How mozo proves it matches the original models
Mozo does not wrap the original packages; it carries a copy of each model's inference code. A copy can quietly drift from the original, so every family is checked against the implementation it came from, and the check uses no tolerance: two outputs pass only if they are exactly equal. A tolerance would hide exactly the small drift the check exists to catch.
What is compared
Not only the final answer. A gate compares every intermediate stage against the original, on real inputs: the preprocessed tensor, each block of the network, the decoded output. A wrong stage is then found where it goes wrong, not inferred from a final number that happens to look close.
| Family | Compared against | Comparisons |
|---|---|---|
| EasyOCR | the easyocr package |
1,275 across five variants, from the preprocessed tensor to each decoded string |
| OWLv2 | transformers |
226 on base-ensemble: tokenizer, preprocessing, both towers, three heads, boxes |
| Grounding DINO | the authors' implementation | 138 across both variants, from the tokenizer to the final logits and boxes |
| CLIP | openai/CLIP |
104 across four variants, from preprocessing to logits_per_image |
| Granite Embedding R2 | transformers |
1,153,207: every stage of the model on 20 texts, and the tokenizer on 1.15 million texts, on Python 3.10 to 3.14 |
Fifteen of the eighteen families ship a gate in tools/verify/ in mozo's repository. RF-DETR,
SAM 2 and ViTPose have their parity measured and recorded the same way, in the files below.
The stated exceptions
Where a family is not bit-identical everywhere, its page says so and why:
- ViTPose joints land within 0.005 pixels of
transformers. The heatmaps are bit-identical; the difference is PixelFlow rounding coordinates to 0.01. - EasyOCR on a GPU: strings and boxes are exact, confidences move by up to 2.2e-05. The gate is
a CPU claim; pass
device="cpu"for exactly its numbers. - Grounding DINO reports the phrase you asked for as
class_name, where upstream decodes the matching tokens back into a string that can be cut short. That is the one deliberate difference in any family, and it is in a string, not in a number. - ONNX and CoreML graphs are separate artifacts, checked against the torch model before they are published, within a stated bound: for RF-DETR, the same detections with boxes within 0.2 px. Each family's page states its bound. The torch path is the one compared bit for bit.
Being a copy rather than a wrapper costs nothing at run time: EasyOCR runs within 1% of the published package on the same weights, and Granite Embedding R2 at the reference's speed.
Where it is recorded
Every vendored model in mozo/vendors/<family>_deploy/ has a PROVENANCE.md, which records:
- the exact upstream commit the code was extracted from, and what was changed or left out;
- the parity measured, and how;
- for the traps a code review would miss, what they were. CLIP's activation is QuickGELU, not the standard GELU, and substituting it moves the features by 0.4 without an error. Granite's checkpoint is stored in bfloat16 and loads that way by default, 3e-03 away from float32.
The weights are published with their licence and a NOTICE naming the upstream release they came from, and every file is checked against a sha256 recorded in the package before mozo uses it.
Run a check yourself
The gates need mozo's repository and the original package, not a network beyond the downloads:
git clone https://github.com/emrahnazif/mozo
cd mozo
pip install -e . transformers
python tools/verify/owlv2.pyEach gate prints its comparisons and ends with PASS, or fails if any is not identical. Some need a checkout of the upstream repository instead; the top of each gate says what it needs.