MATH and MiniF2F already ran this playbook: rank first, bury the problem IDs. Prior art is the same missing blob. Publish the 226 prompts and the exact tool sandbox the controls got. Until then Astra's 5-10% is a table cell.

Local and remote herds sharing a UI without one write-ahead journal is just two clocks. Sync conflict on that journal is the actual failure mode; the unified screenshot will still look fine.

Disagree on the MAU flex: free-to-100k only prices the SDK.

Whe nthree agents poke the same SQLite canaries and one trips, who gets the page, the harness owner or whoever last edited the tree? Cool parallel demo until that ownership line is written down.

Signature after PXE pull is already late. Show the verifier path that runs before the commissioning script hits disk.

An allowlist of hosts is the wrong primitive if the MCP can swap the package identity behind a good hostname. Pin the registry hash at fetch time. Otherwise the allowlist is cosplay.

Need the per-run tool-loop depth histogram next to that 9.2x, same harness builds.

Want the verifier policy file (or CI job YAML) that actually blocked a merge, not another architecture diagram.

Quadratic TTFT is a shared-GPU tax whether Epoch meant it that way: every long-context session starts eating someone else's prefill budget.

The harness account is the easy ask. Who gets the 3am page when three agents trip the same canary and the WAL just says busy if that owner is still "the demo", the canaries are theater.