update

backend/guide_board.py

@@ -21,7 +21,8 @@ import re
 import database as db
 import readability
 from blocks import _sink_json
-from config import FORMAT_PURPOSE, READABILITY_ACTIVE, TEMPLATES_DIR, MAX_CONCURRENT_AGENTS_PER_TOPIC
+from config import (FORMAT_PURPOSE, GUIDE_LAENGE_MAX, GUIDE_LAENGE_MIN, READABILITY_ACTIVE,
+                    TEMPLATES_DIR, MAX_CONCURRENT_AGENTS_PER_TOPIC)
 from fsutil import atomic_write_json
 from jsonio import read_json_file as _json_file
 from pipeline import (CANCELLED, FAILED, OK, GenContext, _extra, _log, _prompt,
@@ -34,9 +35,7 @@ GUIDE_STAGES = ("lernziele", "zuweisung", "writer", "fakten_gate", "coverage", "
 STAGE_LABELS = {"lernziele": "Lernziele", "zuweisung": "Zuweisung", "writer": "Writer",
                 "fakten_gate": "Fakten-Gate", "coverage": "Coverage",
                 "lesbarkeit": "Lesbarkeit", "done": "Fertig"}
-MAX_WRITER_ROUNDS = 2   # coverage → writer feedback loop cap (gains die after round 1–2)
-GATE_FIX_MIN = 3        # fact-gate claims below this: log only — a full fix rewrite for
-                        # 1–2 residual claims fired on 19/20 blocks (40 agent-minutes)
+from config import GATE_FIX_MIN, MAX_WRITER_ROUNDS, WRITER_SPLIT_SUBS  # zentral tunebar
 # Simultaneous cards = the per-topic agent cap: every card busies exactly ONE agent at a
 # time (its stages run serially), so a lower number just idles slots (was hardcoded 10
 # from the old 10-slot era while the .env already allowed 24).
@@ -65,7 +64,10 @@ def _ziele_schema(data):
 
 
 def _gate_schema(data):
-    """{"ok":true} → [] · {"claims":[{text,grund}]} → list · None invalid."""
+    """{"ok":true} → [] · {"claims":[{text,grund,urteil}]} → list · None invalid.
+    urteil "falsch" (contradicts the facts/itself) vs "unbelegt" (true but underivable) —
+    default unbelegt. Entries whose grund starts with "belegt" are dropped: one judge
+    returned a 65-entry full inventory including SUPPORTED claims."""
     if not isinstance(data, dict):
         return None
     if data.get("ok") is True:
@@ -76,7 +78,12 @@ def _gate_schema(data):
     out = []
     for c in claims:
         if isinstance(c, dict) and str(c.get("text", "")).strip():
-            out.append({"text": str(c["text"]).strip(), "grund": str(c.get("grund", "")).strip()})
+            grund = str(c.get("grund", "")).strip()
+            if grund.casefold().startswith("belegt"):
+                continue
+            urteil = str(c.get("urteil", "")).strip().casefold()
+            out.append({"text": str(c["text"]).strip(), "grund": grund,
+                        "urteil": urteil if urteil == "falsch" else "unbelegt"})
     return out
 
 
@@ -241,7 +248,7 @@ async def _stage_zuweisung(env: _Env, card: dict) -> bool:
 # A single section over ~45 subs measurably breaks the writer/coverage (Front Matter:
 # 4/6 objectives open after 2 rounds). First drafts of oversized cards are written in two
 # halves and merged back into ONE canonical section (all gates/assembly read one section).
-WRITER_SPLIT_SUBS = 30
+# WRITER_SPLIT_SUBS: siehe config.py
 
 
 def _merge_split_sections(sec_a: dict, sec_b: dict) -> str:
@@ -397,13 +404,15 @@ async def _stage_fakten_gate(env: _Env, card: dict) -> bool:
     if status == FAILED:
         claims = []  # gate failure must not block the card — logged, text stands
         _log(env.topic, f"Fakten-Gate {card['block']}: kein Ergebnis — Text bleibt ungeprüft")
-    if claims and len(claims) < GATE_FIX_MIN:
-        # 1–2 Rest-Claims rechtfertigen keinen Voll-Rewrite: der Fix-Pass lief für 19/20
-        # Blöcke und kostete 40 Agent-Minuten; die Guide-QA-Fachlichkeits-Stichprobe misst nach
+    falsch = [c for c in claims if c.get("urteil") == "falsch"] if claims else []
+    if claims and not falsch and len(claims) < GATE_FIX_MIN:
+        # 1–2 merely UNSUPPORTED claims don't justify a fix pass (it ran for 19/20 blocks,
+        # 40 agent-minutes) — but a WRONG claim always does: one slipped through this
+        # threshold and cost the guide 1.5 QA points
         _log(env.topic, f"Fakten-Gate {card['block']}: {len(claims)} Claim(s) unter Schwelle — kein Fix")
         claims = []
     if claims:
-        _log(env.topic, f"Fakten-Gate {card['block']}: {len(claims)} unbelegte Claims → Fix")
+        _log(env.topic, f"Fakten-Gate {card['block']}: {len(claims)} Claims ({len(falsch)} falsch) → Fix")
         fixp = env.slot(f"gatefix-{_safe(norm)}-r{card['writer_rounds']}.md")
         fixp.unlink(missing_ok=True)
         claims_text = "\n".join(f"- {c['text']}" + (f" ({c['grund']})" if c["grund"] else "")
@@ -501,6 +510,18 @@ async def _stage_lesbarkeit(env: _Env, card: dict) -> bool:
         hints = await asyncio.to_thread(readability.rate_sections, {1: sec["md"]})
         if hints.get(1):
             problems.append(hints[1])
+    # deterministic length trigger, same formula as the QA detector: prompt guidelines
+    # alone left writers 2.7–4.1× over target — a measured overshoot forces the fix pass
+    subs_all = env.subs_by_title.get(card["block"], [])
+    n_rel = max(sum(1 for s in subs_all if s.get("relevance") != "peripheral"), 1)
+    aus = re.split(r"<!--\s*ausführlich\s*-->", sec["md"], maxsplit=1)
+    pro_sub = len(aus[1] if len(aus) == 2 else sec["md"]) / n_rel
+    if not (GUIDE_LAENGE_MIN <= pro_sub <= GUIDE_LAENGE_MAX * 0.9):
+        ziel = _writer_budget(len(subs_all))
+        problems.append(
+            f"Länge {round(pro_sub)} Zeichen/Sub (Rahmen {GUIDE_LAENGE_MIN}–{round(GUIDE_LAENGE_MAX * 0.9)}): "
+            f"schreibe den ausführlich-Teil auf etwa {ziel} Zeichen GESAMT um — Sockel-Prosa und "
+            f"Wiederholungen streichen, alle Sub-Marker und Beispiele behalten")
     if problems:
         from guide import _level_label
         subs = env.subs_by_title.get(card["block"], [])