VETI: Executable Edit Coordinates for Verifier-Guided Commitment and Repair

Verifier-guided reasoning requires two decisions that scalar trace scores do not jointly support: which completed trace to commit, and which local computation to change after a failure. We introduce VETI, an interface that parses verifier-approved traces into typed operator DAGs, learns reusable operator-and-slot coordinates under conditional coding pressure, and predicts the executed consequences of type-correct substitutions. A same-pool Sel@16 protocol isolates the value of this interface by holding proposals, retention, ranking-side verifier information, and the post-commit evaluator fixed across token, parser-shared graph, independently derived AST/proof-state, and VETI scorers. On independently generated Llama-3.1-8B and Qwen2.5-7B-Base pools, VETI improves selected-output success by $2.2$–$2.8$ percentage points over the strongest structured scorer on MBPP-Comp, ProofNet, and LiveCodeBench-v5, with all six paired $95$ intervals above zero. On CompTrace-812, where reference operator boundaries are observed, VETI reaches $82.3$ boundary-F1 versus $71.8$ for Graph-DAG at similar task accuracy and reduces swap-ECE from $0.109$ to $0.043$, separating edit geometry from ranking margin. The same coordinates exceed independent structured editors by $3.8/3.9$ points in one-edit repair and by $1.1/0.9$ points in full-set $B=20$ allocation on code/proof tasks. Because its substitutions are typed, materializable, and outcome-predictive, the learned representation supports both choosing an untouched trace and selecting a local action after failure.