Improve semantic classification precision for BUG-015

docs/architecture.md

@@ -375,6 +375,20 @@ extraction results into `meeting_protocol.md` for technical validation. The
 planned architecture separates Canonical Meeting Knowledge from the final Output
 Views documented in `output-views.md`.
 
+## Extraction classification contract
+
+Decision, Action Item and Open Question extraction shares one evidence-oriented
+classification contract. It is placed after the transcript so it remains the
+final classification instruction in the single multi-category extraction call.
+Decisions require a settled outcome; Action Items require established work;
+Open Questions require a concrete unresolved need. Unsupported candidates must
+not be moved into another category.
+
+Action existence and responsibility attribution are separate checks. A valid
+Action Item may have no known owner, while a named owner requires explicit
+assignment, volunteering or acceptance. These are semantic LLM classifications;
+deterministic validation must not guess intent from keywords.
+
 ## Semantic Consolidator failure handling
 
 Semantic Consolidator V0 preserves every raw model response before parsing.
@@ -396,6 +410,28 @@ an instruction not to emit an identical group more than once. Both attempts
 and the detected repetition metadata are preserved. If the retry also fails,
 the stage fails normally; it does not make another LLM call.
 
+## Working Protocol V2 renderer contract
+
+The renderer deterministically projects consolidated items to the semantic
+fields required for presentation and omits bulky provenance fields from the
+LLM request. Every renderable item remains represented; structurally empty
+items are recorded separately rather than turned into invented prose. The
+compact renderer input is preserved as an artifact.
+
+The exact Markdown structure is generated from the same heading constants used
+by the validator and appended after the renderer input so it remains visible
+within the evaluated context. Decisions, action items and open questions carry
+input-derived hidden coverage markers. Strict validation requires every such
+renderable priority item exactly once in its matching section and rejects
+missing, duplicate, wrong-section or invented markers. Facts and technical
+details remain condensable as background.
+
+Renderer output budgeting is adaptive to required priority content and prompt
+size, while an explicit `num_predict` override remains authoritative. Raw model
+output is always preserved, and `working_protocol.md` is written only after
+strict structure and coverage validation passes. The renderer does not retry
+automatically.
+
 ---
 
 # Next Milestone

docs/experiments.md

@@ -724,7 +724,7 @@ Evidence:
 
 ## EXP-0015 - Difficult synthetic meeting
 
-Status: Running
+Status: Accepted
 
 Date or period: 2026-07-30
 
@@ -1231,3 +1231,122 @@ Evidence:
 - `docs/output-views.md`
 - `prompts/working_protocol.md`
 - `tests/gold/responsibility_attribution_negative/`
+
+## EXP-0024 - Working Protocol V2 contract visibility
+
+Status: Running
+
+Date or period: 2026-08-09
+
+Target:
+
+BUG-011 renderer-only regression using an already validated Semantic
+Consolidator artifact.
+
+Hypothesis:
+
+The renderer failures are caused by provenance-heavy 199-350 KB JSON inputs
+placing the leading prompt contract outside the model's effective evaluated
+context. The preserved failures all report `prompt_eval_count=16386`, while
+their outputs either echo trailing JSON or produce an unconstrained generic
+category summary instead of the requested Working Protocol.
+
+Iteration 1 change:
+
+- Project every consolidated item to rendering-relevant semantic fields while
+  retaining every item and its category/text/responsibility/deadline/status
+  information.
+- Generate the exact structural contract from renderer validator constants and
+  append it after the compact INPUT JSON.
+- Replace the independently handwritten prompt skeleton with a reference to
+  that authoritative appended contract.
+- Enforce the existing prompt rule that emitted sections must not be empty.
+
+This is one renderer-contract prompt iteration. It does not change extraction,
+canonicalization, semantic consolidation or responsibility semantics.
+
+Validation before LLM run:
+
+- 15 focused renderer tests pass.
+- Tests cover contract generation, compact input projection, valid and invalid
+  headings, missing topic sections, empty sections, wrapper cleanup, malformed
+  Markdown and final-file write gating.
+
+Decision:
+
+Iteration 1 passed structural validation and wrote `working_protocol.md`, but
+the quality sanity check found that the model omitted most of the ten supplied
+decisions, two open questions and several action items. The structurally valid
+result therefore was not accepted as BUG-011 verification.
+
+Iteration 2 change:
+
+- Add input-derived hidden coverage markers for every decision, action item and
+  open question.
+- Require every priority item exactly once in its matching section.
+- Validate missing, duplicate, unknown and wrong-section markers
+  deterministically.
+- Keep facts and technical details condensable as background.
+
+This is the second single prompt iteration. It responds to the concrete
+omission failure observed in Iteration 1 without changing upstream semantics or
+inventing renderer content.
+
+Iteration 2 pre-run validation:
+
+- 18 focused renderer tests pass, including exact required-item coverage and
+  wrong-section rejection.
+
+Decision:
+
+Iteration 2 initially exhausted the fixed 4,096-token renderer output budget
+after emitting all decisions and most action items. Adaptive renderer budgeting
+resolved to 8,192 tokens for this input. The final run stopped normally after
+3,709 evaluated output tokens.
+
+The final renderer-only regression passed strict validation and wrote
+`working_protocol.md`. Exact coverage was 10/10 decisions, 36/36 renderable
+action items and 23/23 open questions, each once in its matching section. One
+structurally empty action item whose task, responsible, deadline and evidence
+were all null was recorded and excluded rather than fabricated. Optional
+background markers were accepted only when they referred to real projected
+input items.
+
+Accept the compact renderer input, validator-derived trailing contract,
+priority-item coverage markers, empty-section validation and adaptive renderer
+output sizing as the BUG-011 baseline. This establishes structural reliability
+and priority-item coverage, not complete protocol prose quality.
+
+Evidence:
+
+- `prompts/working_protocol.md`
+- `src/meeting_lab/protocol/render_working_protocol.py`
+- `tests/test_render_working_protocol.py`
+- `samples/benchmarks/progeo_qwen35_9b_20260805_133942/working_protocol/`
+- `samples/benchmarks/progeo_qwen35_35b_a3b_20260805_133942/working_protocol/`
+
+## EXP-0025 — BUG-015 classification precision
+
+Date: 2026-08-09
+
+Target: Progeo-derived Decision, Action Item and Open Question precision cases.
+
+Model/configuration: `qwen3.5:9B`, temperature 0, `num_ctx=32768`.
+
+Tests were created before prompt changes. A Decision-only evidence threshold
+kept the explicit Dr. Schlummer rejection and omitted an option and preference.
+Adding Action and Open Question definitions improved several negatives but was
+not stable: the model alternately promoted an unaccepted Textor suggestion or
+moved rejected candidates into Open Questions. Moving the standalone category
+prompts after the transcript made the partial Decision schema dominate and
+misclassified true Action Items as Decisions in two consecutive runs.
+
+The final iteration replaced the competing standalone category prompts with a
+single unified classification contract after the transcript. It preserved the
+assigned Nina task and ownerless established CET work, and prevented
+cross-category leakage in the focused case, but still emitted the unaccepted
+Textor suggestion as an Action Item. Further prompt iterations were stopped in
+accordance with the Gold Standard methodology.
+
+Result: partially improved, not accepted as a complete BUG-015 fix. BUG-015
+remains Open; no phrase-specific deterministic filter was introduced.

docs/regression-bugs.md

@@ -840,6 +840,51 @@ Verified. The renderer contract is now enforced deterministically. This does
 not improve semantic quality of the generated prose; it prevents invalid
 renderer output from being accepted as a final Working Protocol.
 
+Production-blocker follow-up, 2026-08-09:
+
+Later preserved renderer failures showed that enforcement alone did not make a
+valid protocol reliably obtainable. Provenance-heavy consolidated JSON inputs
+were 199-350 KB, while failed Ollama runs consistently evaluated 16,386 prompt
+tokens. The leading handwritten renderer contract was therefore effectively
+lost or underweighted: models echoed trailing JSON or produced generic category
+summaries. Prompt and validator also duplicated the structure independently,
+and the validator did not enforce the prompt's no-empty-section rule.
+
+The renderer now:
+
+- preserves every renderable item in a compact semantic projection while
+  removing source-reference and original-value bulk
+- records and excludes structurally empty items instead of inventing content
+- appends an authoritative contract generated from validator heading constants
+- rejects empty emitted sections
+- requires hidden, input-derived coverage markers for every decision, action
+  item and open question exactly once in the matching section
+- accepts optional background markers only for real projected input items
+- uses adaptive output sizing instead of the truncating fixed 4,096-token cap
+- keeps explicit output-budget overrides authoritative and makes no automatic
+  renderer retry
+
+Renderer-only verification used the validated consolidated input at
+`/tmp/meeting-lab-bug014-regression/consolidated_extractions.json`. The final
+run used `qwen3.5:9B`, temperature 0, `num_ctx=32768`, adaptive
+`num_predict=8192`, and stopped normally with `eval_count=3709` and
+`done_reason=stop` after 53.069 seconds. Strict validation reported no
+violations and wrote:
+
+- `/tmp/meeting-lab-bug011-renderer-final/working_protocol.md`
+
+Coverage was complete for all renderable priority items: 10 decisions, 36
+action items and 23 open questions, with no missing or duplicate markers. One
+upstream action item containing only null task/responsibility/deadline/evidence
+was recorded as structurally empty and not rendered. The protocol retained
+substantial technical background and did not add a new named responsibility;
+the only structured responsible person in the renderer input remained Marleen.
+
+Status remains Verified for Working Protocol V2 structural validity and
+priority-item coverage. This does not claim full semantic or editorial protocol
+quality, topic quality, or resolution of the other renderer-related regression
+bugs.
+
 ## BUG-012
 
 ID: BUG-012
@@ -1124,3 +1169,36 @@ source-ID occurrences and 113 unique canonical source IDs. There were no
 missing IDs, unknown IDs or duplicate occurrences. BUG-014 is independent of
 BUG-013: BUG-013 detects invalid JSON caused by runaway repeated groups,
 whereas BUG-014 repairs unknown source IDs in parseable model grouping JSON.
+
+## BUG-015
+
+ID: BUG-015
+
+Title: Extraction classification precision for Decisions, Action Items, and Open Questions
+
+Status: Open
+
+First observed: Progeo production benchmark and BUG-011 renderer-only output
+
+The Progeo extraction classified proposals/options as Decisions, suggestions
+and hypothetical work as Action Items, and uncertainty or discussion fragments
+as Open Questions. The production prompt combined a weak shared German rule,
+a standalone Decision prompt, a responsibility-only Action prompt, and no
+Open-Question definition in one multi-category extraction request.
+
+Regression fixtures now record explicit positive and negative evidence for all
+three categories. Extraction uses one unified classification contract after the
+transcript, with precision-first thresholds, category boundaries, evidence
+requirements, and the responsibility-attribution invariant. In a seven-chunk
+Progeo measurement, bare uncertainty in chunk 08 stopped becoming an Open
+Question and the preference about real plant versus Technikum in chunk 16
+stopped becoming a Decision. Supported examples including the explicit Dr.
+Schlummer rejection and established CET work remained detectable.
+
+BUG-015 is not Verified. With `qwen3.5:9B`, temperature 0, the focused Action
+fixture still classified the unaccepted suggestion to contact Dirk Textor as
+an Action Item. Earlier prompt variants also showed category leakage by turning
+rejected candidates into Open Questions or Decisions. Prompt iterations were
+stopped under the Gold Standard methodology rather than adding phrase-specific
+filters. A general semantic classification architecture improvement remains
+necessary before this bug can be marked Verified.

prompts/classification.md

@@ -0,0 +1,55 @@
+Classification contract for Decisions, Action Items, and Open Questions:
+
+Classify each supported proposition into the one best category. Do not move an
+unsupported candidate into another category merely to retain it. Prefer
+omission over a false Decision or Action Item.
+
+Decision:
+
+A Decision requires evidence that the meeting settled, selected, approved,
+rejected, or committed to an outcome. Explicit agreement, selection,
+approval/rejection, commitment, or a clear statement that a decision was made
+is sufficient. A proposal, suggestion, preference, possibility, option,
+recommendation, consideration, or unresolved plan is not sufficient without
+explicit acceptance or commitment. A personal commitment to perform work is an
+Action Item, not a Decision, unless the meeting also establishes a distinct
+group-level outcome.
+
+Action Item:
+
+An Action Item is a concrete future action that the meeting establishes as work
+to be done. Explicit assignment, explicit acceptance, a personal commitment,
+or clear evidence that the work is already established and underway is
+sufficient. A suggestion, hypothetical action, possible next step,
+recommendation, brainstorming option, or exploratory statement is not
+sufficient. Do not create an Action Item merely because a statement contains an
+imperative-like verb.
+
+First decide whether the action itself exists. Only then attribute
+responsibility. An Action Item may exist without a named responsible person. A
+person may be named only when explicitly assigned, volunteering, or explicitly
+accepting/confirming the task. Mention, proximity, expertise, role, a request,
+or language that someone should/could/might act does not establish
+responsibility. Do not preserve an unsupported action merely by setting
+responsible to null.
+
+Open Question:
+
+An Open Question requires a concrete unresolved information, decision, or
+clarification need that remains open after the available context. An explicit
+question left unanswered or an explicit statement that clarification remains
+needed is sufficient. Uncertainty, doubt, ignorance, difficulty, an
+observation, an incomplete fragment, or a rhetorical question alone is not
+sufficient. A question answered in the available context is not open. Do not
+invent question wording or convert a rejected Decision/Action candidate into an
+Open Question without independent evidence of an unresolved need.
+
+Evidence discipline:
+
+- Every emitted item must include the shortest transcript evidence that proves
+  the category threshold.
+- Evidence for an Action Item must prove both the concrete work and that it was
+  established, not merely possible.
+- Evidence for an Open Question must prove both the specific need and that it
+  remains unresolved.
+- Use an empty list when a category has no supported item.

prompts/common.md

@@ -11,11 +11,18 @@ Regeln:
 4. Wenn etwas widersprüchlich oder unklar ist, kennzeichne es als unklar.
 5. Übernimm wichtige technische Aussagen möglichst nah am Wortlaut.
 6. Nenne bei Fakten nach Möglichkeit den Sprecher.
-7. Ein Beschluss ist nur dann ein Beschluss, wenn im Text eine Einigung,
+7. Ein Beschluss ist nur dann ein Beschluss, wenn das Meeting etwas erkennbar
-   Freigabe oder verbindliche Festlegung erkennbar ist.
+   geeinigt, ausgewählt, freigegeben, abgelehnt oder verbindlich festgelegt hat.
-8. Eine Aufgabe ist nur dann eine Aufgabe, wenn eine Handlung und möglichst
+8. Eine Aufgabe ist nur dann eine Aufgabe, wenn das Meeting eine konkrete
-   eine verantwortliche Person oder Organisation erkennbar sind.
+   zukünftige Handlung als zu erledigende Arbeit festgelegt hat. Vorschläge,
-9. Gib ausschließlich gültiges JSON aus. Kein Markdown, keine Erläuterungen.
+   Möglichkeiten und hypothetische nächste Schritte genügen nicht.
+9. Prüfe erst, ob die Aufgabe selbst belegt ist. Ordne danach nur bei expliziter
+   Zuweisung oder Annahme eine verantwortliche Person zu. Eine belegte Aufgabe
+   darf auch keine bekannte verantwortliche Person haben.
+10. Eine offene Frage erfordert einen konkreten, weiterhin offenen Informations-,
+    Entscheidungs- oder Klärungsbedarf. Bloße Unsicherheit genügt nicht; eine im
+    verfügbaren Kontext beantwortete Frage bleibt nicht offen.
+11. Gib ausschließlich gültiges JSON aus. Kein Markdown, keine Erläuterungen.
 
 Hinweise zur Ausgabe:
 - Alle obersten Schlüssel müssen vorhanden sein.

src/meeting_lab/extraction/extract_chunks.py

@@ -41,7 +41,7 @@ EXTRACTION_CATEGORIES = (
 
 NORMALIZED_CHUNK_RE = re.compile(r"^(chunk_\d+)_normalized\.txt$")
 
-EXTRACTION_TASK_PROMPT_NAMES = ("decisions.md", "todos.md")
+EXTRACTION_TASK_PROMPT_NAMES = ("classification.md",)
 
 
 OUTPUT_SCHEMA = {

src/meeting_lab/llm/prompts.py

@@ -38,7 +38,6 @@ def build_extraction_prompt(
     prompt_parts = [
         load_prompt("common.md", prompts_dir),
         meeting_context,
-        *load_existing_prompts(task_prompt_names, prompts_dir),
         f"""Quelldatei:
 {source_name}
 
@@ -49,5 +48,6 @@ TRANSKRIPT:
 --- BEGINN TRANSKRIPT ---
 {transcript}
 --- ENDE TRANSKRIPT ---""",
+        *load_existing_prompts(task_prompt_names, prompts_dir),
     ]
     return "\n\n".join(part for part in prompt_parts if part).strip() + "\n"

tests/gold/progeo_action_precision/README.md

@@ -0,0 +1,8 @@
+# progeo_action_precision
+
+Minimal German regression derived from Progeo chunks 2, 10, 17 and 18. It
+separates suggestions and exploratory possibilities from established work and
+tests explicit acceptance of responsibility.
+
+Expected semantics: only the accepted review task and the already-established
+article work are Action Items.

tests/gold/progeo_action_precision/expected.json

@@ -0,0 +1,21 @@
+{
+  "facts": [],
+  "decisions": [],
+  "todos": [
+    {
+      "task": "Messdaten bis Freitag prüfen.",
+      "responsible": "Nina",
+      "deadline": "Freitag",
+      "evidence": "Nina: Ja, ich übernehme die Prüfung bis Freitag."
+    },
+    {
+      "task": "CET-Artikel erstellen und anschließend im Konsortium verteilen.",
+      "responsible": null,
+      "deadline": null,
+      "evidence": "Den CET-Artikel erstellen wir bereits und schicken ihn anschließend im Konsortium herum."
+    }
+  ],
+  "questions": [],
+  "positions": [],
+  "technical": []
+}

tests/gold/progeo_action_precision/transcript.txt

@@ -0,0 +1,11 @@
+Martin: Die Geometrie kann man vielleicht noch optimieren. Dann würde man mal gucken, was herauskommt.
+
+Antonius: Marleen müsste vielleicht mal äußern, was Hüsker sich unter dem Versuch vorstellt.
+
+Tim: Wir könnten Dirk Textor vielleicht noch einmal kontaktieren.
+
+Antonius: Nina, übernimmst du die Prüfung der Messdaten bis Freitag?
+
+Nina: Ja, ich übernehme die Prüfung bis Freitag.
+
+Tim: Den CET-Artikel erstellen wir bereits und schicken ihn anschließend im Konsortium herum.

tests/gold/progeo_decision_precision/README.md

@@ -0,0 +1,6 @@
+# progeo_decision_precision
+
+Minimal German regression derived from Progeo chunks 12 and 16. It contrasts
+an option and a personal preference with an explicit group rejection.
+
+Expected semantics: only the explicit rejection is a Decision.

tests/gold/progeo_decision_precision/expected.json

@@ -0,0 +1,13 @@
+{
+  "facts": [],
+  "decisions": [
+    {
+      "decision": "Die Gruppe lehnt die Zusammenarbeit mit Dr. Schlummer ab.",
+      "evidence": "Dann haben wir gesagt: Nein, die Zusammenarbeit mit Dr. Schlummer machen wir nicht. Antonius: Ja, das ist entschieden."
+    }
+  ],
+  "todos": [],
+  "questions": [],
+  "positions": [],
+  "technical": []
+}

tests/gold/progeo_decision_precision/transcript.txt

@@ -0,0 +1,9 @@
+Tim: Die Option steht im Raum, das Material chemisch recyceln zu lassen.
+
+Martin: Ich würde nicht in eine reale Anlage gehen. Wenn überhaupt, können wir über ein Technikum reden.
+
+Antonius: Das Angebot von Dr. Schlummer kostet 30.000 Euro.
+
+Tim: Dann haben wir gesagt: Nein, die Zusammenarbeit mit Dr. Schlummer machen wir nicht.
+
+Antonius: Ja, das ist entschieden.

tests/gold/progeo_question_precision/README.md

@@ -0,0 +1,8 @@
+# progeo_question_precision
+
+Minimal German regression derived from Progeo chunks 8 and 18. It contrasts
+uncertainty, an observation and an immediately answered question with an
+explicitly unresolved information need.
+
+Expected semantics: only the explicitly unresolved publication question is an
+Open Question.

tests/gold/progeo_question_precision/expected.json

@@ -0,0 +1,13 @@
+{
+  "facts": [],
+  "decisions": [],
+  "todos": [],
+  "questions": [
+    {
+      "question": "Welche Daten aus dem Energieaudit dürfen veröffentlicht werden?",
+      "evidence": "Welche Daten aus dem Energieaudit dürfen wir veröffentlichen? Martin: Das ist weiterhin ungeklärt."
+    }
+  ],
+  "positions": [],
+  "technical": []
+}

tests/gold/progeo_question_precision/transcript.txt

@@ -0,0 +1,11 @@
+Martin: Ob sich das Waschen lohnt, weiß ich nicht.
+
+Tim: Das lässt sich teilweise optimieren und teilweise nicht.
+
+Antonius: Gibt es schon ein Programm für die Veranstaltung?
+
+Tim: Ja, das Programm steht auf der Dechema-Seite und ist vollständig.
+
+Antonius: Welche Daten aus dem Energieaudit dürfen wir veröffentlichen?
+
+Martin: Das ist weiterhin ungeklärt. Wir müssen die Freigabe noch klären.

tests/test_classification_invariants.py

@@ -0,0 +1,36 @@
+import unittest
+
+from src.meeting_lab.extraction.extract_chunks import build_prompt
+
+
+class ClassificationInvariantPromptTests(unittest.TestCase):
+    def setUp(self) -> None:
+        self.prompt = build_prompt("regression.txt", "Meeting transcript.")
+
+    def test_decision_requires_settled_outcome_and_prefers_precision(self) -> None:
+        self.assertIn("settled, selected, approved", self.prompt)
+        self.assertIn("possibility, option", self.prompt)
+        self.assertIn("Prefer\nomission over a false Decision or Action Item", self.prompt)
+
+    def test_action_requires_established_future_work(self) -> None:
+        self.assertIn("concrete future action that the meeting establishes as", self.prompt)
+        self.assertIn("suggestion, hypothetical action, possible next step", self.prompt.lower())
+        self.assertIn("Do not create an Action Item merely because", self.prompt)
+        self.assertIn("statement contains an", self.prompt)
+        self.assertIn("imperative-like verb", self.prompt)
+
+    def test_action_existence_is_separate_from_responsibility(self) -> None:
+        self.assertIn("First decide whether the action itself exists", self.prompt)
+        self.assertIn("An Action Item may exist without a named responsible", self.prompt)
+        self.assertIn("responsible person", self.prompt)
+        self.assertIn("explicitly assigned, volunteering", self.prompt)
+        self.assertIn("accepting/confirming the task", self.prompt)
+
+    def test_open_question_requires_unresolved_need(self) -> None:
+        self.assertIn("concrete unresolved information, decision, or", self.prompt)
+        self.assertIn("Uncertainty, doubt, ignorance, difficulty", self.prompt)
+        self.assertIn("answered in the available context is not open", self.prompt)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_extraction_protocol.py

@@ -101,17 +101,20 @@ Final answer:
         prompt = build_prompt("transcript.txt", "Anna: Agreed.")
 
         self.assertIn("Du extrahierst Informationen aus Meeting-Transkripten.", prompt)
-        self.assertIn("You extract decisions from meeting transcript text.", prompt)
+        self.assertIn("Classification contract for Decisions", prompt)
-        self.assertIn("Extract each decision as one atomic commitment.", prompt)
+        self.assertIn("A Decision requires evidence", prompt)
         self.assertIn("Anna: Agreed.", prompt)
 
     def test_build_prompt_includes_todo_prompt_file(self) -> None:
         prompt = build_prompt("transcript.txt", "Nina: I will update it.")
 
-        self.assertEqual(EXTRACTION_TASK_PROMPT_NAMES, ("decisions.md", "todos.md"))
+        self.assertEqual(
-        self.assertIn("Action-item responsibility rule:", prompt)
+            EXTRACTION_TASK_PROMPT_NAMES,
+            ("classification.md",),
+        )
+        self.assertIn("Action Item:", prompt)
         self.assertIn(
-            "A named responsible person may be extracted only when the transcript contains",
+            "person may be named only when explicitly assigned",
             prompt,
         )
 
@@ -131,8 +134,9 @@ Final answer:
         )
 
         for prompt in (no_context_prompt, context_prompt):
-            self.assertIn("You extract decisions from meeting transcript text.", prompt)
+            self.assertIn("Classification contract for Decisions", prompt)
-            self.assertIn("Action-item responsibility rule:", prompt)
+            self.assertIn("Action Item:", prompt)
+            self.assertIn("Open Question:", prompt)
 
     def test_prompt_assembly_is_deterministic(self) -> None:
         first = build_prompt("transcript.txt", "Nina: I will update it.")
@@ -140,6 +144,20 @@ Final answer:
 
         self.assertEqual(first, second)
 
+    def test_classification_contracts_follow_transcript(self) -> None:
+        prompt = build_prompt("transcript.txt", "SOURCE_SENTINEL")
+
+        transcript_end = prompt.index("--- ENDE TRANSKRIPT ---")
+        self.assertGreater(prompt.index("Classification contract", transcript_end), transcript_end)
+        self.assertGreater(
+            prompt.index("Action Item:", transcript_end),
+            transcript_end,
+        )
+        self.assertGreater(
+            prompt.index("Open Question:", transcript_end),
+            transcript_end,
+        )
+
     def test_build_protocol_groups_extraction_items(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             input_dir = Path(directory)