Judgment Layer

A reading-first interface for evaluating AI-generated text

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

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

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

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

  • 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

  • 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

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.