Case Study
Judgment Layer
A reading-first interface for evaluating AI-generated text
Problem
Problem
AI-generated text is fluent, but difficult to evaluate.
Most interfaces optimize for producing better outputs.
They do not help users understand how meaning is constructed, changed, or degraded across versions.
As a result, interaction collapses into acceptance or rejection.
The intermediate space — where interpretation and judgment occur — is largely absent.
Insight
Insight
Improving outputs is not the only problem.
Users need ways to:
- inspect how meaning is produced
- compare versions without losing context
- recognize rhetorical shifts rather than surface differences
Making reasoning visible is a different problem than making output better.
Design Approach
Design Approach
Judgment Layer reframes interaction as reading, not prompting.
The system is built around three principles:
- Keep the text intact
No rewriting or suggestion layer - Preserve context across versions
Multiple passages remain visible simultaneously - Name meaning shifts explicitly
Changes are interpreted, not corrected
Interaction Model
Interaction Model
Two modes support different forms of evaluation:
Passage Mode
Enables close reading of a single text.
- Annotates concept operation within the passage
- Connects segments to conceptual panels
- Supports structured interpretation rather than summary
Compare Mode
Makes meaning shifts visible across versions.
- Displays multiple passages side by side
- Links corresponding segments
- Surfaces rhetorical and structural differences
Key Decisions
Key Decisions
- No optimization layer
- No “improve this” workflow
- No hidden transformations
- No scoring or ranking
The system does not attempt to produce better outputs.
It creates conditions under which outputs can be evaluated.
Implementation
Implementation
- Built with Astro, HTML, CSS, and JavaScript
- Deployed via Cloudflare Pages
- Uses D3 for concept mapping and relationship visualization
- Designed and implemented as a working system, not a mockup
Outcome
Outcome
Judgment Layer demonstrates an alternative model for AI interfaces:
- from generation → evaluation
- from output quality → reasoning visibility
- from passive consumption → active interpretation
It is both a working interface and a design argument about how AI systems should be encountered.