This page records who maintains The Behavioral Layer and how its notes are made. Attribution here is provenance, not promotion.
Steward
The Behavioral Layer is curated by Joel Goldfoot, a product design and research executive. He is the author of How to Lead Design in the AI Era (2025) and the creator of the BiModal Design framework.
An interest worth declaring: he builds systems of this kind professionally, and this site argues that building them is a discipline the industry needs. That overlap is the source of whatever practitioner grounding the notes have, and it is also a reason to read the original arguments here with the author’s stake in mind. Specifics of that work, including the employer, the systems, and any customer, are withheld under confidentiality and never appear on this site.
- blog.goldfoot.com
- bimodal.design
- linkedin.com/in/goldfoot
- How to Lead Design in the AI Era
- ORCID 0009-0007-1843-2089
How this site is made
The notes on this site are produced by a defined process, described here so that any claim can be traced and audited.
Research agents discover primary sources and draft resource notes. Each note must trace its claims to a primary source that was fetched and read. A note that cannot cite a primary source is not published.
Source tiers describe provenance, not endorsement. “Tier 1” means the note rests on the primary artifact itself (the paper, model card, changelog, or repository) rather than secondhand coverage; it is not a peer-review judgment, and preprints are labeled the same way as published work when they are the artifact under discussion.
The standards are written down. A public editorial contract, EDITORIAL.md in the repository, defines what a note must contain and how it must read. Continuous integration enforces that contract mechanically on every proposed change: it checks that each resource note cites a source, that internal links resolve, that frontmatter is valid, that freshness dates are present, and that no term on a confidentiality denylist appears.
Enforcement is not the last step. Publication runs on tiers calibrated to risk, which is this site’s own doctrine applied to itself. Short, source-quoted Signal entries may publish automatically, but only after passing every mechanical check plus an independent verification pass in which each quote is re-confirmed against a fresh fetch of its source, and only after a 48-hour window in which the human editor can veto. Everything analytical (research notes and all longer analysis) and everything in the editor’s own voice (briefings and concept notes) is reviewed and merged by the human editor, Joel Goldfoot, before it publishes. Any failure, flag, or doubt demotes a batch to human review, and a defective automated publication suspends automation entirely until the cause has become a written rule or an automated check.
The tier system was adopted on 2026-08-09. As of that date nothing on this site has ever published automatically: every entry here was merged by hand, and the automated tier is newly operational rather than proven. This paragraph will be wrong the first time a batch publishes on its own, and it will be corrected then.
Changes are proposed as pull requests. That has not been universally true, and the exceptions belong on this page rather than only in the log: a number of commits early in the site’s history, and one Signal batch on 2026-07-28, were pushed directly to the main branch without review. Every change by either route is recorded in the public git history at github.com/jgoldfoot/behavioral-layer, so the provenance of any sentence is inspectable.
The site as its own subject
This site is an instance of its own subject. It is maintained by AI research agents operating under a written behavioral contract, with mechanical enforcement, human oversight, and public provenance. That arrangement, an autonomous system governed by a specified contract and verified by a person, is the thesis of The Behavioral Layer applied to the site that documents it.