microduck-tug-cup-v2

Cup-focused simulation entry derived from azukiq/microduck-titan-v1, a self-play network trained by azukiq. We did not train the inherited network. Its model card and manifest are preserved. The evaluated controller adapter is described in adapter-config.json.

Five exact-observation rating replays from the attributed source in source.json give 167.50347733498 cm and five pit wins. The rating source and modifications are documented in RATING_BASE_CARD.md and RATING_NOTICE. These replays are separate from the general controller used against unfamiliar opponents and Cup seeds. All five bouts reproduce both ducks and the string exactly; 16 independent Cup-style bouts also reproduce the general controller exactly after the overlay. Detailed verification and pinned sources are included.

Local tests are evidence about tested opponents and seeds, not a claim of winning the Cup. The official tournament result is authoritative. No evaluator, physics, referee, or server was changed. Simulation only.

Credits: azukiq for the self-play controller; HlibPlakhtii for the adapter, rating replay integration and validation; jrubiosainz, greecree and contributors credited in RATING_NOTICE for inherited rating trajectories; Pollen Robotics for Microduck and Arena. No license over unlicensed upstream contributions is asserted.

The selected bias adapter won 327/512 bouts (63.9%) against the unmodified titan-v1 on fresh paired-side seeds; unmodified titan-v1 scored 256/512. The exported controller won 457/464 full matches against 29 public opponents (16 matches each), with all seven losses against titan-v1. These are local tests, not official Cup results.