Chart library nine recurring chart types, all built from the same primitives

Lines and intervals.

A chart is a sourced argument, not a decoration. Every figure carries its own title, a one-sentence read, and a three-line source block in Söhne Mono. The plot itself uses two weights of ink, one weight of greyed italic (the said series), and a single signal-red annotation for the gap. No rainbows, no legends floating in the top-right corner of a tile.

01 · paired bars Hold-out accuracy on 350 replicated studies. fig. 1 · lab replication · Q1 26
0 25 50 75 100 hold-out accuracy · % consumer 62 93 financial 58 91 health 65 95 labor 60 92 95% CI gap · 31 pts
stated (focus-group baseline) did (subconscious, held-out) observed gap 95% CI whiskers

Across four domains and 350 published studies, behavioral match of a stated-preference baseline sits around 60%. Subconscious holds 93% [91.5, 94.5] on the same held-out choices — 31 points, 95% CI shown. The gap is the whole company.

Source internal replication, Q1 26
Method choice-level AUC, 5-fold held-out
Sample N=350 studies · 42 journals
Status pre-print · arXiv ref
02 · dot plot, log x Cost per study, seven legacy vendors vs us. fig. 2 · USD, list · 26.04
$500 $5k $50k $500k $5M $50M Kantar $85k Ipsos Nielsen YouGov McKinsey clinical, bespoke DIY panel us us 170× to 16,000× less

Legacy studies span three orders of magnitude, from $18k DIY to $8M clinical. A subconscious run is priced as a fraction of any value above; specific numbers stay between us and the buyer. The vendors on the x-axis are selling labor; we are selling a model.

Source public pricing pages · sales quotes · 2025 procurement disclosures
Method median list, US, 1 typical study
Caveat bespoke work varies; we quote the median
Status verified, 26.04
03 · small multiple, lines Revealed-preference trend, 2019 to 2026, four categories. fig. 3 · index: 2019 = 100
01 · coffee
Said “dark roast.” Bought sweet cold.
said · 101 did · 238 '19 '26
02 · streaming
Said “quality dramas.” Watched reality.
said · 94 did · 214 '19 '26
03 · auto
Said “sedan.” Bought SUV.
said · 74 did · 192 '19 '26

In every category we have tested, stated preference is flat while revealed behavior climbs or diverges sharply. Survey data is a photograph of the lie at equilibrium.

Source Nielsen Scantrack · Netflix Top 10 audit · IHS Polk new-car sales
Method indexed to 2019=100, volume-weighted · Status public
04 · ridgeline · distribution Predicted willingness-to-pay, three counterfactual conditions. fig. 4 · synthetic population n=10,000
$0 $25 $50 $75 $100 ±1σ · baseline · $30–$54 baseline · $42 median price cut · $32 median bundle · $58 median

The shape of the distribution, not just the median, is the answer. A price cut compresses the tail; a bundle pulls the mass rightward. The same population, three counterfactual worlds, on one canvas.

Source simulated run · n=10,000 synthetic respondents
Method latent-utility sampling, conditioned on stated panel
Status illustrative · re-run per client
05 · waterfall · counterfactual decomp Where the incremental $2.4M comes from. fig. 5 · case · fortune 500 beverage
baseline pack price placement cannibal. projected $8.1M +1.1 +0.9 +0.6 -0.2 $10.5M +$2.4M · +29%

The simulator decomposes the projected uplift into its components. Pack, price, placement each contribute separately; a small cannibalization against the flagship does appear, honestly, as a red negative bar.

Source Fortune 500 beverage · anonymized · Q4 25
Method do-calculus over simulator choice graph
Caveat projection, not guarantee
Status replicated in pilot
06 · dumbbell · before / after Five operational metrics, legacy vs subconscious. fig. 6 · procurement audit · set 26.04
legacy vendor subconscious 01 · time to answer 6 weeks 4 minutes 15,000× faster 02 · cost per study $85,000 us 170× cheaper 03 · hold-out accuracy 62% 93% +31 points 04 · panel fraud rate 46% 0% no panel, no fraud 05 · iterations / quarter 2 40+ 20× more shots on goal

Every row a different dimension; every line the same story. The grey circle is what the industry quotes. The ink circle is what a subconscious run delivers. The dumbbell is the pitch.

Source 8 customer procurement audits · 2025
Method matched study scope
Status anonymized composite
07 · stat cards · count-up The headline number, settled — not ticked. fig. 7 · mono · tabular · 72px
93%
alignment with held-out human outcomes
350+
academic studies replicated
5×
prices raised · revenue doubled

A stat is a verdict. It earns one mono numeral at 72px and a single line of source in caption grey. The number settles toward its target on a quart ease — a gauge coming to rest, never a slot machine.

Source internal replication · pricing pilots · 2025–26
Method count-up animates 0→target on viewport entry, staggered 100ms
Status see fig. 1 for the accuracy interval
08 · bar comparison · paired Traditional research versus a subconscious run. fig. 8 · 2-row pairs · no axis
Time to insight
6–8 wk
ours
<5 min
Behavioral accuracy
62%
ours
93%

When the comparison is two numbers, skip the axis. A track, a fill, the value in mono. Grey fill is the legacy quote; ink fill is the run. The fills grow from the left, staggered, so the eye reads them as a sequence — not a wall.

Source 8 customer procurement audits · 2025
Method matched study scope · single ink, single grey
Status anonymized composite
09 · dot matrix · population Hold-out match, one dot per synthetic respondent. fig. 9 · 140 cells · grey / ink only

When the unit is a person, draw the person. 140 cells, ink for a matched choice, faint grey for a miss. No third colour, no shading — the density is the statistic.

Source simulated run · 100 synthetic respondents vs held-out humans
Method one cell per respondent, ink = choice matched · dots fade in, staggered
Status illustrative · 93 of 100 matched
rules · chart doctrine Six constraints, applied to every figure on this page. enforced · DOCTRINE.md§3
01Never more than two series in one chart, unless comparing four or more personas.
02Our line is always var(--color). The comparison is always var(--grey-300), dashed.
03Numbers in Söhne Mono, tabular. Axis labels in Söhne Mono 9–10px.
04No gridlines except a dotted x-every-N. No bounding box. No legend unless the chart needs one.
05Endpoints get a dot and a callout. Never label every point.
06Fills only for area charts showing uncertainty — a ±confidence band at low alpha.
Rules two ink weights plus one grey, dashed "said" line, one red annotation per chart  ·  Every figure carries number, title, read-sentence, three-line source block  ·  Banned rainbow palettes · 3D extrusion · donut charts · gradient fills · axis tick marks on every unit · legends floating inside the plot