UI paired-artifact evidence

Activity-First UI Generation Evidence

A cautious public report on a small UI paired-artifact rerun. It compares committed baseline and JudgmentKit-guided UI artifacts; it is not an MCP pilot status page or a statistically powered benchmark.

The UI generation bottleneck

Generated interfaces often inherit the shape of the implementation material that fed them. Tables become screens, schemas become forms, internal labels become product language, and the user is left translating the interface back into the work they meant to do.

JudgmentKit treats that as a judgment problem before it is a styling problem. The relevant question is whether the screen helps a person understand evidence, make a decision, and leave a useful handoff.

What JudgmentKit changes

JudgmentKit adds an activity-first review layer before UI generation. It asks what work is being supported, who participates, what decision matters, what vocabulary belongs in the surface, and what diagnostic detail should stay out of the primary experience.

Activity fit

Generated UI is judged against the work, not just against the input structure.

Decision support

The surface must make the next judgment or handoff easier to complete.

Disclosure discipline

Implementation detail remains diagnostic unless the activity is setup, debugging, or audit.

Compact activity-first generation flow.
  1. Source briefProduct context and implementation-heavy material enter together.
  2. Activity reviewParticipant, objective, decision, outcome, and vocabulary are named.
  3. HandoffOnly ready guidance advances to UI generation.
  4. Generated UIThe draft is judged against the work, then loops back with findings.

How the evaluation works

The committed UI paired-artifact eval compares raw baseline artifacts with JudgmentKit-guided artifacts for the same UI tasks. Scores use a 0-100 weighted total with 0-5 metric scores underneath. Current hosted MCP release: 0.6.5. Latest UI eval MCP release: 0.6.5. These figures do not claim MCP pilot pass/fail status.

Qualitative paired-artifact evidence from live provider-generated artifacts only; not a statistically powered benchmark.

Desktop and mobile screenshots are captured as visual evidence for review. The screenshots support inspection, but the scoring source is the committed artifact text and deterministic rubric.

Latest committed eval run
2026-07-03 / mcp-0.6.5 / run-002
Current hosted MCP release
0.6.5
Latest UI eval MCP release
0.6.5
UI paired cases
2
UI paired pass rate
100%
UI guided wins
2
UI baseline wins
0
Ties
0
Claim level
repeated_pair_signal

Benchmarks

Score comparison: raw baseline versus JudgmentKit-guided output.
Paired score comparison Bars compare raw baseline and JudgmentKit-guided scores for each committed paired eval case. 0255075100 Refund triage handoff +36.08 Dinner playlist builder +41.65
Raw baseline JudgmentKit guided
Activity-fit terms increase while implementation leakage falls.

Refund triage handoff

Activity terms 1 -> 5
Implementation leaks 1 -> 0

Dinner playlist builder

Activity terms 2 -> 6
Implementation leaks 11 -> 0
Case Raw score Guided score Delta Winner Activity fit Leakage Evidence
Refund triage handoff 55.92 92 +36.08 JudgmentKit guided 1 -> 5 1 -> 0 Raw screenshot · Guided screenshot
Dinner playlist builder 36.82 78.47 +41.65 JudgmentKit guided 2 -> 6 11 -> 0 Raw screenshot · Guided screenshot

Example evidence

The model UI matrices separate the source context from JudgmentKit guidance and Material UI rendering. The visible matrix below uses Support refund triage as a compact example; the full set covers every committed use case.

Generation path
Raw brief
JudgmentKit skill context
Material UI only
JudgmentKit skill + Material UI
Fixture-rendered baseline Scripted renderer paths show the controlled fixture-rendered baseline for each context combination.
Fixture-rendered baseline - Raw brief screenshot source brief + sample case Fixture-rendered baseline - JudgmentKit skill context screenshot source brief + sample case + reviewed handoff + frontend skill context Fixture-rendered baseline - Material UI only screenshot source brief + sample case + Material UI adapter Fixture-rendered baseline - JudgmentKit skill + Material UI screenshot source brief + sample case + reviewed handoff + frontend skill context + Material UI adapter
Gemma 4 via LM Studio lms Local Gemma 4 captures show how a smaller local model responds to the same four context boundaries.
Gemma 4 via LM Studio lms - Raw brief screenshot source brief + sample case
Diagnostic only Gemma 4 via LM Studio lms - JudgmentKit skill context Status: Needs repair before evidence Failed checks: Token provenance failed, Capture quality failed
Gemma 4 via LM Studio lms - Material UI only screenshot source brief + sample case + Material UI adapter Gemma 4 via LM Studio lms - JudgmentKit skill + Material UI screenshot source brief + sample case + reviewed handoff + frontend skill context + Material UI adapter
GPT-5.5 xhigh via codex exec GPT-5.5 captures use extra-high reasoning to show the same matrix with a stronger model path.
GPT-5.5 xhigh via codex exec - Raw brief screenshot source brief + sample case
Diagnostic only GPT-5.5 xhigh via codex exec - JudgmentKit skill context Status: Needs repair before evidence Failed checks: Token provenance failed, Capture quality failed
GPT-5.5 xhigh via codex exec - Material UI only screenshot source brief + sample case + Material UI adapter GPT-5.5 xhigh via codex exec - JudgmentKit skill + Material UI screenshot source brief + sample case + reviewed handoff + frontend skill context + Material UI adapter

Committed use cases

Support refund triage

A support operations manager reviews refund escalation cases and decides approve, policy review, or missing evidence.

Open matrix Manifest JSON

Field service dispatch

A field operations manager assigns, reschedules, or escalates a repair visit using route, parts, and SLA constraints.

Open matrix Manifest JSON

Clinical intake review

An intake coordinator reviews administrative appointment readiness and decides schedule, return missing forms, or escalate insurance verification.

Open matrix Manifest JSON

B2B renewal risk review

A customer success manager reviews renewal risk and decides save plan, executive escalation, or wait for usage evidence.

Open matrix Manifest JSON

Limitations and future work

This report is intentionally narrow. It uses committed paired artifacts and committed model matrix captures. It does not claim broad model behavior, does not call live providers during site build, does not claim MCP pilot pass/fail status, and does not treat visual polish as proof of activity fit.

Future versions can add broader MCP impact runs, more surface types, reviewer agreement, richer interaction probes, and completed walkthrough videos without changing this page structure.

Run data

Implementation details are listed here for audit and reproduction rather than used as primary product language above.

Latest committed HTML report Latest committed JSON report Eval catalog JSON Model matrix index JSON