Add conversation-aware extraction for Telegram sources

When source format is "conversation", inject specialized extraction
rules that prioritize human corrections/pushback as highest-value
content. Fixes null-result on short but high-signal correction
messages. Maps corrections to existing KB claims as challenges.

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

lib/extract.py

@@ -320,6 +320,7 @@ async def _extract_one_source(
     rationale = fm.get("rationale")
     intake_tier = fm.get("intake_tier")
     proposed_by = fm.get("proposed_by")
+    source_format = fm.get("format")
 
     logger.info("Extracting: %s (domain: %s, agent: %s)", source_file, domain, agent_name)
 
@@ -343,6 +344,7 @@ async def _extract_one_source(
         proposed_by=proposed_by,
         prior_art=prior_art,
         previous_feedback=feedback,
+        source_format=source_format,
     )
 
     # 4. Call LLM (OpenRouter — not Claude Max CLI)

lib/extraction_prompt.py

@@ -29,6 +29,7 @@ def build_extraction_prompt(
     proposed_by: str | None = None,
     prior_art: list[dict] | None = None,
     previous_feedback: dict | None = None,
+    source_format: str | None = None,
 ) -> str:
     """Build the lean extraction prompt.
 
@@ -45,6 +46,7 @@ def build_extraction_prompt(
         prior_art: Qdrant search results — existing claims semantically similar to this source.
                    Each dict has: claim_title, claim_path, description, score.
                    Injected as connection candidates for extract-time linking.
+        source_format: Source format hint (e.g. "conversation" for Telegram chats).
 
     Returns:
         The complete prompt string
@@ -131,6 +133,65 @@ Set `contributor_thesis_extractable: true` if you extracted the contributor's th
     else:
         connection_candidates = ""
 
+    # Build conversation extraction section (for Telegram/chat sources)
+    if source_format and source_format.lower() == "conversation":
+        conversation_section = """
+## Conversation Source — Special Extraction Rules
+
+This source is a **conversation between a human domain expert and an AI agent**.
+The extraction rules are DIFFERENT from article sources:
+
+### Who said what matters
+
+- **The human (@m3taversal / contributor)** is the domain expert. Their statements carry
+  authority — especially corrections, pushback, and factual assertions.
+- **The AI agent's responses** are secondary. They are useful for context (what was being
+  discussed) and for confirming when the human's correction landed (look for "you're right",
+  "fair point", confidence drops).
+
+### Corrections are the HIGHEST-VALUE content
+
+When the human says "that's wrong", "not true", "you're wrong", "out of date", or similar:
+
+1. **Extract the correction as a claim or enrichment.** The human is correcting the KB's
+   understanding. This is precisely what the KB needs.
+2. **The correction itself IS the claim.** "Curated launches had significantly more committed
+   capital than permissionless launches" is a testable, disagreeable proposition — extract it.
+3. **Short corrections are HIGH value, not low value.** A 15-word correction that fixes a
+   factual error is worth more than a 500-word article that confirms what we already know.
+   NEVER null-result a conversation just because the human's message is short.
+4. **Map corrections to existing claims.** Search the KB index for claims that the correction
+   challenges. Output as an ENRICHMENT with `type: "challenge"` if the target claim exists.
+
+### Bot LEARNING lines are extraction hints
+
+When the AI agent includes a `LEARNING:` line, it's a pre-extracted correction. Use it as
+a starting point — but reformulate it as a proper claim (the LEARNING line is often too
+casual or too specific to the conversation context).
+
+### Bot CONFIDENCE drops are signals
+
+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.
+
+### Anti-circularity rule
+
+If the AI agent is simply reflecting the human's thesis back (restating what the human said
+in different words), do NOT extract that as a claim sourced from the agent. That's circular.
+Only extract claims that either:
+- Represent the human's ORIGINAL assertion (source it to the human)
+- Introduce genuinely NEW information from the agent's knowledge (source it to the agent + context)
+
+### Retrieval-only conversations → null_result
+
+If the conversation is purely a lookup request ("what is X", "give me a list of Y",
+"what's the market cap of Z") with no analytical content, corrections, or novel claims,
+return an empty extraction (null_result). The dividing line: did the human ASSERT something
+or only ASK something?
+"""
+    else:
+        conversation_section = ""
+
     return f"""You are {agent}, extracting knowledge from a source for TeleoHumanity's collective knowledge base.
 
 ## Your Task
@@ -195,7 +256,7 @@ Single source = experimental at most. Pitch rhetoric or marketing copy = specula
 **File:** {source_file}
 
 {source_content}
-{contributor_directive}{previous_feedback_section}{connection_candidates}
+{conversation_section}{contributor_directive}{previous_feedback_section}{connection_candidates}
 ## KB Index (existing claims — check for duplicates and enrichment targets)
 
 {kb_index}