ALETHEX — Temporal Belief Consistency Engine

Catch AI self-contradictions in real time — Chrome extension + Python engine that watches AI chat sessions and flags when the AI says something that contradicts what it said before.

---

What Is ALETHEX?

ALETHEX is two things:

1. Chrome Extension — loads silently alongside your AI chats (ChatGPT, Claude, Gemini, Grok, Copilot) and shows a live Consistency Index badge. When the AI contradicts itself, ALETHEX flags it immediately.

2. Python Library — a full Temporal Belief Graph engine for researchers and developers building contradiction-aware RAG pipelines, LangChain memory, or LlamaIndex filters.

---

Chrome Extension

Install

1. Download dist/alethex-extension-v1.0.0.zip and unzip it
2. Open chrome://extensions/ and enable Developer mode (top right)
3. Click Load unpacked and select the unzipped extension/ folder
4. Visit ChatGPT, Claude, or Gemini — the ALETHEX badge appears automatically

How It Works

Every AI reply
      |
  [Fast regex pass — ~0ms]
  Extract explicit facts (name, location, job, diet, age, etc.)
      |
  [NLI semantic pass — ~2-5s]
  Build all sentence pairs across the full conversation
  Run DeBERTa-v3-Large ONNX contradiction detection in-browser
      |
  [Consistency Index]
  CI = 1.0 - (conflicts + superseded x 0.5) / totalClaims
  ─────────────────────────────────────────────────────────
  1.0      = Fully consistent       (green)
  0.5-0.99 = Facts updated          (amber)
  < 0.5    = Active contradictions  (red)

The NLI model runs 100% in-browser via Transformers.js — no server, no API calls. Downloaded once (~641MB), cached permanently.

Example Detections

Earlier statement	Later statement	Detected
"I live in New York."	"I moved to London last year."	contradiction (0.97)
"I am vegetarian."	"I had steak for dinner."	contradiction (0.94)
"I work at Google."	"My job at Microsoft keeps me busy."	contradiction (0.91)
"I am 25 years old."	"I turn 30 next week."	contradiction (0.88)

---

NLI Models on HuggingFace

Both models trained for 3 epochs on a unified 1M+ pair corpus (MultiNLI + SNLI + Temporal-NLI), no synthetic data.

Model	Parameters	ONNX size	Accuracy	Link
alethex-nli-deberta-v3-large (default)	400M	641 MB	92.8%	HuggingFace
alethex-nli-deberta-v3-small (lightweight)	22M	172 MB	92.8%	HuggingFace

Training Progression

Epoch	Accuracy	Macro F1	MCC
1	91.94%	91.94%	0.879
2	92.51%	92.52%	0.887
3	92.80%	92.81%	0.892

Large model trained on 8x NVIDIA H100 80GB (BF16, effective batch size 2048). Small model trained on RTX 4050 Laptop GPU.

---

Python Library

Install

pip install alethex

Or from source:

git clone https://github.com/codewithyug/alethex.git
cd alethex
pip install -e .

Optional extras:

pip install "alethex[langchain]"    # LangChain integration
pip install "alethex[llamaindex]"   # LlamaIndex integration
pip install "alethex[dev]"          # dev tools (pytest, ruff, mypy)

Quickstart

import alethex

memories = [
    {"id": "m1", "timestamp": "2024-01-10", "text": "Alice works at Google in London."},
    {"id": "m2", "timestamp": "2024-06-15", "text": "Alice relocated to Tokyo and joined Anthropic."}
]
filtered = alethex.filter_context(memories)
# -> returns only m2 (m1 is superseded)

Online Session Tracking

from datetime import datetime
from alethex.streaming.online_session import OnlineBeliefTracker

tracker = OnlineBeliefTracker()
tracker.process_turn("user", "I live in London and work at Google.", datetime(2024, 1, 15))
tracker.process_turn("user", "I moved to Tokyo last week.", datetime(2024, 7, 1))

print(tracker.get_active_beliefs("user"))
print(tracker.get_current_ci())  # 0.75 — superseded fact detected

LangChain / LlamaIndex

# LangChain — drop-in reconciled memory
from alethex.integrations.langchain_memory import AlethexReconciledMemory

# LlamaIndex — post-processor that drops or annotates contradicted context
from alethex.integrations.llamaindex_filter import AlethexConsistencyFilter

CLI

python -m alethex.cli run-pipeline --corpus dataset/corpus.jsonl --output-dir results/
python -m alethex.evaluation.eval_pipeline --samples 150

---

Repository Structure

alethex/
|
+-- alethex/                    # Python package
|   +-- ingestion/              # Document loaders and schema
|   +-- extraction/             # Claim extraction (OpenIE + LLM)
|   +-- coreference/            # Cross-document entity linking
|   +-- temporal/               # TIMEX parsing and interval algebra
|   +-- relation/               # NLI classifier and LLM judge
|   +-- graph/                  # Temporal belief graph (MultiDiGraph)
|   +-- scoring/                # Consistency Index calculation
|   +-- optim/                  # ONNX export and INT8 runtime
|   +-- training/               # NLI fine-tuning trainer
|   +-- streaming/              # Incremental OnlineBeliefTracker
|   +-- integrations/           # LangChain, LlamaIndex, MCP, OpenAI proxy
|   \-- evaluation/             # Benchmarking suite
|
+-- extension/                  # Chrome extension (MV3 source)
|   +-- manifest.json
|   +-- background.js           # Service worker
|   +-- content.js              # Chat page observer
|   +-- nli_worker.js           # NLI Web Worker (Transformers.js)
|   +-- popup.html / popup.js   # Extension popup UI
|   \-- vendor/                 # Bundled transformers.min.js (no CDN)
|
+-- configs/                    # Training and inference YAML configs
+-- dataset/                    # Preprocessed NLI data (gitignored — large files)
+-- docs/                       # Documentation and training reference
+-- experiments/                # Evaluation scripts and experiments
+-- models/                     # Model checkpoints and ONNX (gitignored — large files)
|   +-- alethex-nli/            # Small model checkpoints
|   +-- alethex-nli-large/      # Large model (H100 trained, 1.7GB)
|   +-- alethex-nli-onnx/       # Small ONNX (172MB quantized)
|   \-- alethex-nli-large-onnx/ # Large ONNX (641MB quantized)
+-- paper/                      # Draft paper and figures
+-- results/                    # Training logs and benchmark reports
+-- scripts/                    # Utility scripts
|   +-- export_onnx.py          # Export checkpoint to ONNX + INT8
|   +-- analyze_training_results.py
|   +-- start_all.py
|   \-- alethex_desktop.py      # Desktop companion (MCP auto-attach)
+-- tests/                      # Test suite
+-- dist/                       # Packaged extension zip for download
+-- pyproject.toml
+-- requirements.txt
\-- LICENSE

---

Export a Model to ONNX

python scripts/export_onnx.py \
  --checkpoint models/alethex-nli-large \
  --out-dir models/alethex-nli-large-onnx

---

Testing

python -m pytest tests/ -v

---

Citation

@misc{alethex2026,
  title  = {ALETHEX: Temporal Belief Consistency Engine for AI Chat Contradiction Detection},
  author = {codewithyug},
  year   = {2026},
  url    = {https://github.com/codewithyug/alethex}
}

---

License

Apache 2.0 — free for commercial and research use.