Cerberus
Start free
Cerberus · a product design case study · april 2026

Legible coaching,
not opaque scoring.

A design exploration of a GEO optimisation product that presumes the writer is smart, and earns every number it shows. Five surfaces, one trust contract.

Scope
6 surfaces · 1 design system
Artefact
High-fidelity HTML prototype
User archetype
Rowan Whelan · cadence.co
Type system
Instrument Serif · Bricolage Grotesque · JetBrains Mono
Palette
Stone neutrals · Champagne accent
The argument

Most tools that score writing tell you what without telling you why. They produce a number, a grade, a traffic light - and leave the writer to infer the underlying model by trial and error. This works for compliance. It fails for craft.

Cerberus is designed for the writer who wants to argue with the machine. Every recommendation carries its cohort, its interval, its provenance. Every override becomes an event in a permanent log. The model is not a black box - it is a set of observations, each editable, each with visible weight.

The five surfaces below are the minimum case to prove this. They share one spine: confidence is always the first thing you see, and the last thing you can edit.

Four principles

The rules the design refuses to break.

01 · confidence

Confidence is a spine, not a decoration.

Every number, every recommendation, every observation declares its tier - high, medium, low - in the same place, the same way. The spine never moves, across all five surfaces.

02 · provenance

"Why this?" is always one click away.

Cohort size, model version, base rate, interval, the specific examples that drove the inference - none of it is buried, none of it is inferred. The receipts are always shown, not assumed.

03 · correction

Disagreeing is a first-class verb.

Override isn't a rare admin action - it's how the product learns. The weight of every correction is visible before it's saved, and stays in the log forever after.

04 · restraint

Champagne is a flourish, not a filler.

Colour is reserved for single moments - a threshold crossed, a scope highlighted, a correction weighted. Otherwise the surface is stone. The less we say, the more each word carries.

Five surfaces

The minimum case.

06
Accountability surface
Every applied recommendation, reconciled against its predicted interval. A calibration chart, a ledger, and three cohort-reliability panels- the model's annual review, written by the user's own work.

The model's reckoning with its own predictions.

Most tools stop predicting once the user has applied the change. Cerberus holds itself to the prediction. When the model is wrong, the ledger says so. When a cohort is thin, the cohort says so. When a pattern is contested, it is downgraded visibly, on the next run.

Calibration: predicted interval vs observed outcome, 30d windows
Ledger: every applied rec, expandable for provenance
Cohort cards: reliable, provisional, contested
Tone: owns its errors, names its thin samples
01
Command surface
The daily page. Visibility score, the citation trend, and the day's 3 top recommendations - each with visible confidence and a one-click Apply. Month-3 and month-1 states toggle from the header.

A dashboard that starts with why you should trust it.

The visibility number is never alone. It carries its 90-day trend, its cohort, its prior month's delta - and always, always, its confidence interval. The dashboard assumes the reader will look, not glance.

Hero metric: citation-worthiness, not vanity score
Trend shape: sparkline with interval ribbon
Cold-start: honest "not enough data yet" state
Recommendations: high / medium / low tiers, never mixed
02
Per-piece surface
The three-column editor: manuscript on the left, scored passages in the middle, inline recommendation and receipts on the right. The user never loses their place in their own text.

Scoring that reads the prose, not the metadata.

Each scored passage is marked in the manuscript itself, anchored to the sentence it concerns. The score is a conversation between the writer and the model, not a verdict handed down from outside the text.

Anchor model: passage-level, not paragraph-level
Recommendation carries: cohort, lift, interval, preview
"Why this": unfurls on the same page, never a modal
Override: reachable in one click from any card
03
Reflective surface · A + B
The "What I've Learned" page - two interchangeable treatments of the same data. Readers toggle between them from the top-right of either page.

Two ways of hearing the same portrait.

The data is identical: seventeen observations about Rowan's voice, audience, goals, and how AI systems quote him. The treatments differ. Editorial is a magazine spread - chapters, rules, a running ticker. Portrait is a letter - second person, one column, narrative prose laced with the same numbers.

Both support: inline correction, evidence reveal
Contested observations: kept, struck, quoted verbatim
Confidence spine: preserved across both treatments
Trade: editorial rewards skim; portrait rewards sit
04
Component gallery
The recommendation card in its three tiers, three variants, and four states. The atomic coaching unit, shown in every room of the product.

One card, many rooms, one shape.

A recommendation appears on the dashboard, inline in the editor, on the tracker, and in the weekly email digest. The affordances compress. The contract - confidence, evidence, predicted lift, override - does not.

Tiers: high (Apply), medium (Preview), low (verify first)
Variants: canonical, inline, digest
States: default, applied, dismissed, flourish
Champagne: used for exactly one state - a tier-crossing
05
Trust contract
The four-frame override sequence: the recommendation in context, the dialog, the corrected state six seconds later, and the receipt visible fourteen days on. The heart of the product.

Disagreeing, all the way through.

The override dialog is the only surface in the product that calculates, in plain numbers, how much weight a user's input will carry before it's submitted. The corrected card is the only element that never vanishes. The log is the only place the user's corrections speak back to them, later.

Dialog: reason chips + sentence + scope + visible weight
Scope: one piece, one pattern, or one whole topic
Corrected card: original struck; correction quoted verbatim
Receipt: appears in the Portrait trail, permanent
Plans

One engine, three volumes.

Cerberus is in free beta — every plan is open at no charge until the beta closes. Paid billing begins after that date.

Free
Free
no card required
  • 10 evaluations / day
  • 1 project
  • no citation testing
Pro
£75/mo
£60/mo billed annually
  • 200 evaluations / day
  • unlimited projects
  • 20 citation tests / day
  • batch scoring, 20 at a time
Agency
£140/mo
£112/mo billed annually
  • 1,000 evaluations / day
  • unlimited projects
  • 50 citation tests / day
  • batch scoring, 50 at a time
Colophon
Set in Instrument Serif for editorial voice, Bricolage Grotesque for UI, JetBrains Mono for metadata. Colours drawn from stone neutrals with a single champagne accent. No gradients, no faux-glass, no ambient glow. Every number in the case study is a plausible figure from a plausible user, not a real one - but the model behind them is honest about what it would and wouldn't know.
cerberus · case study · edition 1 · rowan whelan · cadence.co