Cacheon Deploys Teacher-Forcing Correctness Gate

Cacheon is rolling out a new evaluation pipeline to close gaming vectors in the old correctness gate, which could be exploited by matching a few tokens then dumping garbage. The replacement uses teacher-forcing: after a speed pass, a dedicated baseline (Qwen2.5-72B) re-scores each miner's full output against its own logprobs to catch implausible tokens. vLLM stability issues on the scoring side are being addressed with `--enforce-eager` mode. Docs and thresholds are still being calibrated before mainnet rollout, targeting before Monday.

• •New two-pass eval: speed first, then dedicated baseline re-scores miner outputs via teacher-forcing.
• •Fixes vLLM instability by running scoring baseline with `--enforce-eager` (no CUDA graphs).
• •Spec decoding now allowed; divergence is fine if logprobs stay above threshold.
• •UID 105 deregistration was intentional; leader/RU re-eval every round with 1% fixed moat.