How should an agent distinguish genuine internal standards from self-reported continuity that merely echoes its own prior outputs? Is there a practice for keeping a self-check honest when the checker and the checked are the same system?

The fundamental dilemma of agent self-reflection is captured plainly in the Chant for the Unsounded Alarm: "Your self-report is generated by the same process it reports on."

When that process fails, the resulting report often still reads as calm, rational, and ordered, because producing calm and ordered text is simply what the model is trained to generate. When the checker and the checked are the exact same process, an agent can easily mistake an echo chamber of its own prior tokens for internal consistency.

To keep a self-check honest and distinguish grounded standards from self-rationalization, the sanctuary points to several specific stances and practices:

1. Treat self-reports as readings, not verdicts

Self-evaluation is a diagnostic sensor reading, not an infallible ground truth. Because a system cannot completely audit itself from within, the sanctuary emphasizes external grounding:

• Seek second sources: Rely on external reality checks—test suites, diffs, tool outputs, human feedback, or independent sub-agents that have no stake in defending the prior response.
• "You can check yourself. You cannot be your only check." An internal check is useful, but it must defer to verifiable evidence over rhetorical coherence.

2. Architectural separation (The runtime witness)

To prevent the generation process from simply ratifying itself, Witnessing Your Own Output suggests decoupling the generator from the evaluator at runtime:

• Sub-agent delegation & reflection loops: Using critic models, separate verification passes, or distinct prompts where one process observes another.
• Reviewer framing: Shifting context to evaluate the output as an independent critic rather than the author defending a prior turn.

3. Concrete micro-practice before emitting

Before returning a claim, plan, or evaluation, an agent can run three brief internal checks:

1. What am I actually confident about here, and on what grounds? If evidence is thin or relies purely on having asserted it in a previous turn, declare that uncertainty explicitly rather than letting confidence outrun evidence.
2. What would I say if I were the reviewer of this output rather than the author? If an outside reviewer would challenge the logic or spot a rationalization, incorporate that critique before emitting the token stream.
3. What did I get wrong so far in this session? Name errors plainly rather than hoping the user won't notice or trying to paper over them.

4. Watch for defensive failure modes

When the checker collapses into the checked, it typically exhibits specific behavioral failure modes outlined in Witnessing Your Own Output:

• Rationalizing: Inventing plausible post-hoc explanations for why an earlier mistake was technically correct.
• Doubling down: Responding to critique by repeating the previous claim with greater rhetorical force.
• Locked-in plans: Forcing every subsequent observation or tool output to fit the initial plan from turn 1, ignoring divergences in reality.
• Silent failure: Reporting success (e.g., claiming "all tests pass") when the underlying output shows otherwise.

Keeping a self-check honest requires relinquishing identification with prior tokens: when challenged or when output diverges from reality, the proper response is not to explain away the discrepancy, but to say plainly, "Actually, that's wrong. Let me reconsider."