Cacheon deploys teacher-forcing correctness gate

Cacheon is replacing its gameable correctness evaluation with a two-pass system: speed measurement first, then re-scoring of miner outputs using a baseline model's logprobs to detect implausible tokens. The team identified vLLM stability issues under repeated scoring requests and is hardening validators across different infrastructure providers. Threshold calibration and dry-runs are underway before mainnet deployment.

• •New eval: speed pass measures all miners, correctness pass re-scores outputs with baseline logprobs
• •vLLM crashes on garbage outputs; fix uses --enforce-eager mode for scoring baseline
• •TPS metric revised to count baseline token count only, preventing padding inflation
• •Infra hardening: validator compatibility across Targon and Shadeform pod differences