Verify — multimodal disinformation detection platform (6 modules + chatbot)

.dockerignore

@@ -0,0 +1,44 @@
+# Speed up `docker build` by skipping files the runtime doesn't need.
+
+# Local environments / caches
+.venv/
+venv/
+__pycache__/
+*.pyc
+.pytest_cache/
+.mypy_cache/
+
+# OS / IDE
+.DS_Store
+Thumbs.db
+.idea/
+.vscode/
+
+# Large contributor notebooks (documentation only — not needed at runtime)
+malek/*.ipynb
+rayen/*.ipynb
+youssef/*.ipynb
+yassmine/*.ipynb
+*.ipynb_checkpoints/
+
+# Trained weights (pulled at runtime from HF Dataset via scripts/download_models.py)
+backend/data/io1_resnet50.pth
+backend/data/io1_resnet50_deepfake.pth
+islem/*.pth
+
+# Large generic weights (downloaded fresh in the Dockerfile pre-warm step)
+yolov8m.pt
+
+# Local secrets — never bake into an image
+.env
+.env.local
+backend/data/openai_key.txt
+*.key
+
+# Project documentation files unrelated to runtime
+ios images/
+demo/
+*.pdf
+*.docx
+*.xlsx
+!backend/data/IO6_Base_Reference_V3_FULL.xlsx

.env.example

@@ -0,0 +1,59 @@
+# ─── Verify — Configuration template ───────────────────────────────────────
+# Copy this file to `.env` and fill in only the variables you need to override.
+# `.env` is gitignored — never commit your real keys.
+
+# ─── Module loading ─────────────────────────────────────────────────────────
+# Comma-separated list of modules to load at startup. Empty = all modules with
+# status="active" (i.e. all 6 by default).
+# LOAD_MODULES=io3,io1
+
+# ─── Compute device ─────────────────────────────────────────────────────────
+# FORCE_DEVICE=cpu          # set to "cpu" to disable GPU detection
+# FORCE_DEVICE=cuda
+
+# ─── OpenAI API (optional — used ONLY by the chatbot module, NOT by the 6 verdicts) ───
+# Get your key at https://platform.openai.com/api-keys
+# OPENAI_API_KEY=sk-proj-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
+
+# ─── io1 — Fake Media Detection ─────────────────────────────────────────────
+# IO1_USE_OPENAI=off              # set "on" to delegate the verdict to GPT-4o (default: local)
+# IO1_DEEPFAKE_THRESHOLD=0.60     # p_fake threshold for the ResNet50 ensemble
+# IO1_USE_IMAGE_DEEPFAKE=on       # load Islem's Model X (resnet50_deepfake.pth)
+# IO1_USE_AI_DETECTOR=on          # enable the 2-detector AI consensus (Organika + umm-maybe)
+# IO1_AI_TIER1_PRIMARY=0.95       # tuning of the consensus thresholds (see ai_detector.py)
+# IO1_AI_TIER1_SECONDARY=0.20
+# IO1_AI_TIER2_PRIMARY=0.90
+# IO1_AI_TIER2_SECONDARY=0.40
+
+# ─── io2 — Visual Manipulation / Persuasion ─────────────────────────────────
+# IO2_USE_OPENAI=off
+# IO2_CLICKBAIT_MODEL=valurank/distilroberta-clickbait
+# IO2_ENABLE_TRANSLATION=auto     # set "off" to skip the FR→EN MarianMT translator
+
+# ─── io3 — Image-Caption Coherence ──────────────────────────────────────────
+# CLIP_FINETUNED_PATH=clip_finetuned_coherence.pth   # optional fine-tuned CLIP checkpoint
+# SAM_CHECKPOINT=sam_vit_b_01ec64.pth                # optional Segment-Anything weights
+# YOLO_WEIGHTS=yolov8m.pt
+# WHISPER_MODEL=small
+# ENABLE_SAM=auto
+
+# ─── io4 — Image Tampering Detection ────────────────────────────────────────
+# IO4_USE_OPENAI=off
+# IO4_ELA_QUALITY=90              # JPEG re-save quality used by ELA
+# IO4_THRESHOLD=0.50              # legacy CNN threshold (ignored when forensics path is used)
+
+# ─── io5 — Caption Fidelity ─────────────────────────────────────────────────
+# IO5_CLIP_MODEL=ViT-B-32
+# IO5_FAITHFUL_THRESHOLD=0.60
+# IO5_MISLEADING_THRESHOLD=0.40
+
+# ─── io6 — Cosmetic Ads Fact-Check ──────────────────────────────────────────
+# IO6_USE_OPENAI=off
+# IO6_WHISPER_MODEL=tiny          # tiny|base|small|medium — accuracy/speed trade-off
+# IO6_YOLO_WEIGHTS=yolov8n.pt
+# IO6_ENABLE_MINILM=auto
+# IO6_KB_PATH=backend/data/IO6_Base_Reference_V3_FULL.xlsx
+
+# ─── Narrative layer (post-processing) ──────────────────────────────────────
+# IO_XAI_NARRATIVE=off            # set "off" to disable the AI-written explanation (saves OpenAI calls)
+# IO_XAI_MODEL=gpt-4o-mini

.gitattributes

@@ -0,0 +1,7 @@
+*.png filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.pdf filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.xlsx filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,73 @@
+# ─── Secrets (NEVER commit) ──────────────────────────────────────────────
+.env
+.env.local
+backend/data/openai_key.txt
+*.key
+*.pem
+secrets/
+credentials.json
+google-services.json
+
+# ─── Python ───────────────────────────────────────────────────────────────
+.venv/
+venv/
+env/
+__pycache__/
+*.py[cod]
+*$py.class
+*.egg-info/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+
+# ─── Trained model weights (downloaded via scripts/download_models.py) ────
+# GitHub blocks files > 100 MB; combined model size is ~250 MB+
+*.pth
+*.pt
+*.onnx
+*.h5
+*.keras
+*.bin
+*.safetensors
+
+# Exception: small generic YOLO weights that are useful to ship if small enough
+!yolov8n.pt
+
+# ─── HuggingFace / cache ──────────────────────────────────────────────────
+.cache/
+~/.cache/huggingface/
+
+# ─── OS / IDE ─────────────────────────────────────────────────────────────
+.DS_Store
+Thumbs.db
+.idea/
+.vscode/
+*.swp
+*.swo
+*~
+
+# ─── Jupyter ──────────────────────────────────────────────────────────────
+.ipynb_checkpoints/
+*.ipynb_meta
+
+# ─── Build / dist ─────────────────────────────────────────────────────────
+dist/
+build/
+node_modules/
+
+# ─── Temporary / logs ─────────────────────────────────────────────────────
+*.log
+*.tmp
+tmp/
+/temp/
+.tox/
+
+# ─── Personal scratch files ───────────────────────────────────────────────
+test_local.py
+notes.md
+
+# Contributor notebooks (kept locally — too large for HF Space git, ~40 MB combined)
+malek/*.ipynb
+rayen/*.ipynb
+youssef/*.ipynb
+yassmine/*.ipynb

AI_Investigate_Report_COMPLETE.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5fbe755ecdb47fd2d62dbc779e4013618ee2848eb8957507a6da0dde1587f7f
+size 1920340

DEPLOYMENT.md

@@ -0,0 +1,238 @@
+# Verify — Free deployment guide (HuggingFace Spaces)
+
+This guide deploys the **full Verify platform** (6 detection modules + chatbot + static frontend)
+**on a single free HuggingFace Space**. The jury gets one public URL — e.g. `https://yassminenouisser-verify.hf.space` — that they can hit with their browser to test everything.
+
+> **Total cost: 0 €.** Total wall-clock time the first time: ~30 min (mostly waiting for the Space to build).
+> Cold-start delay for the first request after 48h of inactivity: ~1 min (the persistent storage keeps models warm).
+
+---
+
+## Architecture
+
+```
+            ┌──────────────────────────────────────────────────┐
+            │  HuggingFace Space  (Docker SDK, CPU Basic 16GB) │
+            │                                                  │
+            │   ┌────────────────────────────────────────┐     │
+            │   │  FastAPI (uvicorn :7860)               │     │
+            │   │  ├─ /                  static frontend │     │
+            │   │  ├─ /api/io[1-6]/...   detection ML    │     │
+            │   │  ├─ /api/chat/...      chatbot         │     │
+            │   │  └─ /health, /docs                     │     │
+            │   └────────────────────────────────────────┘     │
+            │                                                  │
+            │   Secrets (encrypted):                           │
+            │     OPENAI_API_KEY                               │
+            │     IO1_RESNET50_URL          ← from HF Dataset  │
+            │     IO1_RESNET50_DEEPFAKE_URL ← from HF Dataset  │
+            └──────────────────────────────────────────────────┘
+                          ▲
+                          │
+            ┌─────────────┴────────────────┐
+            │  HuggingFace Dataset (free)  │
+            │  Verify-weights (private)    │
+            │  ├─ best_ResNet50.pth        │
+            │  └─ resnet50_deepfake.pth    │
+            └──────────────────────────────┘
+```
+
+The Space hosts everything in one container. The two heavy `.pth` weights are stored in a
+companion **HuggingFace Dataset** (also free, unlimited) so the Space repo stays under the
+git-LFS-free 10 MB-per-file limit.
+
+---
+
+## Step 1 — Create a HuggingFace account
+
+Go to **https://huggingface.co/join**, create your free account, verify your email.
+Note your username (e.g. `yassminenouisser`). You'll use it everywhere below.
+
+---
+
+## Step 2 — Upload Islem's `.pth` weights to a HF Dataset
+
+The two ResNet50 checkpoints (94 MB each) cannot live in the Space's git repo. We host them in
+a HuggingFace Dataset.
+
+1. Go to **https://huggingface.co/new-dataset**
+2. Name: `verify-weights`
+3. Visibility: **Private** (the weights are Islem's research artifacts)
+4. License: `other`
+5. Click **Create dataset**
+
+Then upload the files (web UI is the easiest):
+
+1. Open your new dataset → **Files** tab → **Add file** → **Upload files**
+2. Drag-drop `backend/data/io1_resnet50.pth` and `backend/data/io1_resnet50_deepfake.pth`
+3. Commit
+
+The files are now at:
+- `https://huggingface.co/datasets/<your-username>/verify-weights/resolve/main/io1_resnet50.pth`
+- `https://huggingface.co/datasets/<your-username>/verify-weights/resolve/main/io1_resnet50_deepfake.pth`
+
+> ⚠️ **Private dataset**: the Space will need a **read-token** to download these.
+> Generate one at https://huggingface.co/settings/tokens (role: `read`). Copy it; you'll paste
+> it as a Space Secret in step 4.
+
+---
+
+## Step 3 — Create the HuggingFace Space
+
+1. Go to **https://huggingface.co/new-space**
+2. Owner: your username
+3. Name: `verify` (the public URL will be `https://<username>-verify.hf.space`)
+4. License: `apache-2.0`
+5. **SDK: `Docker`** (NOT Gradio/Streamlit — we use our own Dockerfile)
+6. Hardware: **CPU basic — free** (16 GB RAM, 2 vCPU)
+7. Visibility: **Public**
+8. Click **Create Space**
+
+The Space is created empty. Don't push anything yet — first configure the secrets.
+
+---
+
+## Step 4 — Configure Space Secrets (encrypted environment variables)
+
+In your Space, go to **Settings** → **Variables and secrets** → **New secret**. Add these one by one:
+
+| Secret name | Value | Why |
+|---|---|---|
+| `OPENAI_API_KEY` | `sk-proj-...` (your real key) | Needed by the chatbot module and the narrative layer |
+| `IO1_RESNET50_URL` | `https://huggingface.co/datasets/<username>/verify-weights/resolve/main/io1_resnet50.pth` | Where the build script pulls Islem's weights |
+| `IO1_RESNET50_DEEPFAKE_URL` | `https://huggingface.co/datasets/<username>/verify-weights/resolve/main/io1_resnet50_deepfake.pth` | Same, for Model X |
+| `HUGGINGFACE_HUB_TOKEN` | the read-token you generated in step 2 | Lets the build authenticate to download from the private dataset |
+
+Click **Save** after each one. Secrets are encrypted and never visible in the Space's logs or code.
+
+---
+
+## Step 5 — Push the Verify code to the Space
+
+Each HF Space is a git repo. From your local Verify checkout:
+
+```bash
+cd /Users/yassminesmachine/Desktop/template
+
+# 1. Make sure the local repo is up to date (no uncommitted work)
+git status
+git add -A
+git commit -m "Prepare for HuggingFace Space deployment"
+
+# 2. Add HF Space as a remote (use your own username + space name)
+git remote add space https://huggingface.co/spaces/<username>/verify
+
+# 3. Push
+git push space main
+
+# (HF git will prompt for your HuggingFace credentials — username + an access token from
+#  https://huggingface.co/settings/tokens, role 'write')
+```
+
+Once the push completes, HuggingFace starts building the Docker image automatically. You can
+watch the build at **https://huggingface.co/spaces/`<username>`/verify** → **Logs** tab.
+
+> ⏱ **First build takes ~15 min** because the Dockerfile pre-downloads all HuggingFace models
+> (CLIP, TrOCR, Whisper, etc.) inside the image. Subsequent builds are much faster — only the
+> changed layers rebuild.
+
+---
+
+## Step 6 — Verify the deployment
+
+When the build is done and the Space status flips to **Running**:
+
+1. Open `https://<username>-verify.hf.space` in your browser → you should see the Verify homepage.
+2. Click **Verifier** → choose any module → upload an image → click **Analyze**.
+3. The first analysis after a cold start takes ~30-60 s (CPU + first-time module init);
+   subsequent calls are 5-15 s.
+
+Sanity checks:
+
+```bash
+# From your laptop — confirms the API is reachable
+curl https://<username>-verify.hf.space/health | jq .status      # → "ok"
+
+curl https://<username>-verify.hf.space/api                       # → JSON describing the 6 modules + chatbot
+```
+
+---
+
+## Step 7 — Fill in the ESPRIT submission form
+
+Now you can fill in section 2 of the submission PDF:
+
+| Champ | Valeur |
+|---|---|
+| Nom du projet | `Esprit-PI-<TaClasse>-2526-Verify` |
+| Lien GitHub | `https://github.com/<YourGitHub>/Esprit-PI-<TaClasse>-2526-Verify` |
+| Lien de déploiement | **`https://<username>-verify.hf.space`** ✓ |
+| Type de projet | IA |
+| Commande de lancement | `docker build -t verify . && docker run -p 7860:7860 verify` (or local: `uvicorn backend.main:app --port 8000`) |
+| Temps d'installation estimé | < 10 min (local) or instant (deployed link) |
+
+---
+
+## Updating the Space later
+
+After pushing changes to GitHub, push the same commits to the Space:
+
+```bash
+git push origin main         # GitHub
+git push space  main         # HF Space (triggers rebuild)
+```
+
+If you only changed Python code (no new model), the rebuild takes ~2-3 min thanks to Docker
+layer caching.
+
+---
+
+## Troubleshooting
+
+### "Build failed: file too large"
+GitHub-via-LFS isn't enabled by default on HF Spaces. Make sure `.gitignore` is excluding all
+`*.pth`/`*.pt` files except `yolov8n.pt` (it's only 7 MB).
+Check with: `git ls-files | xargs -I{} ls -l {} 2>/dev/null | awk '$5 > 10000000 {print}'`
+
+### "Space running out of memory"
+The default CPU-basic Space has 16 GB RAM, which is enough. If you upgraded the hardware and
+hit OOM, try `LOAD_MODULES=io1,io4` to load fewer modules at startup (set this as a Space variable).
+
+### Chatbot returns "service unavailable"
+The `OPENAI_API_KEY` secret is missing or the key is invalid. Re-set it in **Space Settings** →
+**Variables and secrets**, then restart the Space (top-right menu → **Restart this Space**).
+
+### Cold start is very long (> 3 min)
+The first request after a 48h sleep needs to (a) re-download HF models if persistent storage was
+not enabled, (b) load all PyTorch models into RAM. Enable **Persistent storage** in Space
+settings (free 50 GB tier — was previously paid) to keep the cache between sleeps.
+
+### Frontend says "Service unavailable"
+Open the browser DevTools console. If `fetch('/api/io1/health')` returns 404, the FastAPI static
+mount is broken — check `backend/main.py` and confirm the `app.mount("/", StaticFiles(...))` line
+is present and that `index.html` exists at the repo root.
+
+### Some images come up as REAL when they should be FAKE
+This is the limit of the open-source AI detectors (see `MODELS.md`). The deepfake ensemble +
+2-detector consensus catches ~66% of ThisPersonDoesNotExist images and ~99% of obvious DALL-E /
+Midjourney / Flux images. Re-test with another sample if you hit a model blind spot.
+
+---
+
+## Alternative: deploy WITHOUT HuggingFace Spaces (local Docker only)
+
+If for any reason you cannot use HF Spaces, the project also runs with plain Docker. In your
+local terminal:
+
+```bash
+docker build -t verify .
+docker run -p 7860:7860 \
+  -e OPENAI_API_KEY=sk-... \
+  -e IO1_RESNET50_URL=https://... \
+  -e IO1_RESNET50_DEEPFAKE_URL=https://... \
+  verify
+```
+
+Then open `http://localhost:7860`. The jury will need Docker on their machine — but the
+**ESPRIT acceptance criterion for IA projects** (guide page 8) is precisely that: local launch in
+<20 min with Docker. So this fully passes the rubric even without HF Spaces.

Dockerfile

@@ -0,0 +1,97 @@
+# ─── Verify — single-container deployment for HuggingFace Spaces ────────────
+# Serves the FastAPI backend (6 detection modules + chatbot) AND the static frontend
+# from the same container, on port 7860 (HuggingFace Spaces standard).
+#
+# Pre-downloads the heavy HuggingFace models at BUILD time so the Space cold-start
+# stays under ~60 s instead of ~10 min on first user request.
+#
+# Required Space "Secrets" (set in Space Settings → Variables and secrets):
+#   OPENAI_API_KEY              for the chatbot module (optional but recommended)
+#   IO1_RESNET50_URL            HF dataset URL for Islem's best_ResNet50.pth
+#   IO1_RESNET50_DEEPFAKE_URL   HF dataset URL for Islem's resnet50_deepfake.pth
+#   HUGGINGFACE_HUB_TOKEN       only if the weights dataset is private
+
+FROM python:3.11-slim
+
+# ─── System packages ────────────────────────────────────────────────────────
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        ffmpeg \
+        libgl1 \
+        libglib2.0-0 \
+        build-essential \
+        wget \
+        curl \
+    && rm -rf /var/lib/apt/lists/*
+
+# ─── Layout & user ──────────────────────────────────────────────────────────
+# HF Spaces runs the container as a non-root user with UID 1000.
+RUN useradd -m -u 1000 verify
+USER verify
+WORKDIR /home/verify/app
+
+# Caches inside the user's home so HF Spaces persistent storage (if enabled) keeps them.
+ENV HOME=/home/verify \
+    HF_HOME=/home/verify/.cache/huggingface \
+    TORCH_HOME=/home/verify/.cache/torch \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1
+
+# ─── Python dependencies (cached layer) ─────────────────────────────────────
+COPY --chown=verify:verify backend/requirements.txt backend/requirements.txt
+RUN pip install --user --upgrade pip \
+ && pip install --user -r backend/requirements.txt
+
+ENV PATH="/home/verify/.local/bin:${PATH}"
+
+# ─── Pre-warm the HuggingFace cache during build (saves cold-start time) ────
+# This pulls the heaviest models so the first user request doesn't trigger a 5-minute
+# download. Each line is independent — comment out any model you want to lazy-load.
+RUN python -c "\
+from open_clip import create_model_and_transforms, get_tokenizer; \
+create_model_and_transforms('ViT-B-32', pretrained='openai'); \
+create_model_and_transforms('ViT-L-14', pretrained='laion2b_s32b_b82k'); \
+print('CLIP models cached')" \
+ && python -c "\
+from transformers import AutoTokenizer, AutoModelForSequenceClassification, \
+    AutoModelForImageClassification, AutoImageProcessor, \
+    TrOCRProcessor, VisionEncoderDecoderModel, MarianTokenizer, MarianMTModel; \
+TrOCRProcessor.from_pretrained('microsoft/trocr-base-printed'); \
+VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-base-printed'); \
+AutoTokenizer.from_pretrained('valurank/distilroberta-clickbait'); \
+AutoModelForSequenceClassification.from_pretrained('valurank/distilroberta-clickbait'); \
+AutoImageProcessor.from_pretrained('Organika/sdxl-detector'); \
+AutoModelForImageClassification.from_pretrained('Organika/sdxl-detector'); \
+AutoImageProcessor.from_pretrained('umm-maybe/AI-image-detector'); \
+AutoModelForImageClassification.from_pretrained('umm-maybe/AI-image-detector'); \
+MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-fr-en'); \
+MarianMTModel.from_pretrained('Helsinki-NLP/opus-mt-fr-en'); \
+print('HF transformers cached')" \
+ && python -c "\
+import whisper; whisper.load_model('tiny'); whisper.load_model('small'); \
+print('Whisper cached')" \
+ && python -c "\
+from sentence_transformers import SentenceTransformer; \
+SentenceTransformer('sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2'); \
+print('MiniLM cached')"
+
+# ─── EasyOCR models (download to a known place) ─────────────────────────────
+RUN python -c "\
+import easyocr; easyocr.Reader(['en', 'fr'], gpu=False, verbose=False); \
+print('EasyOCR models cached')"
+
+# ─── Application code ───────────────────────────────────────────────────────
+COPY --chown=verify:verify . .
+
+# ─── Pull Islem's .pth weights (private HF Dataset) ─────────────────────────
+# The download script reads IO1_RESNET50_URL & IO1_RESNET50_DEEPFAKE_URL from env.
+# Failing silently here keeps `docker build` reproducible without the secrets — the
+# runtime will retry on startup using the Space's Secrets.
+RUN python scripts/download_models.py || echo "[build] weights not downloaded (will retry at runtime)"
+
+# ─── Runtime ─────��──────────────────────────────────────────────────────────
+EXPOSE 7860
+ENV PORT=7860
+
+# Use a startup wrapper so we can retry the weights download with the Space secrets
+# (which are NOT available during `docker build` — only at runtime).
+CMD ["bash", "-c", "python scripts/download_models.py || true && uvicorn backend.main:app --host 0.0.0.0 --port ${PORT:-7860}"]

MODELS.md

@@ -0,0 +1,105 @@
+# Verify — Models inventory
+
+Exhaustive list of every model and trained weight file used by the Verify platform.
+Total disk footprint after first run: **~4 Go** (cached in `~/.cache/huggingface/` and `~/.cache/torch/`).
+Combined in-memory footprint when all modules are loaded: **~3 Go RAM**.
+
+Files **with a download URL** = auto-fetched from HuggingFace Hub on first use (no manual setup).
+Files **without** a URL = hosted on HuggingFace Hub by the Verify team (see `scripts/download_models.py`).
+
+## io1 — Fake Media Detection (deepfake & AI-generated)
+
+| Model | Type | Size | Source | Auto-download? |
+|---|---|---|---|---|
+| `io1_resnet50.pth` (Islem's `best_ResNet50.pth`) | ResNet50 fine-tuned for face-deepfake (idx 0=FAKE) | 94 Mo | HuggingFace Hub (Verify Dataset) | ⚙️ `scripts/download_models.py` |
+| `io1_resnet50_deepfake.pth` (Islem's `resnet50_deepfake.pth` / Model X) | ResNet50 + Dropout+Linear head (idx 0=REAL) | 94 Mo | HuggingFace Hub (Verify Dataset) | ⚙️ `scripts/download_models.py` |
+| **MTCNN** (`facenet-pytorch`) | Face detector (3 stages: PNet/RNet/ONet) | 6 Mo | embedded in `facenet-pytorch` pip package | ✅ pip install |
+| **Organika/sdxl-detector** | ViT image classifier for SDXL/diffusion AI images | 86 Mo | [`Organika/sdxl-detector`](https://huggingface.co/Organika/sdxl-detector) | ✅ HuggingFace |
+| **umm-maybe/AI-image-detector** | ViT image classifier (precision-tuned) | 86 Mo | [`umm-maybe/AI-image-detector`](https://huggingface.co/umm-maybe/AI-image-detector) | ✅ HuggingFace |
+
+## io2 — Visual Manipulation / Persuasion
+
+| Model | Type | Size | Source | Auto-download? |
+|---|---|---|---|---|
+| **microsoft/trocr-base-printed** | Vision-Encoder + GPT-2 Decoder (OCR for screen text) | 558 Mo | [`microsoft/trocr-base-printed`](https://huggingface.co/microsoft/trocr-base-printed) | ✅ HuggingFace |
+| **valurank/distilroberta-clickbait** | DistilRoBERTa fine-tuned for clickbait detection (binary) | 330 Mo | [`valurank/distilroberta-clickbait`](https://huggingface.co/valurank/distilroberta-clickbait) | ✅ HuggingFace |
+| **open_clip ViT-B-32** (openai pretrained) | CLIP for zero-shot "clickbait visual style" scoring | 150 Mo | [`openai/clip-vit-base-patch32`](https://huggingface.co/openai/clip-vit-base-patch32) | ✅ open_clip |
+| **Helsinki-NLP/opus-mt-fr-en** | MarianMT FR→EN (so the EN-only clickbait model receives EN) | 300 Mo | [`Helsinki-NLP/opus-mt-fr-en`](https://huggingface.co/Helsinki-NLP/opus-mt-fr-en) | ✅ HuggingFace |
+| **EasyOCR** (EN + FR) | Text detection + recognition (CRAFT + CRNN) | 100 Mo | embedded in `easyocr` pip package | ✅ pip install |
+
+## io3 — Image-Caption Coherence
+
+| Model | Type | Size | Source | Auto-download? |
+|---|---|---|---|---|
+| **open_clip ViT-L-14** (laion2b_s32b_b82k) | CLIP large for image↔text similarity | 890 Mo | [`laion/CLIP-ViT-L-14-laion2B-s32B-b82K`](https://huggingface.co/laion/CLIP-ViT-L-14-laion2B-s32B-b82K) | ✅ open_clip |
+| **YOLOv8m** (`yolov8m.pt`) | Object detection — 80 COCO classes | 52 Mo | [Ultralytics releases](https://github.com/ultralytics/assets/releases) | ✅ ultralytics |
+| **EasyOCR** (EN + FR) | reused from io2 | shared | — | ✅ |
+| **Whisper small** (`small.pt`) | OpenAI Whisper audio transcription | 470 Mo | [`openai-whisper` pip package](https://github.com/openai/whisper) | ✅ pip install |
+| **SAM ViT-B** (optional) | Segment Anything for region scoring | 375 Mo | [`sam_vit_b_01ec64.pth`](https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth) | ⚙️ manual download (else 3×3 grid fallback) |
+
+## io4 — Image Tampering (Photoshop forensics)
+
+| Model | Type | Size | Source | Auto-download? |
+|---|---|---|---|---|
+| **None** (pure model-free forensics) | ELA + Noise residual + JPEG ghost | 0 Mo | implemented in [`backend/modules/io4_photoshop/forensics.py`](backend/modules/io4_photoshop/forensics.py) | ✅ pure NumPy |
+
+## io5 — Caption Fidelity
+
+| Model | Type | Size | Source | Auto-download? |
+|---|---|---|---|---|
+| **open_clip ViT-B-32** (openai) | CLIP for image↔caption + per-phrase scoring | shared with io2 | — | ✅ |
+| **EasyOCR** (EN + FR) | reused from io2 | shared | — | ✅ |
+
+## io6 — Cosmetic Ads Fact-Check
+
+| Model | Type | Size | Source | Auto-download? |
+|---|---|---|---|---|
+| **Whisper tiny** | Audio transcription (rapid version for ads) | 75 Mo | [`openai-whisper` pip package](https://github.com/openai/whisper) | ✅ pip install |
+| **YOLOv8n** (`yolov8n.pt`) | Object detection (lightweight) | 7 Mo | [Ultralytics releases](https://github.com/ultralytics/assets/releases) — versioned in repo | ✅ committed |
+| **EasyOCR** (FR + EN) | reused from io2 | shared | — | ✅ |
+| **paraphrase-multilingual-MiniLM-L12-v2** | Sentence embeddings (semantic match with KB fake claims) | 470 Mo | [`sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2`](https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2) | ✅ HuggingFace |
+| **IO6_Base_Reference_V3_FULL.xlsx** (Yassmine's KB) | Regulatory knowledge base — 25 patterns, 33 fake claims, 145 ingredients, 93 brands | 33 Ko | versioned in `backend/data/` | ✅ committed |
+
+## chatbot — Verify Assistant
+
+| Service | Type | Size | Source | Auto-download? |
+|---|---|---|---|---|
+| **OpenAI gpt-4o-mini** | Topic-restricted assistant via OpenAI API | n/a (remote) | requires `OPENAI_API_KEY` env var | ☁️ remote API |
+
+The chatbot is **the only component that hits an external API**. The 6 detection modules above
+all run 100% locally. On HuggingFace Spaces the key is set as a "Secret" (encrypted, never in the code).
+
+## Narrative layer (cross-module, post-processing)
+
+| Service | Type | Source |
+|---|---|---|
+| **OpenAI gpt-4o-mini** | Rewrites every module's raw result into plain-English narrative | ☁️ remote API (optional — disable with `IO_XAI_NARRATIVE=off`) |
+
+The narrative layer is **optional**: if `OPENAI_API_KEY` is not set, modules return their raw
+local explanation only. Set `IO_XAI_NARRATIVE=off` to disable explicitly.
+
+---
+
+## How to obtain Islem's `.pth` weights
+
+The two ResNet50 checkpoints used by io1 (`io1_resnet50.pth` and `io1_resnet50_deepfake.pth`) are
+not versioned in this repo because each file weighs ~94 MB (GitHub's per-file limit is 100 MB but
+combined size kills clone speed).
+
+They are hosted as a private HuggingFace Dataset by the Verify team. To download them locally:
+
+```bash
+# Either: download via the helper script, with the URLs set in .env or exported
+export IO1_RESNET50_URL=https://huggingface.co/datasets/<verify_user>/verify-weights/resolve/main/best_ResNet50.pth
+export IO1_RESNET50_DEEPFAKE_URL=https://huggingface.co/datasets/<verify_user>/verify-weights/resolve/main/resnet50_deepfake.pth
+python scripts/download_models.py
+
+# Or, on a HuggingFace Space: the Dockerfile pulls them automatically using HUGGINGFACE_HUB_TOKEN
+# (set as a Secret in Space settings → so the file stays private).
+```
+
+## Auto-download summary
+
+After running `pip install -r backend/requirements.txt` and starting the backend once,
+**12 of 14 components self-install** from public sources. Only Islem's 2 `.pth` files need a manual
+step (the download script).

README.md

@@ -0,0 +1,177 @@
+---
+title: Verify
+emoji: 🔍
+colorFrom: red
+colorTo: blue
+sdk: docker
+app_port: 7860
+pinned: false
+short_description: Multimodal disinformation detection — 6 ML modules + chatbot
+license: apache-2.0
+---
+
+# Verify — Plateforme multimodale de détection de désinformation
+
+## Description
+
+**Verify** est une plateforme web qui vérifie l'authenticité de contenus visuels (image, vidéo, texte, publicité)
+en analysant en parallèle **six dimensions** distinctes de la désinformation :
+
+1. **io1 — Fake Media** : détection deepfake (face-swap) + images générées par IA (StyleGAN, DALL-E, Midjourney)
+2. **io2 — Visual Manipulation** : détection des images persuasives/clickbait (texte alarmiste, urgence forcée, mise en scène)
+3. **io3 — Image↔Caption Coherence** : vérifie la cohérence entre une image/vidéo et sa légende
+4. **io4 — Image Tampering** : forensique photoshop (ELA + analyse de bruit + JPEG ghost)
+5. **io5 — Caption Fidelity** : vérifie si une légende décrit fidèlement le contenu d'une image
+6. **io6 — Cosmetic Ads Fact-Check** : audit réglementaire EU 655/2013 sur les publicités cosmétiques
+
+Tous les modules tournent **100% en local** — aucune API externe n'est requise pour produire un verdict.
+Le backend est une seule API FastAPI qui monte les 6 modules ; le frontend est statique (HTML + JavaScript).
+
+## Technologies utilisées
+
+| Couche | Stack |
+|---|---|
+| **Frontend** | HTML5 · CSS3 · Vanilla JavaScript (pas de framework) |
+| **Backend** | Python 3.11 · FastAPI · Uvicorn |
+| **IA / ML** | PyTorch · TorchVision · HuggingFace Transformers · open_clip · Ultralytics YOLO · OpenAI Whisper · EasyOCR · facenet-pytorch (MTCNN) · sentence-transformers |
+| **Connaissances** | Pandas + OpenPyXL (knowledge base Excel pour io6) |
+| **Forensique** | Pure NumPy + Pillow (io4 ELA / noise / JPEG ghost — sans modèle) |
+
+## Prérequis
+
+- **Python 3.11+** (testé sur 3.11.15)
+- **ffmpeg** dans le PATH (utilisé par io3 et io6 pour extraire l'audio des vidéos)
+- **~4 Go RAM disponible** (les 6 modules chargent ~3 Go de modèles en mémoire au démarrage)
+- Premier démarrage : ~3-5 minutes (téléchargement automatique des modèles HuggingFace, ~2 Go)
+- Démarrages suivants : ~30 secondes (tout est caché localement)
+
+> ⚠️ **Modèles pré-entraînés non versionnés** : les poids `.pt`/`.pth` ne sont pas inclus dans le repo
+> (taille > 250 Mo combinée + politique ESPRIT/GitHub). Voir la section **Installation** ci-dessous.
+
+## Installation
+
+```bash
+# 1. Cloner le repo
+git clone https://github.com/USERNAME/Esprit-PI-CLASSE-2526-Verify.git
+cd Esprit-PI-CLASSE-2526-Verify
+
+# 2. Créer l'environnement Python
+python3.11 -m venv .venv
+source .venv/bin/activate          # macOS / Linux
+# .venv\Scripts\activate            # Windows
+
+# 3. Installer les dépendances backend
+pip install -r backend/requirements.txt
+
+# 4. Télécharger les poids ResNet50 d'Islem (io1) — non versionnés
+python scripts/download_models.py
+
+# 5. (Optionnel) Configurer les variables d'environnement
+cp .env.example .env
+# Éditez .env si vous voulez activer le chatbot OpenAI ou ajuster un seuil.
+```
+
+## Lancement
+
+```bash
+# Backend (terminal 1)
+uvicorn backend.main:app --host 0.0.0.0 --port 8000
+
+# Frontend (terminal 2) — au choix :
+python -m http.server 5500          # serveur HTTP statique
+# OU : extension "Live Server" dans VSCode (clic-droit sur index.html → "Open with Live Server")
+```
+
+Ensuite ouvrez votre navigateur sur **http://127.0.0.1:5500/index.html**.
+
+La doc Swagger de l'API est disponible sur **http://127.0.0.1:8000/docs**.
+
+## Variables d'environnement
+
+Voir [`.env.example`](.env.example) pour la liste complète. Les principales :
+
+| Variable | Description |
+|---|---|
+| `LOAD_MODULES=io3,io1` | Limite quels modules sont chargés au démarrage (default : tous). Pratique en dev pour économiser la mémoire. |
+| `FORCE_DEVICE=cpu` | Force le CPU même si un GPU est disponible. |
+| `OPENAI_API_KEY=sk-...` | (Optionnel) Active le module chatbot uniquement. **N'affecte pas les verdicts** des 6 modules d'analyse. |
+| `IO1_DEEPFAKE_THRESHOLD=0.60` | Seuil de décision FAKE pour l'ensemble ResNet50 d'Islem. |
+| `IO4_ELA_QUALITY=90` | Qualité JPEG utilisée par ELA dans io4. |
+| `IO6_WHISPER_MODEL=tiny` | Variante Whisper pour io6 (`tiny` rapide, `small`/`medium` plus précis). |
+
+## Démo
+
+- **Site déployé** : Non disponible (déploiement local uniquement pour cette livraison)
+- **Vidéo de démonstration** : voir [`demo/`](demo/) (à fournir par l'équipe)
+- **Diagrammes d'architecture** : voir [`docs/`](docs/)
+- **Description commerciale** : [`Verify_Description_Commerciale.pdf`](Verify_Description_Commerciale.pdf)
+
+## Performances des modèles
+
+| Module | Backend | Performance | Source |
+|--------|---------|------------|--------|
+| io1 deepfake | ResNet50 ensemble (Islem) + MTCNN | ~97% acc sur FaceForensics | [`islem/`](islem/) |
+| io1 AI-image | Consensus Organika/sdxl + umm-maybe (HF) | ~66% recall sur TPDNE, ~0% FP sur photos réelles | tuned empirically |
+| io2 NLP | DistilRoBERTa fine-tuned clickbait | ~99% acc sur Webis Clickbait 2017 | `valurank/distilroberta-clickbait` |
+| io2 visuel | CLIP ViT-B/32 zero-shot | calibration manuelle | OpenAI CLIP |
+| io3 cohérence | CLIP ViT-L/14 + YOLOv8m fusion | F1 ~0.85 sur dataset interne | [`youssef/`](youssef/) |
+| io4 forensique | ELA + Noise + JPEG ghost (sans modèle) | détection compositing/inpainting localisé | méthode classique |
+| io5 fidelity | CLIP ViT-B/32 + EasyOCR | calibrée sur paires CC3M | sigmoid calibration |
+| io6 cosmétique | Whisper + KB Excel + MiniLM semantic | 25 patterns / 33 fake claims / 145 ingrédients | [`yassmine/`](yassmine/) |
+
+## Structure du projet
+
+```
+Verify/
+├── README.md                       ← ce fichier
+├── .env.example                    ← variables d'environnement
+├── .gitignore
+├── index.html, verifier.html, ...  ← frontend statique
+├── assets/                         ← CSS, JS, images
+├── backend/                        ← API FastAPI
+│   ├── main.py                     ← entrypoint
+│   ├── requirements.txt
+│   ├── shared/                     ← code commun (device, narration XAI)
+│   ├── data/                       ← KB Excel io6 (poids .pth téléchargés via script)
+│   └── modules/
+│       ├── io1_ai_generated/
+│       ├── io2_persuasion/
+│       ├── io3_coherence/
+│       ├── io4_photoshop/
+│       ├── io5_caption_fidelity/
+│       └── io6_cosmetic_ads/
+├── scripts/
+│   └── download_models.py          ← téléchargement automatique des poids
+├── docs/                           ← diagrammes architecture, doc API
+├── demo/                           ← captures et vidéos de démo
+├── islem/, malek/, youssef/, rayen/, yassmine/   ← notebooks et docs des contributeurs
+└── ios images/                     ← jeu d'images de référence pour la démo
+```
+
+## Auteurs
+
+| Nom | Module | Année | Tuteur |
+|---|---|---|---|
+| Yassmine Nouisser  | io6 — Cosmetic Ads Fact-Check + intégration globale | 2025-2026 | *(à compléter)* |
+| Islem              | io1 — Fake Media Detection (deepfake + AI) | 2025-2026 | *(à compléter)* |
+| Malek Tirellil     | io2 — Visual Manipulation Detection         | 2025-2026 | *(à compléter)* |
+| Youssef            | io3 — Image-Caption Coherence               | 2025-2026 | *(à compléter)* |
+| Rayen              | io4 — Image Tampering Detection             | 2025-2026 | *(à compléter)* |
+| Maryem             | (contribution complémentaire)               | 2025-2026 | *(à compléter)* |
+
+Classe : *(à compléter)* — Groupe : *(à compléter)*
+
+---
+
+## Documentation supplémentaire
+
+- [`backend/README.md`](backend/README.md) — détails techniques de l'API
+- [`docs/architecture.md`](docs/architecture.md) — architecture système et flux de données
+- [`docs/api.md`](docs/api.md) — référence des endpoints
+- [`docs/modules.md`](docs/modules.md) — détails par module
+- [`AI_Investigate_Report_COMPLETE.pdf`](AI_Investigate_Report_COMPLETE.pdf) — rapport d'enquête initial
+- [`Verify_Description_Commerciale.pdf`](Verify_Description_Commerciale.pdf) — pitch commercial
+
+## Licence
+
+Projet académique — ESPRIT School of Engineering, année universitaire 2025-2026.

Verify_Description_Commerciale.html

@@ -0,0 +1,358 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8" />
+<title>Verify — Commercial overview</title>
+<meta name="description" content="Verify: Tunisia's visual-information verification platform. Drop in an image, a video or a post — get a clear, reasoned verdict in minutes, with the visual evidence behind it." />
+<link href="https://fonts.googleapis.com/css2?family=Playfair+Display:wght@400;700;800&family=Inter:wght@400;500;600;700&display=swap" rel="stylesheet">
+<style>
+  @page { size: A4; margin: 22mm 18mm; }
+  * { box-sizing: border-box; }
+  body {
+    font-family: 'Inter', -apple-system, sans-serif;
+    color: #1a1a1a;
+    line-height: 1.6;
+    font-size: 11pt;
+    margin: 0;
+  }
+  h1, h2, h3, h4 { font-family: 'Playfair Display', Georgia, serif; color: #052962; margin-top: 0; }
+  h1 { font-size: 32pt; line-height: 1.05; margin-bottom: 6px; font-weight: 800; }
+  h2 { font-size: 18pt; margin-top: 28px; margin-bottom: 10px; padding-bottom: 6px; border-bottom: 2px solid #052962; }
+  h3 { font-size: 13pt; margin-top: 18px; margin-bottom: 6px; color: #052962; }
+  p  { margin: 0 0 10px; }
+
+  .cover {
+    page-break-after: always;
+    text-align: center;
+    padding-top: 60mm;
+  }
+  .cover .kicker {
+    display: inline-block;
+    font-size: 10pt; font-weight: 700; letter-spacing: 2px; text-transform: uppercase;
+    color: #9a3412; background: #ffedd5; padding: 6px 14px; border-radius: 2px;
+    margin-bottom: 24px;
+  }
+  .cover h1 { font-size: 56pt; color: #052962; margin: 0; }
+  .cover .tagline {
+    font-family: 'Playfair Display', Georgia, serif;
+    font-style: italic;
+    font-size: 18pt;
+    color: #4a4a4a;
+    margin-top: 18px;
+  }
+  .cover .meta {
+    margin-top: 80mm;
+    font-size: 10pt;
+    color: #6b6b6b;
+    letter-spacing: 1px;
+    text-transform: uppercase;
+  }
+  .cover hr {
+    width: 60px; border: 0; border-top: 3px solid #ffe500; margin: 20px auto;
+  }
+
+  .lede {
+    font-size: 13pt;
+    font-style: italic;
+    color: #333;
+    border-left: 4px solid #ffe500;
+    padding: 10px 16px;
+    background: #fcfaf2;
+    margin: 16px 0 22px;
+  }
+
+  ul { margin: 6px 0 14px 18px; padding: 0; }
+  li { margin-bottom: 5px; }
+
+  table {
+    width: 100%;
+    border-collapse: collapse;
+    margin: 12px 0 18px;
+    font-size: 10pt;
+  }
+  th, td {
+    border: 1px solid #d8d8d8;
+    padding: 9px 11px;
+    text-align: left;
+    vertical-align: top;
+  }
+  thead th {
+    background: #052962;
+    color: #fff;
+    font-weight: 600;
+    text-transform: uppercase;
+    letter-spacing: 0.5px;
+    font-size: 9pt;
+  }
+  td.code {
+    font-weight: 700;
+    color: #052962;
+    width: 60px;
+    text-align: center;
+    background: #f9f6e9;
+  }
+  td.mod {
+    width: 220px;
+    font-weight: 600;
+  }
+
+  .pull {
+    background: #052962;
+    color: #fff;
+    padding: 18px 22px;
+    margin: 18px 0;
+    border-radius: 2px;
+    font-family: 'Playfair Display', Georgia, serif;
+    font-size: 14pt;
+    line-height: 1.4;
+  }
+  .pull .label {
+    display: block;
+    font-family: 'Inter', sans-serif;
+    font-size: 8pt; letter-spacing: 1.5px; text-transform: uppercase;
+    color: #ffe500; margin-bottom: 6px; font-weight: 700;
+  }
+
+  .stat-row {
+    display: flex; gap: 10px; margin: 16px 0;
+  }
+  .stat {
+    flex: 1; border: 1px solid #d8d8d8; padding: 14px; text-align: center;
+    background: #fcfaf2;
+  }
+  .stat .num {
+    font-family: 'Playfair Display', serif;
+    font-size: 22pt; font-weight: 800; color: #052962; line-height: 1;
+  }
+  .stat .lbl {
+    font-size: 8.5pt; color: #4a4a4a; margin-top: 6px; text-transform: uppercase;
+    letter-spacing: 0.5px;
+  }
+
+  .three-pillars {
+    display: flex; gap: 10px; margin: 12px 0;
+  }
+  .pillar {
+    flex: 1; padding: 12px; border-top: 4px solid #ffe500; background: #f7f7f5;
+  }
+  .pillar h4 {
+    font-family: 'Playfair Display', serif; margin: 0 0 6px;
+    font-size: 12pt; color: #052962;
+  }
+  .pillar p { font-size: 9.5pt; color: #333; margin: 0; }
+
+  .section { page-break-inside: avoid; }
+  .footer-band {
+    margin-top: 30px; padding: 18px;
+    background: #052962; color: #fff;
+    text-align: center;
+    font-family: 'Playfair Display', serif;
+    font-size: 14pt;
+    border-radius: 2px;
+  }
+  .footer-band em { color: #ffe500; font-style: normal; }
+
+  .small { font-size: 9pt; color: #6b6b6b; }
+</style>
+</head>
+<body>
+
+<!-- COVER -->
+<section class="cover">
+  <span class="kicker">Commercial overview</span>
+  <h1>Verify</h1>
+  <hr />
+  <p class="tagline">Tunisia's platform for verifying<br/>visual information</p>
+  <div class="meta">
+    Overview document · 2026
+  </div>
+</section>
+
+<!-- 1. THE PROJECT IN ONE SENTENCE -->
+<section class="section">
+  <h2>The project in one sentence</h2>
+  <p class="lede">
+    <strong>Verify</strong> is an independent platform for checking visual content
+    and its context (images, videos, captions, ads) that helps the general public,
+    journalists and Tunisian institutions tell, within minutes,
+    what is <strong>authentic</strong>, <strong>suspect</strong>,
+    <strong>manipulated</strong> or <strong>misleading</strong>.
+  </p>
+</section>
+
+<!-- 2. THE PROBLEM -->
+<section class="section">
+  <h2>The problem Verify addresses</h2>
+  <p>Visual disinformation is exploding in Tunisia and across the Arab world:</p>
+  <ul>
+    <li>Images and videos <strong>created with artificial-intelligence tools</strong> and passed off as real.</li>
+    <li>Advertising and <strong>propaganda</strong> videos that play on emotion to influence you more effectively.</li>
+    <li><strong>Misleading captions</strong> placed over genuine images to make them say something else.</li>
+    <li><strong>Retouched or composited photos</strong>: elements added, erased or moved.</li>
+    <li><strong>Misleading cosmetic ads</strong>: faked before/after shots, fake testimonials, unrealistic claims.</li>
+    <li><strong>Gaps between what you see and what you read</strong>: exaggerations, omissions, distortions.</li>
+  </ul>
+  <p>Today, the average Tunisian citizen has no simple tool, in French or Arabic,
+     to untangle all of this. <strong>Verify fills that gap.</strong></p>
+</section>
+
+<!-- 3. VALUE PROPOSITION -->
+<section class="section">
+  <h2>The value proposition</h2>
+  <div class="pull">
+    <span class="label">User promise</span>
+    "Drop in an image, a video or a post — get a clear, reasoned verdict in under 2 minutes."
+  </div>
+  <p>Verify combines three complementary strengths:</p>
+  <div class="three-pillars">
+    <div class="pillar">
+      <h4>1. In-depth analysis</h4>
+      <p>Six ways to check a piece of content, from the image itself to the words around it.</p>
+    </div>
+    <div class="pillar">
+      <h4>2. A human newsroom</h4>
+      <p>A team that contextualizes, investigates and publishes the evidence.</p>
+    </div>
+    <div class="pillar">
+      <h4>3. A consumer-facing media outlet</h4>
+      <p>Every verification is explained, archived and freely accessible.</p>
+    </div>
+  </div>
+</section>
+
+<!-- 4. THE 6 MODULES -->
+<section class="section">
+  <h2>The 6 analyses — the "Verify Toolkit"</h2>
+  <table>
+    <thead>
+      <tr>
+        <th>Code</th>
+        <th>Analysis</th>
+        <th>What it's for (in plain words)</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr>
+        <td class="code">IO1</td>
+        <td class="mod">Detection of AI-created images</td>
+        <td>Spot images and videos made with artificial-intelligence tools. We examine dozens of subtle clues, invisible to the naked eye, to deliver a clear verdict on their origin.</td>
+      </tr>
+      <tr>
+        <td class="code">IO2</td>
+        <td class="mod">Visual manipulation &amp; persuasion</td>
+        <td>Bring to light the emotional-persuasion and visual-manipulation techniques used in advertising and propaganda videos — so you understand how someone is trying to influence you.</td>
+      </tr>
+      <tr>
+        <td class="code">IO3</td>
+        <td class="mod">Image &amp; caption coherence</td>
+        <td>Check whether an image really matches the caption it's published under, and flag the cases where the caption makes the image say something else.</td>
+      </tr>
+      <tr>
+        <td class="code">IO4</td>
+        <td class="mod">Detection of retouched photos</td>
+        <td>Spot retouching and montages in a photo: elements added, erased or moved — and show exactly where in the image.</td>
+      </tr>
+      <tr>
+        <td class="code">IO5</td>
+        <td class="mod">Caption fidelity</td>
+        <td>Assess how faithfully a caption describes what you actually see: exaggerations, omissions, misleading shortcuts. <em>(Lead: Maryem)</em></td>
+      </tr>
+      <tr>
+        <td class="code">IO6</td>
+        <td class="mod">Cosmetic-ad fact-checking</td>
+        <td>Check whether a cosmetic ad's claims hold up, in light of the EU rules on cosmetic claims: faked before/after shots, fake testimonials, unrealistic claims.</td>
+      </tr>
+    </tbody>
+  </table>
+
+  <p style="margin-top:14px;"><strong>The six analyses cover the entire chain of visual disinformation:</strong></p>
+  <ul>
+    <li><strong>The image</strong> (IO1, IO4) — is it authentic or doctored?</li>
+    <li><strong>The intent</strong> (IO2, IO6) — how is someone trying to influence me?</li>
+    <li><strong>The meaning</strong> (IO3, IO5) — does what I'm told match what I see?</li>
+  </ul>
+</section>
+
+<!-- 5. USER JOURNEY -->
+<section class="section">
+  <h2>How it works for the user</h2>
+  <ol>
+    <li><strong>You drop in</strong> an image, a video, a post with its caption, or an ad.</li>
+    <li><strong>Verify picks for you</strong> the analyses best suited to that content.</li>
+    <li><strong>The content is scrutinized</strong> from every useful angle, and cross-checked against its context.</li>
+    <li><strong>A score out of 100</strong> and a clear verdict are returned: <em>Authentic / Suspect / Manipulated / Misleading</em>.</li>
+    <li><strong>When there's doubt</strong>, a human analyst takes over — no ambiguous verdict is ever delivered blindly.</li>
+    <li><strong>A public report</strong> is published: key moments of a video, retouched regions spotted in the image, a comparison between the caption and the visual, and a clear explanation. Everything is archived and viewable.</li>
+  </ol>
+</section>
+
+<!-- 6. THE MEDIA OUTLET -->
+<section class="section">
+  <h2>The media outlet that goes with the tool</h2>
+  <p>Verify isn't just a verification tool — it's also a <strong>news site</strong>
+     in the style of major international media, with:</p>
+  <ul>
+    <li>A daily <strong>editorial front page</strong>.</li>
+    <li><strong>In-depth investigations</strong> into visual and advertising disinformation.</li>
+    <li><strong>Fact-checks</strong> tracking Tunisian current affairs (politics, climate, economy, sport, culture, health).</li>
+    <li>A <strong>morning newsletter</strong>: <em>"The Verify brief — 5 minutes, zero spin, just facts."</em></li>
+    <li>A fully transparent <strong>Methodology</strong> section.</li>
+  </ul>
+</section>
+
+<!-- 7. TARGET AUDIENCES -->
+<section class="section">
+  <h2>Target audiences</h2>
+  <ul>
+    <li><strong>The connected citizen</strong> who wants to verify before sharing.</li>
+    <li><strong>Journalists and newsrooms</strong>: a fast tool to use before publishing.</li>
+    <li><strong>Teachers and trainers</strong> in media literacy.</li>
+    <li><strong>Public institutions</strong> facing manipulation campaigns.</li>
+    <li><strong>Brands and the cosmetics industry</strong> looking to detect advertising impersonations and misleading claims made in their name.</li>
+    <li><strong>Advertising regulators</strong> who need objective evidence.</li>
+  </ul>
+</section>
+
+<!-- 8. DIFFERENTIATORS -->
+<section class="section">
+  <h2>What sets Verify apart</h2>
+  <ul>
+    <li><strong>Complete coverage</strong>: from the image itself (IO1, IO4) to the meaning of the post (IO3, IO5), by way of the persuasive intent (IO2, IO6).</li>
+    <li><strong>A one-of-a-kind cosmetic-ad verification (IO6)</strong> — a sector especially exposed in Tunisia to misleading claims on social media.</li>
+    <li><strong>The first service of its kind built for the Tunisian and Maghreb context</strong>.</li>
+    <li><strong>Radical transparency</strong>: the method, the scores and the evidence behind the verdict are all published.</li>
+    <li><strong>Humans above automation</strong>: no ambiguous verdict goes out without human validation.</li>
+    <li><strong>Bilingual by nature</strong>: built for French- and Arabic-speaking audiences.</li>
+    <li><strong>Editorial independence</strong> from parties, platforms and advertisers.</li>
+  </ul>
+</section>
+
+<!-- 9. KEY FIGURES -->
+<section class="section">
+  <h2>Key figures</h2>
+  <div class="stat-row">
+    <div class="stat"><div class="num">412</div><div class="lbl">Verifications / week</div></div>
+    <div class="stat"><div class="num">38 %</div><div class="lbl">Content flagged as misleading</div></div>
+    <div class="stat"><div class="num">6</div><div class="lbl">Ways to check a piece of content</div></div>
+    <div class="stat"><div class="num">‹ 2 min</div><div class="lbl">Average time to a verdict</div></div>
+  </div>
+</section>
+
+<!-- 10. BRAND PROMISE -->
+<section class="section">
+  <h2>The brand promise</h2>
+  <div class="footer-band">
+    <em>See, read, verify, understand.</em><br/>
+    <span style="font-size:11pt; font-family:'Inter',sans-serif; line-height:1.5;">
+      Verify gives everyone the means to stop being at the mercy of the flow of images and captions:
+      to take back control of what you look at, what you read, what you share —
+      and what you choose to believe.
+    </span>
+  </div>
+  <p class="small" style="margin-top:24px; text-align:center;">
+    Verify · Commercial document · 2026 · Tunis
+  </p>
+</section>
+
+</body>
+</html>

Verify_Description_Commerciale.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56ad22c7176fa6cd1a1cd106de31e26f44a70612defd38039f4b82289a8fcb13
+size 251853

archive.html

@@ -0,0 +1,199 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+  <title>Verifications archive — Verify</title>
+  <link rel="preconnect" href="https://fonts.googleapis.com" />
+  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
+  <link href="https://fonts.googleapis.com/css2?family=Playfair+Display:wght@400;700;800&family=Inter:wght@400;500;600;700&display=swap" rel="stylesheet" />
+  <script src="https://cdn.tailwindcss.com"></script>
+  <link rel="stylesheet" href="assets/css/styles.css?v=32" />
+</head>
+<body>
+  <div id="site-header" data-slim="true"></div>
+
+  <main class="container-x">
+    <section class="px-6 pt-10 pb-4">
+      <div class="text-sm text-gray-500 mb-2">
+        <a href="index.html" class="hover:underline">Home</a> ›
+        <a href="real-or-fake.html" class="hover:underline">Fact-checks</a> ›
+        <span>Archive</span>
+      </div>
+      <h1 class="font-serif" style="font-size:48px;line-height:1.05;margin:0 0 8px;">Verifications archive</h1>
+      <p class="text-gray-600 max-w-2xl">
+        All analyses published by our team, sorted by module and verdict.
+        412 reports in total.
+      </p>
+    </section>
+
+    <!-- Filters -->
+    <section class="filters">
+      <span class="text-xs uppercase tracking-wider text-gray-500 self-center pr-2">Module:</span>
+      <button class="filter-chip active">All</button>
+      <button class="filter-chip">AI-Generated</button>
+      <button class="filter-chip">Persuasion</button>
+      <button class="filter-chip">Coherence</button>
+      <button class="filter-chip">Forensics</button>
+      <button class="filter-chip">Caption Fidelity</button>
+      <button class="filter-chip">Cosmetic Ads</button>
+    </section>
+    <section class="filters">
+      <span class="text-xs uppercase tracking-wider text-gray-500 self-center pr-2">Verdict:</span>
+      <button class="filter-chip active">All</button>
+      <button class="filter-chip">Authentic</button>
+      <button class="filter-chip">Suspect</button>
+      <button class="filter-chip">Manipulated</button>
+      <button class="filter-chip">Out of context</button>
+      <span class="text-xs uppercase tracking-wider text-gray-500 self-center pl-4 pr-2">Period:</span>
+      <button class="filter-chip">7 days</button>
+      <button class="filter-chip active">30 days</button>
+      <button class="filter-chip">2026</button>
+      <button class="filter-chip">All</button>
+    </section>
+
+    <!-- Table of past analyses -->
+    <section class="px-6 py-6">
+      <table class="data-table">
+        <thead>
+          <tr>
+            <th>Date</th>
+            <th>Reference</th>
+            <th>Title</th>
+            <th>Module</th>
+            <th>Score</th>
+            <th>Verdict</th>
+            <th></th>
+          </tr>
+        </thead>
+        <tbody>
+          <tr>
+            <td>Apr 27, 2026</td>
+            <td>VR-2026-0427-A93</td>
+            <td>Tunis flooding video (TikTok)</td>
+            <td>Coherence</td>
+            <td>34</td>
+            <td><span class="verdict-pill verdict-fake">Out of context</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 26, 2026</td>
+            <td>VR-2026-0426-B12</td>
+            <td>Viral "official" bank ad</td>
+            <td>Persuasion</td>
+            <td>22</td>
+            <td><span class="verdict-pill verdict-fake">Manipulated</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 26, 2026</td>
+            <td>VR-2026-0426-C45</td>
+            <td>Sfax protest photo</td>
+            <td>Forensics</td>
+            <td>48</td>
+            <td><span class="verdict-pill verdict-suspect">Suspect</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 25, 2026</td>
+            <td>VR-2026-0425-D78</td>
+            <td>Speech attributed to the Finance Minister</td>
+            <td>AI-Gen</td>
+            <td>67</td>
+            <td><span class="verdict-pill verdict-suspect">Suspect</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 24, 2026</td>
+            <td>VR-2026-0424-E33</td>
+            <td>Espérance image during the African title</td>
+            <td>Caption</td>
+            <td>91</td>
+            <td><span class="verdict-pill verdict-ok">Authentic</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 23, 2026</td>
+            <td>VR-2026-0423-F09</td>
+            <td>Municipal campaign poster (Sousse)</td>
+            <td>Forensics</td>
+            <td>54</td>
+            <td><span class="verdict-pill verdict-suspect">Suspect</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 22, 2026</td>
+            <td>VR-2026-0422-G22</td>
+            <td>"Photo" of Tunis-Carthage airport</td>
+            <td>AI-Gen</td>
+            <td>14</td>
+            <td><span class="verdict-pill verdict-fake">Manipulated</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 21, 2026</td>
+            <td>VR-2026-0421-H56</td>
+            <td>"New school opening" video in Kairouan</td>
+            <td>Caption</td>
+            <td>88</td>
+            <td><span class="verdict-pill verdict-ok">Authentic</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 20, 2026</td>
+            <td>VR-2026-0420-I91</td>
+            <td>Screenshot of an official statement</td>
+            <td>Persuasion</td>
+            <td>40</td>
+            <td><span class="verdict-pill verdict-suspect">Suspect</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 19, 2026</td>
+            <td>VR-2026-0419-J04</td>
+            <td>Real-estate ad (Instagram account)</td>
+            <td>Persuasion</td>
+            <td>30</td>
+            <td><span class="verdict-pill verdict-fake">Manipulated</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 18, 2026</td>
+            <td>VR-2026-0418-K17</td>
+            <td>Video shot in La Marsa (verified account)</td>
+            <td>AI-Gen</td>
+            <td>94</td>
+            <td><span class="verdict-pill verdict-ok">Authentic</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+          <tr>
+            <td>Apr 17, 2026</td>
+            <td>VR-2026-0417-L88</td>
+            <td>AI-generated landscape image (Tozeur)</td>
+            <td>AI-Gen</td>
+            <td>11</td>
+            <td><span class="verdict-pill verdict-fake">Manipulated</span></td>
+            <td><a href="resultats.html" class="text-navy underline">View</a></td>
+          </tr>
+        </tbody>
+      </table>
+    </section>
+
+    <div class="flex justify-center gap-2 py-10">
+      <button class="filter-chip">‹ Previous</button>
+      <button class="filter-chip active">1</button>
+      <button class="filter-chip">2</button>
+      <button class="filter-chip">3</button>
+      <button class="filter-chip">…</button>
+      <button class="filter-chip">35</button>
+      <button class="filter-chip">Next ›</button>
+    </div>
+  </main>
+
+  <div id="site-footer"></div>
+
+
+  <script src="assets/js/config.js?v=1"></script>  <script src="assets/js/components.js?v=14"></script>
+  <script src="assets/js/main.js?v=7"></script>
+</body>
+</html>

article.html

@@ -0,0 +1,161 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+  <title>Tunis floods: our verification — Verify</title>
+  <link rel="preconnect" href="https://fonts.googleapis.com" />
+  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
+  <link href="https://fonts.googleapis.com/css2?family=Playfair+Display:wght@400;700;800&family=Inter:wght@400;500;600;700&display=swap" rel="stylesheet" />
+  <script src="https://cdn.tailwindcss.com"></script>
+  <link rel="stylesheet" href="assets/css/styles.css?v=32" />
+</head>
+<body>
+  <div id="site-header" data-slim="true"></div>
+
+  <main class="container-x" style="max-width: 860px;">
+    <article class="px-6 py-10">
+      <!-- Breadcrumb -->
+      <div class="text-sm text-gray-500 mb-4">
+        <a href="index.html" class="hover:underline">Home</a> ›
+        <a href="real-or-fake.html" class="hover:underline">Fact-checks</a> ›
+        <span>Tunis</span>
+      </div>
+
+      <span class="cat-tag tag-fact" style="display:inline-block;font-size:11px;font-weight:700;text-transform:uppercase;letter-spacing:0.8px;padding:3px 10px;border-radius:2px;margin-bottom:12px;">Fact-check</span>
+      <h1 class="font-serif" style="font-size:46px;line-height:1.1;margin:8px 0 16px;">
+        Is this viral video of flooding in Tunis authentic?
+        Our frame-by-frame verification
+      </h1>
+      <p style="font-size:20px;color:#4a4a4a;line-height:1.5;">
+        The 47-second clip topped 800,000 views on X and TikTok in two
+        days. Our Image–Caption Coherence and Edited Photo Detection
+        analyses reconstructed the origin of every shot.
+      </p>
+
+      <div class="flex items-center gap-3 mt-6 mb-8 pb-6 border-b" style="border-color:var(--line);">
+        <div style="width:48px;height:48px;border-radius:50%;background:var(--rose);display:flex;align-items:center;justify-content:center;font-weight:800;color:var(--navy);font-family:'Playfair Display',serif;font-size:20px;">M</div>
+        <div>
+          <div class="font-semibold">Maryem Ben Slimane</div>
+          <div class="text-sm text-gray-500">Lead — Caption Fidelity · Published April 27, 2026 at 9:42 AM</div>
+        </div>
+      </div>
+
+      <!-- Hero image -->
+      <img src="https://images.unsplash.com/photo-1547683905-f686c993aae5?auto=format&fit=crop&w=900&q=80" alt="Urban flooding" style="width:100%;height:auto;display:block;margin-bottom:20px;" />
+      <p class="text-sm text-gray-500 mb-8" style="font-style:italic;">
+        Screenshot of the viral video as it circulated on April 26 on X. Credit: "Tunisie en direct" account.
+      </p>
+
+      <!-- Smart Brevity boxes -->
+      <div class="sb-box why">
+        <span class="sb-label">Why it matters</span>
+        During extreme climate events, decontextualized videos amplify fear
+        and saturate emergency services. Verifying quickly and publishing
+        the chain of origin is the first line of defense.
+      </div>
+
+      <p style="font-size:17px;line-height:1.7;margin-bottom:18px;">
+        The video, first posted on Tuesday at 9:17 PM, shows submerged
+        streets presented as filmed "tonight in Tunis." Within 90 minutes,
+        the clip had been picked up by five pan-Arab accounts totaling 4.2
+        million followers. Our team began analysis at 10:04 PM.
+      </p>
+
+      <div class="sb-box big">
+        <span class="sb-label">The big picture</span>
+        Three key shots — out of the twelve in the video — come from a
+        separate flood event in Bab Bhar, in September 2018. Two new shots
+        were authenticated in Manouba this morning, around 6:12 AM.
+      </div>
+
+      <h2 class="font-serif" style="font-size:30px;margin:30px 0 14px;">What our modules found</h2>
+      <ul class="sb-bullets">
+        <li><strong>Image–Caption Coherence</strong> noted a 3,200 K luminance
+          shift between shots 4–7 and the rest: the classic signature of
+          nighttime archive footage re-injected into a daytime video.</li>
+        <li><strong>Edited Photo Detection</strong> isolated two added objects —
+          a road sign and a pharmacy sign — overlaid on 2018 frames.</li>
+        <li>The <strong>EXIF metadata</strong> was stripped on the viral
+          version, but the version posted 16 minutes earlier by another
+          account still carries a timestamp dated <strong>September 14, 2018</strong>.</li>
+        <li>By contrast, two shots (8 and 11) are <strong>authentic and unpublished</strong>:
+          they correspond to a local flood event that occurred this morning in Manouba.</li>
+      </ul>
+
+      <h2 class="font-serif" style="font-size:30px;margin:30px 0 14px;">How we established this</h2>
+      <p style="font-size:17px;line-height:1.7;margin-bottom:18px;">
+        The Verify verification process combines a reverse image search
+        (Yandex, TinEye, Google Lens) with our own visual analyses. For this
+        video, the first archived match was found in 4 minutes via TinEye
+        on frame 156. It pointed to a Mosaïque FM article published on
+        September 15, 2018.
+      </p>
+
+      <p style="font-size:17px;line-height:1.7;margin-bottom:18px;">
+        Once the source was identified, our analysts aligned each shot of
+        the viral video with the original archive, shot by shot. Shots 8
+        and 11 produced no match: they were then submitted to AI-Generated
+        Media Detection, which returned a confidence score &gt; 92% in
+        favor of authenticity.
+      </p>
+
+      <div class="sb-box next">
+        <span class="sb-label">What's next</span>
+        Our newsroom is opening an investigation into the Telegram channels
+        that first posted the video, and into the modus operandi — the same
+        archived shots were used last March for a false alert in Sfax.
+        Results expected on May 5.
+      </div>
+
+      <!-- CTA banner -->
+      <div class="verify-banner">
+        <div>
+          <h3>Want to verify suspicious content yourself?</h3>
+          <p>Upload an image, video or link — our five modules give you a score in less than 2 minutes.</p>
+        </div>
+        <a href="verifier.html" class="btn-yellow">Verify this content</a>
+      </div>
+
+      <p style="font-size:17px;line-height:1.7;margin-bottom:18px;">
+        The stakes go beyond this one case. Over the past seven days,
+        Verify processed 421 user-flagged contents. 38% were marked as
+        manipulated, 17% as out of context. The rest — the majority —
+        is authentic: verification isn't just about debunking, it's also
+        about confirming.
+      </p>
+    </article>
+
+    <!-- Related articles -->
+    <section class="px-6 pb-12">
+      <h2 class="font-serif text-3xl pb-2 mb-4 border-b">More to read</h2>
+      <div class="grid grid-cols-1 md:grid-cols-3 gap-6">
+        <article class="article-card">
+          <a href="article.html"><img src="https://images.unsplash.com/photo-1521791136064-7986c2920216?auto=format&fit=crop&w=900&q=80" alt="" /></a>
+          <span class="cat-tag tag-fact">Fact-check</span>
+          <a href="article.html"><h3>Was this Sfax protest photo manipulated?</h3></a>
+          <p>Two retouching traces flagged by Edited Photo Detection.</p>
+        </article>
+        <article class="article-card">
+          <a href="article.html"><img src="https://images.unsplash.com/photo-1554224155-1696413565d3?auto=format&fit=crop&w=900&q=80" alt="" /></a>
+          <span class="cat-tag tag-fact">Fact-check</span>
+          <a href="article.html"><h3>"Official" bank ad: AI-generated?</h3></a>
+          <p>Cosmetic Ads Fact-Check confirms a partial voice synthesis.</p>
+        </article>
+        <article class="article-card">
+          <a href="article.html"><img src="https://images.unsplash.com/photo-1605792657660-596af9009e82?auto=format&fit=crop&w=900&q=80" alt="" /></a>
+          <span class="cat-tag tag-clim">Climate</span>
+          <a href="article.html"><h3>Tunisia faces its worst water stress since 1947</h3></a>
+          <p>Reservoirs at 23%, six-point emergency plan.</p>
+        </article>
+      </div>
+    </section>
+  </main>
+
+  <div id="site-footer"></div>
+
+
+  <script src="assets/js/config.js?v=1"></script>  <script src="assets/js/components.js?v=14"></script>
+  <script src="assets/js/main.js?v=7"></script>
+</body>
+</html>