Per-patch differential entropy on standardised residual streams of ViT-style models, treated as a 2-D (or 3-D in video) scalar field on the patch grid.

Headline finding so far. Across vit-base-patch16-224, vit-large-patch16-224, dinov2-base, and clip-vit-base-patch16 (all loaded ungated from HuggingFace, no fine-tuning), per-image transition layer = argmax_L (ΔH-std) is rock-stable (IQR = 0 across 200 tiny-imagenet val images per arch), the transition is sharp (peak/median 4×–21×), and its location is gated by training objective rather than universal: supervised classification ≪ self-supervised / contrastive in L_trans / L_max. Per- patch entropy beats per-patch L2 norm, attention-rollout entropy, and random-init residual entropy at p < 0.001 on outlier-localisation accuracy. The phase transition is robust to entropy-estimator choice (Vasicek / k-NN / KDE all agree within ±1 layer on ViT-B/16) and has a clean geometric correlate: under per-patch standardisation, low-H "outlier" patches sit 8-20× tighter to each other in residual space than the bulk, forming a discrete attractor. On video the transition layer is frame-stationary (std = 0 across all frames in both ViT-B/16 and DINO-v2), but volumetric outlier-tube tracking on a synthetic translating blob fails for ViT-B/16 due to register-token edge dominance (Darcet et al. 2024). Native-3D V-JEPA 2 escapes the spatial register-token artefact at shallow layers (tube IoU = 0.24 at L=2 vs ViT-B/16's 0.00) but injects a temporal-PE pattern that overwhelms content motion at mid-network — when V-JEPA 2 is run on 64 identical frames vs five 64-frame Something-Something v2 eval clips, the constant-content stimulus exhibits 2-3× larger per- tubelet temporal variance in mid-network residuals than every actual moving SSv2 clip.

See results/summary.md for the full read-out and results/ for figures and tensors.

uv sync
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/synth_check.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/cross_arch.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/video_frames.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/video_volume.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/video_volume_bundle.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/baselines.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/estimator_robustness.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/umap_embeddings.ipynb
uv run jupyter nbconvert --to notebook --execute --inplace notebooks/video_vjepa2.ipynb

Total compute < 30 min on a single 4090; mostly forwards (no fine-tuning).

src/eris/ exports:

from eris.extract import (extract_entropy_stack, load_model,
                          extract_entropy_volume, load_video_model)
from eris.estimators import vasicek_entropy, knn_entropy, kde_entropy
from eris.fields import (gradient_2d, laplacian_2d, depth_derivatives,
                         gradient_3d, laplacian_3d,
                         transition_layer, per_layer_stats)
from eris.video import synthesise_translating_blob, load_real_video
from eris.volumetric import extract_outlier_tubes, render_streamtubes_html
from eris.baselines import (l2_norm_field, attention_rollout_field,
                            random_init_entropy_field)

load_model registry covers vit_b16, vit_l16, dinov2_b, clip_b16 out of the box for still images; arch-specific CLS / patch-grid / forward conventions are hidden behind the unified extract_entropy_stack API. For video, load_video_model covers vjepa2_l and vjepa2_g; use extract_entropy_volume(arch, frames) to get a (n_layers, gt, gy, gx) tensor per clip.

• It IS: a clean depth-wise structural finding in trained-ViT residual streams; the visualisation pipeline that surfaced it; a methodological comparison against simpler per-patch baselines.
• It IS NOT: a new mechanism, a free-energy derivation, or a causal probe. The phenomenon (sparse non-Gaussian "outlier features") is in the literature (Kovaleva 2021, Dettmers 2022, Sun 2024). What this work adds is the spatial dimension — outliers live on a 2-D patch grid and emerge in a single-layer phase transition whose timing is gated by training objective.