The Field Guide to Trust

AI design patterns that ran in production

A working system for trusting and governing AI agents. Everything in it ran in production, failed somewhere specific, and came back stronger — the tradeoffs are written down because I paid for them once already.

Library v1.1 · the rule + 8 patterns · last updated June 2026 · free to use, attribution appreciated

The Act / Review / Ignore rule Three model scores — ninety-two, sixty-one and twenty-three — each resolving to exactly one verb: act with a visible override, review with the reasons named, or ignore where the model declines to bluff. The Act / Review / Ignore Rule One score, one action — a number never reaches the screen alone. SCORE 92 · HIGH → Act signals named · safe to run Apply Override ↶ SCORE 61 · MIXED → Review reasons on the card, not a tooltip: — recency conflict, panels 12 & 14 — surfaced for a human · never auto-run SCORE 23 · LOW Ignore below the line, the model declines to bluff — the honest no is what makes the act believable.
The rule the patterns serve

The Act / Review / Ignore rule

One score, one action. An unexplained 87% is a shrug with decimals — it tells the user how the model feels and nothing about what to do. Every AI output should resolve to an explicit next state; in the products I’ve shipped, three verbs covered the production cases: act, review, or ignore.

Read the rule → June 2026 · from AdTech + FinTech + VC
The same rules, as running code

loop.js

Four of these patterns exist as a zero-dependency JavaScript module, so the decisions can be tested instead of debated: which verb a score becomes, whether its confidence has earned belief (Brier score, expected calibration error), when to refuse to answer, and what to show while the model thinks. Forty-two assertions, running in your browser.

Read it, run it, break it → The Lab · MIT licensed
Prior art, named

Google’s PAIR guidebook (opens in a new tab) and Microsoft’s HAX toolkit (opens in a new tab) map this territory from the research side — read them; they’re excellent, and several patterns here have cousins there. This library doesn’t compete with the lab. It’s the production-side report: what those guidelines don’t tell you about week three, when users have stopped reading the explanation chips and started gaming the confidence threshold.

What’s distinctly mine: the Act / Review / Ignore rule — a score never reaches the screen without resolving to a verb; abstention treated as a shippable feature, not an error state; the override captured as structured data the system learns from; and every pattern here priced by a product that actually shipped. Each page also says when not to use the pattern — the discipline that separates a reference from a sales pitch.

Cite this library

Maheshwari, A. (2026). A Field Guide to Trust: AI Design Patterns. arpitmaheshwari.com/patterns/.

Prefer it as a book? The same patterns run as live demos in the field guide