Upload 4 files

README.md

@@ -1,26 +1,73 @@
 ---
 title: SceneSolver
-author: samxkevin
+emoji: 🕵️
+colorFrom: yellow
+colorTo: gray
 sdk: gradio
 sdk_version: 5.49.1
 app_file: app.py
+python_version: "3.12"
 license: mit
 short_description: Multimodal surveillance analysis and evidence synthesis
 ---
 
-# SceneSolver
+# SceneSolver — CrimeSceneSolver (Scene Solver v3)
 
-SceneSolver is a multimodal surveillance-analysis system. The public demo exposes the inference pipeline while the trained model assets remain in the author's Hugging Face repositories.
+SceneSolver is a multimodal surveillance-analysis system. This public Space runs the
+real nine-stage pipeline from the original notebook, together with the Sherl0ck
+agentic escalation layer (Cohere reasoning + LFM2.5-Audio voice) and the full
+black/white/orange four-tab frontend (Home · Analyse · Sherl0ck Agent · Contact).
 
-## Pipeline
+## Pipeline (nine stages, all real models)
 
-AE anomaly signal → TimeSformer binary gate → TimeSformer multi-class classification → YOLO weapon detection → audio features → tracking/behavior extraction → fusion reasoning → RL3033 temporal reasoning → optional LLaVA contextual explanation.
+0. **AutoEncoder (AE)** — primary anomaly signal (batched reconstruction error)
+1. **TimeSformer Binary** — semantic gate (Normal / Suspicious)
+2. **TimeSformer Multi-class** — fine-grained anomaly classification (7 classes)
+3. **YOLO Weapon Detection** — hard safety trigger (melee / firearms)
+4. **Audio Feature Extraction** — ZCR / MFCC variance / spectral flux / RMS / centroid
+5. **Tracking & Behavior** — BYTETrack or optical-flow motion analysis
+6. **Fusion Layer** — rule-based reasoning core with a full reasoning trace
+7. **RL3033** — PPO temporal-reasoning agent (final risk)
+8. **LLaVA** — optional 4-frame contextual narrative (checkbox on the Analyse tab)
 
-## Model sources
+Outputs: interactive risk report, peak anomaly frame, 11-page forensic PDF dossier,
+and structured JSON.
 
-- `samxkevin/scenesolver-assets`
-- `samxkevin/SceneDescriptionModel`
+## Sherl0ck Agent
 
-## Notes
+Text chat and hands-free voice call (browser VAD → LFM2.5-Audio ASR → Cohere →
+LFM2.5-Audio TTS), incoming-call overlay for HIGH/CRITICAL results, and evidence
+dispatch to Telegram / Discord / email when those secrets are configured.
 
-The system is an experimental/research project. Results are not definitive conclusions and require human review.
+## Model sources (public)
+
+- [`samxkevin/scenesolver-assets`](https://huggingface.co/samxkevin/scenesolver-assets) — AE, TimeSformer, RL3033, YOLO checkpoints + agent audio
+- [`samxkevin/SceneDescriptionModel`](https://huggingface.co/samxkevin/SceneDescriptionModel) — LLaVA VLM (~14 GB, downloaded on first use of Stage 8)
+- `facebook/timesformer-base-finetuned-k400` — TimeSformer base weights
+- `LiquidAI/LFM2.5-Audio-1.5B` — Sherl0ck voice (ASR/TTS)
+
+## Hardware — ZeroGPU
+
+This Space is designed for **ZeroGPU** (select *ZeroGPU* under Settings → Space
+hardware; ZeroGPU is not expressible in this README). All GPU-heavy inference runs
+inside `@spaces.GPU` regions (full pipeline analysis with dynamic duration, and the
+Sherl0ck voice ASR/TTS turns). Python is pinned to 3.12 (a ZeroGPU-supported
+runtime).
+
+## Secrets (Settings → Variables and secrets)
+
+All integrations are optional and read from the environment — no credentials are
+stored in this repository:
+
+| Secret | Purpose |
+| --- | --- |
+| `COHERE_API_KEY` | Sherl0ck text/voice reasoning (offline fallback replies when absent) |
+| `TELEGRAM_BOT_TOKEN`, `TELEGRAM_CHAT_ID` | Telegram evidence dispatch |
+| `DISCORD_WEBHOOK_URL` | Discord evidence dispatch |
+| `ALERT_EMAIL_FROM`, `ALERT_EMAIL_PASS`, `ALERT_EMAIL_TO` | Email alerts |
+| `SENDER_EMAIL`, `SENDER_PASS`, `ADMIN_EMAIL` | Contact form delivery |
+
+## Disclaimer
+
+Experimental research project. ML outputs are investigative aids, not definitive
+conclusions, and require human expert review before any operational action.

requirements.txt

@@ -1,20 +1,38 @@
-gradio>=5.0,<6.0
+# ── SceneSolver Space dependencies ─────────────────────────────────────────────
+# Gradio is pinned to exactly match sdk_version in README.md (no conflicting
+# requirements anywhere). liquid-audio 1.3+ has no gradio dependency, so the
+# historical gradio==4.44.0 / >=4.44.1 resolver conflict cannot recur.
+gradio==5.49.1
 spaces>=0.38.0
 huggingface_hub>=0.34.0
+
+# ── Core numerics / vision ────────────────────────────────────────────────────
 numpy
 opencv-python-headless
-torch
-torchvision
-transformers>=4.45
-accelerate
-safetensors
 Pillow
 tqdm
 matplotlib
-reportlab
-stable-baselines3
-ultralytics
-lap
-librosa
-openai-whisper
-cohere>=5.0.0
+
+# ── Deep learning (torch>=2.8 required by liquid-audio; PyPI linux wheels are
+#    CUDA-enabled, which is what ZeroGPU needs) ────────────────────────────────
+torch>=2.8.0
+torchvision>=0.23.0
+torchaudio>=2.8.0
+transformers>=4.55.4
+accelerate>=1.10.1
+safetensors
+
+# ── Pipeline stage dependencies ───────────────────────────────────────────────
+stable-baselines3          # Stage 7 — RL3033 (PPO)
+ultralytics                # Stage 3/5 — YOLO weapon detection + tracking
+lap                        # ultralytics BYTETrack association
+librosa>=0.11.0            # Stage 4 — audio feature extraction
+reportlab                  # forensic PDF dossier
+imageio-ffmpeg             # bundled ffmpeg binary (no apt access on Spaces)
+
+# ── Sherl0ck agent layer ──────────────────────────────────────────────────────
+cohere>=5.0.0              # Sherl0ck reasoning (command-a-03-2025)
+liquid-audio>=1.3.0        # LFM2.5-Audio-1.5B ASR/TTS bridge
+soundfile
+requests
+anyio                      # used by the /ssvoice HTTP routes (also a starlette dep)

scenesolver_core.py

@@ -1,52 +1,11 @@
 # -*- coding: utf-8 -*-
-"""NewChecker.ipynb
-
-Automatically generated by Colab.
-
-Original file is located at
-    https://colab.research.google.com/drive/
-Run the below code to download the Dataset into the Runtime.
-```
-# ---- interactive prompt ----------------------------------------------------
-print("Kaggle dataset access, choose a mode:")
-print("  [1] Download a single video")
-print("  [2] Download the whole dataset (~103GB)")
-print("  [3] Skip download (use an existing local video)")
-
-import threading
-
-choice = {"value": None}
-
-def get_choice():
-    choice["value"] = input("Enter 1, 2, or 3: ").strip()
-
-t = threading.Thread(target=get_choice, daemon=True)
-t.start()
-t.join(timeout=3)
-
-if choice["value"] is None:
-    choice["value"] = "3"
-    print("\n[Auto] 3 seconds elapsed : defaulting to Skip download.")
-
-choice = choice["value"]
-
-KAGGLE_MODE = {
-    "1": "single_file",
-    "2": "full",
-    "3": "skip"
-}.get(choice)
-
-if KAGGLE_MODE is None:
-    raise ValueError(f"Invalid choice: {choice!r}. Expected 1, 2, or 3.")
-# ----------------------------------------------------------------------------
-```
 """
+scenesolver_core.py — Scene Solver v3 : Coherent Reasoning System
+=================================================================
+Faithful extraction of the nine-stage SceneSolver pipeline from the
+authoritative Colab notebook (NewChecker.ipynb), packaged for the
+CrimeSceneSolver Hugging Face Space.
 
-# !pip install -qq reportlab stable-baselines3 ultralytics lap
-
-"""
-Scene Solver v3 : Coherent Reasoning System
-============================================
 Stage 0 : AutoEncoder (AE)           : PRIMARY anomaly signal (batched)
 Stage 1 : TimeSformer Binary         : Semantic gate (anomaly/normal)
 Stage 2 : TimeSformer Multi-class    : Fine-grained anomaly classification
@@ -59,8 +18,22 @@ Stage 8 : LLaVA (EXPERIMENTAL)       : Context explanation (multi-frame)
 
 LOGIC-NORMAL: When Stage 1 = Normal, status is overridden to SAFE,
 Stages 2 and 3 are skipped, and LLaVA runs with neutral prompts.
+
+Model assets (public repositories):
+  - samxkevin/scenesolver-assets   (AE / TFB / TFM / RL / YOLO / audio)
+  - samxkevin/SceneDescriptionModel (LLaVA VLM)
+
+DEPLOYMENT DIFFERENCES vs the notebook (compatibility only — no pipeline
+logic changes):
+  1. Checkpoint downloads are lazy (resolved on first use, cached) instead
+     of running inside the Config class body at import time.
+  2. An ffmpeg binary shim (imageio-ffmpeg) is placed on PATH when the
+     system ffmpeg is missing (HF Gradio Spaces have no apt access).
+  3. Stage 8 (LLaVA, ~14 GB) is released after each analysis so its
+     weights do not stay resident between runs on constrained hosts.
 """
 
+
 # ── Standard library ──────────────────────────────────────────────────────────
 import json
 import os
@@ -137,51 +110,113 @@ log = logging.getLogger("SceneSolver")
 # ══════════════════════════════════════════════════════════════════════════════
 # CONFIGURATION
 # ══════════════════════════════════════════════════════════════════════════════
-from huggingface_hub import hf_hub_download,snapshot_download
+
+# ── Deployment fix: guarantee an ffmpeg binary on PATH ────────────────────────
+# Gradio SDK Spaces cannot apt-install; imageio-ffmpeg ships a static binary.
+def _ensure_ffmpeg_on_path() -> None:
+    import tempfile as _tempfile
+    if shutil.which("ffmpeg"):
+        return
+    try:
+        import imageio_ffmpeg
+        exe = imageio_ffmpeg.get_ffmpeg_exe()
+        bindir = os.path.join(_tempfile.gettempdir(), "scenesolver_bin")
+        os.makedirs(bindir, exist_ok=True)
+        link = os.path.join(bindir, "ffmpeg")
+        if not os.path.exists(link):
+            try:
+                os.symlink(exe, link)
+            except OSError:
+                shutil.copy2(exe, link)
+        os.environ["PATH"] = bindir + os.pathsep + os.environ.get("PATH", "")
+        log.info(f"[FFmpeg] Using bundled binary: {exe}")
+    except Exception as e:
+        log.warning(f"[FFmpeg] Bundled binary unavailable: {e}")
+
+
+_ensure_ffmpeg_on_path()
+
+
+from huggingface_hub import hf_hub_download, snapshot_download
+from functools import lru_cache
+
 HF_ASSET_REPO = "samxkevin/scenesolver-assets"
-AUDIOPATHS = {
-    "CALLER_TUNE": hf_hub_download(
-        repo_id=HF_ASSET_REPO,
-        filename="audio/phone.mp3",
-    ),
-    "XTTS_SPEAKER_WAV": hf_hub_download(
-        repo_id=HF_ASSET_REPO,
-        filename="audio/DoctorVoice_22050.wav",
-    ),
-}
+LLAVA_REPO    = "samxkevin/SceneDescriptionModel"
+
+# Deployment fix: the notebook resolved every checkpoint with hf_hub_download /
+# snapshot_download inside the class body at import time. On a Space that
+# blocks startup on multi-GB downloads, so each asset is now resolved lazily
+# on first access and cached (identical paths, identical repositories).
+_asset_cache: Dict[str, str] = {}
+_asset_lock = threading.Lock()
+
+
+def _hf_asset(filename: str) -> str:
+    with _asset_lock:
+        if filename not in _asset_cache:
+            log.info(f"[HF] Fetching asset: {HF_ASSET_REPO}/{filename}")
+            _asset_cache[filename] = hf_hub_download(
+                repo_id=HF_ASSET_REPO, filename=filename)
+        return _asset_cache[filename]
+
+
+def _hf_llava_snapshot() -> str:
+    with _asset_lock:
+        if "__LLAVA__" not in _asset_cache:
+            log.info(f"[HF] Fetching VLM snapshot: {LLAVA_REPO} (~14 GB, first use only)")
+            _asset_cache["__LLAVA__"] = snapshot_download(repo_id=LLAVA_REPO)
+        return _asset_cache["__LLAVA__"]
+
+
+class _LazyAssets(dict):
+    """dict-compatible mapping whose values download on first access."""
+
+    def __init__(self, resolvers: Dict[str, Any]):
+        super().__init__()
+        self._resolvers = resolvers
+        self._resolved: Dict[str, str] = {}
+
+    def __getitem__(self, key: str) -> str:
+        if key not in self._resolved:
+            self._resolved[key] = self._resolvers[key]()
+        return self._resolved[key]
+
+    def get(self, key: str, default: Any = None) -> Any:
+        if key not in self._resolvers:
+            return default
+        try:
+            return self[key]
+        except Exception as e:
+            log.warning(f"[HF] Asset '{key}' unavailable: {e}")
+            return default
+
+    def __contains__(self, key: object) -> bool:
+        return key in self._resolvers
+
+
+AUDIOPATHS = _LazyAssets({
+    "CALLER_TUNE":      lambda: _hf_asset("audio/phone.mp3"),
+    "XTTS_SPEAKER_WAV": lambda: _hf_asset("audio/DoctorVoice_22050.wav"),
+})
+
 
 class Config:
+    # NOTE: `import spaces` (done first in app.py) keeps torch.cuda.is_available()
+    # True on ZeroGPU via CUDA emulation, so DEVICE resolves to cuda there and
+    # to cpu on CPU-only hosts — the notebook's device idiom is preserved.
     DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    PATHS = {
-        "AAE": hf_hub_download(
-            repo_id=HF_ASSET_REPO,
-            filename="AAE/best_model.pth",
-        ),
-        "AE": hf_hub_download(
-            repo_id=HF_ASSET_REPO,
-            filename="AE/best_autoencoder.pth",
-        ),
-        "RL_AGENT": hf_hub_download(
-            repo_id=HF_ASSET_REPO,
-            filename="RL/best_model.zip",
-        ),
-        "TS_BINARY": hf_hub_download(
-            repo_id=HF_ASSET_REPO,
-            filename="TFB/timesformer_best_binary.pth",
-        ),
-        "TS_MULTI": hf_hub_download(
-            repo_id=HF_ASSET_REPO,
-            filename="TFM/best_model.pth",
-        ),
-        "YOLO": hf_hub_download(
-            repo_id=HF_ASSET_REPO,
-            filename="YOLO/weapon_yolov8n.pt",
-        ),
-        "LLAVA":     snapshot_download(
-            repo_id="samxkevin/SceneDescriptionModel"
-        ),
-    }
-    print("[HF] All SceneSolver assets are available.")
+    PATHS = _LazyAssets({
+        "AAE":       lambda: _hf_asset("AAE/best_model.pth"),
+        "AE":        lambda: _hf_asset("AE/best_autoencoder.pth"),
+        "RL_AGENT":  lambda: _hf_asset("RL/best_model.zip"),
+        "TS_BINARY": lambda: _hf_asset("TFB/timesformer_best_binary.pth"),
+        "TS_MULTI":  lambda: _hf_asset("TFM/best_model.pth"),
+        "YOLO":      lambda: _hf_asset("YOLO/weapon_yolov8n.pt"),
+        "LLAVA":     _hf_llava_snapshot,
+    })
+    DRIVE_ROOT = "/tmp/scenesolver"
+    log.info("[HF] SceneSolver asset registry ready (lazy download on first use).")
+
     SAFE_THRESHOLD_FLOOR    = 0.030
     SAFE_THRESHOLD_LEGACY   = 0.055
     AUDIO_ANOMALY_THRESHOLD = 0.1
@@ -1381,6 +1416,16 @@ class Stage8LLaVA:
             self.model = None
         return self
 
+    def unload(self):
+        """Deployment fix: release the ~14 GB VLM after each analysis so its
+        weights do not remain resident between runs on Space hardware."""
+        if self.model is not None:
+            del self.model
+            self.model = None
+        self.processor = None
+        cleanup_memory()
+        log.info("[Stage8-LLaVA] Model released.")
+
     def _build_prompt(self, role: str, frame_number: int, yolo_summary: str,
                       motion_behaviors: List[str], audio_status: str,
                       audio_score: float, binary_result: str) -> str:
@@ -2875,6 +2920,13 @@ class SceneSolverPipeline:
         self.alert_system = AlertSystem(self.config)
         self.reporter     = EnhancedReportGenerator()
 
+    def release_stage8(self):
+        """Free Stage 8 (LLaVA) weights after an analysis (HF Spaces fix)."""
+        try:
+            self.stage8.unload()
+        except Exception:
+            pass
+
     def _make_output_dir(self, video_path: str, base: str) -> str:
         name = os.path.splitext(os.path.basename(video_path))[0]
         ts   = datetime.now().strftime("%Y%m%d_%H%M%S")
@@ -3013,6 +3065,8 @@ class SceneSolverPipeline:
 
         self._save_outputs(results, out)
         self._print_summary(results)
+        if enable_llava:
+            self.release_stage8()
         return results
 
     def _save_outputs(self, results: dict, out_dir: str):
@@ -3143,6 +3197,7 @@ class SceneSolverPipeline:
         return stream
 
 
+
 # ══════════════════════════════════════════════════════════════════════════════
 # READY MESSAGE
 # ══════════════════════════════════════════════════════════════════════════════
@@ -3159,886 +3214,11 @@ print("  pipeline = SceneSolverPipeline()")
 print("  results  = pipeline.analyze('/path/to/video.mp4')")
 print("=" * 65)
 
-#from google.colab import drive
-#drive.mount('/content/drive')
-
-# pipeline instantiated lazily by the Space app
-# results = pipeline.analyze('/content/drive/MyDrive/solver1/AnomalyDetectionDataset/Explosion/Explosion001_x264.mp4')
-# Without LLaVA (faster):
-# results = pipeline.analyze('/path/to/video.mp4', enable_llava=False)
-
-# !pip install -q gradio
 
 # ══════════════════════════════════════════════════════════════════════════════
-# GRADIO WEB APP  : CrimeSceneSolver  (refactored)
-#
-# Run AFTER Cell 2 (pipeline) and Cell 4 (gradio install).
-# Agent Escalation Layer below this cell for the full experience.
+# SHARED PIPELINE SINGLETON (models are cached across analyses)
 # ══════════════════════════════════════════════════════════════════════════════
+@lru_cache(maxsize=1)
+def get_pipeline() -> SceneSolverPipeline:
+    return SceneSolverPipeline()
 
-import gradio as gr
-import numpy as np
-import json, os, traceback, smtplib, ssl, re, base64
-from email.mime.text import MIMEText
-from email.mime.multipart import MIMEMultipart
-from PIL import Image
-
-# ─────────────────────────────────────────────────────────────────────────────
-# SHARED CONSTANTS  (used by both the web app and the agent layer)
-# ─────────────────────────────────────────────────────────────────────────────
-ADMIN_EMAIL  = "garoldwerner2512@gmail.com"
-SMTP_SERVER  = "smtp.gmail.com"
-SMTP_PORT    = 465
-SENDER_EMAIL = os.environ.get("SENDER_EMAIL", "")
-SENDER_PASS  = os.environ.get("SENDER_PASS",  "")
-
-# Design tokens — pitch black / white / orange-saffron
-C_BG     = "#000000"
-C_ORANGE = "#FF8C00"
-C_RED    = "#FF4500"
-C_GREEN  = "#4CAF50"
-
-# Shared HTML blocks
-NAV_HTML = """
-<div style="background:#000;border-bottom:1px solid #1a1a1a;
-            padding:16px 36px;display:flex;justify-content:space-between;align-items:center;">
-  <span style="font-family:'Roboto Mono',monospace;font-size:.92rem;
-               letter-spacing:.15em;color:#fff;text-transform:uppercase;
-               font-weight:700;-webkit-font-smoothing:antialiased;">
-    CrimeSceneSolver
-  </span>
-  <span style="font-family:'Roboto Mono',monospace;font-size:.6rem;
-               letter-spacing:.1em;color:#555;text-transform:uppercase;">
-    Scene Solver v3
-  </span>
-</div>
-"""
-
-FOOTER_HTML = """
-<div style="border-top:1px solid #0d0d0d;margin-top:16px;">
-  <div style="background:#0d0d0d;border-bottom:1px solid #1a1a1a;padding:14px 32px;">
-    <div style="max-width:860px;margin:0 auto;display:flex;align-items:flex-start;gap:12px;">
-      <svg style="flex-shrink:0;margin-top:2px;" width="14" height="14" viewBox="0 0 24 24"
-           fill="none" stroke="#FF8C00" stroke-width="2">
-        <circle cx="12" cy="12" r="10"/>
-        <line x1="12" y1="8" x2="12" y2="12"/>
-        <line x1="12" y1="16" x2="12.01" y2="16"/>
-      </svg>
-      <p style="font-family:'Roboto Mono',monospace;font-size:.62rem;
-                letter-spacing:.04em;color:#888;line-height:1.7;margin:0;font-weight:500;">
-        <span style="color:#FF8C00;font-weight:700;">DISCLAIMER &mdash; </span>
-        This report is produced by machine learning models and should be treated as an
-        investigative aid, not a definitive conclusion. Human expert review is strongly
-        recommended before any operational action is taken.
-      </p>
-    </div>
-  </div>
-  <div style="text-align:center;padding:18px;font-family:'Roboto Mono',monospace;
-              font-size:.6rem;letter-spacing:.1em;color:#444;text-transform:uppercase;">
-    CrimeSceneSolver &nbsp;&middot;&nbsp; Scene Solver v3 &nbsp;&middot;&nbsp;
-    Confidential &nbsp;&middot;&nbsp; ML outputs require expert review
-  </div>
-</div>
-"""
-
-
-def sec_label(text: str) -> str:
-    """Shared section-label helper used across the web app and agent layer."""
-    return (f"<div style='font-family:\"Roboto Mono\",monospace;font-size:.62rem;"
-            f"letter-spacing:.14em;text-transform:uppercase;color:#ddd;"
-            f"margin-bottom:10px;font-weight:800;-webkit-font-smoothing:antialiased;"
-            f"border-left:3px solid #FF8C00;padding-left:10px;'>{text}</div>")
-
-
-def risk_color(r: float) -> str:
-    if r >= 0.70: return C_RED
-    if r >= 0.40: return C_ORANGE
-    return C_GREEN
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# EMAIL
-# ─────────────────────────────────────────────────────────────────────────────
-def send_contact_email(name: str, email: str, phone: str, message: str) -> bool:
-    if not SENDER_EMAIL or not SENDER_PASS:
-        with open("/tmp/css_contact.log", "a") as f:
-            f.write(f"[CONTACT] {name} | {email} | {phone}\n")
-        return False
-    try:
-        msg = MIMEMultipart("alternative")
-        msg["Subject"] = f"[CrimeSceneSolver] Contact: {name}"
-        msg["From"]    = SENDER_EMAIL
-        msg["To"]      = ADMIN_EMAIL
-        msg.attach(MIMEText(f"Name:{name}\nEmail:{email}\nPhone:{phone}\nMessage:\n{message}\n", "plain"))
-        ctx = ssl.create_default_context()
-        with smtplib.SMTP_SSL(SMTP_SERVER, SMTP_PORT, context=ctx) as srv:
-            srv.login(SENDER_EMAIL, SENDER_PASS)
-            srv.sendmail(SENDER_EMAIL, ADMIN_EMAIL, msg.as_string())
-        return True
-    except Exception:
-        return False
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# PIPELINE WRAPPER
-# ─────────────────────────────────────────────────────────────────────────────
-def run_pipeline(video_path: str) -> dict:
-    # pipeline instantiated lazily by the Space app
-    results  = pipeline.analyze(
-        video_path=video_path, base_output_dir="/tmp/css_webapp",
-        enable_llava=True, enable_rl=True, enable_tracking=True,
-    )
-    s0 = results.get("stage0") or {}
-    s1 = results.get("stage1") or {}
-    s2 = results.get("stage2") or {}
-    s3 = results.get("stage3") or {}
-    s4 = results.get("stage4") or {}
-    s5 = results.get("stage5") or {}
-    s6 = results.get("stage6") or {}
-    s7 = results.get("stage7") or {}
-    s8 = results.get("stage8") or {}
-    fps = float(s0.get("fps", 25.0))
-
-    def ts(fi): return frame_to_timestamp(int(fi), fps)
-
-    risk_score = float(s7.get("final_risk", s6.get("fused_risk", 0.0)))
-    rl_status  = s7.get("status", "")
-    _rl_map    = {"CRITICAL":"CRITICAL","HIGH_RISK":"HIGH","MODERATE":"MODERATE","LOW_RISK":"LOW"}
-    risk_label = (_rl_map.get(rl_status, "") or
-                  results.get("risk_label_final") or
-                  s6.get("risk_level", "LOW"))
-    status    = results.get("status", "UNKNOWN")
-    narrative = (s8 or {}).get("unified_narrative") or (s8 or {}).get("description") or ""
-
-    report = {
-        "overall_status":   status,
-        "final_risk_score": f"{risk_score:.1%}",
-        "risk_level":       risk_label,
-        "dynamic_threshold": f"{results.get('threshold',0):.6f}",
-        "ae_max_error":     f"{s0.get('max_error',0):.6f}",
-        "ae_avg_error":     f"{s0.get('avg_error',0):.6f}",
-        "peak_timestamp":   ts(s0.get("peak_frame_idx",0)),
-        "anomaly_window":   f"{ts(s0.get('anomaly_start_frame',0))} \u2192 {ts(s0.get('anomaly_end_frame',0))}",
-        "stage1_tf_binary": f"{s1.get('classification','?')}  ({s1.get('confidence',0):.1%})",
-        "stage2_anomaly_type": f"{s2.get('class','N/A')}  ({s2.get('confidence',0):.1%})",
-        "stage2_top3":      s2.get("top3", []),
-        "stage2_all_probs": s2.get("all_probabilities", {}),
-        "stage3_weapon":    ("YES — " + str(s3.get("dominant_class","?"))
-                             if s3.get("weapon_flag") else "None detected"),
-        "stage4_audio":     f"{s4.get('status','N/A')}  (score: {s4.get('audio_score',0):.4f})",
-        "stage5_motion":    f"{s5.get('motion_intensity',0):.3f}  |  {', '.join(s5.get('behaviors',[]) or ['STATIC'])}",
-        "stage6_fusion_risk": f"{s6.get('fused_risk',0):.1%}  [{s6.get('risk_level','?')}]",
-        "stage6_reasoning": s6.get("reasoning_trace", []),
-        "stage7_rl_final":  f"{s7.get('final_risk',0):.1%}  [{rl_status}]",
-        "stage7_interpretation": s7.get("interpretation",""),
-        "llava_narrative":  narrative,
-        # raw values for charts
-        "_stage1_conf":     float(s1.get("confidence", 0)),
-        "_stage2_conf":     float(s2.get("confidence", 0)),
-        "_stage2_top3_raw": s2.get("top3", []),
-        "_stage2_all_probs": s2.get("all_probabilities", {}),
-        "_risk_score":      risk_score,
-        "_fused_risk":      float(s6.get("fused_risk", 0)),
-        "_audio_score":     float(s4.get("audio_score", 0)),
-        "_motion":          float(s5.get("motion_intensity", 0)),
-    }
-
-    frames_l  = s0.get("frames", [])
-    scores_l  = s0.get("scores", [])
-    samp_idx  = s0.get("sampled_indices", [])
-    top_frame = None
-    if frames_l and scores_l:
-        psi       = int(s0.get("peak_score_idx", 0))
-        pfi       = min(samp_idx[psi] if psi < len(samp_idx) else 0, len(frames_l) - 1)
-        top_frame = frames_l[pfi]
-
-    def _serial(o):
-        if isinstance(o, np.ndarray): return o.tolist()
-        if isinstance(o, np.integer): return int(o)
-        if isinstance(o, np.floating): return float(o)
-        return str(o)
-
-    jd = {
-        "status": status, "risk_score": round(risk_score, 4), "risk_level": risk_label,
-        "threshold": round(float(results.get("threshold", 0)), 6),
-        "stage0": {"max_error": round(float(s0.get("max_error",0)),6),
-                   "avg_error": round(float(s0.get("avg_error",0)),6),
-                   "peak_timestamp": ts(s0.get("peak_frame_idx",0)),
-                   "anomaly_start": ts(s0.get("anomaly_start_frame",0)),
-                   "anomaly_end":   ts(s0.get("anomaly_end_frame",0))},
-        "stage1": {"classification": s1.get("classification","?"),
-                   "confidence": round(float(s1.get("confidence",0)),4)},
-        "stage2": {"class": s2.get("class","N/A"),
-                   "confidence": round(float(s2.get("confidence",0)),4),
-                   "top3": s2.get("top3",[])},
-        "stage3": {"weapon_flag": bool(s3.get("weapon_flag",False)),
-                   "dominant_class": s3.get("dominant_class","None"),
-                   "weapon_count": int(s3.get("weapon_count",0))},
-        "stage4": {"status": s4.get("status","N/A"),
-                   "audio_score": round(float(s4.get("audio_score",0)),4)},
-        "stage5": {"motion_intensity": round(float(s5.get("motion_intensity",0)),4),
-                   "behaviors": s5.get("behaviors",[])},
-        "stage6": {"fused_risk": round(float(s6.get("fused_risk",0)),4),
-                   "risk_level": s6.get("risk_level","?"),
-                   "reasoning": s6.get("reasoning_trace",[])},
-        "stage7": {"final_risk": round(risk_score,4), "status": rl_status,
-                   "interpretation": s7.get("interpretation","")},
-        "llava_narrative": narrative,
-    }
-
-    pdf = (results.get("pdf_path_local") or results.get("pdf_path")
-           or "/tmp/forensic_report_v3.pdf")
-    if not os.path.exists(pdf): pdf = None
-
-    graph_path = results.get("graph_path", "/tmp/deviation_graph_v3.png")
-    if not os.path.exists(graph_path): graph_path = None
-
-    return {"risk_score": risk_score, "risk_label": risk_label, "status": status,
-            "report": report, "top_frame": top_frame, "json_data": jd,
-            "pdf_path": pdf, "graph_path": graph_path}
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# HTML HELPERS
-# ─────────────────────────────────────────────────────────────────────────────
-def img_to_data_uri(path: str) -> str:
-    """Encode an image file as a base64 data-URI for inline HTML embedding."""
-    if not path or not os.path.exists(path):
-        return ""
-    ext  = os.path.splitext(path)[1].lower().lstrip(".")
-    mime = {"png":"image/png","jpg":"image/jpeg","jpeg":"image/jpeg"}.get(ext,"image/png")
-    with open(path, "rb") as f:
-        encoded = base64.b64encode(f.read()).decode("utf-8")
-    return f"data:{mime};base64,{encoded}"
-
-
-def spinner_html() -> str:
-    return """
-<div style="text-align:center;padding:18px 0;">
-  <div style="display:inline-block;width:40px;height:40px;
-    border:3px solid rgba(255,255,255,0.07);border-top-color:#FF8C00;
-    border-right-color:rgba(255,140,0,0.28);border-radius:50%;
-    animation:arc-spin .75s cubic-bezier(0.6,0.2,0.4,0.8) infinite;"></div>
-  <div style="font-family:'Roboto Mono',monospace;font-size:.64rem;
-    letter-spacing:.12em;color:#FF8C00;margin-top:11px;text-transform:uppercase;font-weight:600;">
-    Processing&hellip; this may take several minutes
-  </div>
-</div>"""
-
-
-def no_result_html() -> str:
-    return ("<div style='text-align:center;padding:60px 20px;color:#333;"
-            "font-family:Charter,serif;font-size:1rem;letter-spacing:.04em;'>"
-            "Upload footage and press ANALYSE to begin.</div>")
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# REPORT HTML BUILDER
-# ─────────────────────────────────────────────────────────────────────────────
-def build_report_html(report: dict, risk_score: float, risk_label: str,
-                      status: str, graph_path: str = None) -> str:
-    col = risk_color(risk_score)
-
-    def numfmt(v: str) -> str:
-        return re.sub(r'([\d]+\.?[\d]*%?)',
-                      r"<span style='font-family:\"Times New Roman\",Times,serif;'>\1</span>", v)
-
-    def row(label: str, value: str, mono: bool = False) -> str:
-        vd    = numfmt(str(value))
-        vfont = "font-family:'Roboto Mono',monospace;font-size:.8rem;" if mono else "font-size:.82rem;"
-        return (f"<tr>"
-                f"<td style='padding:9px 14px;color:#aaa;font-size:.78rem;"
-                f"border-bottom:1px solid #222;white-space:nowrap;width:38%;'>{label}</td>"
-                f"<td style='padding:9px 14px;color:#eee;{vfont}border-bottom:1px solid #222;'>{vd}</td>"
-                f"</tr>")
-
-    def section_hdr(title: str) -> str:
-        return (f"<div style='font-size:.65rem;letter-spacing:.16em;text-transform:uppercase;"
-                f"color:#FF8C00;padding:14px 0 6px 2px;border-bottom:2px solid #FF8C00;"
-                f"margin-bottom:0;font-family:\"Roboto Mono\",monospace;font-weight:800;'>{title}</div>")
-
-    def section(title: str, rows_html: str, extra: str = "") -> str:
-        return (f"<div style='margin-bottom:26px;'>"
-                + section_hdr(title)
-                + f"<table style='width:100%;border-collapse:collapse;'>{rows_html}</table>"
-                + extra + "</div>")
-
-    # Risk badge
-    is_crit    = (status == "CRITICAL")
-    badge_text = f"{'ANOMALY DETECTED' if is_crit else 'NO ANOMALY DETECTED'} ({status})"
-    badge = (f"<div style='text-align:center;padding:28px 20px;margin-bottom:28px;"
-             f"background:{col};border-radius:4px;'>"
-             f"<div style='font-size:3.8rem;font-weight:700;color:#fff;"
-             f"font-family:\"Times New Roman\",Times,serif;line-height:1;'>"
-             f"{risk_score:.1%}</div>"
-             f"<div style='color:rgba(255,255,255,.85);font-size:.75rem;margin-top:10px;"
-             f"letter-spacing:.14em;text-transform:uppercase;font-family:\"Roboto Mono\",monospace;"
-             f"font-weight:600;'>{badge_text} &nbsp;|&nbsp; {risk_label} RISK</div>"
-             f"</div>")
-
-    # Raw chart values
-    stage1_conf = float(report.get("_stage1_conf", 0))
-    stage2_conf = float(report.get("_stage2_conf", 0))
-    top3_raw    = report.get("_stage2_top3_raw", []) or []
-    all_probs   = report.get("_stage2_all_probs", {}) or {}
-    fused_risk  = float(report.get("_fused_risk", risk_score))
-    audio_score = float(report.get("_audio_score", 0))
-    motion      = float(report.get("_motion", 0))
-    PIE_COLORS  = ["#c0392b","#2471a3","#d35400","#6c3483","#117a65","#1a5276","#7d6608"]
-
-    # ── Donut pie (SVG) ──────────────────────────────────────────────────────
-    def donut_pie_svg(probs: dict) -> str:
-        if not probs: return ""
-        import math
-        items  = sorted(probs.items(), key=lambda x: x[1], reverse=True)
-        labels = [i[0] for i in items]
-        sizes  = [max(float(i[1]), 1e-6) for i in items]
-        total  = sum(sizes)
-        cx = cy = R = 90; R = 70; r = 36
-        wedges = ""
-        start  = -90
-        for idx, (label, val) in enumerate(zip(labels, sizes)):
-            sweep = (val / total) * 360
-            end   = start + sweep
-            x1o, y1o = cx + R * math.cos(math.radians(start)), cy + R * math.sin(math.radians(start))
-            x2o, y2o = cx + R * math.cos(math.radians(end)),   cy + R * math.sin(math.radians(end))
-            x1i, y1i = cx + r * math.cos(math.radians(end)),   cy + r * math.sin(math.radians(end))
-            x2i, y2i = cx + r * math.cos(math.radians(start)), cy + r * math.sin(math.radians(start))
-            lg   = 1 if sweep > 180 else 0
-            fill = PIE_COLORS[idx % len(PIE_COLORS)]
-            mid_a = start + sweep / 2
-            lx = cx + (R + r) / 2 * math.cos(math.radians(mid_a))
-            ly = cy + (R + r) / 2 * math.sin(math.radians(mid_a))
-            num_lbl = (f"<text x='{lx:.1f}' y='{ly:.1f}' text-anchor='middle' "
-                       f"dominant-baseline='middle' fill='#fff' font-size='9' font-weight='bold'>{idx+1}</text>"
-                       if sweep > 15 else "")
-            wedges += (f"<path d='M {x1o:.2f} {y1o:.2f} A {R} {R} 0 {lg} 1 {x2o:.2f} {y2o:.2f} "
-                       f"L {x1i:.2f} {y1i:.2f} A {r} {r} 0 {lg} 0 {x2i:.2f} {y2i:.2f} Z' "
-                       f"fill='{fill}' stroke='#000' stroke-width='1'/>{num_lbl}")
-            start = end
-        legend = "".join(
-            f"<div style='display:flex;align-items:center;gap:7px;margin-bottom:5px;'>"
-            f"<div style='width:11px;height:11px;min-width:11px;background:{PIE_COLORS[i%len(PIE_COLORS)]};border-radius:2px;'></div>"
-            f"<span style='font-family:\"Roboto Mono\",monospace;font-size:.62rem;color:#ccc;'>"
-            f"{i+1}. {cls}&nbsp;<span style='font-family:\"Times New Roman\",Times,serif;color:#fff;font-weight:700;'>"
-            f"({v/total*100:.1f}%)</span></span></div>"
-            for i, (cls, v) in enumerate(zip(labels, sizes))
-        )
-        svg = (f"<svg width='180' height='180' viewBox='0 0 180 180'>{wedges}"
-               f"<circle cx='{cx}' cy='{cy}' r='{r-2}' fill='#050505'/>"
-               f"<text x='{cx}' y='{cy}' text-anchor='middle' dominant-baseline='middle' "
-               f"fill='#666' font-size='8'>ANOMALY</text>"
-               f"<text x='{cx}' y='{cy+11}' text-anchor='middle' dominant-baseline='middle' "
-               f"fill='#666' font-size='8'>CLASSES</text></svg>")
-        return (f"<div style='display:flex;gap:16px;align-items:flex-start;padding:16px;"
-                f"background:#050505;border:1px solid #1e1e1e;border-radius:3px;margin-top:6px;'>"
-                f"<div style='flex-shrink:0;'>{svg}</div><div style='flex:1;'>{legend}</div></div>")
-
-    # ── Horizontal bar chart ──────────────────────────────────────────────────
-    def conf_bar_chart(top3: list) -> str:
-        if not top3: return ""
-        BAR_COLORS = ["#c0392b","#d35400","#5d6d7e"]
-        rows = "".join(
-            f"<div style='margin-bottom:10px;'>"
-            f"<div style='display:flex;justify-content:space-between;align-items:center;margin-bottom:4px;'>"
-            f"<span style='font-family:\"Roboto Mono\",monospace;font-size:.68rem;color:#ccc;'>{cls}</span>"
-            f"<span style='font-family:\"Times New Roman\",Times,serif;color:#fff;font-weight:700;font-size:.82rem;margin-left:8px;'>{prob*100:.1f}%</span>"
-            f"</div><div style='background:#1a1a1a;border-radius:2px;height:10px;overflow:hidden;'>"
-            f"<div style='background:{BAR_COLORS[idx]};height:10px;width:{prob*100:.1f}%;border-radius:2px;'></div>"
-            f"</div></div>"
-            for idx, (cls, prob) in enumerate(top3[:3])
-        )
-        return (f"<div style='padding:14px 16px;background:#050505;border:1px solid #1e1e1e;"
-                f"border-radius:3px;margin-top:6px;'>"
-                f"<div style='font-family:\"Roboto Mono\",monospace;font-size:.6rem;"
-                f"letter-spacing:.14em;color:#FF8C00;text-transform:uppercase;margin-bottom:12px;'>"
-                f"Top-3 Anomaly Predictions</div>{rows}</div>")
-
-    # ── Signal gauge bars ─────────────────────────────────────────────────────
-    def risk_gauge(val: float, label: str, color: str) -> str:
-        pct = min(100, int(val * 100))
-        return (f"<div style='margin-bottom:10px;'>"
-                f"<div style='display:flex;justify-content:space-between;align-items:center;margin-bottom:3px;'>"
-                f"<span style='font-family:\"Roboto Mono\",monospace;font-size:.65rem;color:#aaa;'>{label}</span>"
-                f"<span style='font-family:\"Times New Roman\",Times,serif;color:#fff;font-weight:700;font-size:.8rem;'>{val:.1%}</span>"
-                f"</div><div style='background:#1a1a1a;border-radius:2px;height:8px;overflow:hidden;'>"
-                f"<div style='background:{color};height:8px;width:{pct}%;border-radius:2px;'></div>"
-                f"</div></div>")
-
-    # ── Small donut pies for Stage 1 / Stage 2 ───────────────────────────────
-    def small_donut(pct: float, color: str, label: str, size: int = 84) -> str:
-        import math
-        cx = cy = size // 2
-        R  = cx - 8
-        r  = R - 12
-        sweep = pct * 360
-        end_a = -90 + sweep
-        x1,  y1  = cx + R*math.cos(math.radians(-90)),    cy + R*math.sin(math.radians(-90))
-        x2,  y2  = cx + R*math.cos(math.radians(end_a)),  cy + R*math.sin(math.radians(end_a))
-        xi1, yi1 = cx + r*math.cos(math.radians(end_a)),  cy + r*math.sin(math.radians(end_a))
-        xi2, yi2 = cx + r*math.cos(math.radians(-90)),    cy + r*math.sin(math.radians(-90))
-        lg = 1 if sweep > 180 else 0
-        if pct >= 0.999:
-            arc = (f"<circle cx='{cx}' cy='{cy}' r='{R}' fill='none' stroke='{color}' stroke-width='12'/>"
-                   f"<circle cx='{cx}' cy='{cy}' r='{r}' fill='#050505'/>")
-        else:
-            arc = (f"<circle cx='{cx}' cy='{cy}' r='{R}' fill='none' stroke='#1a1a1a' stroke-width='12'/>"
-                   f"<path d='M {x1:.2f} {y1:.2f} A {R} {R} 0 {lg} 1 {x2:.2f} {y2:.2f} "
-                   f"L {xi1:.2f} {yi1:.2f} A {r} {r} 0 {lg} 0 {xi2:.2f} {yi2:.2f} Z' fill='{color}'/>"
-                   f"<circle cx='{cx}' cy='{cy}' r='{r}' fill='#050505'/>")
-        return (f"<div style='text-align:center;'>"
-                f"<svg width='{size}' height='{size}' viewBox='0 0 {size} {size}'>"
-                f"{arc}"
-                f"<text x='{cx}' y='{cy}' text-anchor='middle' dominant-baseline='middle' "
-                f"fill='#fff' font-family='Times New Roman,Times,serif' font-size='11' font-weight='700'>"
-                f"{pct:.0%}</text></svg>"
-                f"<div style='font-family:\"Roboto Mono\",monospace;font-size:.55rem;color:#666;margin-top:2px;'>"
-                f"{label}</div></div>")
-
-    charts_cls = (
-        f"<div style='display:flex;gap:20px;padding:14px 16px;background:#050505;"
-        f"border:1px solid #1e1e1e;border-radius:3px;margin-top:6px;align-items:flex-start;flex-wrap:wrap;'>"
-        + small_donut(stage1_conf, '#FF8C00', 'TF BINARY · Stage 1')
-        + small_donut(stage2_conf, '#c0392b', 'ANOMALY · Stage 2')
-        + f"<div style='flex:1;min-width:200px;'>{conf_bar_chart(top3_raw)}</div>"
-        + (f"<div style='width:100%;margin-top:10px;'>{donut_pie_svg(all_probs)}</div>" if all_probs else "")
-        + "</div>"
-    )
-    charts_sensing = (
-        f"<div style='padding:14px 16px;background:#050505;border:1px solid #1e1e1e;"
-        f"border-radius:3px;margin-top:6px;'>"
-        f"<div style='font-family:\"Roboto Mono\",monospace;font-size:.6rem;letter-spacing:.14em;"
-        f"color:#FF8C00;text-transform:uppercase;margin-bottom:12px;'>Signal Intensities</div>"
-        + risk_gauge(fused_risk,        "Fusion Risk (Stage 6)", risk_color(fused_risk))
-        + risk_gauge(audio_score,       "Audio Score (Stage 4)", "#2471a3")
-        + risk_gauge(min(motion, 1.0),  "Motion Intensity (Stage 5)", "#117a65")
-        + "</div>"
-    )
-
-    # ── Table rows ────────────────────────────────────────────────────────────
-    ae_rows = (row("Overall Status",    report.get("overall_status",""))
-             + row("Dynamic Threshold", report.get("dynamic_threshold",""), mono=True)
-             + row("AE Max Error",      report.get("ae_max_error",""), mono=True)
-             + row("AE Avg Error",      report.get("ae_avg_error",""), mono=True)
-             + row("Peak Timestamp",    report.get("peak_timestamp",""), mono=True)
-             + row("Anomaly Window",    report.get("anomaly_window",""), mono=True))
-
-    cl_rows = (row("TF Binary (Stage 1)",   report.get("stage1_tf_binary",""))
-             + row("Anomaly Type (Stage 2)", report.get("stage2_anomaly_type","")))
-    for i, (cls, prob) in enumerate(report.get("stage2_top3",[]), 1):
-        cl_rows += row(f"  Top-{i}", f"{cls}  ({prob:.2%})", mono=True)
-
-    det_rows = (row("Weapon Detection (Stage 3)", report.get("stage3_weapon",""))
-              + row("Audio Status (Stage 4)",      report.get("stage4_audio",""))
-              + row("Motion (Stage 5)",            report.get("stage5_motion","")))
-
-    fus_rows = row("Fusion Risk (Stage 6)", report.get("stage6_fusion_risk",""))
-    for i, line in enumerate(report.get("stage6_reasoning",[]), 1):
-        fus_rows += row(f"  Step {i}", line)
-
-    rl_rows = (row("RL3033 Final Risk (Stage 7)", report.get("stage7_rl_final",""), mono=True)
-             + row("Interpretation",              report.get("stage7_interpretation","")))
-
-    narr = report.get("llava_narrative","")
-    narr_html = (
-        f"<div style='margin-bottom:26px;'>"
-        + section_hdr("AI Visual Narrative (LLaVA — Stage 8)")
-        + f"<div style='color:#ccc;font-size:.85rem;line-height:1.78;padding:16px 18px;"
-        f"background:#050505;border:1px solid #1e1e1e;border-radius:3px;margin-top:6px;'>{narr}</div>"
-        + "</div>"
-    ) if narr else ""
-
-    ae_graph_html = ""
-    if graph_path:
-        data_uri = img_to_data_uri(graph_path)
-        if data_uri:
-            ae_graph_html = (
-                f"<div style='margin-bottom:26px;'>"
-                + section_hdr("AE Deviation Graph — Temporal Reconstruction Error")
-                + f"<div style='padding:14px;background:#050505;border:1px solid #1e1e1e;"
-                f"border-radius:3px;margin-top:6px;'>"
-                f"<img src='{data_uri}' alt='AE Deviation Graph' "
-                f"style='width:100%;border-radius:2px;display:block;'/>"
-                f"<div style='font-family:\"Roboto Mono\",monospace;font-size:.58rem;color:#555;"
-                f"margin-top:8px;text-align:center;'>"
-                f"Red = reconstruction error &nbsp;&middot;&nbsp; Green dashed = dynamic threshold "
-                f"&nbsp;&middot;&nbsp; Star = peak anomaly frame"
-                f"</div></div></div>"
-            )
-
-    return (
-        f"<div style='background:#0a0a0a;padding:28px 26px;border:1px solid #1e1e1e;border-radius:4px;'>"
-        + badge + ae_graph_html
-        + section("Autoencoder — Stage 0",            ae_rows)
-        + section("Classification — Stages 1 & 2",    cl_rows, charts_cls)
-        + section("Detection & Sensing — Stages 3-5", det_rows, charts_sensing)
-        + section("Fusion Layer — Stage 6",            fus_rows)
-        + section("RL3033 Temporal Reasoning — Stage 7", rl_rows)
-        + narr_html
-        + "</div>"
-    )
-
-
-def process_video(video):
-    if video is None:
-        return (no_result_html(), None, None, None)
-    try:
-        out         = run_pipeline(video)
-        r           = out["risk_score"]
-        report_html = build_report_html(out["report"], r, out["risk_label"],
-                                         out["status"], graph_path=out.get("graph_path"))
-        fi = Image.fromarray(out["top_frame"]) if out["top_frame"] is not None else None
-        return (report_html, fi, json.dumps(out["json_data"], indent=2, default=str), out["pdf_path"])
-    except Exception:
-        tb = traceback.format_exc()
-        return (f"<div style='color:#FF4500;font-family:\"Roboto Mono\",monospace;"
-                f"font-size:.8rem;padding:20px;background:#050505;'>Analysis failed:\n\n{tb}</div>",
-                None, None, None)
-
-
-def handle_contact(name: str, email: str, phone: str, msg: str) -> str:
-    """Shared contact form handler — used in both the base app and agent layer tabs."""
-    if not name.strip() or not email.strip() or not msg.strip():
-        return ("<div style='font-family:\"Roboto Mono\",monospace;font-size:.72rem;"
-                "color:#FF4500;padding:10px 0;font-weight:600;'>"
-                "Name, email, and message are required.</div>")
-    if not re.match(r"^[^@]+@[^@]+\.[^@]+$", email.strip()):
-        return ("<div style='font-family:\"Roboto Mono\",monospace;font-size:.72rem;"
-                "color:#FF4500;padding:10px 0;font-weight:600;'>"
-                "Please enter a valid email address.</div>")
-    send_contact_email(name.strip(), email.strip(), phone.strip(), msg.strip())
-    return ("<div style='font-family:\"Roboto Mono\",monospace;font-size:.72rem;"
-            "color:#4CAF50;padding:10px 0;letter-spacing:.06em;font-weight:600;'>"
-            "&#10003; Message received. We will respond at the provided email.</div>")
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# CSS — pitch black / white / orange palette
-# ─────────────────────────────────────────────────────────────────────────────
-APP_CSS = """
-@import url('https://fonts.googleapis.com/css2?family=Roboto+Mono:wght@300;400;500;600;700&display=swap');
-
-:root {
-  --bg:     #000000;
-  --surface:#080808;
-  --border: #1e1e1e;
-  --white:  #FFFFFF;
-  --dim:    #888888;
-  --orange: #FF8C00;
-  --red:    #FF4500;
-  --mono:   'Roboto Mono', monospace;
-  --serif:  Charter, 'Bitstream Charter', Georgia, serif;
-}
-
-*, *::before, *::after { box-sizing:border-box; margin:0; padding:0; }
-html, body, .gradio-container, .gradio-container > * {
-  background:var(--bg) !important; color:var(--white) !important;
-  font-family:var(--serif) !important;
-}
-.gradio-container {
-  zoom: 1.3 !important;
-}
-* { -webkit-font-smoothing:antialiased; }
-
-::-webkit-scrollbar { width:4px; height:4px; }
-::-webkit-scrollbar-track { background:var(--bg); }
-::-webkit-scrollbar-thumb { background:#333; border-radius:2px; }
-
-@keyframes arc-spin {
-  from { transform:rotate(0deg); } to { transform:rotate(360deg); }
-}
-
-/* TABS */
-.tab-nav {
-  background:var(--bg) !important; border-bottom:1px solid var(--border) !important;
-  padding:0 24px !important; display:flex !important; gap:0 !important;
-  align-items:center !important; justify-content:flex-start !important;
-}
-.tab-nav button {
-  font-family:var(--mono) !important; font-size:.72rem !important;
-  letter-spacing:.13em !important; color:#999 !important;
-  background:transparent !important; border:none !important;
-  border-bottom:2px solid transparent !important; padding:14px 0 !important;
-  min-width:160px !important; width:160px !important; text-align:center !important;
-  text-transform:uppercase !important; transition:color .18s, border-color .18s !important;
-  font-weight:600 !important; white-space:nowrap !important; flex:0 0 160px !important;
-}
-.tab-nav button:hover  { color:#ddd !important; }
-.tab-nav button.selected {
-  color:#FFFFFF !important; border-bottom-color:var(--orange) !important; font-weight:700 !important;
-}
-
-/* GRADIO INTERNALS */
-.gradio-container .block, .gradio-container .panel,
-.gradio-container .form, .gradio-container fieldset,
-.gradio-container .wrap, .gradio-container .container,
-.gradio-container > div { background:var(--bg) !important; border:none !important; }
-
-/* INPUTS */
-input, textarea, select {
-  background:#0d0d0d !important; border:1px solid #2a2a2a !important;
-  border-radius:3px !important; color:#f0f0f0 !important;
-  font-family:var(--serif) !important; font-size:.88rem !important;
-  padding:12px 14px !important; transition:border-color .18s !important; width:100% !important;
-}
-input:focus, textarea:focus {
-  border-color:var(--orange) !important; outline:none !important;
-  box-shadow:0 0 0 1px var(--orange) !important;
-}
-label, .label-wrap span {
-  font-family:var(--mono) !important; font-size:.65rem !important;
-  letter-spacing:.1em !important; text-transform:uppercase !important;
-  color:#ccc !important; font-weight:600 !important;
-}
-
-/* BUTTONS */
-.gr-button {
-  background:var(--white) !important; color:var(--bg) !important;
-  border-radius:3px !important; font-family:var(--mono) !important;
-  font-size:.72rem !important; letter-spacing:.12em !important; font-weight:700 !important;
-}
-.gr-button:hover { background:var(--orange) !important; }
-#ana-btn, #ana-btn button {
-  width:auto !important; max-width:260px !important;
-  min-height:44px !important; height:44px !important; padding:0 30px !important;
-}
-
-/* VIDEO UPLOAD */
-#vid-up { min-height:150px !important; max-height:200px !important; }
-#vid-up .upload-container, #vid-up .wrap {
-  background:#060606 !important; border:1px dashed #2a2a2a !important;
-  border-radius:3px !important; min-height:150px !important; max-height:200px !important;
-}
-#vid-up .wrap:hover { border-color:var(--orange) !important; }
-
-/* OUTPUT AREAS */
-#report-html { background:var(--bg) !important; }
-#kf-img, #kf-img img { border-radius:3px !important; border:1px solid var(--border) !important; width:100% !important; }
-#json-out, #json-out textarea {
-  background:#060606 !important; color:#bbb !important;
-  font-family:var(--mono) !important; font-size:.75rem !important;
-  line-height:1.7 !important; border:1px solid var(--border) !important; border-radius:3px !important;
-}
-#pdf-out { border:1px solid var(--border) !important; border-radius:3px !important; }
-#pdf-out .file-preview { background:#060606 !important; }
-
-.rule { border:none; border-top:1px solid #111; margin:28px 0; }
-.contact-card {
-  background:#080808; border:1px solid #1e1e1e; border-radius:4px; padding:28px 28px 20px;
-}
-.contact-info-block {
-  background:#060606; border:1px solid #222; border-radius:3px; padding:18px 20px; margin-top:16px;
-}
-footer { display:none !important; }
-"""
-
-HOME_HTML = """
-<div style="max-width:900px;margin:0 auto;padding:0 28px 72px;">
-
-  <!-- HERO -->
-  <div style="padding:72px 0 56px;border-bottom:1px solid #111;">
-    <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;letter-spacing:.18em;
-                color:#FF8C00;text-transform:uppercase;margin-bottom:16px;font-weight:700;">
-      Forensic AI Platform &nbsp;&middot;&nbsp; Scene Solver v3
-    </div>
-    <h1 style="font-family:Charter,Georgia,serif;font-size:3.6rem;font-weight:400;
-               color:#fff;line-height:1.08;margin-bottom:22px;letter-spacing:-.01em;">
-      Uncover the Fact.
-    </h1>
-    <p style="font-family:Charter,Georgia,serif;font-size:1rem;color:#666;
-              max-width:560px;line-height:1.78;margin-bottom:40px;">
-      CrimeSceneSolver is an end-to-end forensic intelligence platform.
-      Upload CCTV footage and a nine-stage AI pipeline analyses it, then
-      Sherl𝟬ck, the agentic escalation layer, briefs you by voice or text,
-      dispatches evidence to all channels, and generates a full PDF dossier.
-    </p>
-    <button onclick="(function(){
-        var b=document.querySelectorAll('.tab-nav button,[role=tab]');
-        for(var i=0;i<b.length;i++){var t=(b[i].textContent||'').trim().toLowerCase();
-          if(t.indexOf('analyse')!==-1){b[i].click();return;}}
-        if(b.length>1)b[1].click();
-      })()"
-      style="background:#fff;color:#000;font-family:'Roboto Mono',monospace;font-size:.74rem;
-             letter-spacing:.14em;text-transform:uppercase;padding:14px 36px;border-radius:3px;
-             border:none;cursor:pointer;font-weight:700;transition:background .2s;"
-      onmouseover="this.style.background='#FF8C00'" onmouseout="this.style.background='#fff'">
-      Analyse Footage &rarr;
-    </button>
-  </div>
-
-  <!-- NINE-STAGE PIPELINE -->
-  <div style="padding:56px 0 48px;border-bottom:1px solid #111;">
-    <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;letter-spacing:.16em;
-                color:#FF8C00;text-transform:uppercase;margin-bottom:28px;font-weight:700;">
-      Nine-Stage Forensic Pipeline
-    </div>
-    <div style="display:grid;grid-template-columns:repeat(3,1fr);gap:10px;">
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 0</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">Autoencoder (AE)</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">Primary anomaly signal. Reconstruction error over a dynamic threshold triggers the pipeline.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 1</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">TimeSformer Binary</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">Semantic gate: classifies the clip as Normal or Suspicious. Normal overrides the AE signal to SAFE.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 2</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">TimeSformer Multi-class</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">Fine-grained anomaly type across 7 categories using AE-weighted temporal sampling.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 3</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">YOLO Weapon Detection</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">Hard safety trigger. Melee and firearm detection within the anomaly window. Forces 0.90 risk floor.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 4</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">Audio Feature Extraction</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">ZCR, MFCC variance, spectral flux, RMS, and centroid scored against global thresholds.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 5</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">Tracking &amp; Motion</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">YOLO tracking with Farneback optical flow fallback. Detects FAST_MOVEMENT, CLUSTERING, STATIC.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 6</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">Fusion Layer</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">Explainable rule-based reasoning combining all stage outputs into a single risk score with a full trace.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 7</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">RL3033 Agent</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">PPO reinforcement learning agent. Final risk blends TFB, TFM, Fusion, and RL scores <span style="font-family:'Times New Roman',Times,serif;">(70/20/5/5%)</span>.</div>
-      </div>
-      <div style="padding:20px 18px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.55rem;color:#FF8C00;margin-bottom:6px;font-weight:700;">STAGE 8</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.9rem;color:#fff;margin-bottom:6px;">LLaVA Narrative</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.76rem;color:#555;line-height:1.6;">Four keyframes described in natural language and synthesised into a coherent scene narrative.</div>
-      </div>
-    </div>
-  </div>
-
-  <!-- SHERL0CK AGENT -->
-  <div style="padding:56px 0 48px;border-bottom:1px solid #111;">
-    <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;letter-spacing:.16em;
-                color:#FF8C00;text-transform:uppercase;margin-bottom:28px;font-weight:700;">
-      Sherl𝟬ck — Agentic Escalation Layer
-    </div>
-    <div style="display:grid;grid-template-columns:1fr 1fr;gap:20px;">
-      <div style="padding:26px 22px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:Charter,Georgia,serif;font-size:.95rem;color:#fff;margin-bottom:8px;">Voice Call Interface</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.8rem;color:#555;line-height:1.7;">
-          When a CRITICAL or HIGH risk is detected, Sherl𝟬ck initiates an incoming call.
-          Accept it to speak directly with the agent, Whisper transcribes your speech,
-          Cohere generates an analytical response, and XTTS v2 delivers it as a voice reply.
-          Server-side VAD handles turn detection automatically.
-        </div>
-      </div>
-      <div style="padding:26px 22px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:Charter,Georgia,serif;font-size:.95rem;color:#fff;margin-bottom:8px;">Text Chat Interface</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.8rem;color:#555;line-height:1.7;">
-          A persistent text chat powered by Cohere command-a-05-2026 with the full pipeline
-          JSON pre-loaded as context. Ask about any stage output, request reasoning traces,
-          or say <em style="color:#FF8C00;font-style:normal;">"send report"</em> to dispatch evidence.
-        </div>
-      </div>
-      <div style="padding:26px 22px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:Charter,Georgia,serif;font-size:.95rem;color:#fff;margin-bottom:8px;">Evidence Dispatch</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.8rem;color:#555;line-height:1.7;">
-          One click dispatches the full forensic dossier simultaneously to Telegram,
-          Discord, and Email, including the peak anomaly frame and the 11-page PDF report.
-          CRITICAL alerts auto-dispatch without user action.
-        </div>
-      </div>
-      <div style="padding:26px 22px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:Charter,Georgia,serif;font-size:.95rem;color:#fff;margin-bottom:8px;">11-Page PDF Dossier</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.8rem;color:#555;line-height:1.7;">
-          Auto-generated PDF covering every stage: AE temporal graph, classification charts,
-          weapon detections with bounding boxes, audio feature breakdown, motion tracking,
-          fusion reasoning trace, RL3033 observation space, anomalous frames, and LLaVA narrative.
-        </div>
-      </div>
-    </div>
-  </div>
-
-  <!-- WORKFLOW -->
-  <div style="padding:56px 0 48px;border-bottom:1px solid #111;">
-    <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;letter-spacing:.16em;
-                color:#FF8C00;text-transform:uppercase;margin-bottom:28px;font-weight:700;">End-to-End Workflow</div>
-    <div style="display:grid;grid-template-columns:repeat(4,1fr);gap:24px;">
-      <div>
-        <div style="font-family:'Roboto Mono',monospace;font-size:.6rem;color:#333;margin-bottom:8px;font-weight:700;">01</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.92rem;color:#fff;margin-bottom:6px;">Upload Footage</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.78rem;color:#555;line-height:1.65;">Drop CCTV footage on the Analyse tab and press Analyse Footage.</div>
-      </div>
-      <div>
-        <div style="font-family:'Roboto Mono',monospace;font-size:.6rem;color:#333;margin-bottom:8px;font-weight:700;">02</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.92rem;color:#fff;margin-bottom:6px;">Pipeline Runs</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.78rem;color:#555;line-height:1.65;">All nine stages execute sequentially. Report, frame, JSON, and PDF are produced.</div>
-      </div>
-      <div>
-        <div style="font-family:'Roboto Mono',monospace;font-size:.6rem;color:#333;margin-bottom:8px;font-weight:700;">03</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.92rem;color:#fff;margin-bottom:6px;">Sherl0ck Calls</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.78rem;color:#555;line-height:1.65;">On HIGH or CRITICAL risk, an incoming call overlay appears. Accept to begin the briefing.</div>
-      </div>
-      <div>
-        <div style="font-family:'Roboto Mono',monospace;font-size:.6rem;color:#333;margin-bottom:8px;font-weight:700;">04</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.92rem;color:#fff;margin-bottom:6px;">Dispatch &amp; Close</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.78rem;color:#555;line-height:1.65;">Dispatch evidence to all channels, download the PDF dossier, and end the session.</div>
-      </div>
-    </div>
-  </div>
-
-  <!-- PRICING -->
-  <div style="padding:56px 0 48px;border-bottom:1px solid #111;">
-    <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;letter-spacing:.16em;
-                color:#FF8C00;text-transform:uppercase;margin-bottom:28px;font-weight:700;">Pricing</div>
-    <div style="display:grid;grid-template-columns:repeat(3,1fr);gap:12px;">
-      <div style="padding:26px 22px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;color:#555;margin-bottom:12px;font-weight:700;">BASIC</div>
-        <div style="font-size:2rem;color:#fff;margin-bottom:14px;font-family:'Times New Roman',Times,serif;">Free</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.8rem;color:#555;line-height:2.1;">
-          9-stage pipeline<br>5 analyses / month<br>PDF dossier<br>JSON output
-        </div>
-      </div>
-      <div style="padding:26px 22px;background:#080808;border:1px solid #FF8C00;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;color:#FF8C00;margin-bottom:12px;font-weight:700;">PRO</div>
-        <div style="font-size:2rem;color:#fff;margin-bottom:14px;font-family:'Times New Roman',Times,serif;">&#8377;49/mo</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.8rem;color:#888;line-height:2.1;">
-          Unlimited analyses<br>Sherl0ck voice &amp; text agent<br>Auto-dispatch (Telegram, Discord, Email)<br>LLaVA narratives<br>Priority support
-        </div>
-      </div>
-      <div style="padding:26px 22px;background:#080808;border:1px solid #1e1e1e;border-radius:3px;">
-        <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;color:#555;margin-bottom:12px;font-weight:700;">ENTERPRISE</div>
-        <div style="font-size:2rem;color:#fff;margin-bottom:14px;font-family:'Times New Roman',Times,serif;">Custom</div>
-        <div style="font-family:Charter,Georgia,serif;font-size:.8rem;color:#555;line-height:2.1;">
-          Custom model fine-tuning<br>On-premise deployment<br>Dedicated SLA<br>API access
-        </div>
-      </div>
-    </div>
-  </div>
-
-  <!-- TESTIMONIAL -->
-  <div style="padding:56px 0 24px;">
-    <div style="font-family:'Roboto Mono',monospace;font-size:.62rem;letter-spacing:.16em;
-                color:#FF8C00;text-transform:uppercase;margin-bottom:26px;font-weight:700;">From the Field</div>
-    <blockquote style="font-family:Charter,Georgia,serif;font-size:1.08rem;color:#aaa;line-height:1.76;
-                       border-left:2px solid #FF8C00;padding-left:22px;font-style:italic;margin:0 0 14px;">
-      &ldquo;We closed a complex six-month investigation in under two weeks after deploying
-      SceneSolver. The nine-stage pipeline, Sherl0ck voice briefings, and automated
-      evidence dispatch changed how our forensics team operates entirely.&rdquo;
-    </blockquote>
-    <div style="font-family:'Roboto Mono',monospace;font-size:.66rem;color:#555;font-weight:500;">
-      &mdash; Detective Lara Monroe, Forensics Division
-    </div>
-  </div>
-
-</div>
-"""
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# BUILD APP