Truth Pipe

A local Python system that preserves provider evidence, classifies explicit truth states, and surfaces uncertainty without creating false certainty.

Released

Jun 21, 2026

Status

Active

Source

GitHub

Key features

• Runs as a local FastAPI service and a button-driven dashboard. There is no hosted app.
• GET /state reads one source, a mock fixture or a live Alpaca paper account. GET /state/conflict compares Alpaca with one mock fixture.
• Classifies one snapshot for a missing field, data older than 300 seconds, and a suspicious zero. Precedence picks the truth state. It does not score.
• Returns UNKNOWN on a cold process until a value has been confirmed. The cache is in memory and is written only on CONFIRMED.
• Attention sits after that decision and can change what is surfaced without reopening how trust was decided.

How to use

Python 3.10 or newer. From the repository root. Mock fixtures need no credentials. Install, then start the local service:

pip install -r requirements.txt
uvicorn truthpipe_demo.app:app --reload --port 8000

Open http://127.0.0.1:8000/. Press Confirm before the degraded balance buttons if you want the last confirmed value filled in. A degraded button on a cold engine returns UNKNOWN. For live Alpaca, copy .env.example to .env and set ALPACA_API_KEY, ALPACA_API_SECRET, and ALPACA_BASE_URL. If those are missing, the adapter raises before it sends the request. python run_pipeline_demo.py prints the classifier for four balance fixtures and then Alpaca, and does not apply the attention step.

Related work