cvconform
the correctness and reliability layer for computer vision

---

Computer vision has a hidden reliability crisis. A model is trained once (PyTorch, CUDA, FP32) and then silently transformed — ONNX, TensorRT, CoreML, OpenVINO, TFLite, quantization, operator fusion. Every transformation creates opportunities for the model to stop being itself.

The industry mostly asks "Can this model run?"

cvconform answers the question nobody was asking: "Is this still the same model?" It is differential conformance verification for vision models — execute the same model on multiple runtimes with the same seeded inputs, compare the outputs, and when they differ, know exactly why.

✨ Zero config

cvconform works on a fresh clone with no configuration.

pip install cvconform
cd your-model-repo

cvconform init      # auto-detects your model + contract, writes .cvconform.yaml
cvconform verify    # auto-runs every installed runtime, prints a conform report

It auto-discovers model format (torchscript, onnx, coreml, tflite, tensorrt engine, mlx, openvino), input shape, output names, and a calibration set — from the artifact itself or a registry of known architectures (yolo, sam, detr, resnet, vit, clip, unet, pose, ocr, …), with a sensible heuristic fallback.

Friction demo — measured time to first green check

cvconform init + bare cvconform verify --targets onnx on three fresh repos (Apple M4):

Family	Model	Time	Result
Detection	yolov8n.pt	2.3s	ONNX 100%
Segmentation	unet_seg.pt	4.6s	ONNX 100%
Classification	resnet18.pt	1.4s	ONNX 100%

Install

pip install cvconform                # core: numpy, onnx, onnxruntime, pyyaml
pip install 'cvconform[pytorch]'     # PyTorch reference backend
pip install 'cvconform[coreml]'      # CoreML target (macOS)
pip install 'cvconform[openvino]'    # OpenVINO target
pip install 'cvconform[all]'         # every optional backend

Requires Python 3.10+. On Apple Silicon you get CoreML + MLX paths for free.

Usage

Python API

from cvconform import verify

result = verify(
    model="yolo11.pt",
    reference="pytorch",
    targets=["onnx", "coreml"],
    seed=0,
)

CLI

# Clear a model — auto-discover everything else
cvconform verify

# Explicit
cvconform verify model.pt --reference pytorch --targets onnx,coreml

# Machine-readable report for CI/CD
cvconform verify model.pt --json report.json

# CI gate: non-zero exit if any runtime diverges
cvconform verify --require-conformant

# Pre-commit fast mode
cvconform verify --pre-commit

# Store findings in your regression-memory corpus
cvconform verify --corpus corpus/

Output

VISION CONFORM REPORT
Model:       yolo11.pt
Reference:   pytorch

CONFORMANCE SCORES
  onnx            100.00%
  coreml           96.36%

FINDINGS: none — all targets conformant within policy.

When a runtime does diverge, cvconform tells you why — not "probably a bug", but:

CRITICAL FINDING
Backend:       CoreML
Affected:      Conv+Activation fusion
Cause:         FP16 accumulation drift
Impact:        37 production images produce different classes
Confidence:    96%
Fixes:         1. Disable operator fusion
               2. Force FP32 accumulation
               3. Report upstream issue

Extras

Feature	How
CI gate	--require-conformant / --pre-commit exit non-zero on divergence
Pre-commit hook	one-line install — see .pre-commit-hooks.yaml
GitHub Action	composite action installs → verifies → comments a badge — see .github/actions/verify
Regression memory	every failure becomes a permanent, versioned corpus entry
Failure discovery	hunts for diverging inputs (noise, edges, illumination, blur, compression, extremes)
Failure reduction	shrinks a failing image to a minimal repro

The science behind it

Every result is reproducible and evidence-driven:

• seedable inputs → identical reports for identical seeds
• recorded environment + exact backend versions
• evidence objects: observed X at node Y, magnitude Z, likely mechanism M, confidence C%
• a permanent regression corpus so cvconform gets smarter about how vision models fail in the real world

What "conformance" means (and doesn't)

cvconform verifies deployment conformance: that a compiled artifact (ONNX, CoreML, …) still behaves like your reference model. It answers "is this still the same model?", not "is this model accurate?".

• accurate-but-broken-by-export → detected
• inaccurate-but-faithfully-exported → conformant (that's a calibration problem, out of scope here)

Architecture

loaders/    framework -> VisionGraph IR
ir/         VisionGraph, operators, precision, passes, registry
autodetect/ zero-config model + calibration discovery
config/     .cvconform.yaml read/write
runtimes/   versioned wrappers (onnxruntime, pytorch, coreml, ...)
engines/    differential, comparison, analyze (root cause)
discovery/  input generation + expedition
reduce/     failing-input minimization
corpus/     regression memory
research/   AI research agent
report/     human + JSON conformance reports
cli.py      init / verify / discover

Development

uv venv --python 3.12 .venv
source .venv/bin/activate
python -m pytest

Typed Python with dataclasses, conventional commits, a green pytest run before any PR. See docs/report-schema.md for the machine-readable CI contract and CHANGELOG.md for release history.

License

MIT © David Nichols