diff --git a/AGENTS.md b/AGENTS.md
index 09ff11b..1fc05a7 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -1,10 +1,48 @@
 # PRISM 2026 repository guidance
 
-## Scope
+## Human authorship boundary
 
-This repository is a project publication. Preserve author intent and treat
-scientific claims, methods, interpretation, author order, and credit as author decisions. Flag
-questions rather than silently changing substantive content.
+This project publication and any related blog posts must remain human-authored.
+Agents may operate the publishing machinery, but they must not supply the
+writing, analysis, or scientific reasoning.
+
+- Do not draft, rewrite, paraphrase, polish, translate, or autocomplete
+  publication or blog prose. This includes titles, summaries, body text,
+  captions, appendices, disclosures, contribution statements, and repository
+  prose such as README or CONTRIBUTING text.
+- Do not invent or add scientific claims, methods, results, interpretations,
+  citations, examples, mathematical arguments, or analysis code whose output
+  supports a scientific conclusion.
+- Do not directly apply grammar, style, clarity, citation, or mathematical
+  corrections to authored content. A human author must decide and make the
+  final change.
+- Text or analysis code supplied by a human may be inserted verbatim when the
+  user explicitly directs where it belongs. Do not silently revise it while
+  inserting it.
+- Do not author AI-use disclosures or CRediT statements on anyone else's
+  behalf. Agents may validate author-supplied records and report omissions.
+
+Agents may:
+
+- Compile, render, convert, package, and debug Quarto, Jupyter, LaTeX, HTML,
+  and PDF outputs.
+- Work on CI, configuration, templates, stylesheets, layout, accessibility,
+  and web design without changing authored prose, analyses, or scientific
+  meaning.
+- Run human-authored analysis code to reproduce outputs and report failures
+  without changing the analysis.
+- Inspect human-authored writing and return a checklist of possible grammar
+  problems, awkward sentences, unclear passages, citation problems, or
+  mathematical inconsistencies. Give that checklist in chat or an untracked
+  local review artifact; never commit or push it, and never apply its proposed
+  edits automatically.
+
+Before every commit or push, inspect the complete diff. If it contains
+agent-authored publication or blog prose, scientific analysis, or scientific
+reasoning, remove that content from the change and stop for a human author to
+provide it. Do not transfer AI-written prose between this repository, the
+publication template, or another blog repository. This `AGENTS.md` policy is
+an explicitly authorized exception.
 
 ## Working agreement
 
@@ -13,10 +51,8 @@ questions rather than silently changing substantive content.
   in `src/analysis/`.
 - Do not change the publication theme, the Warm Journal default, brand colors, fonts, or logo unless
   the repository owner explicitly requests a design change.
-- Keep `authors.yml`, the CRediT contribution record, and `ai-use.yml` consistent with the author-
-  approved manuscript metadata.
-- Do not invent author names, affiliations, ORCIDs, contributions, citations, results, or AI-use
-  statements.
+- Keep `authors.yml`, the CRediT contribution record, and `ai-use.yml` consistent with the exact
+  author-approved manuscript metadata.
 - Do not commit session notes, internal design records, private correspondence, credentials, local
   agent artifacts, or generated `_site/` directories.
 - Do not edit `gh-pages` directly. Merges to `main` publish through GitHub Actions.