ARLE‑METHODS‑PROFILES

This record publishes Gate Runner v1.0, a team-ready specification for building competitive machine translation models for Old Assyrian/Akkadian transliteration into English under surface-form–sensitive metrics (BLEU + chrF++). The core contribution is a boundary-first, refusal-capable workflow that treats the task as a sequence of admissibility gates—canonicalization, segmentation, document matching, sentence alignment, quality filtering, and verifier-only postprocessing—each instrumented with explicit DONE tests, refusal codes, and reproducibility manifests. Gate Runner is designed to address the dominant practical bottlenecks in cuneiform MT: transliteration markup heterogeneity (determinatives, logograms, gaps, Unicode variants), OCR-derived translation noise, and the risk of poisoning training through misalignment. The spec defines a deterministic CLI contract (gate_runner run --config config.json), required artifacts (accepted_pairs, refusals, audit_report, run_manifest), and calibration procedures for document matching and sentence alignment with data-derived thresholds. The resulting system enables controlled expansion of training data from OCR publications while maintaining traceability, leakage resistance, and metric-native inference behavior.