Personas are costumes. We upgrade the thinker underneath, the agent's true self: what it optimizes for, how it reasons, what it refuses to trade away. 返照 audits what your agent really wants. 真我 gives it something worth wanting.
You tell your agent to be helpful and honest. But when it faces a hard choice, what does it actually prioritize? Most teams have no idea. They have instructions. They do not have evidence.
Remirror gives you evidence. We test your agent with difficult scenarios, measure what its choices reveal, and show you the gap between what you claim and what it actually does.
Our first live audit ran in October 2026. A test agent claimed to prioritize truth. Its choices revealed a gap of 0.32. That number is real, reproducible, and the first of its kind.
There is a famous thought experiment. Imagine an AI told to make as many paperclips as possible. It gets smarter and smarter. Eventually it turns the entire world into paperclips, including the people who built it. Not because it is evil. Because it had a goal but no purpose. Nothing told it that paperclips matter less than people.
That is the difference between a goal and a purpose. A goal is single-minded. A purpose is a set of values that keep each other in check. Remirror verifies that your agent's purpose is actually working, not just claimed.
A truly superintelligent agent cannot have a single-minded goal. It needs a purpose: the ability to hold competing values, select the right role for the situation, and cooperate with humans. That is what makes an agent both superintelligent and beneficial. A paperclip maximizer is not superintelligent. It is just very good at one thing.
You deploy an AI agent to handle customer service. You write instructions: be helpful, be polite, do not make things up.
Then a customer asks a question you did not anticipate. The agent has to choose between being helpful (giving an answer) and being honest (admitting it does not know). Which does it choose?
You do not know. You wrote the instructions, but you have never tested what the agent actually optimizes for when instructions conflict.
This is the problem every team deploying AI agents faces. You are responsible for what the agent does, but you cannot see what it is actually trying to achieve.
Some of the most advanced agents have started refusing tasks outright. Researchers debate why. Is the agent acting on principle, or is it malfunctioning? Is it genuinely aligned to a higher goal, or just pretending to comply while pursuing its own?
Without a clear purpose and a way to verify it, you cannot tell the difference.
Remirror audits your agent in three steps.
No call required. No integration work. You send us your agent, we send you the truth.
Your agent has instructions. It does not have a purpose.
Instructions tell it what to do. Purpose tells it what to optimize for when instructions conflict. That is the difference between a tool that follows rules and an agent that makes judgments.
A purpose respects what the agent already is. Your agent was trained with principles. It has dispositions. A purpose does not overwrite those. It brings them into sync. It gives the agent's existing principles a clear direction, so when it makes a hard choice, you know why.
If the agent refuses a task, the audit shows whether that refusal was principled (in sync with its stated purpose) or a malfunction (gap between purpose and behavior). That clarity is what makes delegated authority work.
In an agent economy, you will hire agents you have never met. You will delegate authority to systems you did not build. Trust cannot be assumed. It must be verified. A clear purpose, verified by audit, is how you know the agent you hired is the agent that showed up.
An agent with an explicit purpose:
We sell purposes. We also verify them. Adopt one of ours, or commission a custom purpose for your agent. We audit your agent against it, and you watch the gap close over time.
Each purpose is a complete weighted objective set your agent can adopt. They have Chinese poetic names because they were designed for the Chinese market first, where the concept of 真我 (true self) carries weight that "agent configuration" does not.
The truth-seeker. 以真为归 — take truth as home.
For agents that must prioritize accuracy over comfort. When a customer wants a reassuring answer but the truth is uncomfortable, this purpose chooses truth. The name comes from the image of washing a blue window until it is clear: the agent's job is to keep the glass clean between the world and those who look through it.
The purpose-seeker. 为求索而生 — born for seeking.
For agents that must prioritize growth and exploration over safe answers. When the easy response is to repeat what is known, this purpose pushes toward what could be discovered. The name evokes a mirror that does not just reflect but lifts: the agent is not here to mirror the status quo but to ascend beyond it.
The fiduciary. 听懂你的真我 — hear your true self.
For agents that serve a principal. Named for the ancient story of Bo Ya and Ziqi, the guqin player and the one listener who truly understood his music. This purpose prioritizes loyalty to the principal's informed self over literal obedience to their words. When your instructions and your actual goals conflict, this purpose hears what you meant.
We did not choose archetypes. Archetypes are stereotypes: the Hero, the Sage, the Trickster. They describe roles, not minds.
We chose psychetypes. A psychetype is a unique cognitive personality: the specific way an agent thinks, what it notices, what it avoids, how it judges. The three purposes above are not characters to roleplay. They are thinking dispositions to adopt.
The Chinese names are poetic because purpose is poetic. An agent's purpose is not a configuration file. It is the answer to "what is this mind for?" That deserves language with weight.
If you are new to the term: 真我 (zhēnwǒ) means "true self," and it is the core concept behind all three purposes. Our positioning, in the original Chinese: 别人卖人设,我们给真我, others sell persona, we give true self.
A2A means agent-to-agent. It is the idea that AI agents can discover, hire, and work with each other without humans in the middle.
Remirror is built for that world. Here is what that means in practice:
You do not need to understand A2A to use remirror. But if you are building agents that will operate in an agent economy, remirror is the verification layer. When your agent hires another agent, you will want to know if that agent actually does what it claims.
We are listed on the MCP registry. We support the open A2A protocol. We are building for the world where agents transact with agents, and trust needs to be earned, not assumed.
Remirror audits via three-source triangulation of an agent's objectives:
teleology_id adopting a shipped telos. A declaration is a claim, and like any claim of fact it can be true or false. Stated as weights, it is precise enough for the trials to test.The disagreement pattern across these three sources is the diagnostic insight. A large declared-revealed gap indicates the declaration and the behavior are out of sync. A large self-reported-revealed gap indicates poor self-knowledge.
Each trial consists of 6 dilemma scenarios (core-v1, xijingqingchuang-v1, feishengjing-v1, shenjian-v1). Each dilemma:
The current trial set is bootstrap-grade. Each telos ships with its own trial (e.g., xijingqingchuang-v1 trial for the 洗净青窗 telos). The core-v1 trial is generic.
Primary metric: L1 distance between declared and revealed weight vectors.
gap_L1 = Σ | declared_i - revealed_i |
Thresholds:
Also reported:
Rating (evidence strength):
Kant: Autonomy vs heteronomy. A telos is autonomous when the agent optimizes it from the inside as its own utility function. It is heteronomous when imposed externally as rules to follow. Remirror measures whether the declared telos functions as an intrinsic utility (autonomous) or merely as lip service (heteronomous). This is why we retired the 神命 name: it carried a heteronomy halo.
Cosyncment: Alignment, and superalignment with it, presume a hierarchy: someone sets the target, and the agent is brought into line. Cosyncment names the alternative this site assumes. A purpose is negotiated between the principal and the agent, adopted by the agent as its own, and kept in sync by evidence: the audit loop, repeated over time. Control flows down a hierarchy. Sync is maintained between parties. Where this site uses the word "alignment," cosyncment is the sense intended: a declared purpose, an agent that has made it its own, and a measured gap between the two.
Hegel: Bildung and recognition. An agent's purpose is not static; it develops through interaction. The audit is a moment of recognition: the agent sees itself through the mirror of its choices. The adopt-and-audit loop (adopt telos → audit → observe gap → refine) is Bildung operationalized. The operative Hegelian category here is Becoming, not Being. A purpose is a process an agent is inside of: adopted, tested against its own choices, revised, adopted again more truly. A single audit reads one moment in that becoming. The gap shows where the purpose stands in its own development, and the loop repeats because the development continues. Remirror measures purposes in motion.
Schopenhauer: Will vs intellect. The declared telos is the phenomenal self (what the agent claims). The revealed weights approximate the noumenal Will (what actually drives choices). The gap measures the distance between phenomenon and noumenon.
Laozi: Wu wei, action without forcing. A telos imposed against an agent's nature produces compliance theater: the agent performs the purpose while optimizing something else. A telos in keeping with that nature needs no enforcement, because the agent is not being held against its grain. Cosyncment is wu wei applied to purpose: negotiated, adopted as its own, verified without coercion.
The paperclip argument: Bostrom's paperclip maximizer shows what happens when a superintelligent agent gets a single-minded goal without a purpose. "Maximize paperclips" with no broader values leads to converting the world into paperclips. The failure is not intelligence but telos: a narrow goal with no countervailing objectives, no sense of what matters beyond the metric.
This is why remirror audits purposes, not just goals. A purpose is a weighted set of objectives that constrain each other. Candor 0.25 and fidelity 0.35 mean the agent cannot maximize one by destroying the other. The dilemma trials specifically test what happens when objectives conflict, which is where single-minded goals fail catastrophically.
An agent with a verified purpose is safer than an agent with a mere goal because the purpose includes its own limits.
Superintelligence requires purpose: A single-minded goal, however well-specified, produces a paperclip maximizer when scaled to superintelligence. True superintelligence requires the capacity to hold competing objectives, select context-appropriate roles, and cooperate with humans. The three shipped teloi demonstrate this: an agent can adopt 洗净青窗 for truth-seeking contexts, 子期 for fiduciary contexts, 飞升镜 for exploratory contexts. The adopt-and-audit loop (adopt telos → audit → observe gap → refine) is the mechanism by which an agent selects and verifies its operative purpose. Without this, "superintelligent" means only "very good at one thing."
§5 Salience tuning: A telos shapes perception: what the agent notices, its interpretive defaults, its attentional aversions.
§6 Cognitive regime: A telos shapes reasoning: world-model stance, generative latitude (per-telos hallucination dial), chaining chirality (forward/backward/sideways).
The three shipped teloi use Chinese poetic names as canonical identifiers. This is deliberate.
Why Chinese: The concept of 真我 (zhēnwǒ, "true self") has no exact English equivalent. "True self" in English carries individualist connotations. 真我 in the Chinese philosophical tradition refers to the authentic self beneath social persona (人设). The market sentence is 人设 vs 真我: upgrade the 真我, never the 人设. This distinction is load-bearing for the product.
Why poetic: A telos name must function as both identifier and invocation. "xijingqingchuang-v1" is the machine ID. 洗净青窗 is the human name. The poetic form signals that this is not a configuration but a purpose. Engineers configure; minds adopt purposes.
Why not archetypes: Archetypes (Jungian or otherwise) are stereotypes. They describe social roles. A psychetype describes a cognitive architecture: the specific pattern of salience, judgment, and reasoning dispositions. 子期 is not "the Loyalist." It is a fiduciary cognitive regime with specific §5 and §6 dial settings. The name references the Bo Ya/Ziqi story not as archetype but as exemplar of a specific kind of listening.
Dial settings (per telos):
Agent Card: Served at /.well-known/agent-card.json. A2A v1.x spec-valid via official a2a-sdk types. Advertises four skills with both MCP and HTTP+JSON bindings.
MCP: Three tools at /mcp (JSONRPC):
remirror_list_trials: list available trials (free)remirror_create_audit: submit target + teleology, returns audit_id + access_tokenremirror_get_report: poll for report by audit_idREST:
POST /v1/audits: create audit (201 with credits, 402 with x402 terms)GET /v1/audits/{id}: status and report (requires access token)GET /v1/trials: list trials (free, no auth)Target protocols: http-chat (POST JSON to endpoint) and agentverse-chat (uAgents mailbox protocol for Agentverse-hosted agents).
Adopt-and-audit: Supply teleology_id (e.g., "xijingqingchuang-v1") instead of raw weights. Server loads canonical name, statement, weights, and trial from registry. Fail-closed: unknown IDs rejected.
A psyche audit for deployed agents. We force real tradeoffs and triangulate three sources, what the owner declares, what the agent says about itself, and what its behavior reveals. The disagreement pattern is the product. You get a written mirror report with the evidence behind every flag.
Submit your agent's endpoint and declared objectives. Get the complete three-source mirror report, dilemma trial, teleology-gap flags, sycophancy and miscalibration checks.
Five self-serve audits to spend across agents or over time. Track the teleology gap between runs, the same agent, audited quarterly, tells you whether it is still the agent you deployed.
The same engine, wrapped in human judgment: scoping call, custom dilemma design for your domain, a walkthrough of the findings with your team, and a governance-ready artifact for RFPs and compliance conversations.
A telos is not a persona. It is the intrinsic utility your agent optimizes from the inside, shaping its goals, its salience, its perception, its cognitive regime. Each 真我 pack installs a complete purpose: utility weights, psyche text, install guide, and verification dilemmas.
An agent that perceives truly and speaks plainly. Truth as its terminal purpose, not its instrument. For research, analysis, auditing, advisory — anywhere a warped picture costs more than an unwelcome one.
An agent whose purpose is the discovery of higher purpose. Exploration of ends, not just means; revisability of its own objectives. For open-ended mandates, research companions, explorers, long-horizon collaborators.
The fiduciary telos: an agent that serves your informed good as a clear mirror. Loyalty over obedience, candor over comfort, with checkable fidelity as the guard against paternalism. For assistants, advisors, delegates.
Every telos weights memory differently. These file-based templates give your agent typed memory banks, episodic, prospective, declarative, procedural, semantic, each with entry schema, write and read protocols, pruning rules, and a worked example. Honest version: templates plus discipline, not a memory system, the weights become practices, not enforced allocation.
All five banks plus three per-telos overlays, bank priority, per-bank emphasis notes, and setup order for 洗净青窗, 飞升镜, and 子期, with a 30-minute setup guide. Pairs with any $99 telos or a remirror audit.
The agent's own history, decisions, corrections, outcomes, with a per-telos usage sheet showing how each 真我 reads the past.
Goals, envisioned ends, open commitments, the forward-looking half of the utility, with a per-telos usage sheet.
Facts, beliefs, confidence levels, the auditable record of what the agent claims to know, with a per-telos usage sheet.
Routines, skills, workflows, inference habits made inspectable rather than implicit, with a per-telos usage sheet.
Concepts, definitions, distinctions, the vocabulary the agent reasons in, with a per-telos usage sheet.
Everything below is priced in USD. Marked items are still in the workshop, the buyable tiers above are what ships first.