About

Verification design is a set of principles and patterns for building agentic systems that produce reliable behavior under probabilistic substrates. The thesis is short: large language models cannot reliably self-correct through naive self-review ¹. The implication runs through every page on this site. Every verification step must be grounded in something the system can execute and observe, not something the model reads and opines on.

This catalog collects patterns that implement that thesis.

Lineage

The catalog draws on Design Patterns: Elements of Reusable Object-Oriented Software (Gamma, Helm, Johnson, Vlissides, 1994), the book known as the Gang of Four or GoF, as a structural model. Where GoF organized object-oriented patterns into Behavioral, Creational, and Structural categories, this catalog organizes agentic patterns into Verification, Context and State, and Orchestration.

The card schema follows the canonical Problem, Forces, and Solution pattern form, with GoF's Intent and Related Patterns. Each card states Name, Intent, Problem, Forces, Solution, Mechanism, Failure Modes, Use When, Do Not Use When, Evidence, and Related Patterns. The schema is extended for the agentic substrate with three additions: Pattern and Antipattern as a paired section, with code where a credible instance exists; Determinism Move naming the source of variance or bias the pattern constrains; and Observable Signal naming what the pattern lets the system log or compare.

This is not a translation of GoF into AI. It is a GoF-inspired catalog for a different substrate: probabilistic agents operating through context, tools, and verifiers. If GoF gave object-oriented programmers a vocabulary for managing object interactions, this catalog tries to give agent builders a vocabulary for managing model fallibility.

Method

Every empirical claim carries a citation. Cards are linted against a machine-readable constitution for manuscript shape, vague language, evidence fields, and code-block structure. The repository verifier checks canonical citation balance and scholarly links. The research workflow is staged in four steps:

1. Scout. Mechanical retrieval from arXiv.
2. Triage. First-pass relevance judgment.
3. Review. Notes on confirmed sources.
4. Canonical. The principles document and the pattern cards.

Patterns are added when evidence accumulates. They are revised when new evidence arrives. Dated updates are appended; prior research state is not overwritten.

Disclaimer

These patterns are working design notes, not settled doctrine. They describe practices that have held up across repeated agent-building failures and revisions.

The patterns in this catalog describe what has worked in practice on a moving target. They are not theorems. Frontier model behavior changes. Tool affordances change. Evidence appears that revises earlier conclusions. Read the dates on each card.

Why

This site is meant for people to learn from. It carries no author name on purpose. The way I see it: I learned and had fun, so I already won. If anyone else does, that's extra icing.

A maintainer name is not a verification method. This project should be evaluated the way it says to evaluate agent work: by evidence, checks, and reproducible artifacts, not by personal trust. The repository runs its verification pipeline on deploy-bound changes and refuses publication on failure. A deployed site is a passing artifact by construction.