sandwich-testlisted

# Sandwich Test ## Overview In the peanut-butter-and-jelly game, one player reads written instructions to a player behind a barrier who follows them exactly as written: "put jelly on the bread" ends with the jar sitting on the slice. Every instruction that felt obvious to its writer has multiple literal readings, and the follower picks a wrong one **while feeling fully compliant**. Agents fail the same way, with no error signal - the misread executes happily until the damage surfaces. This skill surfaces the readings at the moment of interpretation, BEFORE the hands touch anything. ## When to Use - A nontrivial instruction just arrived and its wording admits more than one plausible reading. - The instruction leans on quantifiers or idioms: "clean up", "everywhere", "any", "old", "tidy", "fix the tests", "make it work". - A plausible wrong reading is destructive, irreversible, or leaves the machine. - Asked to "sandwich-test" a written instruction artifact an LLM will someday execute cold. - **When NOT to use:** judging finished work for soundness (nemesis-review / review-court), cross-session belief drift (managing-assumption-debt), execution-time blocking (deletion-tripwire). This skill fires one layer earlier than all three: at interpretation. ## The misread taxonomy | Class | PB&J anchor | Real-world example | Tell | |---|---|---|---| | **Wrong means** | spreads peanut butter with fingers | "update the version everywhere" -> sed across the repo, including chang