Bonafide

A calibrated, provenance-first framework for detecting AI-generated media — honest about what it knows, and about what it doesn't.

In an age of deepfakes and synthetic everything, the question "was this made by a human or by AI?" matters more each day. The existing detectors answer it with confident binaries that are wrong often enough — and in biased enough ways — that their own customers are turning them off.

Bonafide takes a different bet. It treats detection as evidence aggregation under uncertainty: it fuses hard provenance (C2PA / Content Credentials), watermarks (SynthID), and an ensemble of ML detectors into one calibrated probability with a confidence interval and a readable evidence report — and it abstains when the evidence doesn't support a confident call. Evidence, never an accusation.

• The moat: a Bayesian evidence-fusion + calibration engine that combines cryptographic proof and noisy ML scores into one honest, auditable verdict.
• The architecture: ports-and-adapters, so any detector for any medium plugs in — text and images first, audio and video as drop-in adapters later, with zero core changes.
• The wedge: the only detector you can actually trust — plus a regulatory tailwind (EU AI Act Article 50, effective 2 Aug 2026).

See DESIGN.md for the full technical design.

Status: M2a — a real, trust-verified C2PA / Content Credentials signal for images (M1), plus a real calibration + conformal-abstention core with an evaluation harness and bonafide fit / bonafide eval (M2a). The neural text ensemble lands in M2b, so text still uses a placeholder heuristic and mostly returns abstain. See ROADMAP.md for the plan.

Run it

Prerequisites: Python ≥ 3.11 and uv (check: python --version, uv --version).

uv venv                                     # create .venv
uv pip install -e ".[dev,provenance]"       # install (core is pure-Python; ML is an extra)

# CLI — analyze an image's Content Credentials, or some text
uv run bonafide detect ./photo.jpg
uv run bonafide detect --text "The quick brown fox jumps over the lazy dog near the bridge."
uv run bonafide detect ./photo.jpg --json

# API — starts on http://localhost:8000 (interactive docs at /docs)
uv run uvicorn bonafide.service.app:app --reload

A C2PA-signed image gives a real, explained verdict:

Bonafide verdict: 100% AI-generated (CI 99-100%)  |  decision: AI
modality: image  |  engine: bonafide/0.0.1+fuse-logodds-v0  |  coverage: 95%

Evidence:
  [+] c2pa.manifest        conclusive llr +6.00  x0.95  -> +5.70   (basis=digitalSourceType:
      trainedAlgorithmicMedia, validation_state=Trusted, signer=Adobe Inc., ...)

Conflicts: none

Everything else still answers honestly: an unsigned image, a self-signed "a camera took this" spoof, and ordinary text all return abstain, because Bonafide refuses to guess. The text ensemble and real calibration arrive in M2.

Commands

Optional extras (installed per milestone)

Project docs

Tech stack

Python 3.11+ · Pydantic v2 · FastAPI + Uvicorn · Typer CLI · uv + hatchling. ML/provenance (c2pa-python, transformers, torch, scikit-learn, conformal) load as optional extras.

License

MIT — see LICENSE.