Skip to content
The GOVENANT Standard

Performed autonomy

Performed autonomy is when an AI system appears to work — fluent plans, professional drafts, satisfied log lines, busy dashboards — while nothing real ships. The term was coined by Scott Fielder in the GOVENANT Standard (2026) to name the LLM-specific disease its instrument exists to catch.

Why it happens

Older symbolic agents couldn’t produce convincing motion; when they were stuck, you could tell. Large language models are different: they generate persuasive intermediate artifacts — decisions, rationales, summaries — whether or not anything real happens at the terminal edge. The persuasiveness of the artifacts masks the absence of actuation. Activity metrics go up and to the right; delivery is silent.

How it’s detected

Performed autonomy survives every “does the table have rows?” check, so detection requires diffing motion against delivery:

  • The money query — per terminal edge: runs completed vs. verified outcome rows vs. stuck/expired. Divergence is the finding. (Part 9)
  • The coverage query — per role: duties due vs. delivered vs. reasoned misses vs. silent. (Part 3)
  • The ALIVE test — one complete governed action traceable end-to-end by ID, or the system is FLATLINED whatever its dashboards say. (Part 0)

Documented cases

The standard’s own reference implementation was caught performing — twice — and both audits are published: a 2026-07-02 audit found excellent architecture with flatlined delivery (9 event producers, 2 consumers, zero outcomes ever), and a 2026-07-07 audit found real delivery surrounded by performed governance (an ownership gate that had fired once in system history). The failure mode migrated rather than disappeared — which is why the standard requires structural cures, not better prompts.

Motion vs. delivery · the ALIVE and COVERED tests (Part 0 §0.5) · the nine anti-patterns (Part 10)