S-AI-PDDL: Predicate-Centric Construction of PDDL Domains

Background. Natural-language-to-PDDL systems may produce parser-valid or high-scoring domains while still failing to reconstruct the applicability and transition semantics required by a new planning contract. Performance on an exposed benchmark therefore cannot be treated as evidence of untouched-domain generalization. S-AI-PDDL reframes domain generation as an explicit predicate-centered construction problem in which canonical predicates form the primary intermediate relational units from which Action Contracts and a Canonical Domain Contract are assembled before PDDL serialization. Methods. Given a declared source contract , S-AI-PDDL induces candidate predicates, maps supported candidates into a typed, role-aware, provenance-preserving Canonical Predicate Space, constructs Action Contracts with explicit parameters, positive and negative preconditions, and add/delete effects, and serializes the resulting Canonical Domain Contract into a declared PDDL fragment. Canonicalization is kept distinct from semantic completion and commitment. The empirical protocol separates controlled mechanism tests, exposed source-grounded evaluation, exposed GOLD-informed capacity, target-GOLD-excluded intra-corpus transfer, and strict prospective evaluation. The final campaign also evaluates an optional LLM proposer against frozen symbolic rules and a deterministic S-AI canonical verifier under one-generation, PRE-GOLD sealing constraints. Results. Under the official Text2World/Tarski evaluator, the exposed S-AI-PDDL V7 branch scores 100% Executability, 96.14% Predicate F1, 93.95% Parameter F1, 62.93% Precondition F1, and 62.70% Effect F1 across 101 domains. Proc2PDDL V6.5-C closes the internal 5×60 criterion after exposed GOLD-informed development, while SPAR strict target-GOLD-excluded LODO shows non-zero semantic transfer and the exposed V6.9 cold-start branch also closes 5×60. In strict prospective testing, BookWorld V7.3 recovers the complete action inventory but not the Action Contracts. Across the full OPENPI+ ledger, 55 targets are consumed; after one post-seal out-of-fragment exclusion, 54 targets remain in-fragment and scorable, and none reaches the complete frozen 5×60 criterion. The LLM proposer substantially improves action identity and yields partial semantic overlap in later pilots, while the verifier provides observed prospective discrimination, including one final correct PRE-GOLD rejection. The preregistered external Model Space targets bloxorz and checkers remain untouched because the official author-controlled package is access-blocked at the final checkpoint. Interpretation. The results establish strong exposed reconstruction capacity and several narrower transfer capabilities, but they do not establish complete prospective PDDL generalization. The main unresolved bottleneck is target-independent induction of canonical relations, argument roles, arities, state carriers, precondition polarity, effect transitions, and action granularity. S-AI-PDDL therefore contributes both a predicate-centered construction architecture and an evidence discipline that separates exposed capacity, intra-corpus transfer, prospective generation, proposer quality, verifier discrimination, and future external validation.