20 interactive elements
OUTCOME SIGNAL

01 · PROBLEM

AI interfaces often collapse intent, progress, evidence, reasoning and approval into one opaque chat stream.

PUBLISHED CLAIMS / PRINCIPLES

01

20 interactive elements

02

Evidence and intent stay visible

03

Human control stays visible

02 · APPROACH

01

Discover: retrieve patterns and verify sources

02

Frame: compare problem frames and keep uncertainty visible

03

Prototype: test motion, hierarchy, tools and human checkpoints

04

Validate: inspect contradictions and own the recommendation

PLAY WITH THE SYSTEM

Collect the signalScan evidence before proposing UI.

03 · DECISIONS

01

Designed one connected use case instead of separate AI concept pages

Frame
02

Attached evidence and limits to recommendations

Control
03

Used motion for system state and checkpoints for human judgment

System

04 · OUTCOME

A single playable study of how loading, thinking, evidence, tools, recommendations and approval can work as one accountable AI interface.

Open-source MIT notices are preserved in source. Product framing is Ayodhya’s exploration.

05 · ONE CONNECTED METHOD

AI widens the work. Human judgment closes it.

01

Discover

AI · Retrieve patternsDesigner · Verify sources
02

Frame

AI · Generate alternativesDesigner · Choose the constraint
03

Prototype

AI · Accelerate variantsDesigner · Decide what to prove
04

Validate

AI · Synthesize evidenceDesigner · Own the recommendation

06 · INTERACTIVE USE CASE

Scroll it. Type. Switch. Approve. Replay.

Interface lab running0.0s21 live patterns · local only
01Streaming answer

Clinical Trial Tool carries the strongest published healthcare outcome: the case reports a 40% streamlined workflow and 90% user satisfaction.

Source attached · native case study
02Reasoning trace
Evidence trace
Match the question to case tags and outcome claims.
03Human approval

Use the case-study claim in a recruiter summary?

ActionPublish claimRiskNeeds source label
04Agent task state
✓Evidence gathereddone
✓Contradictions checkeddone
↻Recommendation draftedactive
·Human approvalwaiting
05Prompt contract
GoalFind decision-ready evidence
MustShow source and limit
NeverInvent experience
06Intent composer
search_casesrank_evidenceattach_sources
07Confidence control
72%Review recommended
08Recommendation
Best evidence match

Healthcare depth + quantified outcome claims.

09Before / after
AI gives one opaque answerAI shows answer + evidence + limits
10Escalation rule
If confidence < 60%Stop before action

Show the unresolved choice to a person with the evidence already gathered.

11Loading state

Retrieving evidence2.4s

12Tool chips
✓ search_cases7 results
✓ rank_evidence32ms
↻ attach_sourcesrunning
13Context cards
Healthcare operationsClinical Trial Tool

40% workflow streamlined

Native case study ↗
Enterprise SaaSApplication Automation

40% faster setup

Native case study ↗
AI-native conceptOrbit

Human owns the decision

Native case study ↗
14Sidebar nav
15Command search
⌘K
16Flowchart
User intenttyped or selected
→
Retrievelocal case index
→
Checksource + limit
→
Show answerwith evidence
17Insight cards
02 / 03Approval at the risk boundary
18Code block
01  const answer = retrieve(intent)
02  return attachSources(answer)
03  // never invent experience
19Fine-tune card
Answer confidence64%
Evidence is visible. Human judgment stays in control.
20Selection actions

AI proposes alternatives, but the designer verifies the evidence, chooses the direction and owns the recommendation.

21Generative surface
assembling…

The catalog is fixed - the assembly is generated. Scripted walkthrough, no model calls: the constraint grammar is the design, the surface is the output.