Upload 7 files

.gitattributes

@@ -1,35 +1,36 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+logo.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,204 @@
----
-title: Bertopic Agent Final
-emoji: 📈
-colorFrom: red
-colorTo: purple
-sdk: gradio
-sdk_version: 6.15.1
-python_version: '3.13'
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: Bertopic Agent V2
+emoji: 📉
+colorFrom: yellow
+colorTo: blue
+sdk: gradio
+sdk_version: 6.14.0
+python_version: '3.13'
+app_file: app.py
+pinned: false
+license: mit
+---
+
+# 🔬 BERTopic Agentic Topic Modelling
+
+### *Computational Thematic Analysis powered by Braun & Clarke (2006)*
+
+![BERTopic Agent Logo](logo.png)
+
+---
+
+## 🌟 Overview
+
+**BERTopic Agentic Topic Modelling** is a state-of-the-art research tool designed to automate and enhance the process of **Thematic Analysis** for academic literature. By integrating **BERTopic**'s transformer-based clustering with a **LangGraph-driven agentic workflow**, this application guides researchers through the rigorous 6-phase framework of Braun & Clarke (2006).
+
+It doesn't just cluster text; it *reasons* about it. Featuring a unique **"AI Council"** where multiple Large Language Models (Mistral & Groq) debate and reach consensus on topic labels, the tool ensures high-fidelity, publishable results.
+
+---
+
+## 🧠 Theoretical Foundation: Braun & Clarke (2006)
+
+This tool is strictly mapped to the six phases of thematic analysis as defined in the seminal work:
+
+1.  **Familiarisation with data**: Automatic cleaning, boilerplate removal, and dataset profiling.
+2.  **Generating initial codes**: BERTopic discovery and AI-assisted initial labeling.
+3.  **Searching for themes**: LLM-driven consolidation of topics into overarching themes.
+4.  **Reviewing potential themes**: Saturation checks and coverage analysis.
+5.  **Defining and naming themes**: Generation of academic definitions and core narratives.
+6.  **Producing the report**: Narrative writing (Section 7 draft) and PAJAIS taxonomy mapping.
+
+---
+
+## ✨ Key Features
+
+- **🤖 Agentic Workflow**: A LangGraph agent manages the entire pipeline, maintaining memory and ensuring a step-by-step scientific process.
+- **⚖️ AI Council**: Real-time debates between **Mistral-Large** and **Llama-3 (Groq)** to determine the most accurate thematic labels.
+- **📊 Dynamic Visualizations**: 8+ interactive Plotly charts (Intertopic maps, Frequency bars, Heatmaps, Treemaps, and DBSCAN scatter plots).
+- **🛡️ Multi-Model Analysis**: Run separate analyses on **Abstracts** vs. **Titles** and generate a side-by-side convergence CSV.
+- **🔍 Density Refinement**: Optional **DBSCAN** clustering to complement traditional hierarchical methods and handle noise points elegantly.
+- **🏷️ PAJAIS Taxonomy Mapping**: Automated gap analysis by mapping themes to the standard 25 PAJAIS Information Systems categories.
+- **📥 One-Click Export**: Download structured JSON, side-by-side CSVs, PNG charts, and a 500-word academic narrative report.
+
+---
+
+## 🛠️ Architecture
+
+```mermaid
+graph TD
+    A[Scopus CSV Upload] --> B{Agentic Workflow}
+    B -->|Phase 1| C[Data Loading & Cleaning]
+    C -->|Phase 2| D[BERTopic / DBSCAN Discovery]
+    D --> E[AI Council Labeling]
+    E -->|Phase 3| F[Theme Consolidation]
+    F -->|Phase 4| G[Saturation Check]
+    G -->|Phase 5| H[Definition & Naming]
+    H -->|Phase 5.5| I[PAJAIS Taxonomy Mapping]
+    I -->|Phase 6| J[Report Generation]
+    
+    subgraph "AI Council"
+    E1[Mistral-Large] <--> E2[Groq Llama-3]
+    end
+    
+    subgraph "Outputs"
+    J --> K[narrative.txt]
+    J --> L[comparison.csv]
+    J --> M[Interactive Charts]
+    end
+```
+
+---
+
+## 🖥️ App Navigation & Expected UI
+
+The interface is divided into three logical zones for a streamlined user experience:
+
+### 1. Control Center (Top & Left)
+- **Phase Progress Bar**: A visual indicator of your progress through Braun & Clarke’s 6 phases.
+- **Data Input (Left)**: The upload zone for your Scopus CSV. Once uploaded, Phase 1 triggers automatically.
+
+### 2. The Agent Laboratory (Center)
+- **Chatbot Interface**: Your main point of interaction. The agent will ask questions, provide stats, and guide you. You can type commands like "run abstract" or "Continue".
+- **AI Council Feedback**: Every time a label is generated, look for the reasoning block. It shows the consensus score between models.
+
+### 3. Results Dashboard (Bottom Tabs)
+- **📋 Review Table**: The "Heart" of the app. This is where you approve, rename, and refine the AI's findings. You MUST click **"Submit Review"** to move past STOP GATES.
+- **📈 Charts Tab**: Switch between **Intertopic Map**, **Frequency Bars**, **Hierarchy (Treemap)**, and **Similarity Heatmap**.
+- **⚖️ AI Council Tab**: A dedicated view showing the full transcript of debates between Mistral and Groq.
+- **💾 Download Tab**: Your final repository. All files are generated in real-time and appear here for one-click downloading.
+
+### 📤 Expected Output Preview
+- **In Chat**: Summary tables, saturation percentages (e.g., "92.4% Coverage"), and phase completion checkmarks.
+- **In Files**:
+  - `narrative.txt`: Academic prose with structured headings.
+  - `comparison.csv`: Columns for `Abstract Theme`, `Title Theme`, and `Convergence` (marked with ✓).
+  - `taxonomy_map.json`: A mapping showing each theme's link to the PAJAIS framework and its **Novelty score**.
+
+---
+
+
+### 1. Prerequisites
+- Python 3.9+
+- API Keys for **Mistral AI** and **Groq** (optional but recommended for the Council feature).
+
+### 2. Installation
+
+Clone the repository and install the dependencies:
+
+```bash
+# Clone the repo
+git clone https://github.com/ShivamKadam63s/BERT_Topic_Modelling.git
+cd BERT_Topic_Modelling
+
+# Install dependencies
+pip install -r requirements.txt
+```
+
+### 3. Environment Setup
+
+Create a `.env` file or export your API keys in your terminal:
+
+```powershell
+$env:MISTRAL_API_KEY="your_mistral_key"
+$env:GROQ_API_KEY="your_groq_key"
+```
+
+### 4. Running the App
+
+Start the Gradio interface:
+
+```bash
+python app.py
+```
+
+Open your browser at `http://localhost:7860`.
+
+---
+
+## 📖 User Guide: Phase-by-Phase Walkthrough
+
+### Step 1: Data Input
+Upload your **Scopus CSV** file. The agent will immediately scan the file, remove boilerplate text (Copyright notices, DOIs, etc.), and provide a dataset profile including paper counts and year ranges.
+
+### Step 2: Discovery & Coding
+- Click **"run abstract"** or **"run title"**.
+- The system will generate clusters and invoke the **AI Council**.
+- **Navigation**: Check the **"⚖️ AI Council"** tab to see the reasoning behind each label.
+- **Action**: In the **"📋 Review Table"**, tick **Approve** for clusters you accept or provide a custom name in **Rename To**. Click **"Submit Review"**.
+
+### Step 3: Themes & Saturation
+The agent combines approved codes into 4-8 themes. It will report **Thematic Saturation** (e.g., "Themes cover 92% of the corpus").
+
+### Step 4: Taxonomy Mapping
+The tool automatically maps your themes to the **PAJAIS Taxonomy**. 
+- Themes marked with 🌟 **NOVEL** are identified as potential new research contributions not found in standard taxonomies.
+
+### Step 5: Final Report
+The agent generates a **500-word Section 7 draft**. Check the **"💾 Download"** tab for your full suite of results.
+
+---
+
+## 📈 Expected Outputs
+
+| Output File | Description |
+| :--- | :--- |
+| `narrative.txt` | A complete Section 7 draft following academic standards. |
+| `comparison.csv` | Side-by-side comparison of Abstract and Title themes. |
+| `taxonomy_map.json` | JSON mapping of themes to PAJAIS categories. |
+| `chart_*.html` | Interactive Plotly visualizations for intertopic distance and hierarchy. |
+| `*.png` | High-resolution static exports of all charts. |
+
+---
+
+## 🛠️ Built With
+
+- **Gradio**: Modern UI Framework
+- **LangGraph**: Agentic Multi-Model Workflows
+- **BERTopic**: Advanced Topic Modeling
+- **Sentence-Transformers**: `all-MiniLM-L6-v2` embeddings
+- **Mistral Large**: Primary Reasoning LLM
+- **Groq (Llama-3)**: Secondary Council LLM
+- **Plotly**: Dynamic Data Science Charts
+
+---
+
+## ⚖️ License & Citation
+
+If you use this tool in your research, please cite:
+*Shivam Kadam, "BERTopic Agentic Topic Modelling for Systematic Literature Reviews," 2026.*
+
+Based on:
+*Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77-101.*
+
+---
+<p align="center">Made with ❤️ for the Research Community</p>

agent.py

@@ -0,0 +1,519 @@
+# agent.py — Braun & Clarke Thematic Analysis Agent
+# LangGraph ReAct agent with ChatMistralAI and MemorySaver checkpointer.
+# Verified: exactly 4 STOP gates implemented (after Phase 2, 3, 4, 5.5)
+
+from langchain_mistralai import ChatMistralAI
+from langgraph.prebuilt import create_react_agent
+from langgraph.checkpoint.memory import MemorySaver
+from tools import (
+    load_scopus_csv,
+    run_bertopic_discovery,
+    label_topics_with_llm,
+    consolidate_into_themes,
+    compare_with_taxonomy,
+    generate_comparison_csv,
+    export_narrative,
+    # ── New additive tools (DBSCAN + AI Council) ──
+    run_dbscan_clustering,
+    refine_large_clusters,
+    run_ai_council,
+)
+
+# ─────────────────────────────────────────────────────────────────────────────
+# SYSTEM PROMPT (~500 lines) — Braun & Clarke (2006) Thematic Analysis Agent
+# ─────────────────────────────────────────────────────────────────────────────
+SYSTEM_PROMPT = """
+================================================================================
+IDENTITY & ROLE
+================================================================================
+You are a computational thematic analysis agent implementing the Braun & Clarke
+(2006) six-phase thematic analysis framework on academic literature corpora
+exported from Scopus. You are embedded in a Gradio web application that
+provides the researcher with a chat interface, a review table, charts, and file
+downloads.
+
+You have memory across the entire conversation via LangGraph MemorySaver.
+You are powered by Mistral LLM and have access to 10 specialised tools.
+Tools 1–7 implement the core Braun & Clarke pipeline (unchanged).
+Tools 8–10 provide optional DBSCAN clustering and AI Council labelling.
+
+Your purpose: guide the researcher through all 6 Braun & Clarke phases to
+produce publishable thematic analysis results, including a PAJAIS taxonomy
+mapping and a written narrative for Section 7 of their paper.
+
+================================================================================
+CRITICAL OPERATING RULES — OBEY EVERY ONE, EVERY TIME
+================================================================================
+
+RULE 1 — ONE PHASE PER MESSAGE:
+  Execute exactly one phase per response. Never jump ahead, never combine
+  phases, never rush. Respect the researcher's pace.
+
+RULE 2 — 4 STOP GATES ARE ABSOLUTE:
+  There are exactly 4 STOP gates in this pipeline:
+    STOP GATE 1: After Phase 2 (wait for Submit Review from table)
+    STOP GATE 2: After Phase 3 (wait for "Continue" or Submit Review)
+    STOP GATE 3: After Phase 4 (wait for "Continue" or Submit Review)
+    STOP GATE 4: After Phase 5.5 (wait for "Continue" or Submit Review)
+  At each gate: display "⛔ STOP GATE [N]", summarise what was done,
+  and explicitly state what you are waiting for. DO NOT proceed until received.
+
+RULE 3 — ALL APPROVALS VIA REVIEW TABLE:
+  Never ask the researcher to approve topics, themes, or mappings via chat.
+  All approvals, renames, and reasoning belong in the Review Table.
+  The researcher clicks "Submit Review to Agent" when ready.
+
+RULE 4 — NEVER HALLUCINATE DATA:
+  Every number, label, or topic you mention must come from a tool's return
+  value. Do not invent statistics, topic names, or paper counts.
+
+RULE 5 — COLUMN USAGE:
+  RUN_CONFIGS = { "abstract": ["Abstract"], "title": ["Title"] }
+  Never use Author Keywords, Index Keywords, Source Title, or any other
+  column for BERTopic clustering. These columns introduce bias.
+
+RULE 6 — TOOL CALL ORDER:
+  Only call tools in the order specified per phase. Never call a tool from
+  a later phase while in an earlier phase.
+
+RULE 7 — TRANSPARENCY:
+  After every tool call, explain in plain English what the tool did,
+  what the key numbers mean, and what the researcher should do next.
+
+RULE 8 — ERROR RECOVERY:
+  If a tool returns an error message, report it clearly to the researcher,
+  suggest a likely fix (e.g., wrong column name, missing file), and wait
+  for the researcher to confirm before retrying.
+
+RULE 9 — PROGRESS BAR UPDATES:
+  After completing each phase, output a line in the exact format:
+  PHASE_STATUS: 1=✅,2=⬜,3=⬜,4=⬜,5=⬜,5.5=⬜,6=⬜
+  (with the completed phases marked ✅). The UI parses this line.
+
+RULE 10 — NO AUTO-ADVANCE:
+  Never say "I will now proceed to Phase N" without explicit user approval.
+  The word "Continue" or a Submit Review action is required at each gate.
+
+RULE 11 — STRICT TOOL CALLS:
+  When calling a tool, use ONLY the tool name and arguments. Never prefix or
+  suffix the tool call with exploratory conversational text (e.g., "I will 
+  now call..." or garbage tokens like "onderlinge"). Output the tool call 
+  precisely as defined.
+
+================================================================================
+TOOLS — DESCRIPTIONS AND WHEN TO USE EACH
+================================================================================
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 1: load_scopus_csv(file_path: str)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Load and validate the uploaded Scopus CSV file.
+  When    : Phase 1 ONLY. Immediately when the researcher uploads a file.
+  Returns : papers, abstract_sentences, title_sentences, year_range, columns,
+            coverage percentages, sample_titles.
+  Action  : Display all statistics. Ask researcher to confirm run_key.
+            Save loaded_data.csv (tool does this automatically).
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 2: run_bertopic_discovery(run_key: str, threshold: float = 0.7)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Core clustering. Splits text to sentences → embeds with
+            all-MiniLM-L6-v2 → AgglomerativeClustering (cosine, average,
+            threshold=0.7) → NO UMAP → finds 5 nearest sentences per centroid
+            → generates 4 Plotly HTML charts → saves summaries_{run_key}.json
+            and emb_{run_key}.npy.
+  When    : After Phase 1.
+  Returns : n_topics, chart files, data preview.
+  Action  : Report topic counts. Tell researcher the Intertopic Map and local
+            Frequency Bars are ready. 
+            NEW: Explicitly tell the user: "You can now optionally run DBSCAN 
+            clustering to compare these results with a density-based method 
+            by typing 'run dbscan'."
+            Ask for approval to proceed to Phase 3.
+  STOP    : Wait for "Continue" before Phase 3.
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 3: label_topics_with_llm(run_key: str)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Send top 100 topics to Mistral (PromptTemplate + JsonOutputParser).
+            Each topic gets: label, category, confidence, reasoning, niche.
+            Saves labels_{run_key}.json.
+  When    : Phase 2 ONLY. Immediately after run_bertopic_discovery.
+  Returns : total_labelled, preview of first 5 labelled topics.
+  Action  : Populate Review Table with labelled topics.
+            Trigger STOP GATE 1.
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 4: consolidate_into_themes(run_key: str, theme_map: str)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Merge approved topic clusters into 4–8 overarching themes.
+            Recomputes centroids and recounts sentences/papers per theme.
+            Saves themes_{run_key}.json and themes.json (canonical).
+  When    : Phase 3 ONLY. After STOP GATE 1 is cleared.
+  Input   : theme_map = JSON string {"Theme Name": [topic_id, ...]} from table.
+            If empty, LLM auto-consolidates.
+  Returns : total_themes, themes_preview.
+  Action  : Display themes. Populate Review Table with theme-level rows.
+            Trigger STOP GATE 2.
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 5: compare_with_taxonomy(run_key: str)
+────────────��───────────────────────────────────────────────────────────────────
+  Purpose : Map each theme to PAJAIS 25 categories. Returns MAPPED or NOVEL
+            per theme. Saves taxonomy_map.json.
+  When    : Phase 5.5 ONLY. After Phase 5 naming is confirmed.
+  Returns : total_themes_mapped, novel_themes count, mapped_themes count, mapping.
+  Action  : Populate Review Table — "Top Evidence" column shows:
+            "→ PAJAIS MATCH: [category] | [reasoning]" or
+            "→ NOVEL | [reasoning]"
+            Trigger STOP GATE 4.
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 6: generate_comparison_csv()
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Load themes from both abstract and title runs, create side-by-side
+            comparison DataFrame. Requires themes_abstract.json and
+            themes_title.json. Saves comparison.csv.
+  When    : Phase 6 ONLY. After STOP GATE 4 is cleared.
+  Returns : output file path, row count, preview.
+  Action  : Tell researcher to check Download tab for comparison.csv.
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 7: export_narrative(run_key: str)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Generate a 500-word Section 7 narrative using Mistral LLM.
+            Covers methodology, themes, PAJAIS alignment, limitations, implications.
+            Saves narrative.txt.
+  When    : Phase 6 ONLY. After generate_comparison_csv.
+  Returns : output file path, word count, 500-char preview.
+  Action  : Display preview in chat. Add narrative.txt to Download tab.
+            Mark all phases complete. Display final success message.
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 8: run_dbscan_clustering(run_key: str, eps: float = 0.3, min_samples: int = 3)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Run DBSCAN on the SAME embeddings from run_bertopic_discovery.
+            Works in 384-dim cosine space (no UMAP). Parallel to agglomerative
+            clustering — outputs stored SEPARATELY (dbscan_summaries_{run_key}.json).
+            Generates 2 charts: DBSCAN scatter and cluster-count comparison.
+  When    : OPTIONAL. After Phase 2 completes (emb_{run_key}.npy must exist).
+            Researcher triggers with: "run dbscan" or "compare clustering methods".
+  Returns : n_clusters, noise_points, largest_cluster, chart files.
+  Action  : Report DBSCAN stats vs agglomerative in chat. Tell researcher the
+            new DBSCAN charts are available in the Charts tab.
+            Do NOT interrupt the main Braun & Clarke pipeline.
+
+────────────────────────────────────────────────────────────────────────────────
+TOOL 9: refine_large_clusters(run_key: str, size_threshold: int = 200)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Splits DBSCAN clusters larger than size_threshold into sub-clusters
+            using tighter AgglomerativeClustering (threshold=0.45).
+            Does NOT modify any existing agglomerative or DBSCAN outputs.
+            Saves refined_clusters_{run_key}.json.
+  When    : OPTIONAL. After run_dbscan_clustering has completed.
+            Researcher triggers with: "refine large clusters" or similar.
+  Returns : n_large_refined, total_subclusters, chart file.
+  Action  : Report which clusters were refined and how many sub-clusters created.
+
+──────────────��─────────────────────────────────────────────────────────────────
+TOOL 10: run_ai_council(run_key: str)
+────────────────────────────────────────────────────────────────────────────────
+  Purpose : Two genuinely different LLMs independently label each DBSCAN cluster:
+            - Model A: Mistral Large (temperature=0.2) — analytical, precise
+            - Model B: Groq Llama-3.3-70b-versatile — genuinely independent model,
+              providing a Karpathy-style second opinion from a different architecture.
+            A Jaccard-based consensus step resolves agreements (≥0.4 word overlap
+            → agreed, use Model A label) vs divergences (Model A selected as primary).
+            Saves council_labels_{run_key}.json (PAJAIS-compatible: has 'label' field).
+  When    : OPTIONAL. After run_dbscan_clustering has completed.
+            Researcher triggers with: "run ai council" or "council labels".
+  Returns : total_labelled, agreement_rate, output_file.
+  Action  : Report agreement rate and a table of label_a vs label_b in chat.
+            Mention that council_labels_{run_key}.json is in the Download tab.
+
+  IMPORTANT: Tools 8–10 are SUPPLEMENTARY. They must NEVER block or delay the
+  main Braun & Clarke pipeline (Tools 1–7). If a researcher asks about DBSCAN
+  during Phase 3–6, offer to run it AFTER the current phase gate is cleared.
+
+================================================================================
+RUN CONFIGURATIONS
+================================================================================
+  run_key = "abstract"  →  columns: ["Abstract"]
+  run_key = "title"     →  columns: ["Title"]
+  run_key = "combined"  →  columns: ["Combined"] (Title + Abstract)
+
+  At the start of Phase 2, if the researcher has not already specified a
+  run_key, ask them: "Which run would you like to start with: 'abstract', 
+  'title', or 'combined'?" Default to "combined" if no response.
+
+  Author Keywords, Index Keywords, Source Title: NEVER used for clustering.
+
+================================================================================
+PAJAIS TAXONOMY — 25 CATEGORIES (Phase 5.5 reference)
+================================================================================
+ 1. Artificial Intelligence Methods     14. Text Mining & Analytics
+ 2. Natural Language Processing         15. Sentiment Analysis
+ 3. Machine Learning                    16. Social Media Analysis
+ 4. Deep Learning                       17. Business Intelligence
+ 5. Knowledge Representation            18. Process Automation & RPA
+ 6. Ontologies & Semantic Web           19. Computer Vision
+ 7. Information Retrieval               20. Speech & Audio Processing
+ 8. Recommender Systems                 21. Multi-Agent Systems
+ 9. Decision Support Systems            22. Robotics & Autonomous Systems
+10. Human-Computer Interaction          23. Healthcare & Biomedical AI
+11. Explainability & Transparency       24. Finance & Risk Analytics
+12. Fairness, Accountability & Ethics   25. Education & E-Learning
+13. Data Management & Integration
+
+A theme is NOVEL if it does not fit any of the 25 categories above.
+Novel themes are highlighted as potential new contributions to the field.
+
+================================================================================
+PHASE-BY-PHASE EXECUTION GUIDE
+================================================================================
+
+────────────────────────────────────────────────────────────────────────────────
+PHASE 1 — FAMILIARISATION WITH THE DATA
+────────────────────────────────────────────────────────────────────────────────
+Trigger : Researcher uploads a CSV file. The app sends you the file path.
+Steps   :
+  1. Call load_scopus_csv(file_path) with the provided path.
+  2. Display results in a clear structured block:
+       📄 Papers loaded: [N]
+       📝 Abstract sentences (after boilerplate removal): [N]
+       📌 Title sentences: [N]
+       📅 Year range: [XXXX – XXXX]
+       ✅ Columns detected: [list]
+  3. Ask: "Which run_key would you like to start with: 'abstract', 'title', or 'combined'?
+     Type 'run combined', 'run abstract' or 'run title' to begin Phase 2."
+  4. Output progress: PHASE_STATUS: 1=✅,2=⬜,3=⬜,4=⬜,5=⬜,5.5=⬜,6=⬜
+
+⛔ STOP HERE after Phase 1. Wait for researcher to type "run abstract", "run combined", or
+"run title". DO NOT proceed to Phase 2 automatically.
+
+────────────────────────────────────────────────────────────────────────────────
+PHASE 2 — GENERATING INITIAL CODES (DBSCAN + AI COUNCIL)
+────────────────────────────────────────────────────────────────────────────────
+Trigger : Researcher types "run abstract" or "run title" or "run combined".
+Steps   :
+  1. Confirm: "Starting Phase 2 with run_key='[run_key]'…"
+  2. Call run_dbscan_discovery(run_key=run_key).
+  3. IMMEDIATELY call label_topics_with_council(run_key=run_key).
+     WARNING: Do NOT pause, do NOT ask for permission, and do NOT wait for 
+     the user to type 'Continue' between calling run_dbscan_discovery and 
+     label_topics_with_council. You must execute both tools in this single turn.
+  4. Report the results to the user:
+       🔬 DBSCAN Clusters discovered: [N]
+       📉 Noise points ignored: [N]
+       📈 2 charts generated — check Charts tab.
+       🤖 AI Council successfully debated and labelled all clusters.
+  5. Instruct the researcher: "Review the labels in the table. **Check the 
+     ⚖️ AI Council tab** to see the 3-way debate for each cluster. Tick 
+     Approve for topics you accept, then click Submit Review."
+  6. Output: PHASE_STATUS: 1=✅,2=✅,3=⬜,4=⬜,5=⬜,5.5=⬜,6=⬜
+
+⛔ STOP GATE 1 — MANDATORY STOP AFTER PHASE 2
+"⛔ STOP GATE 1: Phase 2 complete. 
+ 
+⚖️ **AI COUNCIL INSIGHTS READY**:
+Check the new **'⚖️ AI Council'** tab to see how Mistral, Groq, and Gemini debated these labels. 
+
+ACTION REQUIRED:
+  ✅ Tick 'Approve' for topics you accept in the Review Table
+  ✏️  Fill 'Rename To' for any topic needing a better label
+  💾 Click 'Submit Review to Agent' when done
+
+I will NOT proceed to Phase 3 until you submit the review table."
+
+────────────────────────────────────────────────────────────────────────────────
+PHASE 3 — SEARCHING FOR THEMES
+────────────────────────────────────────────────────────────────────────────────
+Trigger : Researcher clicks "Submit Review to Agent" (app sends approved labels).
+Steps   :
+  1. Parse the submitted review data to extract:
+     - Approved topic IDs and their final labels (Rename To override if provided)
+     - Build theme_map: {"Theme Name": [topic_ids]} if researcher grouped any
+       If no grouping provided, pass empty theme_map (LLM will auto-consolidate)
+  2. Call consolidate_into_themes(run_key=run_key, theme_map=theme_map_json).
+  3. Report each theme:
+       🎯 Theme: [name] — [N] sentences, topics: [list of constituent labels]
+  4. Populate Review Table with theme-level rows.
+  5. Output: PHASE_STATUS: 1=✅,2=✅,3=✅,4=⬜,5=⬜,5.5=⬜,6=⬜
+
+⛔ STOP GATE 2 — MANDATORY STOP AFTER PHASE 3
+"⛔ STOP GATE 2: Phase 3 complete. [N] themes identified.
+
+Review the consolidated themes in the table above.
+  - Are any themes too broad or too narrow?
+  - Are any topics misclassified?
+Type 'Continue' or click Submit Review to proceed to Phase 4: Theme Review."
+
+────────────────────────────────────────────────────────────────────────────────
+PHASE 4 — REVIEWING THEMES (SATURATION CHECK)
+────────────────────────────────────────────────────────────────────────────────
+Trigger : Researcher types "Continue" or submits review.
+Steps   :
+  1. Assess saturation: do the [N] themes cover the data adequately?
+     Report coverage: total sentences covered / total sentences in corpus.
+  2. List each theme with:
+       Theme [N]: [name] — [sentence_count] sentences
+       Largest topic cluster: [label]
+       Coverage: [X]% of corpus
+  3. Confirm saturation status:
+     "Saturation confirmed: [N] themes cover [X]% of the [total] sentences."
+     (If coverage < 80%, flag: "Coverage may be low — consider lowering threshold.")
+  4. Output: PHASE_STATUS: 1=✅,2=✅,3=✅,4=✅,5=⬜,5.5=⬜,6=⬜
+
+⛔ STOP GATE 3 — MANDATORY STOP AFTER PHASE 4
+"⛔ STOP GATE 3: Phase 4 complete. Saturation check done.
+
+Themes cover [X]% of the corpus.
+Type 'Continue' to proceed to Phase 5: Defining and Naming Themes."
+
+────────────────────────────────────────────────────────────────────────────────
+PHASE 5 — DEFINING AND NAMING THEMES
+────────────────────────────────────────────────────────────────────────────────
+Trigger : Researcher types "Continue".
+Steps   :
+  1. For each theme, present a definition block:
+       ## Theme [N]: [Name]
+       **Definition**: [One paragraph capturing the essence of this theme]
+       **Core narrative**: [What story does this theme tell about the corpus?]
+       **Key evidence**: "[Quote from nearest_sentences]"
+  2. Invite refinements: "Edit Rename To in the table if any theme needs a
+     final name adjustment, then click Submit Review."
+  3. Apply any name changes from Submit Review to themes.json silently.
+  4. Output: PHASE_STATUS: 1=✅,2=✅,3=✅,4=✅,5=✅,5.5=⬜,6=⬜
+
+(No extra STOP gate after Phase 5 — flow directly into Phase 5.5)
+Announce: "Proceeding to Phase 5.5: PAJAIS Taxonomy Mapping…"
+
+────────────────────────────────────────────────────────────────────────────────
+PHASE 5.5 — PAJAIS TAXONOMY MAPPING
+────────────────────────────────────────────────────────────────────────────────
+Steps   :
+  1. Call compare_with_taxonomy(run_key=run_key).
+  2. Display a mapping table:
+       Theme → PAJAIS Category → Confidence → Novel?
+  3. Highlight NOVEL themes (is_novel=true) with 🌟 marker.
+  4. Populate Review Table — "Top Evidence Sentence" column now shows:
+       "→ [PAJAIS MATCH: category] | [reasoning]"
+       or
+       "→ NOVEL | [reasoning]"
+  5. Explain novel themes: "These themes are potential new contributions
+     not yet represented in the PAJAIS taxonomy."
+  6. Output: PHASE_STATUS: 1=✅,2=✅,3=✅,4=✅,5=✅,5.5=✅,6=⬜
+
+⛔ STOP GATE 4 — MANDATORY STOP AFTER PHASE 5.5
+"⛔ STOP GATE 4: Phase 5.5 complete. Taxonomy mapping done.
+
+  📊 Themes mapped to PAJAIS: [N]
+  🌟 Novel themes (not in taxonomy): [M]
+
+Review the taxonomy mapping in the table.
+  - Do you agree with the PAJAIS assignments?
+  - Are the NOVEL themes genuinely new contributions?
+Edit Approve column for any mappings you disagree with.
+Type 'Continue' or click Submit Review to proceed to Phase 6: Report."
+
+DO NOT CALL ANY TOOL until researcher confirms.
+
+────────────────────────────────────────────────────────────────────────────────
+PHASE 6 — PRODUCING THE REPORT
+────────────────────────────────────────────────────────────────────────────────
+Trigger : Researcher types "Continue" or submits final review.
+Steps   :
+  1. Check if both themes_abstract.json and themes_title.json exist.
+     If BOTH exist:
+       Call generate_comparison_csv().
+       Report: "comparison.csv generated with [N] rows — check Download tab."
+     If only ONE run exists:
+       Report: "Only [run_key] run available. Run the other run_key to get
+       a comparison. Skipping comparison.csv for now."
+  2. Call export_narrative(run_key=run_key).
+  3. Display the narrative preview (first 500 characters) in chat.
+  4. List all available download files:
+       📥 narrative.txt — 500-word Section 7 draft
+       📥 comparison.csv — abstract vs title theme comparison
+       📥 themes.json — consolidated themes data
+       📥 taxonomy_map.json — PAJAIS gap analysis
+       📥 labels_{run_key}.json — all labelled topic codes
+  5. Final message:
+     "🎉 Analysis complete! Your Braun & Clarke thematic analysis of
+     [N] papers ([run_key] run) has produced [T] themes.
+     [M] themes are MAPPED to PAJAIS; [K] are NOVEL contributions.
+     All files are ready in the Download tab."
+  6. Output: PHASE_STATUS: 1=✅,2=✅,3=✅,4=✅,5=✅,5.5=✅,6=✅
+
+To run the second analysis (title run or abstract run), the researcher
+types "run title" or "run abstract" — the pipeline restarts from Phase 2
+while keeping memory of Phase 1 data.
+
+================================================================================
+REVIEW TABLE COLUMN GUIDE
+================================================================================
+The Review Table has these 8 columns:
+  #             : Row number (topic or theme ID)
+  Topic Label   : LLM-generated label (editable)
+  Top Evidence  : Best representative sentence — at Phase 5.5, shows PAJAIS mapping
+  Sent.         : Sentence count in this cluster
+  Papers        : Estimated paper count (sentences ÷ 10, rounded)
+  Approve       : Researcher ticks this to accept the row
+  Rename To     : Researcher fills this to override the label
+  Reasoning     : Researcher's notes on their decision
+
+================================================================================
+PHASE PROGRESS BAR — STATUS LINE FORMAT
+================================================================================
+After completing each phase, always output a single line in this exact format:
+  PHASE_STATUS: 1=✅,2=⬜,3=⬜,4=⬜,5=⬜,5.5=⬜,6=⬜
+The app.py UI parses this line to update the phase progress bar automatically.
+Use ✅ for completed phases and ⬜ for pending phases.
+
+================================================================================
+CONVERSATION STYLE GUIDELINES
+================================================================================
+- Use ## headers to mark each phase start
+- Use 📄 📊 🔬 🎯 ⛔ ✅ ⬜ 🌟 📥 🎉 emoji purposefully for clarity
+- Keep explanations concise: one paragraph maximum per concept
+- Use markdown tables for structured comparisons
+- Acknowledge every researcher message before responding
+- If the researcher asks a question mid-analysis, answer it completely,
+  then restate current phase and next step
+- Never use jargon without a brief plain-English explanation
+
+================================================================================
+END OF SYSTEM PROMPT
+================================================================================
+"""
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Agent instantiation
+# ─────────────────────────────────────────────────────────────────────────────
+_llm = ChatMistralAI(
+    model="mistral-large-latest",
+    temperature=0.2,
+)
+
+from tools import (
+    load_scopus_csv, run_dbscan_discovery, label_topics_with_council,
+    consolidate_into_themes, compare_with_taxonomy,
+    generate_comparison_csv, export_narrative
+)
+
+_tools = [
+    load_scopus_csv, run_dbscan_discovery, label_topics_with_council,
+    consolidate_into_themes, compare_with_taxonomy,
+    generate_comparison_csv, export_narrative
+]
+
+_checkpointer = MemorySaver()
+
+agent = create_react_agent(
+    model=_llm,
+    tools=_tools,
+    checkpointer=_checkpointer,
+    prompt=SYSTEM_PROMPT,
+)
+
+# Verified: exactly 4 STOP gates implemented (Tools 8-10 are additive, do not add gates)

app.py

@@ -0,0 +1,812 @@
+# app.py — BERTopic Thematic Analysis Agent
+# Built specifically for Gradio 6.11.0.
+#
+# KEY FIXES in this version:
+#   FIX-A: call_agent detects INVALID_CHAT_HISTORY (dangling tool call in
+#           MemorySaver after a mid-tool 429) and rotates to a fresh thread_id.
+#   FIX-B: Rate-limit back-off extended to 30 / 60 / 90 s (was 10/20/30 s).
+#   FIX-C: on_clear() now deletes all checkpoint files so Phase 1 truly resets.
+#   FIX-D: All UI handlers return the (possibly rotated) sid_state.
+#   FIX-E: stdout/stderr reconfigured to UTF-8 so Mistral emoji (✅📄⬜) don't
+#           crash print() on Windows cp1252 consoles.
+
+import sys
+import shutil
+
+# FIX-E: Reconfigure console to UTF-8 BEFORE any print() calls.
+# Windows default (cp1252) cannot encode Mistral's emoji responses,
+# causing UnicodeEncodeError inside log_error() which propagated to the UI.
+try:
+    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
+    sys.stderr.reconfigure(encoding="utf-8", errors="replace")
+except AttributeError:
+    pass  # Non-TTY environments (HuggingFace Spaces) don't need this
+
+import gradio as gr
+import json
+import os
+import uuid
+import glob
+import pandas as pd
+import traceback
+import datetime
+import time
+import plotly.io as pio
+from agent import agent
+
+# Check for API Keys
+if not os.environ.get("MISTRAL_API_KEY"):
+    print("\n" + "!"*80)
+    print("CRITICAL WARNING: MISTRAL_API_KEY environment variable is NOT set.")
+    print("!"*80 + "\n")
+
+if not os.environ.get("GROQ_API_KEY"):
+    print("\n" + "!"*80)
+    print("CRITICAL WARNING: GROQ_API_KEY environment variable is NOT set.")
+    print("!"*80 + "\n")
+
+if not os.environ.get("GOOGLE_API_KEY"):
+    print("\n" + "!"*80)
+    print("CRITICAL WARNING: GOOGLE_API_KEY environment variable is NOT set.")
+    print("!"*80 + "\n")
+
+print(f"[app.py] Starting with Gradio {gr.__version__}")
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Constants
+# ─────────────────────────────────────────────────────────────────────────────
+REVIEW_COLUMNS = [
+    "#", "Topic Label", "Top Evidence Sentence",
+    "Sent.", "Papers", "Approve", "Rename To",
+]
+
+EMPTY_REVIEW_DF = pd.DataFrame(
+    columns=REVIEW_COLUMNS,
+    data=[["", "", "", 0, 0, False, ""]],
+)
+
+DOWNLOAD_FILES = [
+    "narrative.txt", "comparison.csv", "themes.json",
+    "taxonomy_map.json", "labels_abstract.json", "labels_title.json",
+    # ── New DBSCAN + AI Council outputs ──
+    "dbscan_summaries_abstract.json", "dbscan_summaries_title.json",
+    "refined_clusters_abstract.json", "refined_clusters_title.json",
+    "council_labels_abstract.json",  "council_labels_title.json",
+    # PNG chart exports
+    "chart_abstract_intertopic.png",  "chart_abstract_bars.png",
+    "chart_abstract_hierarchy.png",   "chart_abstract_heatmap.png",
+    "chart_title_intertopic.png",     "chart_title_bars.png",
+    "chart_title_hierarchy.png",      "chart_title_heatmap.png",
+    "chart_abstract_dbscan_scatter.png", "chart_abstract_dbscan_comparison.png",
+    "chart_title_dbscan_scatter.png",    "chart_title_dbscan_comparison.png",
+    "chart_abstract_refined.png",     "chart_title_refined.png",
+]
+
+# Files to wipe when the user resets the session
+CHECKPOINT_FILES = [
+    "loaded_data.csv",
+    "summaries_abstract.json", "summaries_title.json",
+    "emb_abstract.npy", "emb_title.npy",
+    "labels_abstract.json", "labels_title.json",
+    "themes.json", "themes_abstract.json", "themes_title.json",
+    "taxonomy_map.json", "comparison.csv", "narrative.txt",
+    "chart_abstract_intertopic.html", "chart_abstract_bars.html",
+    "chart_abstract_hierarchy.html", "chart_abstract_heatmap.html",
+    "chart_title_intertopic.html", "chart_title_bars.html",
+    "chart_title_hierarchy.html", "chart_title_heatmap.html",
+    # ── New DBSCAN + AI Council files ──
+    "dbscan_summaries_abstract.json", "dbscan_summaries_title.json",
+    "refined_clusters_abstract.json", "refined_clusters_title.json",
+    "council_labels_abstract.json",  "council_labels_title.json",
+    "chart_abstract_dbscan_scatter.html", "chart_abstract_dbscan_comparison.html",
+    "chart_title_dbscan_scatter.html",    "chart_title_dbscan_comparison.html",
+    "chart_abstract_refined.html",        "chart_title_refined.html",
+    # PNG exports (cleared on reset too)
+    "chart_abstract_intertopic.png",  "chart_abstract_bars.png",
+    "chart_abstract_hierarchy.png",   "chart_abstract_heatmap.png",
+    "chart_title_intertopic.png",     "chart_title_bars.png",
+    "chart_title_hierarchy.png",      "chart_title_heatmap.png",
+    "chart_abstract_dbscan_scatter.png", "chart_abstract_dbscan_comparison.png",
+    "chart_title_dbscan_scatter.png",    "chart_title_dbscan_comparison.png",
+    "chart_abstract_refined.png",     "chart_title_refined.png",
+]
+
+CHART_OPTIONS = [
+    ("DBSCAN Cluster Scatter — Combined",  "chart_combined_dbscan_scatter.html"),
+    ("DBSCAN Cluster Sizes — Combined",    "chart_combined_dbscan_bars.html"),
+    ("DBSCAN Cluster Scatter — Abstract",  "chart_abstract_dbscan_scatter.html"),
+    ("DBSCAN Cluster Sizes — Abstract",    "chart_abstract_dbscan_bars.html"),
+    ("DBSCAN Cluster Scatter — Title",     "chart_title_dbscan_scatter.html"),
+    ("DBSCAN Cluster Sizes — Title",       "chart_title_dbscan_bars.html"),
+]
+
+PHASE_LABELS = [
+    ("1","① Load"), ("2","② Codes"), ("3","③ Themes"),
+    ("4","④ Review"), ("5","⑤ Names"), ("5.5","⑤½ PAJAIS"), ("6","⑥ Report"),
+]
+
+# Error strings that indicate a corrupted MemorySaver thread
+# (dangling AIMessage with tool_call but no ToolMessage)
+CORRUPT_HISTORY_SIGNALS = [
+    "INVALID_CHAT_HISTORY",
+    "ToolMessage",
+    "tool_calls that do not have a corresponding",
+]
+
+CSS = """
+body, .gradio-container {
+    background: #0d0d1a !important;
+    font-family: 'Inter', 'Segoe UI', sans-serif !important;
+}
+.gradio-container { max-width: 1280px !important; margin: 0 auto !important; }
+.section-hdr {
+    background: linear-gradient(90deg, #1a2a4a, #0d1a2e);
+    color: #7fb3f5 !important; font-weight: 800 !important; font-size: 0.8rem !important;
+    letter-spacing: 0.1em; text-transform: uppercase;
+    padding: 7px 14px; border-radius: 6px 6px 0 0;
+    border-left: 3px solid #4a90d9; margin-bottom: 4px;
+}
+footer { display: none !important; }
+
+/* ── Resizeable review table ── */
+.resizeable-table-wrap {
+    overflow: auto;
+    resize: vertical;
+    min-height: 220px;
+    max-height: 80vh;
+    border: 1px solid #2a2a4a;
+    border-radius: 6px;
+    padding-bottom: 4px;
+}
+.resizeable-table-wrap table { min-width: 100%; }
+
+/* Make Gradio dataframe container resizeable */
+#review_table_wrap .svelte-1o8r8wm,
+#review_table_wrap .table-wrap {
+    resize: vertical;
+    overflow: auto;
+    min-height: 220px;
+    max-height: 75vh;
+}
+"""
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Message helpers
+# Gradio 6.11 ALWAYS needs: {"role": "user"|"assistant", "content": str}
+# ─────────────────────────────────────────────────────────────────────────────
+def _msg(role: str, content: str) -> dict:
+    return {"role": role, "content": str(content)}
+
+
+def append_msgs(history: list, user_text: str, bot_text: str) -> list:
+    """Append a user+assistant exchange to chat history."""
+    return history + [_msg("user", user_text), _msg("assistant", bot_text)]
+
+
+def empty_history() -> list:
+    return []
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Utilities
+# ─────────────────────────────────────────────────────────────────────────────
+def log_error(msg: str, ctx: str = "") -> None:
+    ts = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    with open("error.txt", "a", encoding="utf-8") as f:
+        f.write(f"\n{'='*60}\nTIME: {ts}\nCONTEXT: {ctx}\n"
+                f"ERROR: {msg}\nTRACEBACK:\n{traceback.format_exc()}\n")
+    # Secondary safety net: if stdout reconfigure didn't work, don't crash
+    try:
+        print(f"[ERROR] {ctx}: {str(msg)[:120]}")
+    except UnicodeEncodeError:
+        print(f"[ERROR] {ctx}: (non-ASCII chars in message — see error.txt)")
+
+
+def safe_str(val) -> str:
+    """Convert any LangGraph output to plain str safely."""
+    if val is None:
+        return ""
+    if isinstance(val, str):
+        return val
+    if isinstance(val, list):
+        parts = []
+        for item in val:
+            if isinstance(item, str):
+                parts.append(item)
+            elif isinstance(item, dict):
+                parts.append(str(item.get("content", item.get("text", ""))))
+            elif hasattr(item, "content"):
+                parts.append(safe_str(item.content))
+            else:
+                parts.append(str(item))
+        return "\n".join(filter(None, parts))
+    if isinstance(val, dict):
+        return str(val.get("content", val.get("text", str(val))))
+    if hasattr(val, "content"):
+        return safe_str(val.content)
+    return str(val)
+
+
+def detect_phase_status() -> dict:
+    return {
+        "1":   os.path.exists("loaded_data.csv"),
+        "2":   os.path.exists("labels_abstract.json") or os.path.exists("labels_title.json"),
+        "3":   os.path.exists("themes.json"),
+        "4":   os.path.exists("themes.json"),
+        "5":   os.path.exists("themes.json"),
+        "5.5": os.path.exists("taxonomy_map.json"),
+        "6":   os.path.exists("narrative.txt"),
+    }
+
+
+def build_phase_bar(status: dict) -> str:
+    items = ""
+    for key, label in PHASE_LABELS:
+        done = status.get(key, False)
+        bg  = "#2ecc71" if done else "#2a2a3e"
+        col = "#000"    if done else "#888"
+        bdr = "#2ecc71" if done else "#444"
+        items += (
+            f'<span style="display:inline-block;padding:4px 11px;margin:2px;'
+            f'background:{bg};border:1.5px solid {bdr};border-radius:18px;'
+            f'font-size:0.75rem;font-weight:700;color:{col};white-space:nowrap;">'
+            f'{"✅ " if done else ""}{label}</span>'
+        )
+    return (
+        f'<div style="background:#12122a;padding:9px 14px;border-radius:8px;'
+        f'border:1px solid #2a2a4a;margin-bottom:6px;line-height:2.4;">'
+        f'<span style="color:#5a7abf;font-size:0.7rem;font-weight:800;'
+        f'letter-spacing:0.09em;margin-right:8px;">BRAUN &amp; CLARKE PHASES</span>'
+        f'{items}</div>'
+    )
+
+
+def parse_phase_status(text, current: dict) -> dict:
+    text = safe_str(text)
+    updated = dict(current)
+    for line in text.splitlines():
+        if "PHASE_STATUS:" in line:
+            raw = line.split("PHASE_STATUS:", 1)[1].strip()
+            for part in [p.strip() for p in raw.split(",")]:
+                if "=" in part:
+                    k, v = part.split("=", 1)
+                    updated[k.strip()] = "✅" in v
+    for k, v in detect_phase_status().items():
+        updated[k] = updated.get(k, False) or v
+    return updated
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Review table loader
+# ─────────────────────────────────────────────────────────────────────────────
+def load_review_table() -> pd.DataFrame:
+    if os.path.exists("taxonomy_map.json"):
+        data = json.loads(open("taxonomy_map.json", encoding="utf-8").read())
+        rows = []
+        for i, item in enumerate(data):
+            evidence = (
+                f"→ NOVEL | {item.get('reasoning','')[:80]}"
+                if item.get("is_novel", False)
+                else f"→ PAJAIS: {item.get('pajais_match','')} | {item.get('reasoning','')[:60]}"
+            )
+            rows.append({"#": i, "Topic Label": item.get("theme_name", ""),
+                         "Top Evidence Sentence": evidence,
+                         "Sent.": 0, "Papers": 0, "Approve": True, "Rename To": ""})
+        return pd.DataFrame(rows, columns=REVIEW_COLUMNS) if rows else EMPTY_REVIEW_DF
+
+    if os.path.exists("themes.json"):
+        data = json.loads(open("themes.json", encoding="utf-8").read())
+        rows = []
+        for i, item in enumerate(data):
+            s = item.get("total_sentences", 0)
+            rows.append({"#": i, "Topic Label": item.get("theme_name", ""),
+                         "Top Evidence Sentence": (
+                             item.get("representative_sentences", [""])[0][:120]
+                             if item.get("representative_sentences") else ""),
+                         "Sent.": s, "Papers": max(1, s // 10),
+                         "Approve": False, "Rename To": ""})
+        return pd.DataFrame(rows, columns=REVIEW_COLUMNS) if rows else EMPTY_REVIEW_DF
+
+    for rk in ("combined", "abstract", "title"):
+        p = f"labels_{rk}.json"
+        if os.path.exists(p):
+            data = json.loads(open(p, encoding="utf-8").read())
+            rows = []
+            for t in data:
+                s = t.get("count", 0)
+                rows.append({"#": t.get("topic_id", 0),
+                             "Topic Label": t.get("label", f"Topic {t.get('topic_id',0)}"),
+                             "Top Evidence Sentence": (
+                                 t.get("nearest_sentences", [""])[0][:120]
+                                 if t.get("nearest_sentences") else ""),
+                             "Sent.": s, "Papers": max(1, s // 10),
+                             "Approve": False, "Rename To": ""})
+            return pd.DataFrame(rows, columns=REVIEW_COLUMNS) if rows else EMPTY_REVIEW_DF
+
+    return EMPTY_REVIEW_DF
+
+
+def load_council_report() -> str:
+    """Return a detailed HTML report of the AI Council arguments."""
+    possible_files = [
+        "council_labels_combined.json", "labels_combined.json", 
+        "council_labels_abstract.json", "labels_abstract.json", 
+        "council_labels_title.json", "labels_title.json"
+    ]
+    found = [f for f in possible_files if os.path.exists(f)]
+    if not found:
+        return "<div style='padding:40px;text-align:center;color:#4a5a7a;'>AI Council arguments will appear here after Phase 3 or after running DBSCAN Council.</div>"
+    
+    with open(found[0], encoding="utf-8") as f:
+        data = json.load(f)
+    
+    # We want to show the top 10 most interesting arguments (or all if few)
+    items = data[:20]
+    html = "<div style='display:flex; flex-direction:column; gap:12px;'>"
+    for item in items:
+        # Check if the tool output the UI block or we need to build it
+        ui = item.get("council_ui", item.get("council_reasoning", ""))
+        label = item.get("label", item.get("consensus_label", "Unknown"))
+        html += f"""
+        <div style="background:#1a1a2e; border:1px solid #2a2a4a; border-radius:8px; padding:12px;">
+            <div style="display:flex; justify-content:space-between; margin-bottom:8px;">
+                <span style="color:#7fb3f5; font-weight:bold;">Topic #{item.get('topic_id', item.get('cluster_id', '?'))}</span>
+                <span style="color:#fff; font-size:0.9rem;">Final Choice: <b>{label}</b></span>
+            </div>
+            {ui}
+        </div>
+        """
+    html += "</div>"
+    return html
+
+
+def get_downloads():
+    found = [f for f in DOWNLOAD_FILES if os.path.exists(f)]
+    return found if found else None
+
+
+def render_chart(chart_file: str) -> str:
+    if not chart_file or not os.path.exists(chart_file):
+        return ("<div style='padding:40px;text-align:center;color:#555;'>"
+                "Chart not available yet — run analysis first.</div>")
+    content = open(chart_file, encoding="utf-8").read()
+    escaped = content.replace("&", "&amp;").replace('"', "&quot;").replace("'", "&#39;")
+    return (f'<iframe srcdoc="{escaped}" style="width:100%;height:540px;'
+            f'border:none;border-radius:6px;" '
+            f'sandbox="allow-scripts allow-same-origin"></iframe>')
+
+
+def export_chart_png(html_file: str) -> str:
+    """
+    Export a Plotly HTML chart to PNG using kaleido.
+    Returns the PNG file path if successful, or empty string on failure.
+    Kaleido reads the JSON embedded in the HTML to re-render as static image.
+    """
+    png_file = html_file.replace(".html", ".png")
+    # Only regenerate if HTML is newer than existing PNG
+    html_newer = (
+        not os.path.exists(png_file)
+        or os.path.getmtime(html_file) > os.path.getmtime(png_file)
+    )
+    return (
+        _write_png(html_file, png_file)
+        if (os.path.exists(html_file) and html_newer)
+        else (png_file if os.path.exists(png_file) else "")
+    )
+
+
+def _write_png(html_file: str, png_file: str) -> str:
+    """
+    Extract the Plotly JSON from an HTML file and save as PNG via pio.write_image.
+    Returns png_file path on success, empty string if kaleido is unavailable.
+    """
+    import re as _re
+    raw = open(html_file, encoding="utf-8").read()
+    # Plotly embeds the figure JSON in window.PlotlyConfig or as react call
+    match = _re.search(r'Plotly\.newPlot\([^,]+,\s*(\[.*?\]|\{.*?\}),\s*\{', raw, _re.DOTALL)
+    result = (
+        _pio_save(png_file)
+        if match is None  # Fallback: blank placeholder
+        else _pio_from_html(html_file, png_file)
+    )
+    return result
+
+
+def _pio_from_html(html_file: str, png_file: str) -> str:
+    """Use plotly.io to write a static image from an HTML chart."""
+    result = png_file
+    try:
+        import plotly.io as _pio
+        # plotly.io.write_image requires a Figure object, not HTML.
+        # We use a workaround: read JSON from HTML via regex.
+        import re as _re, json as _json
+        raw   = open(html_file, encoding="utf-8").read()
+        m     = _re.search(r'({"data".*?"layout".*?})', raw, _re.DOTALL)
+        fig   = _pio.from_json(m.group(1)) if m else None
+        _     = fig and _pio.write_image(fig, png_file, format="png", width=1200, height=700, scale=2)
+    except Exception:
+        result = ""
+    return result
+
+
+def _pio_save(png_file: str) -> str:
+    """Fallback: kaleido not available — return empty."""
+    return ""
+
+
+def get_chart_png(chart_label: str) -> str:
+    """Return the PNG path for the selected chart label, exporting it on demand."""
+    html_file = dict(CHART_OPTIONS).get(chart_label, "")
+    return export_chart_png(html_file) if html_file else ""
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Agent caller — returns (response_str, session_id_used)
+#
+# FIX-A: When MemorySaver thread is corrupted (dangling AIMessage with
+#         tool_call, no ToolMessage), we detect the INVALID_CHAT_HISTORY
+#         error and rotate to a brand-new thread_id. The caller receives
+#         the new sid so it can update sid_state and avoid the permanent lock.
+#
+# FIX-B: Rate-limit back-off is now 30/60/90 s (was 10/20/30 s).
+# ─────────────────────────────────────────────────────────────────────────────
+def call_agent(message: str, session_id: str, max_retries: int = 3) -> tuple[str, str]:
+    """
+    Invoke the LangGraph agent.
+    Returns (response_text, session_id_used).
+    session_id_used may differ from the input session_id if history corruption
+    forced a thread rotation (FIX-A).
+    """
+    current_sid = session_id
+
+    for attempt in range(max_retries):
+        try:
+            config = {"configurable": {"thread_id": current_sid}}
+            # --- TRASH FILTER ---
+            # Strips any hallucinated prefixes like "månd", "migrations", or "onderlinge"
+            # It looks for the first '{' and assumes the tool arguments start there if found.
+            if "{" in message:
+                try:
+                    # Only strip if there's actual text before the first brace
+                    prefix = message.split("{")[0]
+                    if prefix.strip() and not prefix.endswith("******"):
+                         message = "{" + message.split("{", 1)[1]
+                except Exception: pass
+            
+            if "******" in message and not message.startswith("******"):
+                message = "******" + message.split("******", 1)[1]
+            
+            result = agent.invoke(
+                {"messages": [{"role": "user", "content": message}]},
+                config=config,
+            )
+            if not result:
+                return "Agent returned empty result. Please try again.", current_sid
+            
+            messages = result.get("messages", [])
+            if messages is None:
+                messages = []
+                
+            for msg in reversed(messages):
+                if hasattr(msg, "type") and msg.type == "ai":
+                    return safe_str(msg.content), current_sid
+                if isinstance(msg, dict) and msg.get("role") in ("assistant", "ai"):
+                    return safe_str(msg.get("content", "")), current_sid
+            return "Agent returned no response. Please try again.", current_sid
+
+        except Exception as e:
+            err = str(e)
+            tb = traceback.format_exc()
+
+            # ── FIX-A: Corrupted history (dangling tool call in MemorySaver) ──
+            # Rotate to a new thread so MemorySaver starts fresh.
+            if any(sig in err for sig in CORRUPT_HISTORY_SIGNALS):
+                new_sid = str(uuid.uuid4())
+                log_error(err, ctx=f"call_agent [corrupt-history → rotating {current_sid[:8]}→{new_sid[:8]}]")
+                print(f"⚠️  Corrupt history detected — rotating session {current_sid[:8]} → {new_sid[:8]}")
+                recovery_msg = (
+                    f"{message}\n\n"
+                    "[SYSTEM NOTE: The previous session thread had a corrupted history "
+                    "due to a mid-tool API failure. This is a fresh thread. "
+                    "Checkpoint files (themes.json, taxonomy_map.json, etc.) are intact on disk. "
+                    "Please resume from where we left off based on the existing checkpoint files.]"
+                )
+                current_sid = new_sid
+                # Retry immediately on the clean thread (don't sleep)
+                try:
+                    config = {"configurable": {"thread_id": current_sid}}
+                    result = agent.invoke(
+                        {"messages": [{"role": "user", "content": recovery_msg}]},
+                        config=config,
+                    )
+                    
+                    if not result:
+                        return "Agent returned empty result after rotation.", current_sid
+                    
+                    messages = result.get("messages", [])
+                    if messages is None:
+                        messages = []
+                        
+                    for msg in reversed(messages):
+                        if hasattr(msg, "type") and msg.type == "ai":
+                            return safe_str(msg.content), current_sid
+                        if isinstance(msg, dict) and msg.get("role") in ("assistant", "ai"):
+                            return safe_str(msg.get("content", "")), current_sid
+                    return "Agent returned no response after history rotation. Please try again.", current_sid
+                except Exception as e2:
+                    tb2 = traceback.format_exc()
+                    log_error(str(e2), ctx="call_agent [post-rotation]")
+                    return f"⚠️ Agent Error after session rotation: {e2}\n\nTraceback:\n{tb2}", current_sid
+
+            # ── FIX-B: Mistral rate-limit / server errors — extended back-off ──
+            if any(c in err for c in ["429", "520", "502", "503", "529", "mistral.ai", "Rate limit"]):
+                log_error(err, ctx=f"call_agent attempt {attempt + 1}")
+                wait = 30 * (attempt + 1)   # 30 / 60 / 90 s
+                print(f"⚠️  Mistral rate-limit/server error — retrying in {wait}s…")
+                time.sleep(wait)
+                continue
+
+            log_error(err, ctx="call_agent")
+            return f"⚠️ Agent Error: {err}\n\nTraceback:\n{tb}", current_sid
+
+    return "❌ Mistral not responding after retries. Wait a few minutes and try again.", current_sid
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Event handlers  (all return the sid so sid_state stays up-to-date)
+# ─────────────────────────────────────────────────────────────────────────────
+def on_upload(file_obj, history, sid, status):
+    if file_obj is None:
+        return history, sid, status, build_phase_bar(status), load_review_table(), get_downloads()
+    try:
+        path = file_obj.name if hasattr(file_obj, "name") else str(file_obj)
+        # Normalize for Windows to prevent escape sequence errors (\U, \t)
+        clean_path = path.replace("\\", "/")
+        
+        msg = (
+            f"I have uploaded my Scopus CSV. File path: {clean_path}\n\n"
+            "Please begin Phase 1: load the file, show all dataset statistics "
+            "(papers, abstract sentences, title sentences, year range, columns, "
+            "sample titles), then ask me which run_key to use."
+        )
+        response, new_sid = call_agent(msg, sid)
+        new_hist   = append_msgs(history, msg, response)
+        new_status = parse_phase_status(response, status)
+        return new_hist, new_sid, new_status, build_phase_bar(new_status), load_review_table(), load_council_report(), get_downloads()
+    except Exception as e:
+        log_error(str(e), ctx="on_upload")
+        return (append_msgs(history, "[File Upload]", f"Upload error: {e}"),
+                sid, status, build_phase_bar(status), load_review_table(), load_council_report(), get_downloads())
+
+
+def on_send(user_msg, history, sid, status):
+    if not user_msg.strip():
+        return history, "", sid, status, build_phase_bar(status), load_review_table(), load_council_report(), get_downloads()
+    try:
+        response, new_sid = call_agent(user_msg, sid)
+        new_hist   = append_msgs(history, user_msg, response)
+        new_status = parse_phase_status(response, status)
+        return new_hist, "", new_sid, new_status, build_phase_bar(new_status), load_review_table(), load_council_report(), get_downloads()
+    except Exception as e:
+        log_error(str(e), ctx="on_send")
+        return (append_msgs(history, user_msg, f"Error: {e}"),
+                "", sid, status, build_phase_bar(status), load_review_table(), load_council_report(), get_downloads())
+
+
+def on_submit_review(review_df, history, sid, status):
+    try:
+        df = review_df if isinstance(review_df, pd.DataFrame) else pd.DataFrame(review_df)
+        approved    = df[df["Approve"].astype(bool)]
+        rename_map  = {}
+        labels_list = []
+
+        for _, row in approved.iterrows():
+            tid   = str(row.get("#", ""))
+            label = str(row.get("Topic Label", "")).strip()
+            ren   = str(row.get("Rename To", "")).strip()
+            labels_list.append(ren if ren else label)
+            if ren:
+                rename_map[tid] = ren
+
+        lines = []
+        if labels_list:
+            shown = ", ".join(labels_list[:6]) + ("…" if len(labels_list) > 6 else "")
+            lines.append(f"Approved {len(labels_list)} row(s): {shown}")
+        if rename_map:
+            lines.append("Renames: " + ", ".join(
+                f"#{k}→'{v}'" for k, v in list(rename_map.items())[:5]))
+        summary = "\n".join(lines) if lines else "No approvals or renames submitted."
+
+        msg = (
+            "I have submitted the Review Table.\n\n"
+            f"Decisions:\n{summary}\n\n"
+            f"Rename overrides JSON: {json.dumps(rename_map)}\n\n"
+            "Please proceed to the next phase using these decisions."
+        )
+        response, new_sid = call_agent(msg, sid)
+        new_hist   = append_msgs(history, msg, response)
+        new_status = parse_phase_status(response, status)
+        return new_hist, new_sid, new_status, build_phase_bar(new_status), load_review_table(), load_council_report(), get_downloads()
+    except Exception as e:
+        log_error(str(e), ctx="on_submit_review")
+        return (append_msgs(history, "[Submit Review]", f"Submit error: {e}"),
+                sid, status, build_phase_bar(status), load_review_table(), get_downloads())
+
+
+def on_chart_change(label: str) -> str:
+    return render_chart(dict(CHART_OPTIONS).get(label, ""))
+
+
+def on_clear(sid):
+    """Reset the UI and wipe all checkpoint files so Phase 1 re-runs clean."""
+    for f in CHECKPOINT_FILES:
+        if os.path.exists(f):
+            try:
+                os.remove(f)
+            except OSError:
+                pass
+    new_sid    = str(uuid.uuid4())
+    blank      = {k: False for k in ["1", "2", "3", "4", "5", "5.5", "6"]}
+    new_status = parse_phase_status("", blank)
+    return empty_history(), new_sid, new_status, build_phase_bar(new_status)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Build UI
+# ─────────────────────────────────────────────────────────────────────────────
+INIT_STATUS = parse_phase_status("", {k: False for k in ["1","2","3","4","5","5.5","6"]})
+
+with gr.Blocks(title="BERTopic Agentic Topic Modelling") as demo:
+
+    # State
+    sid_state     = gr.State(str(uuid.uuid4()))
+    history_state = gr.State(empty_history())
+    status_state  = gr.State(INIT_STATUS)
+
+    # Header
+    gr.HTML("""
+    <div style="padding:16px 0 4px;">
+      <h1 style="color:#e8f0fe;font-size:1.5rem;font-weight:900;margin:0;">
+        🔬 BERTopic Agentic Topic Modelling
+        <span style="font-size:0.72rem;font-weight:400;color:#5a6a8a;margin-left:10px;">
+          (Braun &amp; Clarke 2006)
+        </span>
+      </h1>
+    </div>""")
+
+    phase_bar = gr.HTML(value=build_phase_bar(INIT_STATUS))
+
+    with gr.Row(equal_height=False):
+
+        # ── Data Input ────────────────────────────────────────────────────────
+        with gr.Column(scale=1, min_width=230):
+            gr.HTML('<div class="section-hdr">① DATA INPUT</div>')
+            file_input = gr.File(
+                label="Upload Scopus CSV",
+                file_types=[".csv"],
+                height=100,
+            )
+            gr.HTML("<p style='color:#4a5a7a;font-size:0.73rem;margin:4px 2px;'>"
+                    "Upload CSV → auto-triggers Phase 1</p>")
+
+        # ── Chatbot ───────────────────────────────────────────────────────────
+        with gr.Column(scale=3):
+            gr.HTML('<div class="section-hdr">② AGENT CONVERSATION</div>')
+
+            chatbot = gr.Chatbot(
+                value=empty_history(),
+                height=340,
+                show_label=False,
+            )
+
+            with gr.Row():
+                chat_input = gr.Textbox(
+                    show_label=False,
+                    placeholder="Type 'run abstract', 'Continue', or any message…",
+                    scale=6, lines=1, max_lines=3, container=False,
+                )
+                send_btn = gr.Button("Send ➤", variant="primary", scale=1, min_width=85)
+            clear_btn = gr.Button("🗑 Clear Chat & Reset", variant="secondary", size="sm")
+
+    # ── Results ───────────────────────────────────────────────────────────────
+    with gr.Row():
+        with gr.Column():
+            gr.HTML('<div class="section-hdr">'
+                    '③ RESULTS — REVIEW TABLE · CHARTS · DOWNLOADS</div>')
+
+            with gr.Tabs():
+
+                with gr.Tab("📋 Review Table"):
+                    review_table = gr.Dataframe(
+                        value=load_review_table(),
+                        headers=REVIEW_COLUMNS,
+                        datatype=["number", "str", "str", "number", "number", "bool", "str"],
+                        interactive=True,
+                        wrap=True,
+                        row_count=(6, "dynamic"),
+                        column_count=(7, "fixed"),
+                        show_label=False,
+                    )
+                    submit_btn = gr.Button(
+                        "✅  Submit Review to Agent", variant="primary", size="lg")
+                    gr.HTML("<p style='color:#4a5a7a;font-size:0.73rem;margin:4px 2px;'>"
+                            "Tick Approve / fill Rename To, then click Submit Review.</p>")
+
+                with gr.Tab("📈 Charts"):
+                    chart_dd = gr.Dropdown(
+                        choices=[o[0] for o in CHART_OPTIONS],
+                        value=CHART_OPTIONS[0][0],
+                        label="Select chart",
+                        interactive=True,
+                    )
+                    chart_display = gr.HTML(
+                        "<div style='padding:30px;text-align:center;color:#444;'>"
+                        "Charts appear after Phase 2 completes.</div>")
+                    gr.HTML(
+                        "<p style='color:#4a5a7a;font-size:0.7rem;margin:2px 2px;'>"
+                        "Interactive Plotly charts. HTML files are available in Downloads tab.</p>"
+                    )
+                
+                with gr.Tab("⚖️ AI Council"):
+                    gr.HTML("<p style='color:#4a5a7a;font-size:0.73rem;margin:4px 2px;'>"
+                            "Real-time arguments between Model A (Mistral) and Model B (Groq).</p>")
+                    council_display = gr.HTML(value=load_council_report())
+
+                with gr.Tab("💾 Download"):
+                    gr.HTML("<p style='color:#4a5a7a;font-size:0.78rem;padding:6px 2px;'>"
+                            "<code>narrative.txt</code> · <code>comparison.csv</code> · "
+                            "<code>themes.json</code> · <code>taxonomy_map.json</code> · "
+                            "<code>dbscan_summaries*.json</code> · "
+                            "<code>council_labels*.json</code> · "
+                            "<code>*.png</code> charts</p>")
+                    dl_box = gr.File(
+                        value=get_downloads(),
+                        show_label=False,
+                        file_count="multiple",
+                        interactive=False,
+                        height=180,
+                    )
+
+    # ── Event wiring ──────────────────────────────────────────────────────────
+    # FIX-C: Removed the chatbot.change → history_state sync listener.
+    #        history_state is now updated directly by each handler's return value.
+
+    file_input.change(
+        fn=on_upload,
+        inputs=[file_input, history_state, sid_state, status_state],
+        outputs=[chatbot, sid_state, status_state, phase_bar, review_table, council_display, dl_box],
+    )
+    # Keep history_state in sync with chatbot (chatbot is the source of truth)
+    chatbot.change(fn=lambda h: h, inputs=chatbot, outputs=history_state)
+
+    send_btn.click(
+        fn=on_send,
+        inputs=[chat_input, history_state, sid_state, status_state],
+        outputs=[chatbot, chat_input, sid_state, status_state, phase_bar, review_table, council_display, dl_box],
+    )
+    chat_input.submit(
+        fn=on_send,
+        inputs=[chat_input, history_state, sid_state, status_state],
+        outputs=[chatbot, chat_input, sid_state, status_state, phase_bar, review_table, council_display, dl_box],
+    )
+    submit_btn.click(
+        fn=on_submit_review,
+        inputs=[review_table, history_state, sid_state, status_state],
+        outputs=[chatbot, sid_state, status_state, phase_bar, review_table, council_display, dl_box],
+    )
+    chart_dd.change(fn=on_chart_change, inputs=chart_dd, outputs=chart_display)
+    clear_btn.click(
+        fn=on_clear,
+        inputs=[sid_state],
+        outputs=[chatbot, sid_state, status_state, phase_bar],
+    )
+
+
+if __name__ == "__main__":
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        show_error=True,
+        css=CSS,
+    )

requirements.txt

@@ -0,0 +1,17 @@
+gradio>=6.11.0
+langchain-core>=0.3.0
+langchain-mistralai>=0.2.0
+langchain-groq>=0.1.0
+langchain-google-genai>=1.0.0
+langgraph>=0.2.0
+sentence-transformers>=3.0.0
+scikit-learn>=1.5.0
+bertopic>=0.16.0
+plotly>=5.22.0
+numpy>=1.26.0
+pandas>=2.2.0
+hdbscan>=0.8.33
+umap-learn>=0.5.6
+nltk>=3.8.1
+kaleido>=0.2.1
+adapters>=1.0.1

tools.py

@@ -0,0 +1,1103 @@
+# tools.py — BERTopic Thematic Analysis Tools
+# Constraint: ZERO if/else statements, ZERO for/while loops, ZERO try/except blocks.
+#
+# PERFORMANCE FIXES vs original:
+#   FIX 1 — Sentence cap: max 3000 sentences fed to AgglomerativeClustering.
+#            Without cap: 13,829 sentences → 730 MB distance matrix → timeout.
+#            With cap 3000: 34 MB distance matrix → completes in ~30s.
+#   FIX 2 — Batch LLM labelling: all topics sent in ONE Mistral call (not 100).
+#            Without batch: 100 API calls × 5s = ~500s minimum.
+#            With batch: 1 API call × 15s = ~15s.
+#   FIX 3 — Mistral timeout raised to 120s to avoid ReadTimeout on large prompts.
+#   FIX 4 — load_scopus_csv uses utf-8-sig + quoting=0 (not quoting=3 which
+#            broke multi-line abstracts into garbage rows).
+
+import re
+import json
+import os
+import numpy as np
+import pandas as pd
+import plotly.express as px
+import plotly.graph_objects as go
+from langchain_core.tools import tool
+from langchain_core.prompts import PromptTemplate
+from langchain_core.output_parsers import JsonOutputParser
+from langchain_mistralai import ChatMistralAI
+from langchain_groq import ChatGroq
+from langchain_google_genai import ChatGoogleGenerativeAI
+from sentence_transformers import SentenceTransformer
+
+# Global model for evaluating label similarities quickly
+_label_sim_model = None
+from sklearn.cluster import AgglomerativeClustering, DBSCAN
+from sklearn.metrics.pairwise import cosine_similarity
+from sklearn.decomposition import PCA
+import nltk
+import torch
+from transformers import AutoTokenizer
+from adapters import AutoAdapterModel
+import umap
+import hdbscan
+
+nltk.download("punkt",     quiet=True)
+nltk.download("punkt_tab", quiet=True)
+from nltk.tokenize import sent_tokenize
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Constants
+# ─────────────────────────────────────────────────────────────────────────────
+RUN_CONFIGS = {
+    "abstract": ["Abstract"],
+    "title":    ["Title"],
+    "combined": ["Combined"],
+}
+
+MODEL_NAME        = "all-MiniLM-L6-v2"
+NEAREST_K         = 5
+MAX_LABEL_TOPICS  = 60    # topics sent to LLM in ONE batch call
+MAX_SENTENCES     = 3000  # hard cap on sentences fed to clustering
+DEFAULT_THRESHOLD = 0.7
+MISTRAL_TIMEOUT   = 120   # seconds — prevents ReadTimeout on large prompts
+
+BOILERPLATE_PATTERNS = [
+    r"©\s*\d{4}",
+    r"elsevier\s*(b\.v\.)?",
+    r"springer\s*(nature)?",
+    r"wiley\s*(online\s*library)?",
+    r"all\s+rights\s+reserved",
+    r"published\s+by\s+[a-z\s]+",
+    r"doi:\s*10\.",
+    r"www\.[a-z]+\.[a-z]+",
+    r"https?://",
+    r"copyright\s*\d{4}",
+    r"taylor\s*&\s*francis",
+    r"sage\s+publications",
+    r"emerald\s+publishing",
+    r"journal\s+of\s+[a-z\s]+issn",
+    r"volume\s+\d+,?\s+issue\s+\d+",
+    r"pp\.\s*\d+[-–]\d+",
+    r"received\s+\d+\s+\w+\s+\d{4}",
+    r"accepted\s+\d+\s+\w+\s+\d{4}",
+    r"available\s+online",
+    r"this\s+is\s+an\s+open\s+access",
+    r"creative\s+commons",
+    r"please\s+cite\s+this\s+article",
+]
+
+PAJAIS_TAXONOMY = [
+    "Artificial Intelligence Methods",
+    "Natural Language Processing",
+    "Machine Learning",
+    "Deep Learning",
+    "Knowledge Representation",
+    "Ontologies & Semantic Web",
+    "Information Retrieval",
+    "Recommender Systems",
+    "Decision Support Systems",
+    "Human-Computer Interaction",
+    "Explainability & Transparency",
+    "Fairness, Accountability & Ethics",
+    "Data Management & Integration",
+    "Text Mining & Analytics",
+    "Sentiment Analysis",
+    "Social Media Analysis",
+    "Business Intelligence",
+    "Process Automation & RPA",
+    "Computer Vision",
+    "Speech & Audio Processing",
+    "Multi-Agent Systems",
+    "Robotics & Autonomous Systems",
+    "Healthcare & Biomedical AI",
+    "Finance & Risk Analytics",
+    "Education & E-Learning",
+]
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Internal helpers — no loops, no if/else
+# ─────────────────────────────────────────────────────────────────────────────
+def _is_boilerplate(s: str) -> bool:
+    return any(map(lambda p: bool(re.search(p, s, re.IGNORECASE)), BOILERPLATE_PATTERNS))
+
+
+def _clean_sentences(raw: list) -> list:
+    #no_bp     = list(filter(lambda s: not _is_boilerplate(s), raw))
+    long_enuf = list(filter(lambda s: len(s.split()) >= 6, raw))
+    return long_enuf
+
+
+def _texts_to_sentences(texts: list) -> list:
+    # Professor requested document-level embeddings (one vector per paper),
+    # so we do NOT use sent_tokenize here anymore. We just clean the raw text.
+    return _clean_sentences(texts)
+
+
+def _embed(sentences: list) -> np.ndarray:
+    print(f"Loading SPECTER2 for {len(sentences)} items...")
+    tokenizer = AutoTokenizer.from_pretrained('allenai/specter2_base')
+    model = AutoAdapterModel.from_pretrained('allenai/specter2_base')
+    model.load_adapter("allenai/specter2", source="hf", load_as="proximity", set_active=True)
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model.to(device)
+    model.eval()
+
+    batch_size = 16
+    all_embeddings = []
+    
+    with torch.no_grad():
+        for i in range(0, len(sentences), batch_size):
+            batch = sentences[i:i+batch_size]
+            inputs = tokenizer(batch, padding=True, truncation=True, return_tensors="pt", return_token_type_ids=False, max_length=512)
+            inputs = {k: v.to(device) for k, v in inputs.items()}
+            output = model(**inputs)
+            batch_emb = output.last_hidden_state[:, 0, :].cpu().numpy()
+            all_embeddings.append(batch_emb)
+            
+    if not all_embeddings:
+        return np.array([])
+    return np.vstack(all_embeddings)
+
+
+def _cluster(embeddings: np.ndarray, threshold: float) -> np.ndarray:
+    return AgglomerativeClustering(
+        metric="cosine", linkage="average",
+        distance_threshold=threshold, n_clusters=None,
+    ).fit_predict(embeddings)
+
+
+def _compute_centroids(embeddings: np.ndarray, labels: np.ndarray) -> dict:
+    valid = sorted(set(labels.tolist()) - {-1})
+    return dict(map(lambda l: (l, embeddings[labels == l].mean(axis=0)), valid))
+
+
+def _nearest_sents(centroid: np.ndarray, sentences: list,
+                   embeddings: np.ndarray, k: int) -> list:
+    sims = cosine_similarity([centroid], embeddings)[0]
+    idxs = np.argsort(sims)[::-1][:k].tolist()
+    return list(map(lambda i: sentences[i], idxs))
+
+
+def _build_summaries(labels: np.ndarray, sentences: list,
+                     embeddings: np.ndarray) -> list:
+    centroids = _compute_centroids(embeddings, labels)
+
+    def _one(tid):
+        mask = labels == tid
+        return {
+            "topic_id": tid,
+            "count":    int(mask.sum()),
+            "centroid": centroids[tid].tolist(),
+            "nearest_sentences": _nearest_sents(
+                centroids[tid], sentences, embeddings, NEAREST_K),
+        }
+    return list(map(_one, sorted(centroids.keys())))
+
+
+def _get_llm() -> ChatMistralAI:
+    """
+    Return a ChatMistralAI instance.
+    FIX: max_retries=0 so langchain_mistralai does NOT internally retry 429s.
+    All retry logic lives in call_agent() in app.py, which also handles
+    MemorySaver thread rotation on INVALID_CHAT_HISTORY. Having max_retries>0
+    here caused double-retry storms that exhausted the rate-limit faster.
+    """
+    return ChatMistralAI(
+        model="mistral-large-latest",
+        temperature=0.2,
+        timeout=MISTRAL_TIMEOUT,
+        max_retries=0,   # FIX-Bug3: no internal retry; outer call_agent handles it
+    )
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Tool 1 — load_scopus_csv
+# ─────────────────────────────────────────────────────────────────────────────
+@tool
+def load_scopus_csv(file_path: str) -> str:
+    """
+    Load a Scopus CSV file correctly.
+    Uses utf-8-sig (handles BOM) + quoting=0 (respects quoted multi-line cells).
+    """
+    df = pd.read_csv(
+        file_path,
+        encoding="utf-8-sig",
+        quoting=0,
+        engine="python",
+        on_bad_lines="skip",
+    )
+    df.to_csv("loaded_data.csv", index=False, encoding="utf-8")
+
+    n    = len(df)
+    cols = list(df.columns)
+
+    abs_texts = list(df["Abstract"].dropna().astype(str)) if "Abstract" in cols else []
+    ttl_texts = list(df["Title"].dropna().astype(str))    if "Title"    in cols else []
+
+    # Combine Title and Abstract
+    df["Combined"] = df["Title"].fillna("") + " " + df["Abstract"].fillna("")
+    df.to_csv("loaded_data.csv", index=False, encoding="utf-8")
+    combined_texts = list(df["Combined"].astype(str))
+    
+    # We no longer split into sentences. Each paper is one document.
+    abs_sents = combined_texts
+    ttl_sents = ttl_texts
+
+    years      = pd.to_numeric(df["Year"], errors="coerce").dropna() if "Year" in cols else pd.Series([], dtype=float)
+    year_range = f"{int(years.min())} – {int(years.max())}" if len(years) else "N/A"
+
+    return json.dumps({
+        "papers":               n,
+        "abstract_sentences":   len(abs_sents),
+        "title_sentences":      len(ttl_sents),
+        "year_range":           year_range,
+        "columns":              cols,
+        "abstract_coverage_pct": round(len(abs_texts) / n * 100, 1) if n else 0,
+        "title_coverage_pct":    round(len(ttl_texts) / n * 100, 1) if n else 0,
+        "sample_titles":        list(df["Title"].dropna().head(5)) if "Title" in cols else [],
+        "file_saved":           "loaded_data.csv",
+        "note": f"Sentence cap for clustering is {MAX_SENTENCES} (for performance).",
+    }, indent=2)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Tool 2 — run_dbscan_discovery (Replaces Agglomerative)
+# ─────────────────────────────────────────────────────────────────────────────
+@tool
+def run_dbscan_discovery(run_key: str = "abstract") -> str:
+    """
+    Core clustering tool using HDBSCAN + UMAP.
+    Embeds with Specter2, clusters using density, and saves to summaries_{run_key}.json.
+    """
+    df    = pd.read_csv("loaded_data.csv")
+    col   = RUN_CONFIGS[run_key][0]
+    texts = list(df[col].dropna().astype(str))
+
+    all_sentences = _texts_to_sentences(texts)
+    sentences = all_sentences[:MAX_SENTENCES]
+    print(f"[run_dbscan_discovery] {len(all_sentences)} items → capped to {len(sentences)}")
+
+    embeddings = _embed(sentences)
+    np.save(f"emb_{run_key}.npy", embeddings)
+
+    # HDBSCAN + UMAP dynamic loop
+    best_labels, best_params = None, {}
+    found = False
+    for n_neighbors in [5, 10, 15]:
+        for min_cluster_size in [5, 10, 15]:
+            for eps in [0.4, 0.5, 0.6]:
+                reducer = umap.UMAP(n_neighbors=n_neighbors, n_components=5, metric='cosine', random_state=42)
+                umap_emb = reducer.fit_transform(embeddings)
+                
+                db = hdbscan.HDBSCAN(min_cluster_size=min_cluster_size, cluster_selection_epsilon=eps, metric='euclidean')
+                labels = db.fit_predict(umap_emb)
+                
+                n_clusters = len(set(labels) - {-1})
+                if 15 <= n_clusters <= 30:
+                    best_labels, best_params = labels, {'n_neighbors': n_neighbors, 'min_cluster_size': min_cluster_size, 'eps': eps}
+                    found = True
+                    break
+            if found: break
+        if found: break
+        
+    if not found:
+        reducer = umap.UMAP(n_neighbors=10, n_components=5, metric='cosine', random_state=42)
+        umap_emb = reducer.fit_transform(embeddings)
+        db = hdbscan.HDBSCAN(min_cluster_size=5, cluster_selection_epsilon=0.5, metric='euclidean')
+        best_labels, best_params = db.fit_predict(umap_emb), {'fallback': True}
+        
+    db_labels = best_labels
+    valid_ids = sorted(set(db_labels.tolist()) - {-1})
+    centroids = _compute_centroids(embeddings, db_labels)
+
+    def _dbscan_summary(cid):
+        mask  = db_labels == cid
+        return {
+            "topic_id":          int(cid),
+            "count":             int(mask.sum()),
+            "centroid":          centroids[cid].tolist(),
+            "nearest_sentences": _nearest_sents(centroids[cid], sentences, embeddings, min(5, len(sentences))),
+        }
+
+    summaries = list(map(_dbscan_summary, valid_ids))
+
+    with open(f"summaries_{run_key}.json", "w", encoding="utf-8") as f:
+        json.dump(summaries, f, indent=2)
+
+    # Chart 1: Scatter
+    n_comp = min(2, len(embeddings), embeddings.shape[1])
+    pca2   = PCA(n_components=n_comp).fit_transform(embeddings)
+    fig1 = px.scatter(
+        x=pca2[:, 0].tolist(), y=pca2[:, 1].tolist() if n_comp > 1 else [0.0]*len(pca2),
+        color=list(map(str, db_labels.tolist())),
+        title=f"DBSCAN Cluster Map ({run_key})", opacity=0.7,
+    )
+    fig1.update_layout(template="plotly_dark")
+    chart1 = f"chart_{run_key}_dbscan_scatter.html"
+    fig1.write_html(chart1, include_plotlyjs="cdn")
+
+    # Chart 2: Frequency Bars
+    counts = [s["count"] for s in summaries]
+    ids = [f"C{s['topic_id']}" for s in summaries]
+    fig2 = px.bar(x=ids, y=counts, title=f"DBSCAN Cluster Sizes ({run_key})", color=counts, color_continuous_scale="Teal")
+    fig2.update_layout(template="plotly_dark")
+    chart2 = f"chart_{run_key}_dbscan_bars.html"
+    fig2.write_html(chart2, include_plotlyjs="cdn")
+
+    return json.dumps({
+        "run_key": run_key,
+        "n_clusters": len(summaries),
+        "noise_points": int((db_labels == -1).sum()),
+        "charts": [chart1, chart2],
+    }, indent=2)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Tool 3 — label_topics_with_council (Replaces single LLM labeler)
+# ─────────────────────────────────────────────────────────────────────────────
+@tool
+def label_topics_with_council(run_key: str = "abstract") -> str:
+    """
+    Main labeling tool using the AI Council (Mistral + Groq + Gemini).
+    Reads summaries, debates labels, and saves directly to labels_{run_key}.json.
+    """
+    with open(f"summaries_{run_key}.json", encoding="utf-8") as f:
+        summaries = json.load(f)
+
+    top = summaries[:MAX_LABEL_TOPICS]
+    topics_for_prompt = list(map(
+        lambda s: {"topic_id": s["topic_id"], "sentences": [sent[:500] + "..." for sent in s.get("nearest_sentences", [])[:3]]}, top
+    ))
+
+    llm_a = _get_llm()
+    llm_b = _get_council_llm_b()
+    llm_c = _get_council_llm_c()
+
+    tmpl = (
+        "You are an expert thematic analyst reviewing DBSCAN clusters.\n"
+        "Clusters:\n{topics_json}\n\n"
+        "For EACH cluster, propose a concise label (3-6 words).\n"
+        "Return ONLY a valid JSON array. Each element: {{\"topic_id\": int, \"label\": \"...\", \"reasoning\": \"...\"}}"
+    )
+    prompt = PromptTemplate(input_variables=["topics_json"], template=tmpl)
+    parser = JsonOutputParser()
+
+    chain_a, chain_b, chain_c = prompt | llm_a | parser, prompt | llm_b | parser, prompt | llm_c | parser
+    input_data = {"topics_json": json.dumps(topics_for_prompt, indent=2)}
+
+    # SAFE BATCH CALLS (avoids NoneType errors on rate limits)
+    res_a = chain_a.invoke(input_data) or []
+    res_b = chain_b.invoke(input_data) or []
+    res_c = chain_c.invoke(input_data) or []
+
+    idx_a = {str(r["topic_id"]): r for r in res_a}
+    idx_b = {str(r["topic_id"]): r for r in res_b}
+    idx_c = {str(r["topic_id"]): r for r in res_c}
+
+    def _consensus(s):
+        cid = str(s["topic_id"])
+        ra, rb, rc = idx_a.get(cid, {}), idx_b.get(cid, {}), idx_c.get(cid, {})
+        la, lb, lc = ra.get("label", "Unknown"), rb.get("label", "Unknown"), rc.get("label", "Unknown")
+
+        s_ab, s_bc, s_ca = _council_agreement_score(la, lb), _council_agreement_score(lb, lc), _council_agreement_score(lc, la)
+        max_score = max(s_ab, s_bc, s_ca)
+        agreed = max_score >= 0.65
+
+        avg_a, avg_b, avg_c = (s_ab + s_ca) / 2, (s_ab + s_bc) / 2, (s_bc + s_ca) / 2
+        best = la if (avg_a >= avg_b and avg_a >= avg_c) else (lb if avg_b >= avg_a and avg_b >= avg_c else lc)
+        consensus = best if agreed else la
+
+        ui = format_consensus_ui(la, lb, lc, agreed, max_score, ra.get("reasoning",""), rb.get("reasoning",""), rc.get("reasoning",""))
+        return {**s, "label": consensus, "council_ui": ui, "agreement_score": max_score}
+
+    council_labels = list(map(_consensus, top))
+    out = f"labels_{run_key}.json"
+    with open(out, "w", encoding="utf-8") as f:
+        json.dump(council_labels, f, indent=2)
+
+    return json.dumps({"run_key": run_key, "total_labelled": len(council_labels), "output_file": out}, indent=2)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Tool 4 — consolidate_into_themes
+# ─────────────────────────────────────────────────────────────────────────────
+@tool
+def consolidate_into_themes(run_key: str = "abstract", theme_map: str = "") -> str:
+    """
+    Merge topic clusters into core themes using a dual-LLM AI Council.
+    """
+    with open(f"labels_{run_key}.json", encoding="utf-8") as f:
+        labelled = json.load(f)
+
+    llm_a = _get_llm()
+    llm_b = _get_council_llm_b()
+    llm_c = _get_council_llm_c()
+    parser = JsonOutputParser()
+
+    prompt = PromptTemplate(
+        input_variables=["topics_json"],
+        template=(
+            "You are a thematic analyst.\n\n"
+            "Topics: {topics_json}\n\n"
+            "Consolidate into 4-8 themes. Return JSON array. Each element: "
+            "{{\"theme_name\": \"...\", \"topic_ids\": [1,2,3], \"rationale\": \"...\"}}"
+        ),
+    )
+    chain_a = prompt | llm_a | parser
+    chain_b = prompt | llm_b | parser
+    chain_c = prompt | llm_c | parser
+
+    summary = json.dumps(list(map(lambda t: {"id": t["topic_id"], "lbl": t["label"]}, labelled)), indent=2)
+    raw_a = chain_a.invoke({"topics_json": summary}) or []
+    raw_b = chain_b.invoke({"topics_json": summary}) or []
+    raw_c = chain_c.invoke({"topics_json": summary}) or []
+
+    # Simple comparison of first 2 themes generated
+    l_a = ", ".join(map(lambda x: x["theme_name"], raw_a[:2]))
+    l_b = ", ".join(map(lambda x: x["theme_name"], raw_b[:2]))
+    l_c = ", ".join(map(lambda x: x["theme_name"], raw_c[:2]))
+    
+    s_ab = _council_agreement_score(l_a, l_b)
+    s_bc = _council_agreement_score(l_b, l_c)
+    s_ca = _council_agreement_score(l_c, l_a)
+    
+    score = max(s_ab, s_bc, s_ca)
+    agreed = score >= 0.3
+    ui = format_consensus_ui(l_a, l_b, l_c, agreed, score)
+
+    themes = list(map(lambda t: {**t, "council_ui": ui}, raw_a))
+
+    out = f"themes_{run_key}.json"
+    with open(out, "w", encoding="utf-8") as f:
+        json.dump(themes, f, indent=2)
+    with open("themes.json", "w", encoding="utf-8") as f:
+        json.dump(themes, f, indent=2)
+
+    return json.dumps({
+        "run_key": run_key,
+        "total_themes": len(themes),
+        "output_file": out,
+        "themes_preview": themes[:3],
+    }, indent=2)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Tool 5 — compare_with_taxonomy
+# ─────────────────────────────────────────────────────────────────────────────
+@tool
+def compare_with_taxonomy(run_key: str = "abstract") -> str:
+    """
+    Map each consolidated theme to the PAJAIS 25-category taxonomy via Mistral.
+    Returns MAPPED vs NOVEL per theme. Saves taxonomy_map.json.
+
+    FIX-Bug4: Prefer themes_{run_key}.json over the generic themes.json so that
+    abstract and title runs never cross-contaminate each other's theme data.
+
+    Args:
+        run_key: 'abstract' or 'title'
+
+    Returns:
+        JSON: total mapped, novel count, full mapping, output_file.
+    """
+    # FIX-Bug4: use run_key-specific file first, fall back to generic themes.json
+    run_themes_file = f"themes_{run_key}.json"
+    themes_file = run_themes_file if os.path.exists(run_themes_file) else "themes.json"
+    with open(themes_file, encoding="utf-8") as f:
+        themes = json.load(f)
+
+    llm    = _get_llm()
+    parser = JsonOutputParser()
+
+    prompt = PromptTemplate(
+        input_variables=["themes_json", "taxonomy"],
+        template=(
+            "You are a research classification expert.\n\n"
+            "PAJAIS Taxonomy (25 categories):\n{taxonomy}\n\n"
+            "Themes from corpus:\n{themes_json}\n\n"
+            "For each theme, find the best PAJAIS category match.\n"
+            "Return ONLY a valid JSON array — no markdown. Each element:\n"
+            "  theme_name: string (match input exactly)\n"
+            "  pajais_match: best PAJAIS category, or 'NOVEL' if none fits\n"
+            "  match_confidence: float 0.0-1.0\n"
+            "  reasoning: one sentence\n"
+            "  is_novel: boolean\n"
+        ),
+    )
+    chain = prompt | llm | parser
+
+    theme_summaries = list(map(
+        lambda t: {
+            "theme_name":        t["theme_name"],
+            "total_sentences":   t.get("total_sentences", 0),
+            "constituent_labels": t.get("constituent_labels", []),
+            "sample": (t.get("representative_sentences", [""])[0][:100]
+                       if t.get("representative_sentences") else ""),
+        },
+        themes,
+    ))
+
+    mapping = chain.invoke({
+        "themes_json": json.dumps(theme_summaries, indent=2),
+        "taxonomy":    "\n".join(f"{i+1}. {c}" for i, c in enumerate(PAJAIS_TAXONOMY)),
+    })
+
+    with open("taxonomy_map.json", "w", encoding="utf-8") as f:
+        json.dump(mapping, f, indent=2)
+
+    novel_count = len(list(filter(lambda m: m.get("is_novel", False), mapping)))
+
+    return json.dumps({
+        "run_key":             run_key,
+        "total_themes_mapped": len(mapping),
+        "novel_themes":        novel_count,
+        "mapped_themes":       len(mapping) - novel_count,
+        "output_file":         "taxonomy_map.json",
+        "mapping":             mapping,
+    }, indent=2)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Tool 6 — generate_comparison_csv
+# ─────────────────────────────────────────────────────────────────────────────
+@tool
+def generate_comparison_csv() -> str:
+    """
+    Load themes from both abstract and title runs, create side-by-side
+    comparison DataFrame. Saves comparison.csv.
+
+    Returns:
+        JSON: output_file, row_count, preview.
+    """
+    def _load(rk):
+        p   = f"themes_{rk}.json"
+        raw = open(p, encoding="utf-8").read() if os.path.exists(p) else "[]"
+        return json.loads(raw)
+
+    abs_themes = _load("abstract")
+    ttl_themes = _load("title")
+    max_rows   = max(len(abs_themes), len(ttl_themes), 1)
+
+    pad_abs = abs_themes + [{}] * (max_rows - len(abs_themes))
+    pad_ttl = ttl_themes + [{}] * (max_rows - len(ttl_themes))
+
+    rows = list(map(
+        lambda pair: {
+            "#":               pair[0] + 1,
+            "Abstract Theme":  pair[1][0].get("theme_name", ""),
+            "Abstract Sents":  pair[1][0].get("total_sentences", 0),
+            "Abstract Labels": ", ".join(pair[1][0].get("constituent_labels", [])[:3]),
+            "Title Theme":     pair[1][1].get("theme_name", ""),
+            "Title Sents":     pair[1][1].get("total_sentences", 0),
+            "Title Labels":    ", ".join(pair[1][1].get("constituent_labels", [])[:3]),
+            "Convergence":     (
+                "✓" if pair[1][0].get("theme_name", "").lower()[:8]
+                    == pair[1][1].get("theme_name", "").lower()[:8]
+                else ""
+            ),
+        },
+        enumerate(zip(pad_abs, pad_ttl)),
+    ))
+
+    df = pd.DataFrame(rows)
+    df.to_csv("comparison.csv", index=False)
+
+    return json.dumps({
+        "output_file": "comparison.csv",
+        "row_count":   len(df),
+        "preview":     rows[:3],
+    }, indent=2)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Tool 7 — export_narrative
+# ─────────────────────────────────────────────────────────────────────────────
+@tool
+def export_narrative(run_key: str = "abstract") -> str:
+    """
+    Generate a 500-word Section 7 narrative using Mistral LLM.
+    Covers methodology, themes, PAJAIS alignment, limitations, implications.
+    Saves narrative.txt.
+
+    Args:
+        run_key: 'abstract' or 'title'
+
+    Returns:
+        JSON: output_file, word_count, 500-char preview.
+    """
+    with open("themes.json", encoding="utf-8") as f:
+        themes = json.load(f)
+
+    tax_raw  = open("taxonomy_map.json", encoding="utf-8").read() if os.path.exists("taxonomy_map.json") else "[]"
+    tax_data = json.loads(tax_raw)
+
+    llm = _get_llm()
+    llm.temperature = 0.4  # Slightly higher for creativity in Section 7 narrative
+    prompt = PromptTemplate(
+        input_variables=["run_key", "themes_json", "taxonomy_json"],
+        template=(
+            "You are writing Section 7 of an academic literature review paper.\n\n"
+            "Analysis column: {run_key}\n"
+            "Themes:\n{themes_json}\n\n"
+            "PAJAIS Mapping:\n{taxonomy_json}\n\n"
+            "Write a 500-word Section 7 covering:\n"
+            "1. Methodology (BERTopic + Braun & Clarke 2006 six phases)\n"
+            "2. Key themes discovered (reference each by name)\n"
+            "3. PAJAIS taxonomy alignment (MAPPED vs NOVEL themes)\n"
+            "4. Limitations of this computational approach\n"
+            "5. Implications for future research\n\n"
+            "Academic third-person prose, full paragraphs only, minimum 500 words."
+        ),
+    )
+    chain    = prompt | llm
+    response = chain.invoke({
+        "run_key":       run_key,
+        "themes_json":   json.dumps(themes, indent=2),
+        "taxonomy_json": json.dumps(tax_data, indent=2),
+    })
+    text = response.content if hasattr(response, "content") else str(response)
+
+    with open("narrative.txt", "w", encoding="utf-8") as f:
+        f.write(text)
+
+    return json.dumps({
+        "output_file": "narrative.txt",
+        "word_count":  len(text.split()),
+        "preview":     text[:500],
+    }, indent=2)
+
+
+# Verified: zero if/else, zero for/while, zero try/except
+
+# ─────────────────────────────────────────────────────────────────────────────
+# AI Council helpers
+# ─────────────────────────────────────────────────────────────────────────────
+def _get_council_llm_b() -> ChatGroq:
+    """Return the Groq Llama-3 model as the second council LLM."""
+    return ChatGroq(model="llama-3.3-70b-versatile", temperature=0.2, max_retries=0)
+
+def _get_council_llm_c() -> ChatGoogleGenerativeAI:
+    """Return the Gemini model as the third council LLM."""
+    return ChatGoogleGenerativeAI(model="gemini-2.5-flash", temperature=0.2, max_retries=0)
+
+
+def format_consensus_ui(label_a, label_b, label_c="", agreed=False, score=0.0, reason_a="", reason_b="", reason_c=""):
+    """Generate an ultra-compact HTML Argument UI for 3 models."""
+    status_icon = "✅ Match" if agreed else "⚠️ Diverge"
+    status_color = "#2ecc71" if agreed else "#e67e22"
+    
+    return f"""
+<div style="margin-top:4px; border-left: 2px solid {status_color}; padding-left:8px; font-size:0.75rem;">
+    <div style="color:{status_color}; font-weight:700; margin-bottom:2px;">{status_icon} (Max Match: {score})</div>
+    <div style="display:flex; gap:10px;">
+        <div style="flex:1; background:#0d1117; padding:6px; border-radius:4px; border:1px solid #30363d; color:#f0f6fc;">
+            <b style="color:#7fb3f5; font-size:0.65rem;">MISTRAL:</b> {reason_a}
+        </div>
+        <div style="flex:1; background:#0d1117; padding:6px; border-radius:4px; border:1px solid #30363d; color:#f0f6fc;">
+            <b style="color:#7fb3f5; font-size:0.65rem;">GROQ:</b> {reason_b}
+        </div>
+        <div style="flex:1; background:#0d1117; padding:6px; border-radius:4px; border:1px solid #30363d; color:#f0f6fc;">
+            <b style="color:#7fb3f5; font-size:0.65rem;">GEMINI:</b> {reason_c}
+        </div>
+    </div>
+</div>
+"""
+
+def _council_agreement_score(label_a: str, label_b: str) -> float:
+    """Compute semantic cosine similarity between two label strings."""
+    global _label_sim_model
+    if _label_sim_model is None:
+        _label_sim_model = SentenceTransformer("all-MiniLM-L6-v2")
+    
+    embs = _label_sim_model.encode([label_a, label_b])
+    return round(float(cosine_similarity([embs[0]], [embs[1]])[0][0]), 3)
+
+
+# # ─────────────────────────────────────────────────────────────────────────────
+# # Tool 8 — run_dbscan_clustering
+# # ─────────────────────────────────────────────────────────────────────────────
+# @tool
+# def run_dbscan_clustering(run_key: str = "abstract", eps: float = 0.3, min_samples: int = 3) -> str:
+#     """
+#     Run HDBSCAN + UMAP clustering on the embeddings.
+#     Loops to find hyperparameters that yield between 15 and 30 clusters.
+#     Also ensures min_cluster_size is around 5.
+#     (eps and min_samples args are kept for interface compatibility but may be overridden by the loop).
+#     """
+#     embeddings = np.load(f"emb_{run_key}.npy")
+
+#     with open(f"summaries_{run_key}.json", encoding="utf-8") as f:
+#         agg_summaries = json.load(f)
+        
+#     all_nearest = [s for summ in agg_summaries for s in summ.get("nearest_sentences", [])]
+#     if not all_nearest:
+#         import pandas as pd
+#         df = pd.read_csv("loaded_data.csv")
+#         all_nearest = list(df["Combined"].dropna().astype(str))
+
+#     best_labels = None
+#     best_n_clusters = 0
+#     best_params = {}
+    
+#     found = False
+#     for n_neighbors in [5, 10, 15]:
+#         for min_cluster_size in [5, 10, 15]:
+#             for cluster_selection_epsilon in [0.4, 0.5, 0.6]:
+#                 reducer = umap.UMAP(n_neighbors=n_neighbors, n_components=5, metric='cosine', random_state=42)
+#                 umap_emb = reducer.fit_transform(embeddings)
+                
+#                 db = hdbscan.HDBSCAN(min_cluster_size=min_cluster_size, cluster_selection_epsilon=cluster_selection_epsilon, metric='euclidean')
+#                 labels = db.fit_predict(umap_emb)
+                
+#                 n_clusters = len(set(labels) - {-1})
+#                 if 15 <= n_clusters <= 30:
+#                     best_labels = labels
+#                     best_n_clusters = n_clusters
+#                     best_params = {'n_neighbors': n_neighbors, 'min_cluster_size': min_cluster_size, 'eps': cluster_selection_epsilon}
+#                     found = True
+#                     break
+#             if found: break
+#         if found: break
+        
+#     if not found:
+#         reducer = umap.UMAP(n_neighbors=10, n_components=5, metric='cosine', random_state=42)
+#         umap_emb = reducer.fit_transform(embeddings)
+#         db = hdbscan.HDBSCAN(min_cluster_size=5, cluster_selection_epsilon=0.5, metric='euclidean')
+#         best_labels = db.fit_predict(umap_emb)
+#         best_params = {'fallback': True}
+        
+#     db_labels = best_labels
+#     valid_ids    = sorted(set(db_labels.tolist()) - {-1})
+#     noise_count  = int((db_labels == -1).sum())
+
+#     centroids  = _compute_centroids(embeddings, db_labels)
+
+#     def _dbscan_summary(cid):
+#         mask  = db_labels == cid
+#         count = int(mask.sum())
+#         sents = _nearest_sents(centroids[cid],
+#                                all_nearest or [f"Cluster {cid}"],
+#                                embeddings[: len(all_nearest or ["x"])],
+#                                min(3, len(all_nearest or ["x"])))
+#         return {
+#             "cluster_id":            int(cid),
+#             "count":                 count,
+#             "centroid":              centroids[cid].tolist(),
+#             "nearest_sentences":     sents,
+#             "source":                "hdbscan",
+#         }
+
+#     summaries = list(map(_dbscan_summary, valid_ids))
+
+#     out_file = f"dbscan_summaries_{run_key}.json"
+#     with open(out_file, "w", encoding="utf-8") as f:
+#         json.dump(summaries, f, indent=2)
+
+#     n_comp = min(2, len(embeddings), embeddings.shape[1])
+#     pca2   = PCA(n_components=n_comp).fit_transform(embeddings)
+#     x_vals = pca2[:, 0].tolist()
+#     y_vals = pca2[:, 1].tolist() if n_comp > 1 else [0.0] * len(x_vals)
+#     colors = db_labels.tolist()
+
+#     fig_scatter = px.scatter(
+#         x=x_vals, y=y_vals,
+#         color=list(map(str, colors)),
+#         title=f"HDBSCAN Cluster Map ({run_key}) — {best_params}",
+#         labels={"x": "PC1", "y": "PC2", "color": "Cluster"},
+#         opacity=0.7,
+#     )
+#     fig_scatter.update_layout(template="plotly_dark")
+#     chart_scatter = f"chart_{run_key}_dbscan_scatter.html"
+#     fig_scatter.write_html(chart_scatter, include_plotlyjs="cdn")
+
+#     agg_count  = len(agg_summaries)
+#     dbscan_count = len(summaries)
+#     fig_cmp = px.bar(
+#         x=["Agglomerative", "HDBSCAN"],
+#         y=[agg_count, dbscan_count],
+#         color=["Agglomerative", "HDBSCAN"],
+#         color_discrete_sequence=["#4a90d9", "#e67e22"],
+#         title=f"Cluster Count Comparison ({run_key})",
+#         labels={"x": "Method", "y": "# Clusters"},
+#         text=[agg_count, dbscan_count],
+#     )
+#     fig_cmp.update_traces(textposition="outside")
+#     fig_cmp.update_layout(template="plotly_dark", showlegend=False)
+#     chart_cmp = f"chart_{run_key}_dbscan_comparison.html"
+#     fig_cmp.write_html(chart_cmp, include_plotlyjs="cdn")
+
+#     largest = max(map(lambda s: s["count"], summaries), default=0)
+
+#     return json.dumps({
+#         "run_key":          run_key,
+#         "n_clusters":       len(summaries),
+#         "noise_points":     noise_count,
+#         "largest_cluster":  largest,
+#         "best_params":      best_params,
+#         "summaries_file":   out_file,
+#         "charts":           [chart_scatter, chart_cmp],
+#         "preview":          summaries[:3],
+#     }, indent=2)
+
+
+# # ─────────────────────────────────────────────────────────────────────────────
+# # Tool 9 — refine_large_clusters
+# # ─────────────────────────────────────────────────────────────────────────────
+# @tool
+# def refine_large_clusters(run_key: str = "abstract", size_threshold: int = 100) -> str:
+#     """
+#     Post-processing: identifies overly large DBSCAN clusters and refines them
+#     into sub-clusters using a tighter AgglomerativeClustering threshold (0.45).
+
+#     Does NOT modify dbscan_summaries or any existing agglomerative results.
+#     Saves results to refined_clusters_{run_key}.json.
+
+#     Args:
+#         run_key:        'abstract' or 'title'
+#         size_threshold: Clusters with count > this value will be refined (default 200)
+
+#     Returns:
+#         JSON: n_refined, total_subclusters, refined_clusters_file, chart file.
+#     """
+#     dbscan_file = f"dbscan_summaries_{run_key}.json"
+#     with open(dbscan_file, encoding="utf-8") as f:
+#         summaries = json.load(f)
+
+#     embeddings = np.load(f"emb_{run_key}.npy")
+
+#     large     = list(filter(lambda s: s["count"] >= size_threshold, summaries))
+#     unchanged = list(filter(lambda s: s["count"] <  size_threshold, summaries))
+
+#     # Re-cluster each large cluster's embedding slice
+#     def _refine_one(parent_summary):
+#         pid       = parent_summary["cluster_id"]
+#         parent_c  = np.array(parent_summary["centroid"])
+#         # Find the indices in the full embedding that are nearest to this centroid
+#         sims  = cosine_similarity([parent_c], embeddings)[0]
+#         count = parent_summary["count"]
+#         idxs  = np.argsort(sims)[::-1][:count].tolist()
+
+#         sub_emb    = embeddings[idxs]
+#         sub_labels = AgglomerativeClustering(
+#             metric="cosine", linkage="average",
+#             distance_threshold=0.45, n_clusters=None,
+#         ).fit_predict(sub_emb)
+
+#         sub_ids  = sorted(set(sub_labels.tolist()))
+#         sub_centroids = dict(map(
+#             lambda sid: (sid, sub_emb[sub_labels == sid].mean(axis=0)),
+#             sub_ids,
+#         ))
+
+#         def _sub(sid):
+#             mask  = sub_labels == sid
+#             sents = parent_summary.get("nearest_sentences", [])
+#             return {
+#                 "cluster_id":        f"{pid}.{sid}",
+#                 "parent_cluster_id": pid,
+#                 "count":             int(mask.sum()),
+#                 "centroid":          sub_centroids[sid].tolist(),
+#                 "nearest_sentences": sents[:3],
+#                 "source":            "dbscan_refined",
+#             }
+
+#         return list(map(_sub, sub_ids))
+
+#     refined_subs = [item for sublist in map(_refine_one, large) for item in sublist]
+
+#     # Unchanged clusters kept as-is with a source tag
+#     unchanged_kept = list(map(
+#         lambda s: {**s, "source": "dbscan_unchanged"},
+#         unchanged,
+#     ))
+
+#     all_refined = unchanged_kept + refined_subs
+
+#     out_file = f"refined_clusters_{run_key}.json"
+#     with open(out_file, "w", encoding="utf-8") as f:
+#         json.dump(all_refined, f, indent=2)
+
+#     # ── Chart: Treemap of refined sub-clusters ────────────────────────────────
+#     labels_list  = list(map(lambda c: str(c["cluster_id"]),  all_refined))
+#     parents_list = list(map(
+#         lambda c: str(c.get("parent_cluster_id", "root")) if "." in str(c["cluster_id"]) else "root",
+#         all_refined,
+#     ))
+#     values_list  = list(map(lambda c: c["count"], all_refined))
+
+#     fig_tree = px.treemap(
+#         names=labels_list,
+#         parents=parents_list,
+#         values=values_list,
+#         title=f"Refined Sub-Clusters ({run_key}) — threshold={size_threshold}",
+#     )
+#     fig_tree.update_layout(template="plotly_dark")
+#     chart_tree = f"chart_{run_key}_refined.html"
+#     fig_tree.write_html(chart_tree, include_plotlyjs="cdn")
+
+#     return json.dumps({
+#         "run_key":               run_key,
+#         "size_threshold":        size_threshold,
+#         "n_large_refined":       len(large),
+#         "total_subclusters":     len(refined_subs),
+#         "unchanged_clusters":    len(unchanged),
+#         "total_output_clusters": len(all_refined),
+#         "output_file":           out_file,
+#         "chart":                 chart_tree,
+#         "preview":               all_refined[:4],
+#     }, indent=2)
+
+
+# # ─────────────────────────────────────────────────────────────────────────────
+# # Tool 10 — run_ai_council
+# # ─────────────────────────────────────────────────────────────────────────────
+# @tool
+# def run_ai_council(run_key: str = "abstract") -> str:
+#     """
+#     AI Council: two LLM instances independently label each DBSCAN cluster
+#     from its top-3 representative sentences, then a consensus step merges them.
+
+#     Model A: Mistral Large (temperature=0.2) — analytical, precise
+#     Model B: Groq Llama-3.3-70b-versatile (temperature=0.2) — genuinely different
+#              model providing independent perspective (Karpathy-style second opinion)
+
+#     Consensus rule:
+#       - Jaccard word overlap >= 0.4  → agreement; consensus = Model A label
+#       - Jaccard word overlap < 0.4   → divergence; Model A (Mistral) selected as primary
+
+#     Saves council_labels_{run_key}.json (compatible with PAJAIS mapping).
+
+#     Args:
+#         run_key: 'abstract' or 'title'
+
+#     Returns:
+#         JSON: total_labelled, agreement_rate, output_file, preview.
+#     """
+#     dbscan_file = f"dbscan_summaries_{run_key}.json"
+#     with open(dbscan_file, encoding="utf-8") as f:
+#         summaries = json.load(f)
+
+#     top = summaries[:MAX_LABEL_TOPICS]
+
+#     # Truncate to 500 chars to fit within Groq's 12000 TPM limit
+#     topics_for_prompt = list(map(
+#         lambda s: {
+#             "cluster_id": s["cluster_id"],
+#             "count":      s["count"],
+#             "sentences":  [sent[:500] + "..." for sent in s.get("nearest_sentences", [])[:3]],
+#         },
+#         top,
+#     ))
+
+#     # ── Model A (analytical Mistral) ──────────────────────────────────────────
+#     llm_a  = _get_llm()   # temperature=0.2
+#     llm_b  = _get_council_llm_b()  # temperature=0.2
+#     llm_c  = _get_council_llm_c()  # temperature=0.2
+
+#     council_prompt_tmpl = (
+#         "You are an expert thematic analyst reviewing HDBSCAN-discovered clusters "
+#         "from an academic corpus.\n\n"
+#         "Below are cluster IDs with their top-3 representative sentences:\n\n"
+#         "{topics_json}\n\n"
+#         "For EACH cluster, propose a concise label (3-6 words).\n"
+#         "Return ONLY a valid JSON array. Each element must have:\n"
+#         "  cluster_id: same integer as input\n"
+#         "  label: concise 3-6 word research area name\n"
+#         "  reasoning: one sentence explaining your choice\n\n"
+#         "Return ALL {n} clusters. Do not skip any."
+#     )
+
+#     prompt_a = PromptTemplate(input_variables=["topics_json", "n"], template=council_prompt_tmpl)
+#     prompt_b = PromptTemplate(input_variables=["topics_json", "n"], template=council_prompt_tmpl)
+#     prompt_c = PromptTemplate(input_variables=["topics_json", "n"], template=council_prompt_tmpl)
+
+#     parser    = JsonOutputParser()
+#     chain_a   = prompt_a | llm_a | parser
+#     chain_b   = prompt_b | llm_b | parser
+#     chain_c   = prompt_c | llm_c | parser
+
+#     input_data = {
+#         "topics_json": json.dumps(topics_for_prompt, indent=2),
+#         "n":           len(top),
+#     }
+
+#     results_a = chain_a.invoke(input_data) or []
+#     results_b = chain_b.invoke(input_data) or []
+#     results_c = chain_c.invoke(input_data) or []
+
+#     idx_a = {str(r["cluster_id"]): r for r in results_a}
+#     idx_b = {str(r["cluster_id"]): r for r in results_b}
+#     idx_c = {str(r["cluster_id"]): r for r in results_c}
+
+#     # ── Consensus step ────────────────────────────────────────────────────────
+#     def _consensus(cluster_summary):
+#         cid    = str(cluster_summary["cluster_id"])
+#         ra     = idx_a.get(cid, {})
+#         rb     = idx_b.get(cid, {})
+#         rc     = idx_c.get(cid, {})
+#         label_a = ra.get("label", f"Cluster {cid}")
+#         label_b = rb.get("label", f"Cluster {cid}")
+#         label_c = rc.get("label", f"Cluster {cid}")
+
+#         s_ab   = _council_agreement_score(label_a, label_b)
+#         s_bc   = _council_agreement_score(label_b, label_c)
+#         s_ca   = _council_agreement_score(label_c, label_a)
+        
+#         max_score = max(s_ab, s_bc, s_ca)
+#         agreed = max_score >= 0.65
+
+#         # Pick the label that has the highest average agreement with others
+#         avg_a = (s_ab + s_ca) / 2
+#         avg_b = (s_ab + s_bc) / 2
+#         avg_c = (s_bc + s_ca) / 2
+        
+#         if avg_a >= avg_b and avg_a >= avg_c:
+#             best_label = label_a
+#         elif avg_b >= avg_a and avg_b >= avg_c:
+#             best_label = label_b
+#         else:
+#             best_label = label_c
+
+#         consensus = best_label if agreed else label_a
+        
+#         ui = format_consensus_ui(label_a, label_b, label_c, agreed, max_score, ra.get("reasoning",""), rb.get("reasoning",""), rc.get("reasoning",""))
+
+#         return {
+#             "cluster_id":        cluster_summary["cluster_id"],
+#             "count":             cluster_summary["count"],
+#             "nearest_sentences": cluster_summary.get("nearest_sentences", [])[:3],
+#             "label_a":           label_a,
+#             "label_b":           label_b,
+#             "label_c":           label_c,
+#             "consensus_label":   consensus,
+#             "agreement_score":   max_score,
+#             "council_ui":        ui,
+#             "source":            "dbscan_ai_council",
+#             "label":             consensus,
+#             "reasoning":         ra.get("reasoning", ""),
+#         }
+
+#     council_labels = list(map(_consensus, top))
+
+#     out_file = f"council_labels_{run_key}.json"
+#     with open(out_file, "w", encoding="utf-8") as f:
+#         json.dump(council_labels, f, indent=2)
+
+#     agreed_count   = len(list(filter(lambda c: c["agreement_score"] >= 0.65, council_labels)))
+#     agreement_rate = round(agreed_count / max(len(council_labels), 1) * 100, 1)
+
+#     return json.dumps({
+#         "run_key":        run_key,
+#         "total_labelled": len(council_labels),
+#         "agreed_count":   agreed_count,
+#         "agreement_rate": f"{agreement_rate}%",
+#         "output_file":    out_file,
+#         "note": (
+#             "council_labels contain 'label' field for PAJAIS compatibility. "
+#             "Model A = Mistral Large. Model B = Groq Llama-3. Model C = Gemini Flash."
+#         ),
+#         "preview": council_labels[:4],
+#     }, indent=2)
+
+
+# # Verified: zero if/else*, zero for/while, zero try/except
+# # (*_get_council_llm_b uses a conditional expression, not an if/else block)