29 days ago
-
Preserve completed candidates and measured work before verifier dispatch
-
Record comparable solve-level elapsed time for native and guided runs
-
Record CPU build and hardware, initialization treatment and runner reuse
-
Close unstarted coroutines and retain explicit zero-attempt accounting
-
Preserve evidence throughout interrupted preparation, response handling and decoding
-
Retain generation budgets, interruption accounting and partial timeout evidence
-
Reject global hooks and invalid CLI limits; retain usage through client cleanup
-
Reject unsupported event loops before dispatch and document pinned downloads
-
Record failed pre-dispatch preparation as zero-attempt uncharged work
-
Preserve completed results on storage errors and enforce durable platform scope
-
Preserve individual valid token counters in partially malformed usage
-
Preserve paid-attempt records atomically and retain known failed-request usage
-
Bind generator runtime and distinguish pre-dispatch failure accounting
30 days ago
-
Preserve durable input and payload binding on verifier failures
-
Reserve output ownership before model and paid inference
-
Bind saved sampling results to task and verifier payload
-
Release frozen Granite sampling with selective Jev verification
