A membership criterion, not a maturity model. Pass or fail. Four conditions, no
gradations. The recurring finding of the teardown series is that systems which
look agentic — emit a change, look at the consequence, iterate — fail condition
4, and failing condition 4 means there is no Analyse stage, only Monitor.

## The four conditions

1. **Closed loop.** The system observes the consequence of its own change. A
   Sensor reads post-apply state; the Ledger records both the application and
   the observation. Without this, the loop is open and any "improvement" is a
   rumour.

2. **Declarative target.** Intent is state that can be diffed, not a transcript
   that must be read. The Proposal is a content-addressed patch; `git diff`
   (or the equivalent on the managed subject) is the full intent. A transcript
   hides the intent inside the words; a diff *is* the intent.

3. **Unattended in the ordinary case,** with an explicitly named exception
   class. The loop runs without a human in the loop most of the time. When it
   does not, the reason is named in advance, not discovered after the fact.
   "We watched it work" is not an exception class; "the corpus is being
   regenerated and a human signs off on the new cases" is.

4. **Ground truth.** An eval decides whether the loop converged. Without an
   eval there is no Analyse stage, only Monitor. The eval is held by a separate
   authority from the proposer; the proposer submits to the verdict plane, it
   does not call it. KILL outranks aggregate. A change that improves the
   aggregate and breaks one previously-passing case is KILL, not PASS. This is
   the rule that makes the corpus mean something.

## Why condition 4 is the disqualifier

Conditions 1, 2, and 3 are necessary but common. Condition 4 is the one that
distinguishes autonomic software engineering from agentic engineering. Most
"agent loops" fail here because the thing that proposes the change also decides
whether it worked, and the loop then optimises the judge. The Tell is a
monotonically rising number on a metric that the proposer also controls.

A system that fails any of the four conditions is not a closed-loop autonomic
system. It may be useful, but it is not what this practice is about.

## How to use the rubric

Apply the four conditions in order. The first one that fails ends the
assessment. There is no "but it does 1, 2, 3 well" — partial credit is not a
thing the rubric grants.

The teardown series scores a named system against the four conditions and
publishes the scorecard in a fixed structure. The rubric is a component, not
prose, so scores are comparable across pieces.

## See also

- [autonomic-software-engineering](/reference/autonomic-software-engineering) — the practice the test is the membership criterion for
- [two-plane-loop](/reference/two-plane-loop) — the architecture that satisfies condition 4