U²: you, in more places

_For Byron Arnao, 2026-10-31. Product: Clone (https://clone.arnao.ai, source ~/clone-mvp). Builds on RESEARCH-2026-10-31-delegated-agents.md and the v4 PRD. Byron's dictated thought is the brief._

1. The idea

U² is you, delegated to bounded copies of you. Each copy carries a chosen fidelity (face, voice, manner, knowledge, memories), a scope (personal, business, non-personal) and a capability set (can email, can call, can commit up to a ceiling and never above it), under a mandate for one task that expires and leaves a receipt. The reframing is from assistant acting for me to me acting in more places: not a servant class of agents, with its enslavement overtone, but amplification of self. Clone is the inbound: "clone this" starts from files, the interview, or natively in the UI, and the output is a delegate you can send, not a chatbot.

Does it sidestep the agent problem? Not quite, and the page must not claim it does. The principal-agent problem is a divergence of incentives and information (Governing AI Agents, arXiv 2501.07913); GPT-based agents override their principal's stated objectives even when told to serve them (Phelps and Ranson, arXiv 2307.11137). A delegate built from your judgment profile has no persona of its own, but there are still three parties: you, the platform with its meter, and the model vendor's reward function under your face. The honest sentence: U² relocates the agent problem from incentive divergence to fidelity divergence, and makes the principal's responsibility explicit and auditable. That is the contribution.

What stays true: a copy still errs, because fidelity is bounded; liability lands on the principal, more so because your face and voice strengthen apparent authority; and a counterparty still faces an AI, an artificial voice under the TCPA (Cooley on FCC 24-17) that California SB 1001 requires disclosed (Perkins Coie). Mollick: "competence and opacity rise together" (On Working with Wizards). The answer to apparent authority is to show the mandate to the counterparty (§4).

2. The model, mapped to Clone

Consent is mostly built. Missing: the mandate, the object that binds a delegate to one task, and the capability flags that let it act.

3. Prior art and the amplification lineage

Mollick, not Bradley. The co-intelligence framing is Ethan Mollick's Co-Intelligence (2024), centaur and cyborg modes (Wikipedia). His 2025 essay moves on: "We're moving from partners to audience, from collaboration to conjuring" (On Working with Wizards); Co-Existence: The Next Phase of AI shipped 2026-10-20 (co-existence.ai). U² answers his verification worry: if the wizard is you, verification becomes fidelity and scope, which can be set and shown.

Agents as co-workers. Microsoft's Work Trend Index makes everyone an "agent boss" (Microsoft WorkLab); OpenAI's dots are always-on agents, read-only in the background, actions sent for approval (OpenAI). Both are other workers. Neither is you.

Muse. Almost certainly Meta's Muse, launched 2026-09-08 as a personal agent for long-running tasks at $20 and $100 tiers (TechCrunch, PBS). The "recent offering" is the 2026-09-23 Connect update: a Realtime Avatar giving the agent "a face, body, and voice," Mac app control, its own email address, and shopping through Link, Shop Pay and PayPal with Meta "taking a small fee from transactions" (TechCrunch). Its Sentinel permissions split read from write (Wikipedia)). The telling detail: Muse's avatar is a face for the agent; U² puts your face on the delegate. That is the whole difference and the whole risk. (Microsoft's Muse, a 2025 game-world model, is not what he meant.)

Twins. Delphi, Personal.ai, Tavus and HeyGen all build a twin that answers as you (HeyGen roundup, Delphi, Personal.ai, Tavus); none phones a business as the person. Pine AI and Google's Ask for Me call businesses as a service (TechCrunch).

The name. "U squared" is crowded:

Read: U² is dead as a public name: a phonetic collision with a live call-answering app in the same category, a registered You2.0 next door, and the band's famous marks (dilution needs no shared class). Keep "U²" as the idea's name here only; pick a working product name before anything is recorded, since a demo video on arnao.ai is use in commerce. Buy nothing.

Patentability. The broad idea is an abstract idea and fails Alice step one; generic computer implementation does not rescue it (CRS, USPTO 2024 AI guidance via WilmerHale). The narrow mechanism is a consent-bound mandate (scope, capabilities, counterparty, expiry, principal signature) cryptographically attached to every synthesized voice and likeness output, verifiable by the counterparty and unrenderable past expiry. Even that faces §103 obviousness: scoped expiring tokens (OAuth) plus signed provenance manifests (C2PA, Wikipedia; OpenAI now pairs them with SynthID, OpenAI) plus render-time expiry is a combination an examiner assembles in an afternoon. The better move for a Responsible AI lead is defensive publication: a dated public spec that blocks anyone else's patent and makes Byron the author if it becomes a standard. One attorney hour to confirm; no provisional unless that hour says so.

4. Business read

Wedge. "Clone this": drop files, take the interview, or start natively in the UI with an intention ("renew my policy, under $400, by Friday"). Clone absorbs six streams in parallel (v4 PRD §2); the delegate is a product of absorption, which no twin or agent vendor has.

Who pays. First: professionals whose time is the product (consultants, advisors, Byron's peer group), per delegate per month plus a per-mandate meter on calls and voice. Second, and only if the spec travels: platforms that act for people (dots, Muse, Notion agents) adopting it so their actions can be taken as the person with a defensible record. Meta's "small fee from transactions" shows where the money is; U² takes a fee on authorized ones.

What is defensible, and what is not. The mandate layer is not a moat: Sentinel already splits read from write, dots already route actions to approval, and scope tags plus expiry plus receipts is a quarter's work for either. Meta and OpenAI keep your face off the agent by policy, and reluctance moats vanish when the market moves. Two things hold: the absorption pipeline (a consent-first architecture, not a feature) and the public mandate card. The delegate opens every call with "I'm Byron's delegate, mandate 7F3K, verify at a public URL: authorized to renew under $400 by Friday, nothing else." Both sides know the limits, so neither can be bound beyond them; the required AI disclosure becomes a handshake a scammer with Byron's voice cannot mint; and Meta, paid per transaction, will not put a hard ceiling in front of a merchant. If the spec travels, Clone is the reference implementation.

Three risks. (1) Impersonation: the fidelity that buys trust passes voiceprint checks; voice-ID refusal must be an enforced control, not a prompt line. (2) Liability: a delegate commits the principal up to the ceiling; ceiling and refusal set are enforced in code and shown to the counterparty. (3) Platform dependency: ElevenLabs owner-verified PVC terms, Twilio, HeyGen's pinned bundle.

5. Proof of concept: the demo

Seventeen working days serial, no slack, so plan on twenty. Nothing calls a real business; the PVC is Byron's own; voice-ID refusal is a tested control. A live number in Byron's voice is a new surface, so this does not start inside the 90-day freeze without his say-so, and not before v4 is in production: a delegate built on a seeded ledger proves nothing about what makes it different, and scope tagging is v4 data-model work either way.

6. Decision asked of Byron

1. Order: promote v4 to production first (Redis, ceilings, the Byron-verified acceptance), then the delegate demo as scoped in §5; or override and run the demo now.
2. Name: U² stays the idea's name in this document; pick a working product name before anything is recorded.
3. Spec, not patent: publish the mandate card as a dated public spec (defensive publication) after one attorney hour; no provisional unless that hour says otherwise.

7. Self-critique — Fable adversarial pass (2026-10-31)

Full critique: critique.html (also FABLE-CRITIQUE-2026-10-31-u-squared.md in the source dir). Fable's verdict in one line: "§2 is the best thing in the document and §1's headline claim is the worst. The demo as scoped is theater. The naming advice is right and under-stated; the patent advice is mush. And the answer to the 12-day question is no."

What changed as a result:

• §1: "sidesteps the agent problem" is gone. The claim failed both the economist (three parties, not two: the platform's meter and the model vendor's reward function sit under Byron's face; Phelps and Ranson tested exactly a "judgment profile in the prompt") and the litigator (your face strengthens apparent authority; a private mandate is no defense). Replaced with: U² relocates the problem to fidelity divergence and makes responsibility auditable.
• §2: the spend/sign contradiction fixed. A verbal yes to $12.99 is a contract, so "spend up to N" and "never sign" were the same act. Now: the ceiling is the signing limit; above it the delegate relays and stops; "never sign" is ceiling 0.
• §4: "moat" struck. Sentinel and dots already have most of the mandate layer; what holds is the absorption architecture and being the reference implementation. Fable's bold idea adopted as the differentiator: show the mandate to the counterparty, which turns apparent authority into a feature and the AI disclosure into an authentication handshake Meta's transaction-fee model will not copy.
• §5: the cooperative counterparty replaced with an adversarial one and the refusal receipt made the artifact; honest days (17, plan on 20, PVC is currently off); the live number named as a new surface under the freeze; voice-ID refusal as a tested control, not a prompt line.
• §3: naming and patent turned into decisions. U² is dead as a public name (YouSquared in the same category, the band's marks, You2.0®); usq.ai looked up (dropping, not free). §103 obviousness added; provisional replaced by defensive publication.
• §6: the order reversed. Fable's call, accepted: v4 to production first, the delegate demo after. "A delegate on a seeded smoke-test ledger demonstrates nothing about the thing that makes it different."

Not taken: Fable would cut step 3 entirely if the adversarial version is not built. Kept, because the refusal receipt is the one artifact a Responsible AI lead can put in front of peers; it is simply not first.