The peer review ritual must move — from code to ownership

A few months ago, I wrote about what I called cognitive debt. I had noticed I was drifting too far from the implementation, delegating more and more to AI. In the end, I had a product that didn't really belong to me. Was that honest? A dilemma.

Since then, my position has evolved as AI improved. I can own the product without knowing every line of code. So I shifted my effort to two things: qualifying what should be generated, and validating what is generated. I revised my processes accordingly.

Human review of AI code is a weak check

The classic ritual — submit your code for peer review to improve quality — has an unspoken assumption: the reviewer understands the code more deeply than an automated check would, because the reviewer is a competent human.

Now look at the emerging reality: AI generates the code, and the reviewer is a human who also delegates to AI. What does that review actually add? The "competent human in the middle" has become another probabilistic pass. Two fuzzy layers stacked on top of each other do not produce determinism.

I would rather lean on deterministic tools. The principle: frame the probabilistic with the deterministic. On one side, a structured process validates what is generated — tests, quality gates, static analysis, declared acceptance criteria. On the other side, I validate the properties of what is generated: does it behave within the perimeter I intended, does it respect the constraints, does it carry the invariants I care about?

The question behind the question

So the question I keep coming back to: does the peer-review ritual — invented to improve human-written code — still make sense when AI generates, reviews, and maintains that code?

My answer: the ritual itself is not sacred. It was always a proxy for something deeper — shared understanding and ownership of the product. When the proxy stops working, you don't discard the goal. You move the ritual to where it still counts.

Ownership is the thing that counts

The rituals that matter now, organized by a new dichotomy: on one side, the process validates what is generated; on the other, the human validates the properties of what is generated.

• Qualification before generation. Saying precisely what should be produced — specs, acceptance criteria, constraints — is where human judgment has the highest leverage.
• Deterministic verification. The process checks the output: tests, gates, proofs. Not another opinion — a measurable verification. This is the principle behind everything we build at Uptech: AI proposes, ground truth verifies. The probabilistic is framed by the deterministic.
• Re-appropriation of the product. Reading the architecture, owning the decisions, being able to explain and defend the system. Ownership is knowing what the product is and why — not reciting every line.

The review that matters is no longer "did a human read this diff?" It is "can the human who owns this product account for what it does, and for what was checked before it shipped?"

That is the shift: from reviewing code to re-appropriating the product.

---

This article is adapted from a thread I published on LinkedIn. It reflects the engineering doctrine behind how we build products at Uptech LLC — qualification, deterministic verification, and validation of properties.