What does Design-Token Algebra conclude, in numbers, about “The 50-Millisecond First Impression: Exposure, Complexity, Prototypicality and the Distance to a Sale”?

Answer: The 50-millisecond research is real and its findings are sturdy: judgements of a page's visual appeal form at very short exposures and agree closely across raters and across exposures. But 50 milliseconds on a screen has to be registered in frames, and a 75 Hz screen cannot show it at all. Consistency comes mostly from panel size: 40 raters push the correlation between exposures to 0.927 but cannot push past the validity ceiling set by a shared halo. Complexity and prototypicality are the two parts of a first impression the group can pull directly, the first through token caps and the second through family layout skeletons. From attractive to sold runs a chain of three discounted inferences; even when each link is strong, only a correlation of 0.180 survives at the sales end, and testing it takes thousands to tens of thousands of visitors, whereas a first-impression test needs a few dozen to a hundred raters. So first-impression tests guard the first screen against slipping back, conversion tests answer the business question, and the two are registered and reported separately. Chapter 6 turns to how tokens for thirty sites are released and how drift is detected, where first-impression ratings return as a weekly monitoring measure.

Scope of the numbers: All brands, sites, colours and ratings are invented, every calculation comes from the book's programs, and legal matters are described only in general terms, not as legal advice. Each hypothesis states what would refute it.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.927, 40, 50

50 ms: single-rater ICC 0.177; raters for reliability 0.8 / 0.919 / 42
500 ms: single-rater ICC 0.232; raters for reliability 0.8 / 0.914 / 30
assumed corr(A, T)0.40; ceiling for corr(page mean, T) with the halo = 0.333
panel corr(50 ms mean, 500 ms mean) corr(500 ms mean, T) (median of 200 seeded panels)
50.616 0.276
100.762 0.317
200.862 0.330
400.927 0.338
sample corr(A, T) in these 30 pages0.532; corr(A + H, T): 0.349
Show the full program
# Chapter 5. p51: Consistent is not the same as right: panel size, reliability and a shared halo.
# The 30 Hengyu home pages of the chapter 1 sample (B01-K07). Every number below is an assumption chosen
# to show the arithmetic, not an estimate from any study. Each page has a latent appeal A (sd 0.6
# points) and a latent trustworthiness T (standardised); corr(A, T) = 0.40 by construction. A rating on
# the book's 1-7 scale is   rating = 4 + A + H + rater bias (sd 0.4) + noise,
# where H is a halo every rater shares (vividness of the hero image, sd 0.4, unrelated to T), and the
# noise sd is 1.5 at 50 ms and 1.25 at 500 ms (1.25 is the rater sd assumed in chapter 1). Seed 5151.
# Prints: single-rater reliability (ICC), Spearman-Brown panel reliability, the 50 ms vs 500 ms
# correlation of panel means for several panel sizes, and how well panel means track T.
import math
import random

rng = random.Random(5151)
SITES = 30
R_AT = 0.40
NOISE = {50: 1.5, 500: 1.25}
SD_A, SD_BIAS, SD_HALO = 0.6, 0.4, 0.4


def corr(x, y):
    mx, my = math.fsum(x) / len(x), math.fsum(y) / len(y)
    sxy = math.fsum((a - mx) * (b - my) for a, b in zip(x, y))
    sx = math.sqrt(math.fsum((a - mx) ** 2 for a in x))
    sy = math.sqrt(math.fsum((b - my) ** 2 for b in y))
    return sxy / (sx * sy)


Z = [rng.gauss(0, 1) for _ in range(SITES)]
A = [SD_A * z for z in Z]
T = [R_AT * z + math.sqrt(1 - R_AT ** 2) * rng.gauss(0, 1) for z in Z]
H = [rng.gauss(0, SD_HALO) for _ in range(SITES)]
signal = SD_A ** 2 + SD_HALO ** 2
for ms, sd in NOISE.items():
    icc = signal / (signal + SD_BIAS ** 2 + sd ** 2)
    need = [math.ceil(r / (1 - r) * (1 - icc) / icc) for r in (0.8, 0.9)]
    print("{:3d} ms: single-rater ICC {:.3f}; raters for reliability 0.8 / 0.9: {} / {}".format(
        ms, icc, need[0], need[1]))
print("assumed corr(A, T) = {:.2f}; ceiling for corr(page mean, T) with the halo = {:.3f}".format(
    R_AT, R_AT * SD_A / math.sqrt(signal)))
print("panel  corr(50 ms mean, 500 ms mean)  corr(500 ms mean, T)   (median of 200 seeded panels)")
for n in (5, 10, 20, 40):
    c1, c2 = [], []
    for _ in range(200):
        bias50 = [rng.gauss(0, SD_BIAS) for _ in range(n)]
        bias500 = [rng.gauss(0, SD_BIAS) for _ in range(n)]
        m50 = [math.fsum(4 + A[s] + H[s] + b + rng.gauss(0, NOISE[50]) for b in bias50) / n
               for s in range(SITES)]
        m500 = [math.fsum(4 + A[s] + H[s] + b + rng.gauss(0, NOISE[500]) for b in bias500) / n
                for s in range(SITES)]
        c1.append(corr(m50, m500))
        c2.append(corr(m500, T))
    c1.sort()
    c2.sort()
    print("{:5d}  {:30.3f}  {:20.3f}".format(n, c1[100], c2[100]))
print("sample corr(A, T) in these 30 pages: {:.3f}; corr(A + H, T): {:.3f}".format(
    corr(A, T), corr([a + h for a, h in zip(A, H)], T)))

Single source and citation

Fan, Y. (2026). Design-Token Algebra across Thirty Sites, Perceptually Uniform Colour Spaces and the Millisecond Science of Brand-Authority First Impressions (Version 0.1.0), Chapter 5: The 50-Millisecond First Impression: Exposure, Complexity, Prototypicality and the Distance to a Sale. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-140 · v0.1.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026140,
  author    = {Fan, Yafu},
  title     = {Design-Token Algebra across Thirty Sites, Perceptually Uniform Colour Spaces and the Millisecond Science of Brand-Authority First Impressions},
  year      = {2026},
  publisher = {5D Library},
  edition   = {0.1.0},
  note      = {DOI pending}
}

Other chapter answers in this book