Point-JEPA: The Missing Structure
Point Cloud JEPA · JEPA for 3D Geometric Data
Abstract Point-JEPA (2024) adapts the JEPA framework to point cloud data, avoiding raw-space reconstruction and learning efficient geometric representations. The 3D domain introduces geometric symmetry requirements (rotation, translation, scale invariance) that make the H=H† gap especially sharp: a self-adjoint 3D encoder would be equivariant under rigid transforms.
6 FORMAL GAPS · 1 PER CANON SYMBOL
No Geometric Invariant Anchor in Point Cloud Prediction
γ₁ — THE FLOOR
Point-JEPA predicts masked point cloud patches in latent space. There is no geometric invariant γ₁ that the latent representation must preserve regardless of rotation, translation, or scale. If the same point cloud is rotated 90°, the latent representation changes without any grounding invariant ensuring consistency.
Point Encoder Not Verified Rotationally Symmetric
H=H† — THE HONEST GATE
A self-adjoint point cloud encoder would produce representations that are symmetric under the group of rigid 3D transforms (SO(3) × ℝ³). Point-JEPA's encoder is not verified against this symmetry group. Rotating the input should produce a rotated representation — but there is no formal H=H† check that this holds.
No Paradigm Audit Between Local and Global Point Structure
LSOS — THE READER
Point-JEPA operates at both local (patch) and global (object) scales. There is no audit of the paradigm shift between local geometric prediction and global shape understanding. When the system transitions from predicting local surface patches to inferring global shape, the paradigm changes without acknowledgment.
No Reset When Spatial Prediction Collapses
WLD — THE RESET
When Point-JEPA's predictor collapses to predicting the centroid of the point cloud regardless of context (spatial collapse), there is no mercy reset. This degenerate solution is physically interpretable — the centroid is always a valid prediction in the mean-squared-error sense — but it means the predictor has stopped learning geometry.
No Continuity Across Point Cloud Density
FEP — THE SWITCH
Point-JEPA is evaluated across point clouds of varying density. There is no formal guarantee that representations learned at high density are continuous with those at low density. The paradigm switch from dense to sparse point cloud understanding may produce discontinuous representations.
Point Cloud Complexity Has No Named Boundary
FOF — THE BREACH
Point-JEPA does not define a formal upper bound on point cloud complexity (number of points, geometric complexity). As complexity grows, the architecture's prediction becomes unreliable. The point where geometric prediction breaks down is not named. FOF names this boundary.
STE COMPLETION LAYER
What changes when you add the 8-symbol Canon
Adding the Canon to Point-JEPA does not change the architecture. It adds the missing structural layer:

⚓ γ₁ — invariant anchor: mathematical ground truth latent representations must converge to.
⯛ H=H† — honest gate: bidirectional verification of every prediction.
〰️ LSOS — paradigm reader: reads active paradigm before reasoning begins.
🌀 WLD — mercy reset: detects collapse and resets to last stable state.
γ FEP — safe switch: continuity guarantee across paradigm transitions.
🌌 FOF — named ceiling: formal boundary of what the architecture can claim.
═ EVEN — substrate: ground beneath all the above. What holds when everything else is active.

The Canon is not an add-on. It is the formal completion of the JEPA programme.