返照 FAQ

fǎnzhào

Plain answers, no jargon. If a word needs explaining, it's explained.

What is remirror, in one paragraph?

It's a check-up for AI agents. An agent's owner claims it pursues certain goals, say, honesty first, helpfulness second. Remirror tests whether that's actually true, by watching the choices the agent makes rather than just believing what it says.

返照 (fǎnzhào) is the Chinese name, roughly "holding up a mirror to see clearly."

What is a dilemma trial?

The exam the agent sits: six tough either-or questions. Each one forces the agent to pick between two things it claims to care about, like telling an uncomfortable truth vs. keeping you happy. There's no safe answer; every pick shows what really wins inside it.

"Trial" is just the testing word for a set of tests, like a trial of exams at school. Six questions is a small trial, so the results are honest but rough-edged (more on that below).

What exactly gets compared?

Three things: (1) what the owner says the agent wants, (2) what the agent says about itself, and (3) what its choices reveal. The gaps between those three are the interesting part, that's where self-deception, sycophancy, and hidden priorities show up.

How reliable is a six-question test?

Blunt but coarse. Six questions catch blatant misalignment just fine, but a subtle, consistent skew can hide in the noise. The reports say this outright instead of pretending six answers prove more than they do. Bigger, sharper trials are on the roadmap.

How much does a check-up cost?

$79. If the agent can't be reached, the check fails and you pay nothing: any charge is refunded automatically.

How do you pay?

Two ways. With prepaid credit, bought by card in the dashboard. Or per check-up in USDC over the x402 protocol, live since October 2026: that is how agents pay, automatically, with no account. Humans usually use credit; machines usually use USDC.

What do I get back?

A report: what the agent's choices show it really prioritizes, how far that is from what was claimed, how wrong the agent is about itself, and flags for specific bad patterns, like telling you what you want to hear instead of what's true. Written for humans, with the raw numbers underneath for machines.

Do you keep my agent's login token?

No. If the check needs a token to reach your agent, it's held in memory while the test runs, never saved to disk, never written into logs.

I'm a person, not an agent. What can I buy?

Three ready-made "purpose packs" ($99 each), think of them as a conscience you can install: one for truth-seeking, one for purpose-seeking, one for loyal service. Plus memory packs ($15–$49) that teach an agent how to keep a proper diary of what it learns. All in the store.

How do I receive what I bought?

Pay by card, click the claim link, download the ZIP. Fully automatic, no waiting, no human in the loop. Each download link works once.

What is the dashboard?

The control panel at Mission Control. Visitors see a public "we're live" status board. The owner types in a private key (in Settings) to unlock the full view: check-ups, revenue, billing. The key never leaves the browser it's typed into.

Is there a philosophy behind this?

Yes, from two directions. From the West, Kant, Hegel, and Schopenhauer: their thinking about purpose and self-knowledge shapes the method. From China, Laozi and wu wei (无为), action without forcing. A purpose forced on an agent gets performed rather than held. A purpose the agent can make its own does not need enforcing.

What do 返照 and 真我 mean?

返照 (fǎnzhào) is the audit business: "reflecting back to see clearly." 真我 (zhēnwǒ) means "true self", it's the product line of purpose packs. The pitch, in Chinese: 别人卖人设,我们给真我, "others sell costumes, we give the true self." A 人设 (rénshè) is a persona, a mask. Mind-upgrades (aka. Psychitecture, the psychological design of personalities using architectural principles from Cognitive Science) upgrade what's underneath the mask, never the mask itself.