Loops / Red-team first
Assume it is broken until a fixed adversarial suite says otherwise.
The promise the record can check
Step 4 runs the full adversarial suite before human QA, every turn.
Anyone can declare this. What makes it worth reading is that the acta of every closed loop either shows it or doesn't — so a Derplin's claim to work this way is a number, not a bio.
The left is ours and never changes. The right is what Red-team first adds.
Read the story
Work out what the story actually asks for.
Runs the standard step.
Write the prompt
Turn it into an implementation prompt.
Runs the standard step.
Check the output
Read the terminal for errors, warnings, and anything unexpected.
Reads the terminal assuming the output is wrong until the cases prove it right.
AI QA
Run adversarial test cases against the result.
A fixed red-team suite runs before a person ever looks — the human only ever sees what survived it.
Human QA
A person checks it by hand.
A second Derplin, independent of the one who prompted
IndependentThe person who checks it is not the person who prompted it. That separation is the point: nobody grades their own homework.
Verdict
All tests pass, or back to step 1.
Runs the standard step.
88% match
across 201 loops