update

backend/config.py

@@ -70,6 +70,11 @@ SUB_VARIANT_COS = 0.90
 # subs are statements: true covers measure 0.27–0.38 while a wrong hit measured 0.76. The
 # embedding stage only backs up the lexical one (catches „Line Breaks (Soft)" 0.888).
 SEED_COVER_COS = 0.80
+# Sub duplicate CANDIDATE floor for the judge paths (in-block consolidation band hint,
+# cross-block stage, QA detector): the bulk of real paraphrase duplicates measures 0.75–0.90
+# (Markdown: 50 pairs in the band, 4 above) — below every auto-merge threshold, so an LLM
+# judge decides. Candidates only; a merge still needs judge unanimity.
+SUB_DUP_KANDIDAT_COS = 0.75
 
 # Umbrella grouping (block granularity level 2, step "Blocks-Gruppierung", AFTER the filter):
 # collapse sibling DEFINITIONS that are components of ONE umbrella concept (TM model:
@@ -134,6 +139,11 @@ CRAWL_MIN_CHARS = 400                                            # too little te
 QUELLE_RELEVANZ_CHUNK = 12       # pages per rater package (small, since a snippet ships per page)
 QUELLE_RELEVANZ_SNIPPET = 800    # body characters per page in the prompt (URL is the primary signal)
 
+# QA gate: after the inventory phase an automatic QA run scores the blocks; below the
+# threshold the flow PAUSES before board 2 burns tokens (frontend offers force-continue).
+QA_GATE_NOTE = 9.5   # 0 = gate off; quota-based, so the tolerated finding count scales with topic size
+QA_GATE_LLM = True   # include the LLM samples (Echtheit/Dubletten) in the gate run
+
 # Inline evidence for judge agents: corpus excerpts go INTO the prompt instead of letting
 # every judge re-search the source folder (measured: ~10 tool turns/judge, 82 % of the
 # run's tokens were cache reads from those loops).
@@ -153,6 +163,8 @@ TIMEOUTS = {
     "content_check":  (300, 10),   # content exam per block in the package
     "subblock":       (400, 15),  # finder round — p95 measured 124 s (was 900+45n)
     "subblock_check": (300, 15),  # judge decides contested subblocks in the chunk
+    "konsolidierung": (600, 25),  # consolidation judge sees ALL subs with key points — 585 s
+                                  # (subblock_check at n=19) produced 15 % timeouts
     "level":             (300, 10),  # classify subblocks per chunk
     "level_check":       (300, 10),  # judge decides contested levels in the chunk
     "relevance":          (300, 10),  # subblocks relevant/peripheral per chunk