Review Is the Spine, Not a Feature on Top

The Argument That Made It Structural

The instinct after measuring a generation defect is to swap generators. Try a different system, a different version, a newer release. That instinct was followed here and it failed in an instructive way: newer releases got worse at this specific defect under explicit awareness-raising, which is the opposite of the expected direction and is exactly what you would predict if the defect belongs to the arrangement rather than to the writer.

So the conclusion is not "all systems are equally good." It is narrower and more useful: changing the writer changes the flavor of the defect, not its presence, because every writer working from a source produces output shaped by that source.

What Single-System Dependence Actually Costs

The real cost is not generation quality, and locating it correctly is what turns review from a nice-to-have into the spine of the design. Depending on one system costs you four specific things.

No decorrelated reviewer. Self-review normalizes exactly the patterns that self-generation produced. A different system fails differently, which is the entire mechanism by which it catches what you cannot.

No comparison point. With one writer you cannot tell "this is hard" from "this writer is bad at this."

No regression baseline. When quality shifts, you cannot say whether the work changed or the tool did.

Hostage to one trend line. Whatever direction one vendor moves, you move with it, and you find out afterward.

Which Makes Review Load-Bearing

If review is the only mechanism that supplies decorrelation, then it cannot be the optional final step that gets skipped under deadline. It has to be a stage with a contract: a fresh instance, the artifact read cold from disk, a fixed request template rather than author-composed prose, and a verdict that carries its rejections. Each of those exists to protect the decorrelation from being quietly spent.

The fixed template matters more than it looks. If the author writes the review request, the author chooses what the reviewer looks at — and an author choosing the evidence has converted a review into a presentation without anybody deciding to.

Locate what your single point of dependence actually costs before designing around it. Here the naive answer was "quality" and the true answer was "decorrelation", and only the true answer tells you what to build — which is not a better generator but a reviewer that fails differently from you.

Code

What the review contract fixes, and what each fix protects·text

PROPERTY                     PROTECTS AGAINST
---------------------------  -----------------------------------
fresh instance               the author's context normalizing
                             the author's own patterns

different system preferred   correlated blind spots; two writers
                             with the same habits agree for the
                             same wrong reason

artifact read COLD from      the author's summary standing in
disk                         for the artifact - the summary is
                             the belief under test

fixed request template,      evidence curation: an author who
not author-composed          writes the request chooses what is
                             looked at

verdict carries a            an empty round being recorded as a
rejection log                clean one

one run per round, and       a resumed round handing the reviewer
never a resumed thread       the context the boundary exists to
                             deny

Exercise

Identify your single largest dependence — one tool, one person, one vendor — and write down four costs of it that are not quality. Then ask which of those four your current process has any mechanism for. Most teams have none, because the cost was filed under quality and quality felt acceptable.

Hint

The regression-baseline cost is the one nobody has a mechanism for and the one that bites hardest over time. Without a second source, a slow degradation in your dependence is indistinguishable from a slow increase in the difficulty of your work, and both feel like the work getting harder.

The Argument That Made It Structural

What Single-System Dependence Actually Costs

Which Makes Review Load-Bearing

Code

External links

Exercise

Progress

Comments 0