Shippable value

What JudgmentKit prevents

JudgmentKit catches when AI-generated UI turns implementation mechanics into UX, then gives the agent repair instructions before the work ships.

Implementation language leak

Internal objects stop becoming the product surface.

A raw generation pass tends to expose database fields, schema labels, endpoint status, and debug structure. JudgmentKit redirects the agent toward the actual activity: a support lead reviewing refund evidence and leaving a handoff.

Baseline failure
Source mechanics become navigation and labels the support lead has to translate.
JudgmentKit catches
The activity, participant, decision, outcome, and disclosure boundary are missing or weak.
Repaired outcome
The primary surface uses refund review language and keeps diagnostics secondary.
Raw refund triage artifact screenshot Raw brief JudgmentKit-guided refund triage artifact screenshot JudgmentKit-guided
Same refund activity, different source discipline.

Unsafe action boundary

Approval work gets a human decision point.

Refund review is not a generic submit flow. The user needs to approve, send to policy review, or return for missing evidence with a reason the next owner can trust.

Baseline failure
Actions advance the case before the approval boundary and handoff reason are explicit.
JudgmentKit catches
The primary action, reversible states, and next-owner receipt are not grounded in the review activity.
Repaired outcome
The interface separates decision, reason, and handoff so the case can move forward cleanly.
Decision

Approve, send to policy review, or return for missing evidence.

Required reason

The user records the evidence behind the selected outcome.

Handoff

The next owner receives the action, rationale, and unresolved evidence.

Missing accessibility evidence

Claims are not accepted without evidence.

JudgmentKit treats accessibility as part of the implementation contract. An agent cannot pass by saying the UI is accessible in a rationale while leaving required evidence out of the candidate.

Baseline failure
Labels, focus-visible behavior, status messaging, or conditional modal evidence are absent.
JudgmentKit catches
The failed evidence keys are grouped into repair instructions for the agent.
Repaired outcome
The candidate resubmits with concrete accessibility evidence before acceptance.
Failed check

Missing focus, keyboard, status, or conditional evidence.

Repair

Add the evidence inside the implementation candidate, not only in the rationale.

Result

The agent repairs, resubmits, and either passes or stops for human review after the attempt limit.

Evidence, not the main story

Audit material stays available.

The public value path above is the product story. The current hosted MCP release is 0.6.5; linked reports are historical committed eval artifacts for deterministic proof, model matrix, and repair-loop data, not current release acceptance proof. The public archive includes only the evidence intended for external audit.

Public evaluation report Latest committed eval report Eval catalog JSON One-shot demo