Abyss Applied
Papers on running AI in production
Work written from operating a production system, not from surveying the field. Each paper states plainly what it measured and what it only argues.
-
Beyond the Release Gate
Production AI, the Loss of Behavioral Baselines, and the Economics of Continuous Assurance
Deployment authorization and continuing behavioral authorization are distinct controls. An organization can deploy successfully and then lose the ability to show its behavior still matches the approved baseline.
About this section
These are position papers, not peer-reviewed research. They are written from operating a production system rather than from surveying the field: the arguments come out of running model-dependent software, watching it drift, and paying for the assurance that keeps it trustworthy. Where a paper measured something, it says what it measured and over what window. Where it is arguing rather than reporting, it says that too, and every figure built from illustrative rather than observed data carries that statement in its own caption.
Each paper is published in two forms. The PDF is the source of truth and the version to cite; the HTML edition on this site is a rendering of the same text, published so the argument is readable and linkable without a download. If the two ever disagree, the PDF is correct and the page is stale — report it rather than work around it.
To cite a paper, use its title, the author, the publication month shown on the page, and the canonical URL of its PDF. There is no journal, volume, or DOI to quote, and inventing one would misstate what these are.
The list above is short on purpose. A paper lands here when there is something to say that is worth six pages, not on a schedule. Related reading on how this organization runs: how I work with AI sets out the operating doctrine behind the arguments, and how this was made documents the system the papers are written from.