extraction quality: trust hierarchy + verified tagging + telegram review endpoint

Three fixes for conversation-sourced claim quality:

1. Trust hierarchy in extraction prompt: bot-generated numbers are
   flagged as unverified context, not evidence. Directional claims
   are extractable but specific figures require external verification.
   Prevents laundering bot guesses into the KB as evidence.

2. Conversation-sourced claims tagged with verified: false and
   source_type: conversation in frontmatter. Downstream consumers
   (Leo, dashboard) can filter/flag these for verification.

3. GET /api/telegram-extractions endpoint for daily spot-checking.
   Shows recent Telegram-sourced PRs with claim titles, status,
   merge rate, and eval issues. Quick review surface.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

diagnostics/dashboard_routes.py

@@ -1109,6 +1109,79 @@ async def handle_pr_lifecycle(request):
         conn.close()
 
 
+# ─── GET /api/telegram-extractions ───────────────────────────────────────
+
+async def handle_telegram_extractions(request):
+    """Review surface for Telegram conversation extractions.
+
+    Shows recent PRs sourced from Telegram conversations with claim titles,
+    status, and source info. Designed for quick daily spot-checking.
+
+    Query params:
+        days (int): lookback window (default 7, max 90)
+    """
+    conn = request.app["_get_conn"]()
+    try:
+        days = min(int(request.query.get("days", "7")), 90)
+        day_filter = f"-{days}"
+
+        # Find PRs from Telegram sources (source_path contains 'telegram' or submitted_by is @m3taversal via bot)
+        rows = conn.execute(
+            """SELECT p.number, p.agent, p.domain, p.tier, p.status,
+                      p.created_at, p.merged_at, p.description, p.source_path,
+                      p.submitted_by, p.branch, p.eval_issues, p.leo_verdict
+               FROM prs p
+               WHERE (p.source_path LIKE '%telegram%' OR p.source_path LIKE '%futardio%')
+               AND p.created_at > datetime('now', ? || ' days')
+               ORDER BY p.number DESC""",
+            (day_filter,),
+        ).fetchall()
+
+        prs = []
+        for r in rows:
+            desc = r["description"] or ""
+            claim_titles = [t.strip() for t in desc.split("|") if t.strip()] if desc.strip() else []
+
+            issues = None
+            if r["eval_issues"]:
+                try:
+                    issues = json.loads(r["eval_issues"]) if isinstance(r["eval_issues"], str) else r["eval_issues"]
+                except (json.JSONDecodeError, TypeError):
+                    pass
+
+            prs.append({
+                "number": r["number"],
+                "agent": r["agent"],
+                "domain": r["domain"],
+                "tier": r["tier"],
+                "status": r["status"],
+                "created_at": r["created_at"],
+                "merged_at": r["merged_at"],
+                "claim_titles": claim_titles,
+                "source_path": r["source_path"],
+                "submitted_by": r["submitted_by"],
+                "eval_issues": issues,
+                "leo_verdict": r["leo_verdict"],
+            })
+
+        # Summary stats
+        merged = sum(1 for p in prs if p["status"] == "merged")
+        closed = sum(1 for p in prs if p["status"] == "closed")
+        open_prs = sum(1 for p in prs if p["status"] == "open")
+
+        return web.json_response({
+            "days": days,
+            "total": len(prs),
+            "merged": merged,
+            "closed": closed,
+            "open": open_prs,
+            "merge_rate": round(merged / len(prs) * 100, 1) if prs else 0,
+            "prs": prs,
+        })
+    finally:
+        conn.close()
+
+
 # ─── Registration ──────────────────────────────────────────────────────────
 
 def register_dashboard_routes(app: web.Application, get_conn):
@@ -1125,3 +1198,4 @@ def register_dashboard_routes(app: web.Application, get_conn):
     app.router.add_get("/api/trace/{trace_id}", handle_trace)
     app.router.add_get("/api/growth", handle_growth)
     app.router.add_get("/api/pr-lifecycle", handle_pr_lifecycle)
+    app.router.add_get("/api/telegram-extractions", handle_telegram_extractions)

lib/extract.py

@@ -215,7 +215,7 @@ def _parse_extraction_json(text: str) -> dict | None:
         return None
 
 
-def _build_claim_content(claim: dict, agent: str) -> str:
+def _build_claim_content(claim: dict, agent: str, source_format: str | None = None) -> str:
     """Build claim markdown file content from extraction JSON."""
     today = date.today().isoformat()
     domain = claim.get("domain", "")
@@ -265,6 +265,9 @@ def _build_claim_content(claim: dict, agent: str) -> str:
         lines.append(f"scope: {scope}")
     if sourcer:
         lines.append(f'sourcer: "{sourcer}"')
+    if source_format and source_format.lower() == "conversation":
+        lines.append("verified: false")
+        lines.append("source_type: conversation")
     lines.extend(edge_lines)
     lines.append("---")
     lines.append("")
@@ -401,7 +404,7 @@ async def _extract_one_source(
         filename = Path(filename).name  # Strip directory components — LLM output may contain path traversal
         if not filename.endswith(".md"):
             filename += ".md"
-        content = _build_claim_content(c, agent_lower)
+        content = _build_claim_content(c, agent_lower, source_format=source_format)
         claim_files.append({"filename": filename, "domain": c.get("domain", domain), "content": content})
 
     # Build entity file contents

lib/extraction_prompt.py

@@ -178,6 +178,26 @@ casual or too specific to the conversation context).
 When the AI agent drops its confidence score after a correction, that CONFIRMS the human
 was right. Low confidence (0.3-0.5) after pushback = strong signal the correction is valid.
 
+### Trust hierarchy for numbers and specifics
+
+**CRITICAL:** Neither the human NOR the AI agent should be treated as authoritative sources
+for specific numbers, dates, dollar amounts, or statistics UNLESS they cite a verifiable
+external source (on-chain data, official announcements, published reports).
+
+- **Bot-generated numbers are ALWAYS unverified.** When the AI agent says "$25.6M committed
+  capital" or "15x oversubscription" — these are the bot's best guess, NOT verified data.
+  NEVER extract bot-generated numbers as evidence in a claim.
+- **Human-asserted numbers are ALSO unverified** unless they cite a source. "It raised $11.4M"
+  from the human is a claim about a number, not proof of the number.
+- **Extract the DIRECTIONAL insight, not the specific figures.** "Curated launches attracted
+  significantly more committed capital than permissionless launches" is extractable.
+  "$25.6M vs $11.4M" is not — unless the conversation cites where those numbers come from.
+- **If specific figures are important to the claim, flag them.** Add a note in the claim body:
+  "Note: specific figures cited in conversation require verification against on-chain data."
+
+The goal: capture WHAT the human is asserting (the mechanism, the direction, the pattern)
+without laundering unverified numbers into the knowledge base as if they were evidence.
+
 ### Anti-circularity rule
 
 If the AI agent is simply reflecting the human's thesis back (restating what the human said