Sandbox
01
A decade of designing financial products taught me that the gap between trust and trustworthiness is the gap where products fail their users. Trust is a feeling. Trustworthiness is a property of the system. When a user's confidence drifts above what the system can support, the gap closes later, suddenly, at someone's expense.
Designing AI products is the most consequential place this gap exists right now. The work is not making the assistant more capable. It is deciding which parts of the system's state to surface at decision points where silent resolution would erode trust, and which to leave below the interface. Correctness and calibrated reliance are different metrics. Optimizing for one does not produce the other.
This is the design problem I have been organized around for ten years.
02
Four working principles. Each one shows up in a specific decision when I am in the work.
01
Making trust track capability is not an abstract principle. It shows up in timing, typography, and copy, and it requires the same detail attention as any other craft surface.
Stale point estimate → Range with data provenance
02
Where a system fails tells you more about its behavior than where it succeeds. I start by mapping the edge cases, the model's confidence gaps, and the recovery paths, then work backward into the happy path.
Hidden risk → System state visible at the decision point
03
The interesting behavior in AI products only shows up at runtime. Static screens hide the design problems that matter, how the model handles ambiguity, where it gets confident, where it should defer. I build working prototypes with the Anthropic API and treat them as my first design artifact.
Identical transaction rows → Ambiguity surfaced and flagged
04
Every project is a series of hand-off decisions. What do I frame, what do I let Claude draft, what do I review and revise, what do I refuse to hand off at all. The decisions are specific to the task, not a blanket rule. Claude is strong at synthesis, at generating specific variations of a pattern I already understand, at scaffolding code I would otherwise write by hand, at classifying inputs I have defined the taxonomy for. It is not strong at deciding what matters, at noticing when a task is the wrong task, at knowing when to stop. Those I keep.
HandoffLog artifact
Component does not exist yet. Build it to replace this placeholder.
03
Anthropic API prototype that classifies content across eight cognitive dimensions and recommends against gaps. Working code.
github.com/royamonty/cognitive-curatorFull-stack scanner with model routing across Haiku and Sonnet. Z-score detection, wallet clustering, two-pass economic impact scoring.
[PLACEHOLDER: one-line description of a current experiment.]
04
January 2026
On the difference between conceptual frameworks generated through reading and specific beliefs generated through building.
August 2025
Stuart Russell's misspecification framework applied concretely to a visual QA workflow I built.
May 2025
A workplace observation on why the friction in AI interfaces does real epistemic work.