What does Dense Data Workbenches conclude, in numbers, about “Preattentive Salience: Let the Anomalies Jump Out, and Only the Anomalies”?
Answer: Preattentive features are the least costly force in a back-office table: one red cell on a screen is seen without searching. But the force can be spent. Colour is the strongest and must carry a back-up channel, and red and green must not be paired; length suits size, not anomalies; orientation and closure are weaker and suit the second tier. p33 shows that pale fills sit only between 1.10 and 1.24 against white paper and are visible through hue alone, so the top tier needs a dark fill with white text. Conditional formatting has three tiers: red only for rules with a positive predictive value of at least 0.5, amber for 0.1 to 0.5, neutral marks for everything else. At Harbourbridge's base rate of 0.0150, a red rule needs a positive likelihood ratio of 65.7; the rules in Chapter 1's p10 reach only single digits. Colouring by band takes red from 435 rows a day to fewer than 3, at the price that half of the anomalies must be caught by means other than colour, which is the account Chapter 4 settles.
Scope of the numbers: The book is not accounting, audit or legal advice. Marked boxes separate published evidence, runnable calculations, sources and the author's hypotheses, each stating what would refute it; the source studied is the reporting system of the Tang finance minister Liu Yan, given in paraphrase.
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: 0.0150, 0.5, 10, 65.7
| prevalence 0.0150; prior odds 1 | 65.67 |
| target PPV LR+ needed max false-positive rate at sensitivity 0.9 / 0.7 / 0.5 | |
| 0.10 | 7.3 0.1234 / 0.0959 / 0.0685 |
| 0.25 | 21.9 0.0411 / 0.0320 / 0.0228 |
| 0.50 | 65.7 0.0137 / 0.0107 / 0.0076 |
| 0.75 | 197.0 0.0046 / 0.0036 / 0.0025 |
| red rows per screen if a rule meets PPV 0.5 at sensitivity 0.7 | 0.525 |
| p10 rules: sensitivity, false-positive rate, LR+, implied PPV at this prevalence | |
| R1 difference not zero | 0.689 0.120 5.74 0.080 |
| R2 status unchanged 48h | 0.567 0.100 5.67 0.079 |
| R4 new customer | 0.222 0.080 2.78 0.041 |
| R6 note not empty | 0.300 0.040 7.50 0.103 |
Show the full program
# Chapter 3. p31: How strong must a rule's evidence be before it may paint a cell red? Odds form of Bayes.
# Shared sample (p10): 1,200 rows a day, 18 real anomalies, prevalence 0.015, 25 rows a screen.
# posterior odds = prior odds x LR+, with LR+ = sensitivity / false-positive rate. For a target PPV the rule
# needs LR+ >= ((1 - prev) / prev) x (PPV / (1 - PPV)). The rule settings at the bottom are copied from p10
# (probability of firing on a normal row, and on each anomaly type weighted by p10's daily counts).
from fractions import Fraction
ROWS, ANOM, PER_SCREEN = 1200, 18, 25
PREV = Fraction(ANOM, ROWS)
COUNTS = {"A1": 6, "A2": 3, "A3": 4, "A4": 2, "A5": 3}
P10_RULES = [("R1 difference not zero", 0.12, {"A1": 1.0, "A3": 1.0, "A2": 0.3, "A4": 0.3, "A5": 0.3}),
("R2 status unchanged 48h", 0.10, {"A1": 0.5, "A2": 0.5, "A3": 0.5, "A4": 0.5, "A5": 0.9}),
("R4 new customer", 0.08, {"A1": 0.2, "A2": 0.2, "A3": 0.2, "A4": 0.4, "A5": 0.2}),
("R6 note not empty", 0.04, {"A1": 0.3, "A2": 0.3, "A3": 0.3, "A4": 0.3, "A5": 0.3})]
def lr_needed(ppv):
return (1 - PREV) / PREV * ppv / (1 - ppv)
if __name__ == "__main__":
prior = PREV / (1 - PREV)
print(" prevalence {:.4f}; prior odds 1 : {:.2f}".format(float(PREV), float(1 / prior)))
print(" target PPV LR+ needed max false-positive rate at sensitivity 0.9 / 0.7 / 0.5")
for ppv in (Fraction(1, 10), Fraction(1, 4), Fraction(1, 2), Fraction(3, 4)):
need = lr_needed(ppv)
fprs = " / ".join("{:.4f}".format(float(Fraction(s, 10) / need)) for s in (9, 7, 5))
print(" {:10.2f} {:12.1f} {}".format(float(ppv), float(need), fprs))
print(" red rows per screen if a rule meets PPV 0.5 at sensitivity 0.7: {:.3f}".format(
float(PER_SCREEN * PREV * Fraction(7, 10) * 2)))
print(" p10 rules: sensitivity, false-positive rate, LR+, implied PPV at this prevalence")
for name, p0, p1 in P10_RULES:
sens = sum(COUNTS[k] * p1[k] for k in COUNTS) / ANOM
lr = sens / p0
post = float(prior) * lr
print(" {:<24s} {:6.3f} {:6.3f} {:6.2f} {:7.3f}".format(name, sens, p0, lr, post / (1 + post)))Single source and citation
Fan, Y. (2026). Cognitive Unloading of Dense Data Workbenches, Conditional-Format Salience and Zero-Training Error-Proof Form Engineering (Version 0.1.0), Chapter 3: Preattentive Salience: Let the Anomalies Jump Out, and Only the Anomalies. 5D Library. DOI pending
DOI pending · monograph page goes live after publication · 5DL-B-138 · v0.1.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026138,
author = {Fan, Yafu},
title = {Cognitive Unloading of Dense Data Workbenches, Conditional-Format Salience and Zero-Training Error-Proof Form Engineering},
year = {2026},
publisher = {5D Library},
edition = {0.1.0},
note = {DOI pending}
}Other chapter answers in this book
- Chapter 1 · Method: Three Losses, Four Metrics, Evidence Grades and the Liu Yan Reporting System
- Chapter 2 · Scan Paths and the Skeleton of a Table: Column Order, Alignment, Frozen Panes and Density
- Chapter 4 · Alarm Thresholds: Signal Detection, ROC, the Cost Matrix and a Per-Screen Alarm Cap
- Chapter 6 · Zero-Training Error-Proof Forms: Constraints, Defaults, Inline Validation, Confirm-or-Undo and Batch Guards
- Chapter 7 · The Workbench Playbook: Reconciliation, Orders, Stock and Customer Records, and the Distillation Spec
- Chapter 8 · Practice: Rebuilding the Harbourbridge Reconciliation Back Office from Diagnosis to Redesign