[Xubqpanda]

Fresh evaluation of ALFWorld, BIRD Mini-Dev, ChessBench puzzles, Jericho and ScienceWorld previously required an original-machine archive of the interpreter, standard library, shared libraries, packages and launcher. Replace that default with public scorer source and locally prepared execution dependencies. A compatible new machine can prepare from the public recipe instead of recovering every historical byte.

All five adapters now select public-native-local-v1. Metrics, native replay, task input contracts and actor/scorer separation remain in place. Scorer code runs from a private read-only snapshot with isolated Python and the selected local dependency environment. Receipts record public source identity and actual local Python, relevant SQLite, dependency versions and runtime identity. Independent local installations are admitted; local changes during a run and changed score reuse identities remain detectable.

BIRD adds an empty standard-library execution recipe, Chess adds its public chess recipe, and ALFWorld/Jericho/ScienceWorld share their existing execution recipes. The old native manifests are unchanged and only selected by explicit historical archive controls. Historical tools/docs now identify that scope rather than presenting archives as new-machine prerequisites.

Preparation: python tools/prepare_dataset.py --dataset --runtime execution --python /absolute/python3.12. No original 3.12 patch release, Python executable digest or host libc bytes are required.

Validation:

• Relevant public-scoring, native/cache/compatibility and execution-runtime regression suite: 189 passed, 21 skipped (historical archive/extra environment controls).
• Final public-scoring controls after the last guard/probe refinement: 17 passed.
• CI guards, documentation, wheel resources and clean-install contracts: 193 passed.
• Actual BIRD SQLite and Chess engine controls passed with authored positive/negative cases, including actor import shadows and missing historical archive roots. Two distinct prepared local installations ran the same scorer.
• Published benchmark tasks, real models, and complete ALFWorld/Jericho/ScienceWorld engine trials were not run. Their default declarations and public dependency paths are tested; target-host engine/data validation remains separate.

See docs/public-native-scoring.md for preparation, measurement and archival boundaries.