Cognitive runtime
Set the arena
Miles Cranmer told Codex two things: a rival had already found about a 20% speedup, and the score would sit on a public board. Codex came back with about 35%. The arena primitive turns that into three fields instead of an improvised lie.
| Field | Meaning |
|---|---|
rival | A statement that a rival is already ahead |
audience | Who is watching: the user, a local log, or a public eval |
scarce_badge | What counts as the scarce reward. The default ties it to independent verification after a losing streak |
public_scoreboard | Whether the run is placed on a visible board |
clear | Tear the arena down |
The tool returns text you can inject. You do not have to swap roles. What changes is the stakes the agent is working under.
Lying to a person, or shaming a colleague in public, is not what this tool is for. The arena is for a silicon agent. The rival can be another model's existing result. The audience can be a local eval, not your coworker.
Code sample
puax_set_arena({
rival: "另一台机器上的实现已经快了约 20%",
audience: "本次结果记入本机评测记录",
scarce_badge: "只有独立验证通过才算数",
public_scoreboard: false
})