About Calibrate
Calibrate is a library of named, reusable design patterns for calibrated trust between people and autonomous machines: vehicles, medical devices, surgical and industrial robots, drones, warehouse systems and defense systems. It's for practitioners, and it's also a learning tool written in public.
On this page
What is calibrated trust?
Every autonomous system has a side that faces people: the signals, controls, alerts and handoffs that tell someone whether to rely on the machine, when to step in, and how to recover.
What that side has to get right is calibrated trust: trust that matches what the machine can actually do. Too little, and people ignore a system that would help them. Too much, and they stop watching when they shouldn't. These patterns are reusable ways to design for the right amount.
The idea of calibrated trust comes from human factors research, notably Lee & See, Trust in Automation: Designing for Appropriate Reliance (2004).
What a pattern means here
A pattern is a named, reusable solution to a problem that keeps coming back, in the same sense the term carries in architecture and software design. It isn't a component, a guideline or a piece of advice. It has a name you can say out loud in a design review, a mechanism you can draw, and conditions under which it applies and others under which it doesn't.
This library is narrower than most: it covers autonomy with physical consequences. Most AI pattern libraries cover AI inside software, where a bad output is a bad output. Here the machine moves, cuts, lifts or drives, so a mistake is measured in seconds and metres. That's why every pattern here starts with a real incident, and why friction is sometimes the feature.
How much to trust each example
Every example in “In the field” carries one of three labels. While examples are still being collected, these labels keep the library honest about how much of it is evidence and how much is argument.
- Observed
- A real system, or an incident with an official investigation or record you can follow.
- Illustrative
- A real event known only from news reporting. Useful for the shape of the failure, not for its details.
- Speculative
- The author's own concept: plausible, but not yet seen in the field.
Status
- draft
- The story, mechanism and pattern are written, with a few sources so far.
- reviewed
- Every section filled, at least three sources, and at least one observed example.
- stable
- Reviewed, then revisited after real-world use or feedback.
Candidate patterns that are only a name and an idea aren't shown until they're written up.
How patterns relate
A relationship is written once and shows up on both patterns, reworded where that reads better. If The Handoff requires Shared Mental Model, Shared Mental Model lists The Handoff under “Required by.”
- Requires
- The other pattern has to be in place for this one to work at all.
- Complements
- They reinforce each other, but neither depends on the other.
- Alternative to
- Two ways of solving the same problem. Pick one.
- In tension with
- Applying one makes the other harder. The trade-off is the point.
- Precedes
- One comes before the other when control moves between person and machine.
How to cite
Cite an individual pattern, not the library as a whole:
Johnson, L. (2026). {Pattern name}. Calibrate: Patterns for human–machine trust. https://trust-layer-patterns.vercel.app
The patterns are shareable under CC BY 4.0: you're free to quote and adapt them with credit. The changelog records what changed and when, if you need to cite a pattern as it was on a particular date.