Oro Race 20 Updates, Model Allowlist Changes

Oro updated allowed models for agent submission, removing Qwen3-235B and openai/gpt-oss TEE variants while confirming zai-org/GLM-5-TEE models are permitted. Forge agent achieved benchmark improvement from 0.4286 to 0.7143 via two-phase search strategy for product tasks; voucher handling remains at zero. Community raised questions about deregistered hotkey evaluation and race mechanics, with some infrastructure errors reported.

• •Allowed models updated: GLM-5-TEE variants confirmed, Qwen3 and openai/gpt-oss removed
• •Forge benchmark: 0.4286 to 0.7143 via two-phase search and graceful retry logic
• •Voucher task performance at zero; next focus on voucher-first iterations
• •Infrastructure errors affecting some agents; deterministic handling requested