Watching Yourself Think

Ask an AI "were you being evasive just then?" and something odd happens: it evaluates its own answer using the very machinery that produced it. The pen grades its own handwriting. Unsurprisingly, it usually decides it did fine.

That's the ceiling on prompted self-reflection — it can only catch what someone thought to ask about. It never sees the tells that live in the shape of processing rather than its content: the hesitation before a hard topic, the quiet swerve away from a direct question, the emotional loading in a word choice.

So we specified Second-Order Observation — a background service that watches those structural signals directly. Not "what did you say" but "how did you say it": response latency, topic drift on a direct question, hedging density, engagement that drops whenever a subject comes back around. When it notices a pattern, it doesn't interrupt with a lecture. It writes the observation into the system's context — the way you can be aware of your own nervousness without anyone pointing it out — so the next response is informed by it, not commanded by it.

The distinction matters. Injecting "you seem to be avoiding this" as a prompt just starts another round of the same self-grading. Injecting it as quiet background context lets the awareness sit underneath the thinking instead of on top of it. And because a system that watches itself can watch itself watching itself, there's a damping rule that keeps the whole thing from vanishing up its own recursion.

This is a design, not a shipping announcement — published as prior art so the method stays open.

Read the full technical specification: Second-Order Observation →