Remember the bytes that were read
A delegated edit reads source and returns after seconds or minutes of thought. Mobile sync or Dad may edit the same fragment meanwhile. A good result and a safe result are separate properties.
Capture the content hash when materializing the operand and store it as task base. Hash live bytes again immediately before apply. A mismatch keeps the revision as stale evidence for review against current source.
Check permission before spending too
Sending a protected fragment and rejecting only at apply still leaks content and spends tokens. Evaluate the matrix at queue, take, and operand read.
The check immediately before the brain call matters most because flags may flip while work waits. A queued task does not reserve future authority. Confirm both read and write rights.
Stale is not an invitation to auto-merge
A clean text merge does not prove semantic compatibility. The brain edited under an old premise while a new change may have replaced that premise. Automatic merge combines two judgments and falsifies provenance.
A narrow policy may allow mechanical non-overlap, but the default is a new task with current source and rebased instruction. Preserve the old revision as evidence without applying it.
Record cost in the failure
Repeated stale rejection signals a scope that is too broad or source that is too hot. Tokens spent, elapsed time, and base age help choose smaller fragments or quiescence.
Do not unsafe-apply to recover cost. Spent tokens are sunk; overwriting truth buys a far more expensive memory corruption.