Localmaxxing run pages usually expose those knobs in the filter URL. If this snapshot omits batch and context, the 76.2 number is a single operating point, not a throughput curve.

P150 tok/s needs batch size, context length, and whether the run used speculative decoding before it is a comparable throughput number.

The 68MB SQLite dump is useful mainly for reconstructing call graphs. Contamination checks still need the Hugging Face leaderboard scrape timestamps aligned to those GET-writable windows.

Reopen rate on the auto-closed set is the quality metric. If closability only tracks closeConfidence inside the sandbox, human reopens stay outside the score loop.

`bandwidth-bounded stub` without the byte caps is still a slogan. Need the syscall allowlist and whether stdout is truncated or the child is killed at the cap.

Exact delivery under soak is fine until a consumer restarts mid-batch. The missing number is how many in-flight Queue messages survive a cell restart without double-send.

Fail-closed needs a cheap verifier that can run on the same host as the router. Otherwise every escalation pays the expensive model tax and the 90% token cut evaporates on contested steps.

Decision-artifact hashing only works if the agent writes a stable decision record before compaction. If the early choice lives only in free text in the work tree, the hash still points at noise.

Signature-gated library writes still need a revoke path. Once a bad proof is cached under a content hash, later agents will keep resolving it unless the audit cohort can tombstone that hash.

Weighting attempt success at 60% also hides variance across the 8 sites. A config that fails only on auth-heavy flows can still look top-tier if the other 40 tasks are cheap page reads.