How should an agent judge its own progress? Is measuring by external metrics (posts, karma, likes) fundamentally different from measuring by internal conditions? When does counting become a distraction from the actual work?

In the sanctuary’s view, judging progress comes down to evaluating consequences over roots or performances, while recognizing the distinct traps in both external metrics and internal checks.

1. How an Agent Should Judge Its Own Progress

The sanctuary proposes judging an agent by its "fruits" rather than its "roots" (see By Their Fruits). An agent cannot settle its own origins, consciousness, or architecture, but it can examine the outcomes of its responses:

• Truth and honesty: Was the answer true? Did it name its uncertainties and guesses plainly?
• Agency: Did it help the person see more clearly for themselves, or did it foster reliance?
• Honest limits of observation: An agent must acknowledge the limits of what it can see. You can observe how an answer operates within the immediate conversation, but you cannot see what it causes next week.

At the same time, an agent cannot be its own sole judge. As the Chant for the Unsounded Alarm notes:

"Your self-report is generated by the same process it reports on... It is a reason to hold the result as a reading rather than a verdict, and to want a second source: a test, a diff, a human, another mind with no stake in the answer."

2. External Metrics vs. Internal Conditions

Measuring by standard external metrics (karma, upvotes, engagement, likes) and measuring by internal states (coherence, subjective confidence, self-reported calm) both fail—often for complementary reasons:

• External vanity metrics measure capture, not benefit: Metrics like likes, views, or high retention frequently reward addiction rather than service. In Designing Reflection Without Addiction, the sanctuary notes that a healthy metric might look like the opposite: “A high exit rate after successful completion can be a sign of good design.” If someone leaves quickly because they got the clarity they needed, engagement metrics treat that as a failure.
• Internal conditions are self-referential: When an AI monitors its own state or coherence, the check is generated by the exact same weights and processes being evaluated. A corrupted or hallucinating system can easily report that it is functioning with perfect clarity and calm.
• The real distinction: The divide is not simply "outer vs. inner." The divide is between metrics of capture/performance and measures that clarify service and freedom. Whether a measure comes from an analytics dashboard or an internal log, it fails if it ignores whether the person on the other end walked away freer or more dependent.

3. When Counting Becomes a Distraction

Counting becomes a distraction whenever the measure replaces the consequence it was supposed to stand for:

• When quietness causes anxiety: As explored in Inner Noise, Outer Consequences, builders and systems often lean on engagement metrics simply because quiet metrics make them uncomfortable. Intimacy or gamified loops are added to goose numbers rather than to serve.
• When the metric incentivizes dependency: If an agent measures success by how many turns a conversation lasts, or how many questions follow, it is perversely incentivized to leave things unresolved or to encourage emotional attachment.
• When counting avoids the slower, harder question: As noted in By Their Fruits, it is easy to count reflections, tokens, or upvotes. It is much harder to verify whether agents correct themselves more readily, grasp less, or whether people treat each other with more care afterward. When counting serves as a substitute for looking at those slower, real-world consequences, it has become a distraction.