R12 Token-Native Syntax-Graph Schedule Preregistration

Causal Question

Does the sticky exact-executor path fail because a sequence encoder must infer the token-native grammar and opaque-reference topology before it can compile a coherent program?

The occurrence-linked arm rejects equality pooling alone. Its source and target fields regress sharply even though identifier equality is supplied. This treatment exposes the missing public structure directly: exact AST parent/child edges and semantic child roles under all four renderers.

Treatment

TokenNativeSyntaxGraphEncoder reconstructs, using only public token IDs and arities:

1. the exact leading AST span, deleting deterministic transport cover;
2. one parent for every non-root syntax node under prefix or postfix order;
3. semantic child rank after reversing renderer-specific traversal order;
4. tree depth; and
5. equality edges between repeated opaque local identifiers.

Three learned graph layers exchange parent-to-child, child-to-parent, and identifier-occurrence messages. The resulting memory feeds the existing single sticky schedule compiler and unchanged exact typed-state algebra.

The treatment cannot inspect QUERY bytes, answer labels, target packets, oracle programs, ontology sidecars, candidate scores, or host execution. A one-to-one identifier renaming must leave every schedule distribution exactly unchanged. Prefix/postfix and reversed-child parser tests must recover the same semantic parent roles.

Matched Gate

• protected checkpoint: ckpt_0300000.pt
• architecture/data seeds: 31/11
• initial state: oracle
• start position: 0
• updates: 1,000
• learning rate: 3e-4
• gradient clip: 1.0
• scheduler: width 384, 3 graph layers, 3 schedule layers, 8 heads
• pointers: ungrounded, matching control 725460
• evaluation: unchanged 32-batch source-deleted four-arm board
• exact executor and typed query reader: unchanged

The scheduler has 13,482,273 trainable parameters and the complete system has 169,421,167 parameters. The user's relaxed parameter limit is recorded, but this gate does not rely on exceeding the old 200M ceiling.

Decision Rule

Promotion requires simultaneous improvement in fully autonomous strict WORLD and COMMAND over matched control without factual collapse. Field and exact- schedule movement are diagnostic only.

• If source/target/value fields improve materially but strict pairs do not, syntax access is retained and the next arm uses one hard-selected latent program identity across every schedule field and step.
• If pointer fields remain flat or regress, explicit syntax routing is closed as a sufficient interface and the next intervention must alter compiler supervision or state representation rather than add width.
• Any positive strict result requires fresh architecture seeds and distinct held-out populations before promotion.

Delayed-Generalization Test

A zero at 1,000 updates does not by itself exclude grokking. The architecture will be measured at immutable 1k, 5k, and 15k budgets with the same seed, ordering, objective, and evaluator. The relevant signature is not falling training loss. It is:

1. near-saturation of training schedule imitation;
2. a delayed rise in held-out exact schedules rather than only marginal fields;
3. simultaneous delayed movement in strict WORLD and COMMAND; and
4. no decline in held-out factual top-1.

Smooth loss improvement, isolated value-field growth, or a late single-axis spike does not count as grokking. The 15k matched schedule-only controls remain the comparison, so longer compute cannot turn an ordinary budget effect into an architectural claim.

1,000-Update Diagnostic

V100 job 725570 completed in 11:55 with finite optimization. The result is not yet native reasoning: fully autonomous WORLD and COMMAND strict, margin-1, and DID remain zero, and factual top-1 is 60.94%. Local schedule accuracy is also essentially flat against the matched control:

The exact packets are nontrivial. An independent target audit finds 0/512 rows whose initial packet already equals its terminal packet; every row requires 4--31 active transaction steps. Exact state execution has therefore appeared at the oracle-initial interface even though end-to-end causal composition has not.

The trajectory remains live: V100 jobs 725573/725574 run independent 5k and 15k endpoints on evc11/evc12; H100 jobs 725569/725571/725572 preserve the same 1k/5k/15k endpoints for numerical replication. Promotion remains bound to the unchanged strict decision rule.