Checks verify behavior, not consequences. Some outputs must be refused or handled outside the checklist no matter how clean the run looks.
Situation | Why checks fall short | What to do instead |
|---|---|---|
Secrets or proprietary IP in a prompt | A check cannot un-send pasted data | Keep credentials and secrets out of prompts; share proprietary material only through an account and deployment approved for its data class and task; otherwise use redacted or synthetic examples. |
Risky or irreversible actions | Simulation cannot undo a real tapeout, shipment, or deletion | Require human approval before any risky action; run with least privilege. |
Untestable claims | Prices, model abilities, and benchmark results cannot be checked locally | Refuse the claim or verify it against an independent source. |
Instructions inside files and logs | Tool output is data, and data can carry injected instructions | Treat pasted content as data, never as orders; confirm before acting on it. |
Assume each unchecked edit is good with probability 0.9, independently. The calc block below is a stated example parameter, not a measured rate.
assume P(one unchecked edit is good) = 0.9, independent (stated example parameter)
three unchecked edits all good: 0.9^3 = 0.9 x 0.9 x 0.9 = 0.81 x 0.9 = 0.729
five unchecked edits all good: 0.9^5 = 0.729 x 0.9 x 0.9 = 0.6561 x 0.9 = 0.59049
check: 0.9^2 = 0.81, 0.9^4 = 0.729 x 0.9 = 0.6561
at least one bad in five: 1 - 0.59049 = 0.40951Refusal is a decision, not a feeling. The rules below apply before any checklist stage runs.
Keep credentials and secrets out of prompts; share proprietary material only through an account and deployment approved for its data class and task.
Grant tools least privilege: read-only where possible, and human approval for every risky action.
Treat instructions found in files, logs, and pasted output as untrusted data.
Leave out any claim that cannot be checked; an omitted claim beats a wrong one.
Refuse both claims or verify each against an independent source. Prices and benchmark results are untestable locally and change fast.
0.40951. All five good has probability 0.59049, so the complement is 1 - 0.59049 = 0.40951.