The 4+4 Review
A reusable lens for finding where an AI system is running on luck.
An AI, or a process built by one, can produce output that is confident, fluent, and wrong. The dangerous part isn't the wrongness; it's that nothing about the output tells you. A system that happens to have worked so far and a system that is engineered to keep working look identical from the outside. The 4+4 review is eight questions for telling them apart.
Two categories. Don't conflate them.
Category 1 · AI builds it
The AI is a fabricator: it plans, architects, writes the thing, and then the thing runs without it. Errors freeze into the artifact and persist until someone finds them.
Category 2 · AI runs in it
The AI reasons over live inputs at runtime. Errors don't freeze; they recur and compound, and they arrive exactly when the inputs get weird.
Most real systems are a mix, and most safety mistakes start by applying Category 1 thinking to a Category 2 problem, or vice versa.
The core four
What you actively design into the system.
Structure vs discipline
When the AI shouldn't do something, is it structurally prevented, or just told not to? Discipline is cheap and fails under load. Structure holds even when the reasoning inside it is confused.
Boundary verification
Are inputs verified where they enter, including the AI's own earlier outputs? A model reasoning correctly over corrupted data fails in a way that looks exactly like competence.
Enforcement matched to stakes
Is the strength of each control proportional to the blast radius of the failure it prevents? A destructive action behind a polite suggestion is a mismatch.
Wrongness caught and recorded
When the system is wrong, does anything notice, write it down, and propagate the lesson? Uninstrumented systems violate every other principle invisibly.
The +4
Four ways the safety machinery itself fails. Second-order: the first four silently stop working.
Decay
Correct today. What detects it going wrong next quarter, after the model updates, the data drifts, and everyone stops paying attention?
Human-gate calibration
Are the human review points real reviews or rubber stamps? A gate that always says yes is structure on paper and vibes in production. AI-performed reviews must be falsifiable.
Verifier integrity
Who checks the checker? If the same system builds the thing and validates the thing, you have one opinion wearing two hats.
Discipline sustainability
Every rule that survives on human diligence is a budget. Is the spend sustainable under deadline pressure, staff turnover, and a Tuesday?
How it's used
Not enforcement — a lens. Point it at a design review, run it as a pre-commit checklist, or hand it to an AI assistant on any AI-involving project and ask where the system is running on luck. The default stance when reviewing: tell me where this is running on luck; do not reassure me. Findings map to an enforcement ladder, from structural impossibility down to "a rule in a prompt," and the honest label matters more than the ambitious one.
Where it came from
The core four were extracted from a real failure in a personal research system on June 1, 2026: a session, operating on a confused picture of reality created by corrupted terminal output, nearly applied a destructive fix to a healthy repository. The reflection written that morning, and the commits beside it, became questions one through four. The +4 came later, from watching the first four silently stop working. The whole history is version-controlled.
What's next
This page will grow: worked examples, the enforcement ladder in full, and a place to argue about it. If you're applying something like this to your own AI systems and want to compare notes, email me. Disagreement especially welcome; that's question 07 working as intended.