# Taste is part of correctness.

Why a technically valid output can still be wrong, and how I combine automated gates with human acceptance.

- **Canonical URL:** https://vidhata.me/notes/taste-is-part-of-correctness
- **Author:** Vidhatanand V. (Vid)
- **Role:** Fractional CTO and AI Systems Architect

**Published:** 01 Sep 2026 | **Reading time:** 10 minute read

## Why I hold this position

I have seen systems pass every mechanical check and still fail the moment a person looks at the result. The file exists. The schema validates. The layout does not overflow. The render completes. Yet the diagram teaches the wrong relationship, the interface hides the next action, or the visual feels lifeless enough that nobody wants to use it.

That is why I treat taste as part of correctness. Not taste as personal decoration, and not taste as an excuse to ignore evidence. I mean the human judgement required to decide whether an output is clear, appropriate, coherent, and worth putting in front of real people.

## Mechanical validity protects only what it can measure.

A geometry validator can prove that labels do not overlap. It cannot prove that the diagram explains the right concept. A browser test can prove that a button is visible. It cannot prove that the page makes the next decision obvious. An audio pipeline can prove that loudness is within range. It cannot prove that the delivery feels human.

I still want those validators. They remove entire categories of preventable failure and make review more efficient. The mistake is letting a passing gate claim more than it measured. Every automated assertion needs an evidence boundary, especially in visual, narrative, and generative work.

## The human bar should be explicit.

Taste becomes useful to a team when it is translated into acceptance criteria. Is the hierarchy clear at a glance? Does the image support the story rather than fill space? Can a founder understand the tradeoff without learning the internal vocabulary? Does motion explain a change, or merely announce that the interface can move?

These questions are not perfectly objective, but they are discussable. A review can name what feels wrong, connect it to the product intent, and decide whether the output needs another iteration. That is stronger than either extreme: pretending taste can be fully automated or treating it as an unchallengeable opinion.

## Visual acceptance belongs in the delivery loop.

For interface work I want screenshots at the breakpoints people will actually use. I check text measure, hierarchy, image treatment, touch targets, empty states, focus, hover, and what happens when content is longer than the design expected. Portrait layouts often reveal assumptions hidden by a wide desktop. A technically responsive grid can still feel cramped, repetitive, or directionless.

For generated visuals I keep the same discipline. Labels, geometry, identity, and meaning deserve direct inspection. If the output invents text, changes a face, or creates a plausible but false relationship, a successful generation call is not success. The system should fail the acceptance gate and preserve the reason.

## Craft earns trust before the user reads the architecture.

People experience technical quality through the product surface. A clear state, a useful explanation, a considered transition, and an honest error message tell them the system has been thought through. Inconsistent hierarchy and generic components communicate the opposite, even when the backend is strong.

The lesson is not to polish everything forever. It is to know which human judgement the product cannot outsource. Automate what can be measured. Make the acceptance bar visible. Review the real output in context. When the result passes the tests and the human bar, correctness becomes something a person can actually feel and use.

## What I take into the next build

- State exactly what each automated gate proves and what it cannot prove.
- Turn taste into reviewable criteria connected to product intent.
- Inspect real outputs at real breakpoints, not only isolated components.
- Preserve human rejection as evidence, even when the pipeline passed.

## Continue reading

[All Fractional CTO field notes](https://vidhata.me/notes)

## Work with Vid

Start with the actual technical pressure: [bring the problem](https://vidhata.me/hire).