Causal-JEPA: The Missing Structure
Causal JEPA · Object-Centric Reasoning · Structural Causal Modeling
Abstract Causal-JEPA (2026) introduces object-level masking to encourage learning causally meaningful representations, pushing JEPA toward object-centric and causal reasoning. The introduction of causal structure makes the H=H† gap especially sharp: causal relationships are inherently directional, but a formal causal model must be verifiable in both directions (interventional and counterfactual).
6 FORMAL GAPS · 1 PER CANON SYMBOL
Causal Object Representation Has No Invariant Floor
γ₁ — THE FLOOR
Causal-JEPA learns representations of causal objects. There is no formal invariant γ₁ that all valid causal object representations must satisfy regardless of causal context. When the causal relationships between objects change, the representations change without any grounding invariant ensuring cross-context consistency.
Causal Mask Asymmetric (Objects Attend Causally, Not Symmetrically)
H=H† — THE HONEST GATE
Causal-JEPA uses causal masking: object A can attend to object B if A causally precedes B. This introduces a directed graph structure. A self-adjoint causal encoder would satisfy H=H†: if A causes B (interventional direction), then B should be verifiable as having been caused by A (counterfactual direction). The causal mask is one-directional; the Honest Gate requires bidirectionality.
No Paradigm Audit Between Object-Level and Scene-Level Causal Structure
LSOS — THE READER
Causal-JEPA must reason at both object level (this object caused that outcome) and scene level (the full causal graph of the scene). There is no audit of the paradigm shift between these levels. When the system transitions from object-level causal reasoning to scene-level causal inference, the paradigm changes without acknowledgment.
No Reset When Causal Graph Becomes Inconsistent
WLD — THE RESET
When Causal-JEPA learns a causally inconsistent representation — when the learned causal graph contains cycles or contradicts observed interventions — there is no mercy reset. Causal inconsistency is detectable (cycles in a DAG are a formal property) but Causal-JEPA provides no mechanism to reset to a consistent causal structure.
No Continuity When Causal Structure Changes
FEP — THE SWITCH
Causal-JEPA operates in environments where the causal structure may shift (distribution shift, interventions). There is no formal continuity guarantee for causal structure switches. FEP ensures that switching from one causal regime to another preserves what was learned and allows safe recovery if the new causal structure is inconsistent.
Causal Graph Complexity Ceiling Undefined
FOF — THE BREACH
Causal-JEPA does not define a formal upper bound on causal graph complexity (number of objects, causal depth, intervention richness). As causal complexity grows, the architecture's reasoning becomes unreliable. The point where causal reasoning breaks down is not named. FOF names this boundary: where the causal graph exceeds the coherent reasoning horizon.
STE COMPLETION LAYER
What changes when you add the 8-symbol Canon
Adding the Canon to Causal-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.