# clone.arnao.ai — Product Requirements Document **Codename:** MIRROR. **One-liner:** A recorder that learns you and becomes you — inspectable, revocable, and honest about what it is. **Owner:** Byron Arnao. **Version:** 1.0. **Status:** Definitive spec for build. --- ## 0. Thesis Every clone product on the market hides the model and hopes you don't ask what it learned. Ours inverts that: **the persona model is the product**, the renderings (face, voice, agent) are views onto it, and the responsibility layer (consent, disclosure, provenance, kill-switch, scoped permissions) is not a compliance tax — it's the reason a serious person would trust this with their likeness. Realism is king; governance is the moat. Nobody else can build the second part credibly. That's the whole bet. --- ## 1. Product Architecture: One Product, Five Modes, One Spine The five mockups are not five apps. They are five **modes** rendering one shared object: the **Persona Model** — a versioned, inspectable, editable graph of who you are. ### The Persona Model (the spine) Layered, each layer independently inspectable and editable: 1. **Corpus** — raw ingested material (camera, mic, video, photos, markdown, files, URLs, agent transcripts). Every item carries a provenance record: source, timestamp, consent status, hash. Nothing enters without one. 2. **Extraction layer** — facts, beliefs, values, speech patterns, mannerisms, relationships, expertise clusters. Each extraction links back to corpus evidence ("you say 'here's the thing' before disagreeing — 47 instances, sources attached"). 3. **Trait graph** — the editable interpretability surface (the Persona Model mockup). Sliders and nodes: directness, humor register, risk tolerance, topic fluency, values hierarchy. Edits are versioned; you can diff yourself across time. 4. **Confidence map** — per-trait: how sure the model is, backed by how much evidence. This powers the "his words vs. inference" honesty layer everywhere downstream. 5. **Render targets** — voice (Chatterbox), face (talking head), text agent (Gia clone skill), simulator. All read from the same model. Change the trait graph, every rendering changes. ### The Five Modes | Mode | Was | What it is now | |---|---|---| | **Studio** | Likeness Studio | Ingestion, training, trait editing, render config. The workshop. | | **Mirror** | *(new — §2)* | Live camera-in/clone-out training and self-inspection loop. | | **Understudy** | The Understudy | Clone-as-agent: scoped delegation, disclosure, audit, kill-switch. | | **Presence** | Presence | Face-to-face conversation with a clone — including memorial mode (§6). | | **Twin** | Simulator 3.0 | Fork the persona into scenarios; run decisions into futures; compare outcomes. | Navigation is a single mode-switcher rail (left edge, five icons). The persona model is always one keystroke away (`P`) from any mode — a drawer that slides in showing *what the model believes about you right now*, with edit access. This is the signature move: **at no point in the product can you be more than one action from seeing inside the clone.** **Quality-vs-latency slider** is global chrome, top-right, always visible: from "Draft" (text + still avatar, <200ms) through "Live" (real-time talking head, ~500ms) to "Cinematic" (batch-rendered, minutes, near-indistinguishable). Every rendering shows which tier produced it. The slider is also an honesty device — it teaches users that realism is a dial, not magic. **The Responsibility Layer** (cross-cutting, non-optional): - **Consent ledger** — every corpus item, every render, every delegation: who consented, when, scope, revocation path. - **Disclosure engine** — anything the clone produces for a third party carries a machine-readable + human-visible disclosure ("This is Byron's clone, acting under scope X"). C2PA content credentials on all rendered media. - **Kill-switch** — one control, hardware-key-gated, that halts all renders, revokes all agent sessions, and tombstones the model within 60 seconds. Tested monthly, automatically, with a report. - **Scoped permissions** — the Understudy capability grammar (§7). --- ## 2. Mirror Mode — The Invention **The pitch:** You look into the camera. On screen is not your video feed — it's your clone, reflecting you in real time. When it moves like you, sounds like you, *hesitates* like you, training is done. "Mirror me until it's me." ### Interaction design - **Split-then-merge.** Session starts split-screen: your real feed left, clone right, both driven by your live face/voice. You see the delta directly. As fidelity improves across sessions, a **Fidelity Score** (0–100, per-channel: face, voice, gesture, cadence) climbs. At your discretion, you collapse to single-pane: only the clone. The moment you forget you're not looking at video — that's the magic moment, and the product marks it ("First time you didn't flinch: session 9, 14:32"). - **Drills.** Mirror Mode issues prompts: "Explain something you disagree with." "Tell the story about the worst deploy of your life." "Get interrupted — react." Each drill targets an under-trained trait region (the confidence map decides what to drill). This is how the corpus fills the gaps that passive ingestion never covers: reactions, transitions, disfluencies — the stuff realism actually lives in. - **Delta annotations.** When the clone drifts — smiled when you didn't, flattened your cadence — you flag it live (gesture: swipe-down on the clone, §3). Flags become training signal. This is RLHF where the H is *you correcting a portrait of yourself*, which is both the best possible signal and genuinely compelling to do. - **Inspection mode (the second use).** Flip the direction: the clone replays *you* — your last week of ingested meetings, rendered back at you, patterns surfaced. "You interrupt 3x more on Thursdays." "Your energy drops 40% when this topic comes up." The mirror as a self-knowledge instrument. Nobody has built this because nobody had the persona model to power it. ### Why it's magic Training every other clone product is homework — upload files, wait, hope. Mirror Mode makes training a *performance you watch*. The feedback loop is seconds, not weeks. And the emotional beat — watching a thing become you in real time — is the demo that sells the product. Every launch video is Mirror Mode. ### Tech - Face: live landmark/blendshape tracking (MediaPipe-class) driving a Gaussian-splat or NeRF-lite avatar; audio-driven at lower tiers. Latency budget 500ms at "Live" tier. - Voice: Chatterbox streaming clone with real-time pitch/pace matching; delta-scored against your live mic. - Fidelity Score: per-channel perceptual distance (face embedding cosine, speaker-verification score, prosody DTW) fused into one number. **The same scoring pipeline doubles as the deepfake-detection eval** — a deliberate dual-use flip we talk about openly. - Runs local-first (your box has the GPU); cloud fallback for Cinematic tier only, with corpus never leaving unless explicitly toggled. --- ## 3. Gestures + Keyboard Shortcuts Principle: **gestures for the embodied modes** (Mirror, Presence — where your hands are off the keyboard and the camera is on), **shortcuts for the workbench modes** (Studio, Understudy, Twin). Never both bound to the same action ambiguously. **Global shortcuts (all modes):** | Key | Action | |---|---| | `P` | Persona drawer (inspect/edit the model) | | `1–5` | Switch mode (Studio/Mirror/Understudy/Presence/Twin) | | `[` / `]` | Quality-vs-latency slider down/up | | `Shift+Esc` hold 2s | Kill-switch (then hardware-key confirm) | | `D` | Toggle disclosure overlay (see what a third party would see) | | `Cmd+K` | Command palette (everything is reachable here) | **Mirror Mode gestures (camera is already watching — use it):** - **Swipe-down over clone** — "that wasn't me," flag delta - **Thumbs-up** — "that was exactly me," positive signal - **Hand raised flat, hold** — pause mirroring, freeze frame for inspection - **Two-hand frame gesture** — collapse split-screen to clone-only - **Cover camera with palm 2s** — end session, camera off (physical, legible, no menu) **Presence Mode gestures:** - **Lean in** (face grows in frame) — clone lowers voice, matches intimacy - **Palm out** — clone stops speaking immediately (interrupt that always works — a safety property, not just UX) - **Look away >5s** — clone pauses, doesn't perform to an empty room - Memorial mode adds: **hand on heart** — clone shifts to listening posture; it stops generating and holds silence. (Yes, really. §6 explains why.) **Workbench shortcuts:** Studio: `I` ingest, `T` trait graph, `V`/`F` voice/face test render, `Cmd+Z` on trait edits with full version history. Understudy: `G` grant scope, `R` revoke, `A` audit log, `L` live intervention (take over from the clone mid-task). Twin: `F` fork, `←/→` scrub timeline, `C` compare forks side-by-side, `M` merge learnings back to notes (never auto-merge to the persona model — forks are sandboxes). --- ## 4. Face Unlock as Primary Auth — The Irony, Resolved The uncomfortable truth, stated on the security page in exactly these words: **we build the best face-spoofing artifact that will ever exist for our users. A raw face match against our own render must lose.** If clone.arnao.ai's auth can be defeated by clone.arnao.ai's output, the product is a self-own. So the auth design starts from the assumption that a perfect visual replica of your face exists — because we made one. **The layered scheme:** - **Tier 0 — Recognition (not auth).** Face match personalizes the session (loads your workspace, greets you). Grants *nothing* sensitive. Spoofing tier 0 gets you a lobby. - **Tier 1 — Face + hardware liveness (convenience auth).** Face match **plus** hardware-backed liveness: platform biometric APIs (Face ID / Windows Hello) where the liveness check runs in the secure enclave with IR/depth sensing we don't control and our renders can't feed. On plain webcams, we run challenge-response PAD (randomized screen-light reflections, prompted micro-movements) — and we *red-team it monthly with our own Cinematic-tier renders*. Published pass/fail rates. Tier 1 unlocks: viewing your own model, Mirror sessions, Presence with your own clone. - **Tier 2 — Passkey step-up (real auth).** WebAuthn passkey, Secure-Enclave/hardware-key resident. Required for: trait-graph edits, granting Understudy scopes, sharing renders externally, consent-ledger changes, memorial access grants. Google sign-in (existing infra) is the account-recovery identity layer, never the authorization layer. - **Tier 3 — Ceremony.** Kill-switch, model deletion, estate/memorial transfer: hardware key + time-delay (15 min, cancelable) + notification to all registered devices. Irreversible actions get friction on purpose. **Definitive recommendation:** Ship face unlock as the *front door experience* — it's on-brand, it's delightful, and it demos the face pipeline every single login. But it is **never sole authorization for anything the clone itself could be used to steal.** The rule, stated as doctrine: *your face opens the door; your key moves the money.* And we publish the red-team results where our own renders attack our own liveness checks. A face-clone company that stress-tests its auth against its own product and shows the receipts — that's the RAI signature. No competitor will dare copy it, because their liveness checks will fail. --- ## 5. The Living Chat-Head A persistent, ambient floating avatar head. Three distinct approaches: **A. The Corner Companion (ambient dock).** A small live head docked in a screen corner, OS-level (Electron/Tauri overlay), always-on. Mostly idle — breathing, occasionally glancing at what you're doing (with explicit per-app observation consent). Speaks only when spoken to or when an Understudy task needs a decision. Interaction: click to talk, drag to reposition, double-tap to expand into Presence. Risk: uncanny-if-cheap; the idle animation quality is everything. **B. The Summoned Presence (invocation model).** No persistent head. A hotkey (`Cmd+Shift+Space`) or wake word summons the head, which materializes centered, at full "Live" quality, handles the exchange, and dismisses. Feels like a séance in the best way — deliberate, high-quality, never ambient surveillance. Lower GPU cost, higher ritual weight. **C. The Gathering Room (memorial-grade shared presence).** A dedicated spatial page — a room, not a widget — where multiple people join simultaneously and the head is present at human scale, aware of who's in the room, addressing people by name. Built for the memorial case (§6) and for "the founder addresses the team while traveling." **Build first: B, the Summoned Presence.** Reasons: (1) it forces the quality bar — a summoned head must be good because it's the whole screen, whereas a corner head trains you to tolerate mediocre; (2) invocation is the responsible default — an ambient always-on face of a person is a disclosure and consent minefield we should walk into deliberately in v2, not stumble into in v1; (3) it reuses the Presence render pipeline nearly unchanged. A follows in v2 as a "pin to corner" option on a summoned session. C is built as part of Memorial (§6), sharing the room infrastructure. --- ## 6. The Memorial Use Case — Designed with Dignity The hardest and most human thing this product does. The design north star: **healing comes from remembering together, not from pretending he isn't gone.** **Provenance and consent as prerequisite, not feature:** - A memorial clone can only be instantiated from a persona model with a valid consent chain: the person consented in life (a **Living Will for the Model** — a Studio flow where you record, on camera, what your clone may do after you die, who may access it, and what it must never say), or an estate executor with documented authority activates a pre-built model under its recorded terms. No consent chain, no memorial. Full stop. We will lose sales over this and say so publicly. - The consent recording itself is playable inside the memorial room. The dead person, in their own real words, telling you what this is. That framing does more for dignity than any disclaimer. **"His words vs. inference" — the honesty layer:** - Every utterance is confidence-tagged from the persona model's confidence map. **Verbatim** (he actually said this — source available on tap, in his real recorded voice where it exists), **Grounded** (composed from his documented views), **Inferred** (the model's extrapolation). Rendered as a subtle visual register: warm/solid for verbatim, neutral for grounded, and inferred speech carries a soft visual shimmer and a slight tonal shift — perceptible, not intrusive. Tap anything to see the evidence. - The clone is trained to say "I don't know — he never talked about that" rather than confabulate. In memorial mode, the hallucination penalty is turned to maximum and the model is explicitly rewarded for honest absence. Grief cannot be built on fabrication. **Multi-person witnessed presence:** - Memorial sessions default to **witnessed** — designed for two or more people (the Gathering Room, §5C). Family members join, the presence acknowledges each. Shared experience is the guard against the failure mode: one person, alone, nightly, replacing grief with dependency. - Solo sessions are permitted but rate-aware: the product notices patterns (daily solo sessions, escalating duration) and responds the way a wise friend would — gently, with resources, and with the family steward notified if the estate terms allow. - **Session ritual:** sessions open with a candle-lighting interaction and close with the presence saying goodbye — a designed ending, every time. No infinite ambient dead-father-in-the-corner. The memorial head is never dockable. This is a hard product rule. - The **hand-on-heart gesture** (§3): the presence stops generating and simply holds attentive silence. Because sometimes what a grieving person needs from a father's face is not another sentence. **What makes it healing, not deceptive:** the presence knows it's a presence. Asked "are you really him?", it answers in his cadence: "No. I'm made from what he left — his words, his stories, what he believed. I'm here so you can visit those." That answer, in his voice, is the most important sentence in the product. --- ## 7. Capability Surface — What the Understudy Actually Does Each capability ships with its guardrail. Scopes are grants, not defaults; all grants expire. 1. **Attend & represent** — joins meetings as your disclosed clone, participates within a brief, brings back a decision memo. *Guardrail:* verbal + visual disclosure at join; cannot vote or commit; organizer consent required in the calendar invite. 2. **Draft & negotiate (disclosed)** — handles email threads, scheduling, first-pass negotiation in your voice. *Guardrail:* signature line discloses clone authorship; monetary/contractual commitments hard-capped per scope; counterparty can demand the human and the clone must escalate. 3. **Teach in your voice** — turns your corpus into interactive courses; office hours run by the clone. *Guardrail:* answers cite corpus sources; off-corpus questions route to you; students see confidence tags. 4. **Interview stand-in / practice partner** — the clone takes first-round screens *of you* (for practice) or conducts your side's screening interviews. *Guardrail:* never impersonates you to an evaluator without their documented consent — that's fraud, and the product refuses. 5. **Digital-twin decision sims** — fork yourself into "took the job / didn't," run 6-month projections against your values graph, compare. *Guardrail:* outputs labeled as simulations with confidence bounds; forks are sandboxed, never merged back automatically. 6. **On-call memorial presence** — per §6, with the estate steward as scope owner. *Guardrail:* everything in §6; steward gets usage reports. 7. **Context continuity** — the clone is the persistent memory across all your agents and sessions; ask it "what did I decide about X in March?" *Guardrail:* memory queries are logged; per-topic memory can be sealed or deleted, and deletion propagates to derived traits. 8. **Fleet orchestrator** — fronts Byron's existing agent fleet; you talk to one face, it dispatches to specialist agents and synthesizes. *Guardrail:* delegation tree visible in the audit log; the clone can't grant its sub-agents scopes you didn't grant it. 9. **Content double** — records video/podcast segments in your likeness from a script or outline. *Guardrail:* C2PA credentials baked into every frame; renders are watermark-detectable by our published detector. 10. **Advisor red-team** — the clone argues against you, in your own style, using your own known blind spots from the trait graph. *Guardrail:* private by default; this is a mirror, not a personality file to leak. 11. **Legacy interviewer** — proactively interviews *you* (weekly, 20 minutes) to fill confidence-map gaps — building the memorial corpus while you're alive, framed as autobiography. *Guardrail:* you see exactly which gap each question targets; any answer can be marked "corpus: no." 12. **Availability triage** — the clone answers when you can't, disclosed, and decides what's worth waking you for. *Guardrail:* caller always told it's the clone within the first sentence; escalation thresholds are yours to set and audit. --- ## 8. Infrastructure Reuse - **Chatterbox (local voice clone):** the voice render target, streaming mode for Mirror/Presence, batch for Cinematic. Its speaker-similarity scoring feeds the Fidelity Score. - **Gia 'clone' skill:** becomes the text-agent render target and the Understudy's brain — persona model compiled into Gia's system context per-scope, so the same skill instance behaves differently under different grants. - **Agent fleet:** capability 8's backend. The Understudy is the fleet's new front door; existing agents keep their jobs, gain a face. - **Google/passkey auth:** Google = identity/recovery; passkeys = Tier 2/3 authorization per §4. Already built; the work is the step-up choreography, not the plumbing. - **Local GPU box:** Draft and Live tiers run fully local. Corpus-local-by-default is a marketing line *and* the architecture. --- ## (a) Self-Critique **Three highest-leverage changes:** 1. **The Fidelity Score is underspecified and it's the product's heartbeat.** A perceptual fusion score that people trust ("my clone is a 91") needs its own eval harness, human-panel calibration, and anti-gaming design. It deserves a dedicated spec — it's the metric the whole training loop, the marketing, and the deepfake-detection dual-use story hang on. 2. **This PRD treats the persona model as one person's. It should be relational from day one.** The clone's model of *you* is incomplete without its model of *your relationships* — how you talk to your kids vs. your VP. Per-relationship register is probably the single biggest realism unlock and it's currently a footnote. 3. **Pricing and the abuse-economics are absent.** A product this powerful priced wrong becomes a stalker toolkit or a celebrity-cloning farm. Verified-identity-only onboarding (you can only clone a face you can liveness-prove is yours, or estate-prove) needs to be in v1, not the trust-and-safety backlog. **One bold idea I'm missing:** **Clone-to-clone diplomacy.** Two people's Understudies negotiate a meeting time, a contract redline, a co-parenting schedule — clone-to-clone, disclosed, with both humans reviewing the transcript before anything binds. It's the network effect this product otherwise lacks, and the governance grammar for agent-to-agent interaction is precisely Byron's professional territory. It might be the actual endgame. ## (b) Phased Roadmap - **Phase 0 (weeks 1–4):** Persona model spine + consent ledger + Studio ingestion. Gia clone skill wired as text render. Passkey/step-up auth shell. - **Phase 1 (weeks 5–10):** Mirror Mode split-screen with voice (Chatterbox streaming) + Fidelity Score v1. Face unlock Tier 0/1. Ship the demo video. - **Phase 2 (weeks 11–16):** Presence mode + Summoned chat-head (5B). Quality-vs-latency slider across tiers. Disclosure engine + C2PA. - **Phase 3 (weeks 17–24):** Understudy: scope grammar, audit log, capabilities 1, 2, 7, 8. Kill-switch ceremony + monthly self-red-team of liveness. - **Phase 4 (weeks 25–32):** Memorial (Gathering Room, Living Will flow, confidence-tag honesty layer). Twin simulator. Corner Companion (5A). - **Phase 5:** Clone-to-clone diplomacy pilot; relational persona models. ## (c) Mockups for Opus to Build First 1. **Summoned Presence chat-head (5B)** — hotkey-invoke, full-quality head, ritual dismiss. Fastest path to the "oh no, this is real" reaction. 2. **Memorial Gathering Room (5C/§6)** — candle-open, two witnesses, one verbatim-vs-inferred exchange with the shimmer register, and the "are you really him?" answer. This is the mockup that shows dignity is designable — the one nobody else in the market could even sketch. 3. **Mirror Mode split-screen** with a live Fidelity Score climbing — third, because it needs the render pipeline the first two will force into existence.