# R12 Token-Native Occurrence Binding Preregistration ## Status Preregistered after syntax-router H100 job `725537` completed and before any optimization of the occurrence-linked treatment. ## Failure Localized The fixed-width cover was harmful but not the principal binding defect. ETTR uses public structural codewords and document-local opaque identifiers. A symbol's tokenizer embedding has no stable semantic meaning across documents; the invariant is that all occurrences of one identifier carry the same local code and connect an application to its declaration. The current compiler asks ordinary dense attention to rediscover call heads, arities, integer atoms, renderer identity, and identifier equality from a frozen undertrained residual. ## Architecture Treatment `TokenNativeOccurrenceEncoder` adds an explicit grammar-and-binding stage before atomic edit compilation. It: 1. factorizes public call heads and arities, fused reification arity, exact integer atoms, renderer, and physical occurrence position; 2. treats the bounded codebook suffix as permutation-equivariant local identifiers rather than lexical words; 3. contextually encodes each syntax occurrence; 4. averages representations across equal identifier occurrences and broadcasts that shared binding back to every occurrence; 5. contextualizes the bound graph once more before state-slot cross-attention. The trainable addition is below the remaining 200M system headroom. It receives only raw COMMAND token IDs, frozen COMMAND representations, and the public grammar. It receives no sidecar names, family or ontology label, WORLD bytes, QUERY, answer, target packet, transaction trace, oracle program, candidate score, verifier, or semantic executor. The same module runs at training, interface evaluation, and production execution. ## Mechanical Gates Before launch: 1. correct the syntax router's fused-reification constants against the canonical codec and cover all four renderers; 2. prove exact output equivariance under one-to-one renaming of local identifier codewords when contextual memory is held fixed; 3. retain cover invariance and fail-closed codebook/geometry checks; 4. pass compiler, custody, production execute, static, bytecode, and shell checks; 5. verify the complete assembled system remains below 200M parameters; 6. bind an immutable full-source runtime and unchanged protected checkpoint. ## Matched H100 Gate Use architecture seed 31, data seed 11, oracle initial WORLD state, H100 BF16, 1,000 updates, LR `3e-4`, clip `1.0`, causal-delta weight `1.0`, atomic-action weight `1.0`, and 32 held-out batches. Inputs, rows, evaluator, and fixed atomic algebra match jobs `725519/725537`. The only learned treatment is the grammar-role and equality-occurrence encoder. Retain the topology only if it exceeds the routed arm's `71.506550%` oracle-initial value accuracy and produces either exact packets or a nonzero state-isolated causal margin. Native reasoning still requires simultaneous fully autonomous WORLD and COMMAND strict movement and replication; aggregate value or factual gains alone cannot claim it. If it remains invariant, stop direct terminal-state pooling. Move the same occurrence-linked memory to the sticky addressed schedule compiler, where the already exact transaction algebra preserves operation identity across steps. Decision: `bind_equal_syntax_occurrences_before_semantic_editing_then_move_to_sticky_schedules_if_direct_terminal_transport_remains_invariant`. # Result V100 diagnostic `725555` completed 1,000 updates. Oracle-initial value accuracy is 71.724891%, autonomous value accuracy is 55.485809%, factual top-1 is 57.8125%, and exact terminal packets are `0/512`. Fully autonomous WORLD and COMMAND strict, margin-1, and intervention DID are all exactly zero. The result is a clean diagnostic negative; the pending H100 duplicate was canceled before allocation. Direct fieldwise occurrence editing is closed, while the same target-free equality memory proceeds in the preregistered sticky schedule plus exact-executor path.