From 1e0c1cd788e088e0e776a3e3c6d302f3afdfed30 Mon Sep 17 00:00:00 2001
From: m3taversal <m3taversal@gmail.com>
Date: Thu, 16 Apr 2026 12:12:15 +0100
Subject: [PATCH] Write enrichments as file modifications; strengthen
 correction extraction

Two changes:
1. extract.py: Enrichments now modify existing claim files by appending
   evidence sections. Previously enrichment-only extractions were
   discarded as null-result even when they contained valuable challenges.
2. extraction_prompt.py: Corrections should produce BOTH a claim (the
   corrected knowledge) AND an enrichment (linking to what it corrects).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 lib/extract.py           | 33 +++++++++++++++++++++++++++++++--
 lib/extraction_prompt.py |  8 ++++++--
 2 files changed, 37 insertions(+), 4 deletions(-)

diff --git a/lib/extract.py b/lib/extract.py
index cb68682..ae32e63 100644
--- a/lib/extract.py
+++ b/lib/extract.py
@@ -432,8 +432,8 @@ async def _extract_one_source(
             )
         claim_files = kept_claims
 
-    if not claim_files and not entity_files:
-        logger.info("No valid claims/entities after validation for %s — archiving as null-result", source_file)
+    if not claim_files and not entity_files and not enrichments:
+        logger.info("No valid claims/entities/enrichments after validation for %s — archiving as null-result", source_file)
         await _archive_source(source_path, domain, "null-result")
         return 0, 0
 
@@ -471,6 +471,35 @@ async def _extract_one_source(
         fpath.write_text(ef["content"], encoding="utf-8")
         files_written.append(f"entities/{domain}/{ef['filename']}")
 
+    # Write enrichments as modifications to existing claim files
+    for enr in enrichments:
+        target = enr.get("target_file", "")
+        evidence = enr.get("evidence", "")
+        enr_type = enr.get("type", "extend")  # confirm|challenge|extend
+        source_ref = enr.get("source_ref", source_file)
+        if not target or not evidence:
+            continue
+        # Find the target claim file in the worktree (search domains/)
+        target_stem = target.replace(".md", "")
+        found = None
+        for domain_dir in (worktree / "domains").iterdir():
+            candidate = domain_dir / f"{target_stem}.md"
+            if candidate.exists():
+                found = candidate
+                break
+        if not found:
+            logger.debug("Enrichment target %s not found in worktree", target)
+            continue
+        # Append enrichment evidence to the claim file
+        existing = found.read_text(encoding="utf-8")
+        label = {"confirm": "Supporting", "challenge": "Challenging", "extend": "Extending"}.get(enr_type, "Additional")
+        enrichment_block = f"\n\n## {label} Evidence\n\n**Source:** {source_ref}\n\n{evidence}\n"
+        found.write_text(existing + enrichment_block, encoding="utf-8")
+        rel_path = str(found.relative_to(worktree))
+        if rel_path not in files_written:
+            files_written.append(rel_path)
+        logger.info("Enrichment applied to %s (%s)", target, enr_type)
+
     if not files_written:
         logger.info("No files written for %s — cleaning up", source_file)
         await _git("checkout", "main", cwd=str(EXTRACT_WORKTREE))
diff --git a/lib/extraction_prompt.py b/lib/extraction_prompt.py
index 797f4d8..ae633f9 100644
--- a/lib/extraction_prompt.py
+++ b/lib/extraction_prompt.py
@@ -156,12 +156,16 @@ When the human says "that's wrong", "not true", "you're wrong", "out of date", o
 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.
+   capital than permissionless launches" is a testable, disagreeable proposition — extract it
+   AS A CLAIM, not just an enrichment. If the correction states something specific enough to
+   disagree with, it's a claim. Extract it even if it's only one sentence.
 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.
+   challenges. Output BOTH a new claim (the corrected understanding) AND an enrichment
+   (type: "challenge") targeting the existing claim. The enrichment links the correction
+   to what it corrects; the claim captures the corrected knowledge as a standalone proposition.
 
 ### Bot LEARNING lines are extraction hints