vgc-ma-heuristic-static-eval

vgc-ma-heuristic-static-eval

All 120 declared native panels are available as separately frozen evaluation leagues. Each submitted policy runs one complete selected panel per episode. Panel scores retain their native aggregation; this release defines no score across different panels. The fleet.json index records every panel, population, metric, and fixed reference roster. Qualification and coverage must be established separately for every selected panel.

VGC-Bench heuristic cross-play

This regulation contains 120 variants and 480 complete battles: three native heuristic opponents, prefix and held-out team pools, pool sizes 1/4/16/64, and five native pool seeds. Each variant retains the author's four battles per run, totaling twenty battles per condition. The original LLMPlayer system prompt, team preview, two-slot masks, Python integer parser, fallbacks and native win-rate scorer remain unchanged. Both regulations together cover 960 battles. Shared pool-seed components group correlated battles; incomplete populations cannot produce full means.

The pinned CPU reconstruction uses Python 3.13.5, Torch 2.12.0, Transformers 4.57.6, stable-baselines3 2.8.0 and Node 24.14.1. The declared downstream generation profile is temperature zero and 512 output tokens, not the original Llama generation defaults. Battles run serially on a private local server with open team sheets and native turn limits. Suite seeds label repetitions; native pool seeds select teams. Showdown's actual random seed, complete input log, raw outcomes, exact provider responses, native fallback counts and both player replays are retained. Malformed responses can win through native fallbacks; their reward remains native and their invalid-response count remains separate. Provider or engine failures are unscored.

This is the complete declared heuristic league, not the paper's eleven-agent trained-policy roster. The released model repository supplies only behavior-cloning seed one; missing trained opponents and their original feature indexes remain a separate paper-comparison gate. No paper-score parity or Qwen performance is claimed by offline controls.