What They Ask, What They Fear
"Your AI work is interesting, but have you actually designed for model behavior?"
That's the clean version. You'll also hear it as: "How is this different from building any product that happens to use an API?" Or the gentler probe after you present Agentic Labs: "What's your experience working directly with ML teams on model outputs?"
The surface question is about frontier-lab credentials. The underlying fear is simpler and more damaging: She'll treat AI as a feature to be skinned rather than a material with properties. The interviewer wants to know whether you understand that AI systems behave probabilistically, that their outputs require a different design contract than deterministic software, and that the interface between an agent and a human is the hardest unsolved design problem in the field.
That last part is where your reframe lives. It's a strong one. But it only works if you say the true thing first.
The Grain of Truth
You have not worked inside a frontier lab. You have not sat with a model-behavior team at Anthropic or OpenAI tuning how Claude or GPT responds to ambiguous prompts. You have not designed the RLHF interface that shapes model personality or built evaluation frameworks that determine whether outputs meet a safety threshold before deployment.
State this directly when asked. No hedge, no premature pivot. The moment you appear to be claiming frontier-lab experience you don't have, you lose the room. The reframe earns its credibility from the honesty that precedes it.
"I haven't worked inside a frontier lab. My AI work is at the application layer, where agent systems meet real users and real decisions. That's a different problem, and it's the one that's underserved right now."
One sentence of acknowledgment. Then move forward.
Why Explanation Doesn't Build Trust
The reframe is genuine. Here's the research that grounds it as a design claim.
Chen, Liao, Vaughan, and Bansal (2023) studied two types of AI explanations across decision-making tasks. Feature-based explanations, the kind most XAI interfaces ship today ("here's why the model made this prediction"), actually increased overreliance on incorrect AI predictions. When the AI was wrong and participants had initially been right, feature-based explanations dropped their accuracy to 41% on one task. Participants who received example-based explanations instead scored 75%. Roughly half the participants preferred the explanations that made them perform worse.
Buçinca, Malaya, and Gajos (2021) arrived at the same conclusion from a different direction. Cognitive-forcing interventions, designs that required users to engage analytically before seeing the AI's suggestion, reduced overreliance compared to standard explainable-AI approaches. But participants gave the lowest subjective ratings to the designs that helped them most. People preferred the less demanding systems. They performed worse with them.
Together, those findings point somewhere specific: the standard transparency playbook (show the user why the AI decided what it decided) can make the trust problem worse. More explanation fails users. They need designed calibration points, friction at the right moments, and an interaction architecture that helps them build an accurate mental model of when to trust and when to override.
That's a design problem, and the research identifies the gap your portfolio proves you've been filling.
Your Three Evidence Layers
All citable from junochen.com, with one access caveat flagged below.
Layer 1: Agentic Labs — Built Systems With Real Autonomy Thresholds
Your homepage now lists five systems built with Claude Code. Three carry the strongest evidence because each involves an agent making autonomous judgments that a human then acts on:
Brand Pulse continuously reads Reddit and X, scores mentions by sentiment and urgency, produces an hourly brand-health score plus a weekly agent-written narrative. The design problem you solved: when should a human trust the sentiment score enough to act on it versus investigate the underlying mentions?
Retail Velocity audits accounts daily and produces a ranked list with zero-to-one-hundred scores. The design problem: what does a field team need to see to trust an agent's account prioritization over their own judgment?
Carrier IQ runs carrier quote workflows in parallel, returns quotes with confidence scores, and turns eight carrier outputs into a broker decision surface. The design problem: where does agent autonomy end and broker judgment begin?
Each required you to design the autonomy threshold, the human-system operating contract that determines what the agent handles, what the human verifies, and where the handoff lives. That threshold is the specific capability the Chen and Buçinca research says the industry needs and doesn't have enough of. Lead with Carrier IQ in technically literate rooms. The insurance domain forces real-stakes trust calibration. An agent returning a quote with a confidence score is making a claim about its own reliability. The interface presenting that claim to a broker making a coverage decision for a client is doing exactly the work the research says matters.
Layer 2: "Trust Is the New Interface" — The Published Framework
Your essay lays out a five-handoffs framework: Intent-Setting, In-Progress, Output Review, Decision Gate, and Loop Feedback. It introduces the Watch→Verify→Delegate trust ladder: users first watch the agent work, then spot-check its outputs and build a reliability model, then gradually delegate as evidence accumulates.
The two dimensions interact. A user at the Watch stage needs full evidence display and manual confirmation at every Decision Gate. A user who has reached Delegate needs only exception flagging at that same gate, because the prior stages have built the reliability model that justifies lighter touch. The framework names the handoffs and the trust levels, then goes further: it specifies how the design of each handoff changes depending on where the user sits on the ladder. That's what makes it a system.
The essay names confidence signals, evidence trails, provenance, and gates at loop boundaries as the mechanisms that make the ladder climbable.
"I've published a framework for how trust develops between humans and autonomous systems. It's called the five-handoffs model, and every system I've built since is an implementation of it."
That sentence establishes thesis. It pulls the interviewer into the framework itself and away from a credential audit.
Layer 3: Agentic Rebuild Codas in Alibaba and Thermo Fisher
Both your Alibaba and Thermo Fisher case studies include agentic codas showing how the original product architectures would evolve with agent capabilities. The Alibaba coda maps natural-language chat to a structured procurement brief, with an agent scanning and filtering results and trust signals evaluated before autonomous transaction completion. The Thermo Fisher coda describes five agents running continuously across exceptions, document gaps, forecast drift, and partner signals, with one human gate remaining for batch QA release because regulatory signature is irreplaceable.
That Thermo Fisher boundary matters. The coda doesn't hand everything to the agent. It identifies the specific point where human oversight is non-negotiable and designs the system around that boundary. That judgment, knowing where to draw the line, is the design capability these companies are hiring for.
Both full cases are currently behind a client-side access gate. Labs and the trust essay are fully public. If you cite the codas verbally, that works. If someone tries to verify by visiting the URL, they'll hit a login page. Either remove the gate or be prepared to walk through the codas live with screenshots. Do not cite evidence an interviewer can't verify.
The Market Is Converging on Your Frame
The NIST AI Risk Management Framework describes human-AI configurations ranging from fully autonomous to fully manual, stating that AI systems may "make decisions autonomously, defer decisions to a human expert, or serve as an additional opinion for a human decision-maker." NIST also notes that presenting AI-system information to humans is complex because "humans perceive outputs and explanations differently according to preferences, traits, and skills." That finding mirrors Chen et al. almost exactly. Your trust ladder is a design implementation of the graduated spectrum NIST describes.
EU AI Act Article 14 lists the human oversight capabilities high-risk AI systems must support: understand capacities and limitations, monitor operation, detect anomalies, remain aware of automation bias, correctly interpret outputs, decide to override or reverse, and intervene or interrupt. Map that list against your five-handoffs framework and the correspondence is nearly one-to-one.
Use this lightly. The human-system operating contract you've been designing predates and exceeds compliance, but when the largest regulatory bodies in the world independently arrive at the same problem definition, that validates the problem space.
"The governance world and the product world are converging on the same question: what does the human need to see, verify, and override when an agent acts? I've been designing that operating contract across five published systems."
Tier-Specific Deployment
AI-native startups (Anthropic, OpenAI, Suno). The objection is sharpest here. OpenAI has a Model Designer role that explicitly asks candidates to "understand, predict, and design model behavior." You are not that candidate. But their Product Design Manager posting asks for someone who builds "clear, cohesive experiences that build trust" and rethinks UI patterns for interactions "native to AI." That's your problem space. Even Suno's Head of Product Design posting asks for a hands-on leader who balances speed, experimentation, craft, and emotional resonance across AI-driven products. It doesn't mention model behavior, trust calibration, or agent autonomy at all. Even inside frontier companies, most design roles are application-layer roles. Lead with the trust essay and Carrier IQ. Expect the interviewer to test your craft altitude on agent interaction patterns. Be ready to sketch a trust-calibration flow live.
Growth-stage platforms. The objection is softer, usually framed as "How do you think about AI in the product?" They want to know you won't bolt AI onto existing flows as a feature. Lead with the five-handoffs framework as a product-thinking tool, then ground it in whichever lab system maps closest to their domain. The Buçinca finding (users prefer the designs that make them perform worse) is a strong talking point here because it reframes AI design as a behavioral-design problem, which growth-stage product leaders understand intuitively.
Enterprise companies. The objection rarely surfaces explicitly. Enterprise buyers care about whether you can navigate organizational complexity around AI adoption. Lead with the Thermo Fisher coda: five agents, one human gate, regulatory boundary preserved. The trust ladder maps directly to enterprise change management. You don't ask an organization to delegate to AI on day one. You design the Watch→Verify→Delegate progression into the rollout itself.
Confidence Rating
Reframe strength: High confidence. The research foundation is strong. The evidence layers are published and verifiable (with the access caveat on two cases). The market-validation signals are directionally clear. The reframe works because it doesn't deny the gap. It names which problem matters more and provides proof you've been working on it.
Risk factor: Moderate. In a room with someone who has actually designed model behavior at a frontier lab, the reframe may land as "she's good at the application layer but doesn't understand the model layer." You cannot fully neutralize this. What you can do is show that your interaction-layer work already accounts for model-level properties. Carrier IQ's confidence scores are a claim about model reliability surfaced at the decision point. The trust ladder's Watch stage depends on the user observing output variance and building a mental model of when the agent is reliable. Name those connections explicitly: "The confidence scores in Carrier IQ are a model-reliability signal. The design question is how to present that signal so a broker calibrates trust accurately rather than either over-relying or ignoring it. That's the Chen et al. problem." If the room is deeply technical, ask a question that shows genuine engagement with model-level constraints: "How are you thinking about surfacing calibration quality to users when the model's own confidence estimates are unreliable?" That question demonstrates you understand that confidence scores can lie, which is a model-behavior insight applied at the interaction layer.
Quick-Reference Card
Scan this before you walk in.
Surface question: "Have you designed for model behavior inside a frontier lab?"
Underlying fear: She'll treat AI as a feature to skin, not a material with properties.
Grain of truth: "I haven't worked inside a frontier lab. My AI work is at the application layer, where agent systems meet real users and real decisions."
Bridge to reframe: "The research now shows that explanation doesn't automatically build trust. The harder design problem is calibrating when humans should trust, verify, or override an agent's judgment. That's the problem I've been designing for."
Lead evidence by tier:
- AI-native → Carrier IQ + trust essay + be ready to sketch live
- Growth-stage → Five-handoffs framework + domain-matched lab
- Enterprise → Thermo Fisher coda (five agents, one human gate)
Key phrases:
- "I design the human-system operating contract: what the agent handles, what the human verifies, where the handoff lives."
- "The governance world and the product world are converging on the same question. I've been designing that operating contract across five published systems."
Do not: Claim frontier-lab equivalence. Cite Alibaba/Thermo Fisher unless you've confirmed the access gate is removed. Lead with governance vocabulary. Reference TinyFish.
Do: Name the research (Chen et al., Buçinca et al.) if the room is technically literate. State the grain of truth before the reframe. Let the published framework carry the argument.
The industry is off through the Fourth. Use the quiet days to resolve the access gate on Alibaba and Thermo Fisher and reread the trust essay with fresh eyes. H2 conversations accelerate mid-July.
- OpenAI's Model Designer role: The posting explicitly asks candidates to "understand, predict, and design model behavior," making it the clearest public example of what frontier-lab design credibility actually looks like and where the boundary sits between that role and application-layer design leadership.
- Backdoor references are accelerating: The Wall Street Journal reported that companies like Zapier now require hiring managers to have at least one live reference conversation before extending an offer, with some conducting as many as ten calls for senior roles, which means every claim in this reframe needs to survive repetition by someone who is not you.
- NIST on explanation complexity: NIST's AI Risk Management Framework notes that humans perceive AI outputs and explanations differently according to preferences, traits, and skills, reinforcing the Chen et al. finding that the same transparency feature can help one user and harm another.
- Suno's design-leadership language: Suno's current Head of Product Design posting asks for speed, experimentation, craft, and emotional resonance across AI-driven products but never mentions model behavior, trust calibration, or agent autonomy, suggesting that even AI-native companies frame most design-leadership roles at the application layer.