# DINO detector mechanics Run `python3 examples/dino_detr.py` from the repository root with Python 3.9+. The default generator filename is `dino_detr.py`. It prints deterministic JSON to stdout, writes no files and needs only the standard library. Run `node --test test/dinoDetr*.test.mjs` for mathematical and Python parity checks. This reference demonstrates mixed top-K anchor selection with independent learned content, positive/negative corner perturbations, directional denoising masks, logit-space box refinement and a two-layer scalar gradient graph. Regions `queries`, `denoising`, `mask`, `refinement`, `gradients` are extracted by the repository synchronizer. Its parser uses `# region: name` / `# endregion: name`. Root integration owns generated `src/data/dinoDetrCode.mjs` and public downloads. The source is an authored illustration, not copied training code and not a checkpoint-backed detector. Fixed perturbations are examples of the official corner-noise parameterization. They do not implement random label corruption, variable-sized batches, padding or dynamic DN group allocation. Negatives receive background classification supervision, not a box reconstruction target. The mask returns **allowed** booleans, the inverse of PyTorch's blocked mask. Rows read columns. Matching queries cannot read denoising queries; the reverse is allowed, and denoising groups cannot read each other. The scalar graph instantiates the dependency structure of paper Eq. 2 with an authored tanh offset head and half squared error. It holds the stopped reference fixed during finite differences. It is neither a full autodiff engine nor an empirical gradient trace from DINO training. Only the additional adjacent-loss path is isolated; decoder hidden-state gradient paths are omitted. These numbers cannot establish AP, convergence, robustness or runtime. Primary references: [DINO v4 §§3.3–3.5, Eq. 2](https://arxiv.org/html/2203.03605v4), [official CDN preparation](https://github.com/IDEA-Research/DINO/blob/main/models/dino/dn_components.py), [transformer](https://github.com/IDEA-Research/DINO/blob/main/models/dino/deformable_transformer.py) and [prediction heads](https://github.com/IDEA-Research/DINO/blob/main/models/dino/dino.py). DINO detector is ICLR 2023, preprint 2022, distinct from self-supervised DINO. No inference, training, checkpoint download or timing experiment is performed.