downstream-instruction-framing
Downstream-Instruction Framing Discipline
Problem
LLMs default to authoritative downstream instructions: "Approved Decisions — Do Not Re-Litigate," "Execute the following steps," "These are final." This is usually wrong.
The downstream reader often has evidence the author lacks: target-repo access, tests, and local rules. Command framing can make them silently follow a plan the code contradicts.
Test that asymmetry per decision: who can access the evidence that bears on it? Downstream often knows the target code and tests better; upstream may know legal rules, security policy, product intent, or external facts unavailable in the worktree.
Evidence and decision authority are separate. Better evidence licenses a revisit request, not the right to change scope, values decisions, or explicit user constraints.