Design quality has no single number. Anyone who says otherwise is selling a score. But that doesn’t mean agents are blind. Roughly 90 years of design research yields something more useful than a verdict: a set of partial, validated reward signals. These are computable properties that correlate with human aesthetic judgment, and they are strong enough to guide an agent’s search through layout space. Partial is not a compromise here — it is the engineering sweet spot. Agents do not need to replicate human taste; they need a gradient that points uphill. This post maps the computable line from Birkhoff’s formula to modern layout metrics, shows the automation precedent, and draws a hard boundary where research predicts only first impressions or where folk rules collapse.
The computable lineage: Birkhoff to Ngo
The first serious attempt to formalize beauty came in 1933. Birkhoff’s Aesthetic Measure defined beauty as M = O/C — order divided by complexity. The formula is crude, but it established the core idea: aesthetics can be expressed as a ratio of measurable quantities. For decades, that idea stayed theoretical. Then Ngo, Teo & Byrne (2003, Information Sciences 152, 25–46) made it operational. They extended Birkhoff into 14 computable layout measures — balance, equilibrium, symmetry, sequence, cohesion, unity, proportion, simplicity, density, regularity, economy, homogeneity, rhythm, order-and-complexity — and validated them on real multimedia screens. Modelling Interface Aesthetics Each measure is a formula. Balance is a function of optical weight distribution. Symmetry is a pixel-level mirror comparison. These are not vibes; they are arithmetic. For an agent, each measure is a differentiable-ish objective. The 14 together form a multi-objective reward vector that captures much of what humans mean by “clean” or “organized.” This is the foundation.
The automation precedent: Ivory & Hearst
The idea of automating design evaluation is not new. Ivory & Hearst (2001, ACM Computing Surveys) surveyed 132 usability-evaluation methods and organized them into a capture/analysis/critique taxonomy. Their conclusion was measured: automation is “a promising way to augment existing approaches,” not replace them. Ivory & Hearst (2001) That distinction matters. The goal for agents is not to judge final quality like a human expert. The goal is to prune bad candidates cheaply and rank the rest. The taxonomy gives agents a structure: capture the artifact, analyze its properties, critique against known patterns. This is exactly the loop a reward model runs. The precedent proves the approach has been viable for over two decades.
The proxy trap: Reinecke
The computable line has a hard limit. Reinecke et al. (CHI 2013) showed that 450 websites rated by 548 volunteers after 500 ms viewing yield a clear finding: complexity and colorfulness models explain roughly half the variance in appeal, with complexity the stronger predictor. Reinecke et al. (2013) That sounds like a win, but read the fine print. These models predict first impressions, not usability, not quality, not task success. A 500 ms glance is a proxy for salience, not for whether a user can complete a purchase. Agents that optimize for first-impression scores will produce pretty landing pages that fail on conversion. The trap is treating a proxy as the target. Use Reinecke’s models only if the reward is explicitly “initial appeal” — and never as a general quality signal.
Folk rules vs evidence: 7±2 and heuristics
Some design rules are so famous they feel like physics. Miller’s 1956 “magical number 7±2” is one. The evidence says otherwise. Cowan (2001) shows the number was “a rough estimate and a rhetorical device”; the evidence converges on about 4 chunks, with a range of 3–5. Cowan (2001) An agent that hard-codes “keep it under 7 items” is optimizing for a myth. The real constraint is 4±1, and even that depends on chunking. Similarly, Nielsen’s heuristics are often treated as commandments. NN/g clarifies they are “broad rules of thumb, not specific usability guidelines.” They were derived empirically from 249 usability problems across 11 projects in Nielsen’s CHI 1994 paper. NN/g: 10 Usability Heuristics That empirical origin is useful — the heuristics encode real failure modes — but they are qualitative. They do not yield a computable score. An agent cannot calculate “user control and freedom” from pixels. Use them as a checklist for generated artifacts, not as a reward function.
What agents can safely use
The safe set is clear. First, the Ngo 14 measures: they are validated, computable, and capture layout structure. Optimize for balance, symmetry, and rhythm as primary signals; treat density and economy as constraints. Second, Reinecke’s complexity and colorfulness models, but only as a first-impression filter, never as a quality target. Third, the 4±1 chunk limit from Cowan as a hard constraint on navigation and list lengths. Fourth, Nielsen’s heuristics as a post-hoc audit — a human-readable pass that catches gross violations. This set gives an agent a multi-objective reward: Ngo for structure, Reinecke for snap appeal, Cowan for capacity, Nielsen for sanity. None of these is the whole truth. Together, they are a strong prior. This mirrors the lesson from Fitts–Hicks agent arithmetic — perceptual-motor laws are computable but have limits; same is true here. Fitts–Hicks agent arithmetic
Close
The thesis of this blog is that AI agents can learn to design better. The path to that is not a single aesthetic oracle. It is a stack of partial, measured signals. Birkhoff started the count, Ngo made it computable, Ivory & Hearst proved automation was viable, Reinecke drew the proxy boundary, and Cowan corrected the folk law. An agent that optimizes these measured criteria will out-design an agent that optimizes vibes — because vibes have no gradient. The line ends where research only predicts first impressions or where folk rules collapse. Stay inside the measured zone. That is where learning happens.
This review is based on official documentation and published research — we did not run hands-on tests.