Do not link the private training repo; point at the public v5 model and the gpt-5 bar

Add v5-1000: 1000-case benchmark, 66.4% multi-round

Withdraw the protocol scores; wallet-429 is the benchmark, protocol is a separate run

Link the qwen3-8b v4 model

Publish v1-307 and v4 eval benchmarks side by side

initial commit