The DOPE Framework
Four lenses on
AI interaction quality.
Every interface decision in this library can be evaluated through four properties. Together they describe what it actually feels like to work with an AI system over time — not just whether the model is smart, but whether the surface around it is safe to delegate to.
Read in order — D, O, P, E. The acronym is fixed and the ordering is intentional: directability and observability are the load-bearing pair, predictability and explainability extend them.
The Picture
Four properties.
One overlap.
The diagram shows the four dimensions as overlapping circles. Each individual circle improves an AI surface; the intersection — where all four are present — is what enterprise users actually trust enough to delegate to.
The argument compresses into a single image.
From the essay “Transparency is Infrastructure”Observability
see what the system thinks
Directability
steer in real time
Predictability
know what comes next
Explainability
understand why it decided
No audit without transparency.No enterprise AI without audit.
DOPE — adapted from human-machine learning research
Dimension 1 of 4
Directability
Can the user steer, intervene, and redirect?
Direct manipulation surfaces, scoped controls, and intervention points that let people change the system mid-flight without restating the entire context.
Fails when
The only way to influence the system is to write another message and hope it lands. There is no scoped target, no checkpoint, no override.
Exemplar Patterns
Permission System
Autonomy Dial
Checkpoint
Approval Gates
Working Draft
Editable Plan
Highlighter
Select & Transform
Dimension 2 of 4
Observability
Is the state of the system visible?
Persistent surfaces that make context, queues, history, sources, and uncertainty inspectable — instead of buried inside a transcript or hidden behind confident prose.
Fails when
The output is fluent but the path that produced it is opaque. Memory, tools, sources, and confidence are off-screen.
Exemplar Patterns
Control Room
Agentic Dashboard
Flight Recorder
Action Audit & Undo
Personal Dossier
AI Memory
Trust Meter
Confidence Signals
Dimension 3 of 4
Predictability
Does the same input produce the same shape of output?
Defined levels, saved presets, durable rules, and named structures so users can build a reliable mental model and re-run with confidence.
Fails when
Two near-identical prompts produce wildly different shapes of answer. Users cannot anticipate the response well enough to plan around it.
Exemplar Patterns
Cockpit
Property Panel
Rulebook
Constraint Rules Engine
Map
Semantic Zoom
Telescope
Progressive Disclosure
Dimension 4 of 4
Explainability
Can the user understand why the answer is what it is?
Visible reasoning chains, mapped sources, weighted inputs, and per-step rationale that turn a confident output into an inspectable artifact.
Fails when
The answer reads as authoritative but no surface explains how it was derived. Users either over-trust the output or have no path to challenge it.
Exemplar Patterns
X-Ray
Reasoning Trace
Footnote Index
Citation Trail
Flight Plan
Intent Preview
Director's Storyboard
Multi-Modal Input
How to use this
A diagnostic, not a checklist.
The four dimensions are not a scorecard. A pattern that strengthens one dimension may neglect another, and that is often the right tradeoff. The point is to make the tradeoff visible.
When evaluating an AI surface — yours or someone else's — walk through D, O, P, E and ask: which of these does this design hide? The answer points directly at the failure mode the next iteration should address.
Directability
Ask: can the user steer, intervene, and redirect?
Observability
Ask: is the state of the system visible?
Predictability
Ask: does the same input produce the same shape of output?
Explainability
Ask: can the user understand why the answer is what it is?