Family AI Provisioning → Cornerstone — the family's own governed AI

Status: CONCEPT (Sandon, 2026-07-29). Product named + public page live (2026-07-30). Not built. A strategic direction + a potential hardware licensing deal, captured so it's a shareable artifact.

Product name: Cornerstone (chosen 2026-07-30) — the load-bearing foundation of a home; echoes the "Bedrock for families" metaphor directly. Public product page + waitlist live at /cornerstone (server page, scoped OG, painterly hero; client waitlist form → POST /api/cornerstone/interest, route, cornerstone_interest migration 602). The waitlist is deliberately the first thing built: it's the instrument that validates the plan's #1 open question (families' willingness to pay is unproven) before any hardware spend — and the tier radio (software-now vs. the home box) measures which tier to sequence first. Surfaced in Navbar "More", footer, and the /features catalog. Admin review at GET /api/cornerstone/admin/interest. Deal-structure reminder: any Cicero involvement is licensing only, never partnership/equity (see business model below) — the public page names no vendor.

The one-line: enterprises adopted Bedrock/proxy layers over Claude not for the model but for the governance — visibility, policy, audit, cost control over employee AI. Families face the identical need with their kids and AI, and there is no consumer product for it. Family Wall is the natural console; a local hardware partner (Matt / Cicero) is what makes it defensible.

Where it grows to (2026-07-30): the operating system for the home. Governed AI is the first app, not the product. The same home hardware becomes the family's private data vault (photos + documents), a fault-tolerant multi-device mesh, encrypted off-site backup, and — for the fully sovereign — a self-hostable local HiveJournal. See "The family data layer" below. The prize is the trusted layer a family's whole digital life runs on, owned by them, in their house — the one position the cloud incumbents structurally can't take.

The insight

"Parental controls for the AI age" is a real, unserved category. Kids are using AI; parents have no visibility, no policy, no data ownership, and no product that gives it to them without handing a vendor even more of the family's private life.

The enterprise analogy is exact. Companies didn't put a proxy in front of Claude because they distrust the model — they did it because they needed:

• Visibility — what's being asked, by whom.
• Policy — what's allowed, for which role/age.
• Audit + data control — where the conversations live, who owns them.
• Cost control — a governed, metered gateway.

A family needs the same four things, pointed at their kids instead of employees. Nobody sells it. Family Wall is a credible console because it's already the family's shared, PIN-gated trust surface, with kid-chat + parental-oversight primitives in place (see family-kid-chat-core, family-kid-safety-core).

What it is

A family AI gateway — family members' AI conversations route through a family-controlled proxy — plus a management layer on the Family Wall giving parents:

1. Visibility — age-graded (see the design law below), not a surveillance feed.
2. Policy — allowed topics, time windows, per-child settings, escalation rules.
3. Data ownership — conversations stored in the family's own space, not a vendor's cloud. This is the load-bearing differentiator.

It is a provisioning / governance layer, not a new model — "Bedrock for families."

The Matt / Cicero dovetail — the defensible part

The proxy needs somewhere to run and store family data.

• Cloud-only recreates the surveillance problem — a vendor holds the family's conversations. That's the thing families are (rightly) wary of.
• On local family hardware (Cicero's box — storage + on-device AI), the family's AI and data stay physically in the home, under the family's control. That is the honest version of "the family owns its data," and no big player will ever offer it — their entire model is your data in their cloud.

So the split writes itself:

Together: the family's own private, governed AI — run from the wall, stored at home. It also gives the Sophia / cicero.sh local-NLU thread its use case, and its personal/household-only license fits a family product exactly, where it never fit a SaaS one (see the Sophia licensing gate).

Matt's reply (2026-07-30) — integration confirmed, three specs are ours to answer

Matt replied to the dovetail pitch. Substance:

• His NLU is done and he's confident in it. GitHub + an open-source, reproducible test suite land "in a day or two" — which is exactly what de-risks the Sophia licensing gate's "verify the benchmark before shipping" step: we run his suite ourselves rather than trust a number. Wait for it, run it, then proceed on the near-term path.
• "What you said slides in perfectly with my tech." The Sophia-as-decide- locally + Bedrock-for-families framing fits his engine; he wants our implementation details to build against.
• His end-to-end direction: integrate soniqo/speech-core
  • bundle micro wake word → a fully local STT → wake → NLU → TTS pipeline, ~1.2 GB CPU, no GPU. The standalone NLU alone is ~160 MB.
• The moat thesis, his words: everyone's chasing transformer/probabilistic intent; nobody wants deterministic NLU because it's hard — the field is wide open. That's the same bet our owned-hardware plan already makes.

His three questions — our answers (pending Sandon's confirm)

He asked three implementation questions. They're ours to answer, and the answers are path-dependent — which is the useful part, because it forces the two-path split to be explicit:

The honesty correction this forces on our own plan

Near-term, "decide-locally" on glasses is really "decide-on-our-server" — because Mentra runs the app in our cloud (Railway), not on the frames. That's still a real win (faster than a frontier round-trip, our infra not a third party's LLM cloud, most data stays our side, deterministic on intent), but it is not the "conversations physically stay in the home" story. That story only becomes true on Matt's box. Keep the two claims distinct — the mesh's own honesty discipline (attributed, never overclaimed) applies to our own architecture too: server-side deterministic NLU now, genuinely-local family AI on the box.

Net effect on build order: the glasses NLU seat (apps/glasses/src/sophia.ts, x86-64, text-in) can proceed as soon as his suite verifies — independent of the box. The box stays the endgame (step 3 below), now with a concrete substrate (his 1.2 GB local STT+NLU+TTS pipeline) instead of a hand-wave.

The design law: loving oversight, not surveillance

This is the whole ballgame, and it's the same consent-first line the mesh keeps hitting (Lovio, Lighthouse, kid-chat-safety, "read-only for the dead").

"Parents can read their kid's AI chats" is legitimate for a 7-year-old and corrosive for a 15-year-old. The version families trust rather than resent is:

• Age-graded visibility that decays toward autonomy. Full transparency for little kids; for teens, the parent sees that the AI is being used and any flagged concerns, not a transcript.
• The AI flags a genuine safety concern (self-harm, grooming, dangerous instructions) rather than parents reading everything. Oversight is exception- based, not ambient.
• The child can see what's shared with whom. No covert monitoring — the same "attributed, never covert" rule as the audio-honesty standard, one surface over.

Get this wrong and it's a creepy product families uninstall. Get it right and it's the one they trust with the thing they trust least.

Architecture sketch

• Gateway/proxy — family AI requests route through a family-scoped endpoint (policy + logging + metering applied here).
• Decide-locally, escalate-when-needed — a local model (on Cicero hardware, or a small hosted model) handles most turns and the intent pre-filter; a frontier model (Claude) is called only for the hard asks. This keeps most data local and sidesteps most of the frontier-reseller cost/ToS weight.
• Family-owned conversation store — home hardware (privacy-max) or a family-encrypted cloud space (software-first tier).
• Wall console — parents configure policy + see the (age-graded) oversight surface from the Family Wall they already have.

The family data layer — vault, device mesh, encrypted backup (2026-07-30)

Cornerstone starts as governed AI, but the same home hardware that keeps the family's AI at home is the natural home for the family's whole private data life — a private cloud that is actually theirs. Four layers, added over Sandon's ideation (2026-07-30):

1. The vault — host the family's photos + documents. The home box isn't just AI memory; it's where the family's photos and documents live, on hardware they own. This is the honest competitor to iCloud / Google Photos on the one axis the big players can't match: it's physically yours, in your house.

2. The device mesh — more than one box, fault-tolerant. A single box is a single point of failure (the obvious objection to a home vault: "what if it breaks?"). The answer isn't the cloud — it's a second box. Multiple Cornerstone devices form a local mesh that replicates across each other, so a device failure loses nothing. Even better geo-distributed — one at home, one at a grandparent's — which also seeds the Cairn/legacy story.

3. Encrypted cold backup — a paid add-on. Beneath the device mesh, an off-site last resort: client-side-encrypted backup to cold storage (AWS S3 Glacier Deep Archive / Backblaze B2 / similar). Cheap cold storage makes it a low-margin-cost add-on revenue line with obvious value (fire, theft, both boxes lost).

4. The keystone — WE can never read it. The encryption keys are held by the family, never by us or the storage vendor; we store only ciphertext. This is non-negotiable and is the same honesty line as the mesh's other standards (audio-honesty, honest-scaffold): if we (or AWS) can read the backup, it is just another vendor cloud and the entire "you own your data" promise is a lie. Client-side / family-held-key encryption is what makes the add-on true. (It also lowers our liability on the photos-of-minors surface: we can't moderate what we can't read — a real tension to work through with counsel, and the reason the family photo layer is counsel-gated.)

5. The sovereignty tier — sync a packaged HiveJournal to the box, run fully offline. For the family that doesn't want to rely on us either: a self-contained build of HiveJournal that runs entirely on the Cornerstone device(s), no cloud dependency on our servers at all. The ultimate "you own it" — even the software is theirs, local, inspectable. We stop being a service they trust and become software they run. (Ties to the existing self-host / platform-independence thread.)

Layered: hot (the box, in the house) → resilient (the device mesh, survives a failure) → cold (encrypted off-site backup, survives everything) → sovereign (the whole system runs locally, survives us). No big player offers any of this, because their whole model is your data in their cloud. The Cicero (Matt) hardware — licensing only — is the substrate for layers 1–2.

The thesis this adds up to: the operating system for the home

Each layer alone is a feature; together they're a platform. Governed AI + the family data vault + photos & documents + a fault-tolerant device mesh + encrypted backup + a self-hostable local build = Cornerstone as the operating system for the family's digital home. The AI is the first app, not the whole product. The strategic prize isn't "parental controls for AI" — it's the trusted layer every family's digital life runs on, owned by them, in their house. That is a generational position, and it's the one the cloud incumbents structurally cannot take, because taking it would mean giving families their data back.

Business model

• Family Wall Plus is the obvious home for the software tier (see FAMILY_WALL_MONETIZATION.md). "Keep your kids' AI safe + your family's data in your own home" is a strong emotional + privacy pitch and a natural premium.
• Hardware bundle / privacy-max tier — the Cicero local box as an upsell once the software proves appetite.
• Enterprises pay for AI governance because of compliance; families' willingness to pay is unproven — this is the #1 thing to validate before hardware spend.
• Deal structure with Cicero: licensing only. Engage Matt/Cicero purely as a licensor of the NLU (an OEM/license deal, consistent with the household-only license) — not a partnership, equity, or joint-entity arrangement. This is the cleaner IP/liability structure regardless, and it's a firm constraint: if any plan drifts toward co-ownership, steer it back to a license.

Build order (software-first — do NOT gate v1 on hardware)

Owner update (2026-07-30): market validation is NOT a gate — build it. Sandon is building Cornerstone for his own family first; real marketing comes once it works for them. So willingness-to-pay is no longer a precondition for building (conviction, not proof). The gates that REMAIN are real and unchanged: (a) Matt's NLU verified via his open-source suite before the local tier, and (b) counsel / COPPA review before opening it to OTHER families' children — own-family use (the owner as the parent, his own kids) needs no third-party-consent gate; public launch does. Build now, behind a flag; launch to non-owner minors after counsel.

Key realization (2026-07-30): the engine already exists as kid-chat. HiveJournal already ships the governed-family-AI substrate: family_kid_chat (per-child policy — age_band, blocked_topics, daily_cap), a local-first safety classifier (family-kid-safety-core.localSafetyVerdict, explicitly a "Sophia-style local NLU" seam), an age-banded safety-first system prompt (kidSystemPrompt), and a message/flag audit (family_kid_chat_messages). Cornerstone is the governance CONSOLE + the design-law over that, not a rebuild.

1. Age-graded oversight console — BUILT 2026-07-30. The design-law layer over kid-chat: cornerstone-oversight-core (pure, golden-tested — visibility decays toward autonomy: full transcript for littles, concerns-only for teens; safety flags reach the parent at EVERY age), getFamilyOversight (composes kid-chat data), GET /api/cornerstone/overview, console at /dashboard/cornerstone.
2. Family-owned (encrypted) conversation store + fuller policy UI — next: bring more of the kid-chat config (enable, topics, caps) under the Cornerstone console; own-store framing.
3. Cicero local-hardware tier — the privacy-maximal dovetail (Matt = licensing only, per the business model). The endgame; gated on his verified NLU.

Open questions

• Which model(s)? Local-only (weaker, fully private) vs. frontier-proxied (stronger, ToS/cost weight) vs. the hybrid above. Recommend hybrid.
• Reseller shape — proxying Claude for families has a licensed-reseller flavor (like Bedrock); clarify terms before it's a product.
• Willingness to pay — validate before any hardware commitment no longer a build gate (owner conviction, 2026-07-30; dogfood own-family first, market after).
• Regulatory — kids' data + AI is a COPPA / child-safety surface; the age-graded design helps but counsel is needed before launch.

Connections across the mesh

• OWNED_HARDWARE_HORIZON.md — Sophia / cicero.sh local decide-locally NLU; the licensing gate whose household-only EULA fits here.
• FAMILY_WALL_MONETIZATION.md — Family Wall Plus, the software tier's home.
• Lovio / JQ Bridge — the "loving oversight" pattern this inherits.
• family-kid-chat-core / family-kid-safety-core — the existing kid-chat + safety primitives to build the oversight surface on.