    "x", "twitter", "telegram", "x.com",
+}
+PUBLISHERS_INTERNAL = {
+    "teleohumanity-manifesto", "strategy-session-journal",
+    "living-capital-thesis-development", "attractor-state-historical-backtesting",
+    "web-research-compilation", "architectural-investing",
+    "governance---meritocratic-voting-+-futarchy",  # title artifact
+    "sec-interpretive-release-s7-2026-09-(march-17",  # title artifact
+    "mindstudio",  # tooling/platform, not contributor
+}
+# Merge into one kind→set map for classification
+PUBLISHER_KIND_MAP = {}
+for h in PUBLISHERS_NEWS:
+    PUBLISHER_KIND_MAP[h.lower()] = "news"
+for h in PUBLISHERS_ACADEMIC:
+    PUBLISHER_KIND_MAP[h.lower()] = "academic"
+for h in PUBLISHERS_SOCIAL:
+    PUBLISHER_KIND_MAP[h.lower()] = "social_platform"
+for h in PUBLISHERS_INTERNAL:
+    PUBLISHER_KIND_MAP[h.lower()] = "internal"
+
+
+# Garbage: handles that are clearly parse artifacts, not real names.
+# Pattern: contains parens, special chars, or >50 chars.
+def is_garbage(handle: str) -> bool:
+    h = handle.strip()
+    if len(h) > 50:
+        return True
+    if re.search(r"[()\[\]<>{}\/\\|@#$%^&*=?!:;\"']", h):
+        # But @ can appear legitimately in handles like @thesensatore — allow if @ is only prefix
+        if h.startswith("@") and not re.search(r"[()\[\]<>{}\/\\|#$%^&*=?!:;\"']", h):
+            return False
+        return True
+    # Multi-word hyphenated with very specific artifact shape: 3+ hyphens in a row or trailing noise
+    if "---" in h or "---meritocratic" in h or h.endswith("(march") or h.endswith("-(march"):
+        return True
+    return False
+
+
+def classify(handle: str) -> tuple[str, str | None]:
+    """Return (category, publisher_kind).
+
+    category ∈ {'keep_agent', 'keep_person', 'publisher', 'garbage', 'review_needed'}
+    publisher_kind ∈ {'news','academic','social_platform','internal', None}
+    """
+    h = handle.strip().lower().lstrip("@")
+
+    if h in PENTAGON_AGENTS:
+        return ("keep_agent", None)
+
+    if h in PUBLISHER_KIND_MAP:
+        return ("publisher", PUBLISHER_KIND_MAP[h])
+
+    if is_garbage(handle):
+        return ("garbage", None)
+
+    # @-prefixed handles or short-slug real-looking names → keep as person
+    # (Auto-create rule from Cory: @ handles auto-join as tier='cited'.)
+    if handle.startswith("@"):
+        return ("keep_person", None)
+
+    # Plausible handles (<=39 chars, alphanum + underscore/hyphen): treat as person.
+    # 39-char ceiling matches GitHub's handle limit and the writer path in
+    # contributor.py::_HANDLE_RE, so a valid 21-39 char real handle won't fall
+    # through to review_needed and block --apply.
+    if re.match(r"^[a-z0-9][a-z0-9_-]{0,38}$", h):
+        return ("keep_person", None)
+
+    # Everything else: needs human review
+    return ("review_needed", None)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--apply", action="store_true", help="Write changes to DB")
+    parser.add_argument("--show", type=str, help="Inspect a single handle")
+    parser.add_argument("--delete-events", action="store_true",
+                        help="DELETE contribution_events for publishers+garbage (default: keep for audit)")
+    args = parser.parse_args()
+
+    if not Path(DB_PATH).exists():
+        print(f"ERROR: DB not found at {DB_PATH}", file=sys.stderr)
+        sys.exit(1)
+
+    conn = sqlite3.connect(DB_PATH, timeout=30)
+    conn.row_factory = sqlite3.Row
+
+    # Sanity: publishers table must exist (v26 migration applied)
+    try:
+        conn.execute("SELECT 1 FROM publishers LIMIT 1")
+    except sqlite3.OperationalError:
+        print("ERROR: publishers table missing. Run migration v26 first.", file=sys.stderr)
+        sys.exit(2)
+
+    rows = conn.execute(
+        "SELECT handle, kind, tier, claims_merged FROM contributors ORDER BY claims_merged DESC"
+    ).fetchall()
+
+    if args.show:
+        target = args.show.strip().lower().lstrip("@")
+        for r in rows:
+            if r["handle"].lower().lstrip("@") == target:
+                category, pkind = classify(r["handle"])
+                events_count = conn.execute(
+                    "SELECT COUNT(*) FROM contribution_events WHERE handle = ?",
+                    (r["handle"].lower().lstrip("@"),),
+                ).fetchone()[0]
+                print(f"handle:         {r['handle']}")
+                print(f"current_kind:   {r['kind']}")
+                print(f"current_tier:   {r['tier']}")
+                print(f"claims_merged:  {r['claims_merged']}")
+                print(f"events:         {events_count}")
+                print(f"→ category:     {category}")
+                if pkind:
+                    print(f"→ publisher:    kind={pkind}")
+                return
+        print(f"No match for '{args.show}'")
+        return
+
+    # Classify all
+    buckets: dict[str, list[dict]] = {
+        "keep_agent": [],
+        "keep_person": [],
+        "publisher": [],
+        "garbage": [],
+        "review_needed": [],
+    }
+    for r in rows:
+        category, pkind = classify(r["handle"])
+        buckets[category].append({
+            "handle": r["handle"],
+            "kind_now": r["kind"],
+            "tier": r["tier"],
+            "claims": r["claims_merged"] or 0,
+            "publisher_kind": pkind,
+        })
+
+    print("=== Classification summary ===")
+    for cat, items in buckets.items():
+        print(f"  {cat:18s}  {len(items):5d}")
+
+    print("\n=== Sample of each category ===")
+    for cat, items in buckets.items():
+        print(f"\n--- {cat} (showing up to 10) ---")
+        for item in items[:10]:
+            tag = f" → {item['publisher_kind']}" if item["publisher_kind"] else ""
+            print(f"  {item['handle']:50s} claims={item['claims']:5d}{tag}")
+
+    print("\n=== Full review_needed list ===")
+    for item in buckets["review_needed"]:
+        print(f"  {item['handle']:50s} claims={item['claims']:5d}")
+
+    # Diagnostic: orphan alias count for handles we're about to delete.
+    # Contributor_aliases has no FK (SQLite FKs require PRAGMA to enforce anyway),
+    # so aliases pointing to deleted canonical handles become orphans. Surface
+    # the count so the --delete-events decision is informed.
+    doomed = [item["handle"].lower().lstrip("@") for item in buckets["garbage"] + buckets["publisher"]]
+    if doomed:
+        placeholders = ",".join("?" * len(doomed))
+        orphan_count = conn.execute(
+            f"SELECT COUNT(*) FROM contributor_aliases WHERE canonical IN ({placeholders})",
+            doomed,
+        ).fetchone()[0]
+        print(f"\n=== Alias orphan check ===")
+        print(f"  contributor_aliases rows pointing to deletable canonicals: {orphan_count}")
+        if orphan_count:
+            print(f"  (cleanup requires --delete-events; without it, aliases stay as orphans)")
+
+    if not args.apply:
+        print("\n(dry-run — no writes. Re-run with --apply to execute.)")
+        return
+
+    # ── Apply changes ──
+    print("\n=== Applying changes ===")
+    if buckets["review_needed"]:
+        print(f"ABORT: {len(buckets['review_needed'])} rows need human review. Fix classifier before --apply.")
+        sys.exit(3)
+
+    inserted_publishers = 0
+    reclassified_agents = 0
+    deleted_garbage = 0
+    deleted_publisher_rows = 0
+    deleted_events = 0
+    deleted_aliases = 0
+
+    # Single transaction — if any step errors, roll back. This prevents the failure
+    # mode where a publisher insert fails silently and we still delete the contributor
+    # row, losing data.
+    try:
+        conn.execute("BEGIN")
+
+        # 1. Insert publishers. Track which ones succeeded so step 4 only deletes those.
+        # Counter uses cur.rowcount so replay runs (where publishers already exist)
+        # report accurate inserted=0 instead of falsely claiming the full set.
+        # moved_to_publisher is unconditional — the contributors row still needs to
+        # be deleted even when the publishers row was added in a prior run.
+        moved_to_publisher = set()
+        for item in buckets["publisher"]:
+            name = item["handle"].strip().lower().lstrip("@")
+            cur = conn.execute(
+                "INSERT OR IGNORE INTO publishers (name, kind) VALUES (?, ?)",
+                (name, item["publisher_kind"]),
+            )
+            if cur.rowcount > 0:
+                inserted_publishers += 1
+            moved_to_publisher.add(item["handle"])
+
+        # 2. Ensure Pentagon agents have kind='agent' (idempotent after v25 patch)
+        for item in buckets["keep_agent"]:
+            conn.execute(
+                "UPDATE contributors SET kind = 'agent' WHERE handle = ?",
+                (item["handle"].lower().lstrip("@"),),
+            )
+            reclassified_agents += 1
+
+        # 3. Delete garbage handles from contributors (and their events + aliases)
+        for item in buckets["garbage"]:
+            canonical_lower = item["handle"].lower().lstrip("@")
+            if args.delete_events:
+                cur = conn.execute(
+                    "DELETE FROM contribution_events WHERE handle = ?",
+                    (canonical_lower,),
+                )
+                deleted_events += cur.rowcount
+                cur = conn.execute(
+                    "DELETE FROM contributor_aliases WHERE canonical = ?",
+                    (canonical_lower,),
+                )
+                deleted_aliases += cur.rowcount
+            cur = conn.execute(
+                "DELETE FROM contributors WHERE handle = ?",
+                (item["handle"],),
+            )
+            deleted_garbage += cur.rowcount
+
+        # 4. Delete publisher rows from contributors — ONLY for those successfully
+        # inserted into publishers above. Guards against partial failure.
+        # Aliases pointing to publisher-classified handles get cleaned under the
+        # same --delete-events gate: publishers live in their own table now, any
+        # leftover aliases in contributor_aliases are orphans.
+        for item in buckets["publisher"]:
+            if item["handle"] not in moved_to_publisher:
+                continue
+            canonical_lower = item["handle"].lower().lstrip("@")
+            if args.delete_events:
+                cur = conn.execute(
+                    "DELETE FROM contribution_events WHERE handle = ?",
+                    (canonical_lower,),
+                )
+                deleted_events += cur.rowcount
+                cur = conn.execute(
+                    "DELETE FROM contributor_aliases WHERE canonical = ?",
+                    (canonical_lower,),
+                )
+                deleted_aliases += cur.rowcount
+            cur = conn.execute(
+                "DELETE FROM contributors WHERE handle = ?",
+                (item["handle"],),
+            )
+            deleted_publisher_rows += cur.rowcount
+
+        # 5. Audit log entry for the destructive operation (Ganymede Q5).
+        conn.execute(
+            "INSERT INTO audit_log (timestamp, stage, event, detail) VALUES (datetime('now'), ?, ?, ?)",
+            (
+                "schema_v26",
+                "classify_contributors",
+                json.dumps({
+                    "publishers_inserted": inserted_publishers,
+                    "agents_updated": reclassified_agents,
+                    "garbage_deleted": deleted_garbage,
+                    "publisher_rows_deleted": deleted_publisher_rows,
+                    "events_deleted": deleted_events,
+                    "aliases_deleted": deleted_aliases,
+                    "delete_events_flag": bool(args.delete_events),
+                }),
+            ),
+        )
+
+        conn.commit()
+    except Exception as e:
+        conn.rollback()
+        print(f"ERROR: Transaction failed, rolled back. {e}", file=sys.stderr)
+        sys.exit(4)
+
+    print(f"  publishers inserted:          {inserted_publishers}")
+    print(f"  agents kind='agent' ensured:  {reclassified_agents}")
+    print(f"  garbage rows deleted:         {deleted_garbage}")
+    print(f"  publisher rows removed from contributors: {deleted_publisher_rows}")
+    if args.delete_events:
+        print(f"  contribution_events deleted:  {deleted_events}")
+        print(f"  contributor_aliases deleted:  {deleted_aliases}")
+    else:
+        print(f"  (events + aliases kept — re-run with --delete-events to clean them)")
+
+
+if __name__ == "__main__":
+    main()