v0.1.0 — Initial public release

First public release of ghostfail — a Python library for detecting silent AI failures.

What it does

Runs three heuristic detectors on every request-response turn, optionally escalates to an LLM judge, persists results to SQLite, and alerts via webhook or log:

• Confidence trap — claim extraction + source grounding catches responses that sound confident but contradict source material
• Drift — semantic similarity + keyword decay catches multi-turn conversations that silently wander off-topic
• Mismatch — entity overlap + TF-IDF catches responses that answer the wrong question

Scoring

Composite = 0.75 × highest + 0.25 × second-highest, with single-flag dampening (exempt for source-grounded confidence detector) to reduce noisy individual-detector false positives.

Quality

• 179 tests passing, 0 warnings
• 30-case calibration suite: 75% precision / 90% recall / 81.8% F1
• Recovery sweep for dropped and partial evaluations
• Webhook retry with pending-thread tracking on shutdown
• Prometheus-style + 'metrics()' + export with operational alert callbacks
• Flask dashboard with Bearer token auth (HMAC compare)

Install

+ '```bash' +

Clone for now — PyPI package coming later

git clone https://github.com/aevryone/ghostfail
cd ghostfail
pip install -e .

Optional extras

pip install -e '.[anthropic]' # LLM judge
pip install -e '.[embeddings]' # sentence-transformers for drift
+ '```' +

See the README for usage, config, and the full detection pipeline walkthrough.