Four Additions

Our October 9 resource refresh brings the catalog to 209 entries. These four additions cover behavior evaluation, execution records, human approval, and financial AI assurance:

  • TraceDance: a benchmark and paper that turn deployment traces into behavior-specific evaluation tasks. Code remains unreleased. Paper.
  • Vestige: a local MCP server with append-only, hash-chained memory and offline-checkable write receipts. Its recorded execution relationships support failure diagnosis. Project.
  • Impri: a self-hostable inbox for human approvals, edits, and rejections of proposed agent actions. Agents must integrate returned decisions into their execution flow. Project.
  • FAIAS: Wonder Attah's independently authored financial AI assurance framework, with 109 review questions across nine domains. Framework.

Two Corrections

  • sofagent: the description now records 28 registered rules, of which 17 are enabled by default. Project.
  • PrefixGuard: the paper link now points explicitly to v1, which matches the catalog's retained title and reported results. Pinned paper.

The date above records a catalog update. Original project publication dates remain in the linked sources. Future substantive catalog changes will appear as resource-update summaries in our RSS / Atom feed.