Sovereign AI Architecture Spec / v1.0
CC BY-SA 4.0 · Open SpecSovereign AI
Architecture Spec
A citable, openly-licensed reference for building sovereign AI infrastructure. One stack, five planes, two products — from the edge of a learner's cracked phone screen, up through cognition, community, market, and sovereignty, and back. For the conceptual framing, see the AI Sovereignty Framework.
The shape we're building toward: an ARPANET for AI
The whole stack
One stack, five planes.
HumaneFrame (the generative half — pedagogy, AI runtime, skill verification) and the Diné Talent Marketplace (the allocative half — supply, demand, matching) share one membrane: the Personal Capability Record.
Roadmap
The sovereignty arc.
Phase 01
Pilot
Cloud-gateway Saki + Aspire Ability marketplace
Prove the thesis. Real Navajo learners. Real jobs. Real verified competencies. Cloud-hosted intelligence with sovereign data ownership at the edge.
Phase 02
Lab
Locally-tuned LLM at Diné College / NTU
Move the model. Domain-specialized LoRAs trained on Nation-authored corpora. DGX Spark deployment. Bellman-equation routing for community-aware decisioning.
Phase 03
Mesh
Chapter House mesh — sovereign by infrastructure
Move the infrastructure. Local compute at Chapter Houses. Offline-first. Federated. The Nation operates its own AI without dependency on any external provider.
Technical reference
The deep stack — expand to explore.
Each section below is a self-contained chapter of the architecture. Open what's relevant to you; the rest stays out of your way.
System map
The whole picture — four planes
Alastair's space, Saki's intelligence, the labor market, and the infrastructure beneath. Red dots mark architectural decisions still being worked through.
System map
The whole picture — four planes
Alastair's space, Saki's intelligence, the labor market, and the infrastructure beneath. Red dots mark architectural decisions still being worked through.
Every competency signal: L0 self-attested → L1 inferred → L2 assessed → L3 endorsed → L4 validated (OBv3)
Saki adapts its register to who is in the seat. Same agent, five postures.
Every claim Saki records carries a provenance tier. L4 alone is portable across institutions.
- L0Self-attestedAlastair claims the skill.
- L1InferredSaki infers from behavior.
- L2AssessedSaki probes with rubric tasks.
- L3EndorsedA mentor or peer vouches.
- L4Validated (OBv3)Issued as a verifiable credential.
Numbered red dots throughout the map are tappable. They expand on key architectural mechanisms.
- 1System 2 — System 2 leverages the SNBH vector for routing and adjustments.
- 2Artifacts + docs — Artifacts are versioned (along with metadata) to maintain history.
- 3Validated (OBv3) — Federated implementation and crosswalking across issuers.
- 4Librarian — Librarian spins up and maintains domain, profession, and career repositories that are dynamically updated.
- 5Collaborative — Peers are matched for collaborative learning.
- 6Parent — Parents are also learners — they can work on their own learning while also functioning as Alastair's mentor.
- 7Prof. domains — Domains are managed by the Librarian (governed by fine-tuned LoRA).
User journey
Alastair's path — seven steps
From first conversation with Saki to a lifetime of accumulating verified competencies.
User journey
Alastair's path — seven steps
From first conversation with Saki to a lifetime of accumulating verified competencies.
Step 01 — Explore
Alastair opens HumaneFrame on her phone. She meets Saki. She talks about what she's interested in — jobs at the Navajo Housing Authority, remote work, project coordination. Saki doesn't present a dropdown. She listens. The SNBH state vector begins initializing.
Core infrastructure
The 5-tier stack
Edge → Engagement → Intelligence → Data → Safety Net. Click any layer for components and tech.
Core infrastructure
The 5-tier stack
Edge → Engagement → Intelligence → Data → Safety Net. Click any layer for components and tech.
The entry point. Built to meet the learner exactly where they are, regardless of hardware or connectivity.
Progressive Web App
Mobile-first design for cracked Android screens, with "Progressive Escalation" to desktop for complex tasks.
Local-First / CRDT Sync
IndexedDB local storage ensures the system works entirely offline in broadband dead-zones, syncing immutably to the cloud only when a connection is stable.
Fluid Multimodality
Edge-computed Speech-to-Text (WASM Whisper-lite) allows learners to bypass the cognitive load of typing and speak naturally about their lived experiences.
How the system extracts signal from the noise of human experience.
Saki — The Orchestrator
The user-facing agentic partner. Saki engages in a multimodal "Diagnostic Dialogue" — no dropdowns.
The Broken Thing Sandbox
Replaces multiple-choice testing. Saki provisions secure, ephemeral environments to assess executive functioning, resilience, and problem-solving under productive frustration.
Calibration Handshake
The learner explicitly sets their friction/empathy baseline at "Minute One."
The proprietary cognitive engine behind Saki. This is the technical moat.
System 1 — Sensorium
Fast associative LLM that ingests real-time signals and generates candidate responses.
System 2 — Security Enclave
Deterministic routing backend. Calculates the Murasaki Bellman Equation to enforce the Generative Friction Constraint. Applies the "Ezekiel Penalty" to mathematically block cognitive surrender.
System 3 — Hyperagent
Self-referential evolutionary loop that monitors efficacy and autonomously improves coaching strategies — strictly bound by the equity objective function.
The infrastructure of learner sovereignty. We store the messy truth and export the market standard.
Schema-on-Need
Ingests unstructured, authentic human reality. The Librarian Agent parses at inference time to tag with machine-readable Rich Skill Descriptors.
Personal Capability Record
A learner-owned, portable, immutable ledger built on 1EdTech open standards. It belongs to the learner, not the institution.
Labor Market Grounding
Hydrates the system with Employer Reference Objects using real-time, localized job postings to ensure all learning maps to actual demand.
The ultimate fail-safe. AI should not solve human crises.
Triage — Rule-Based
When the SNBH emotional load vector spikes, the system halts learning and triggers escalation. Explicitly deterministic — generative LLMs are documented as dangerous in crisis contexts.
Routing — RAG Directory
Queries HSDS-compliant directory of vetted, local community resources. Integrates Gravity Project FHIR SDOH standards. 90–96% precision vs general-purpose LLM hallucination.
Accompaniment — Human Weavers
A trained, compensated Navajo Nation community navigator receives warm handoff from Saki and stays through resolution.
Epistemological core
SNBH state vector
s⃗ₜ = [cₜ, eₜ, pₜ, kₜ]ᵀ — the live four-axis vector Saki routes through. Optimization target is Hózhó, not task completion.
Epistemological core
SNBH state vector
s⃗ₜ = [cₜ, eₜ, pₜ, kₜ]ᵀ — the live four-axis vector Saki routes through. Optimization target is Hózhó, not task completion.
Saki reduces friction when overwhelmed; adds Socratic challenge when coasting.
Trauma and exhaustion are first-class signals. Triggers Weaver Protocol when threshold exceeded.
High-dimensional. Grounds logic in the learner's actual organization, geography, and culture.
Tracks relational connections. Actively moves learner from isolation toward kinship.
Multi-agent system
The hyperagent swarm
Four agents — Mapper, Designer, Instructor, Provocateur — that compose Saki's behavior.
Multi-agent system
The hyperagent swarm
Four agents — Mapper, Designer, Instructor, Provocateur — that compose Saki's behavior.
Learning theory
The pedagogical framework behind the swarm
How learning theory is formalized into computable constructs and routed to agents.
Learning theory
The pedagogical framework behind the swarm
How learning theory is formalized into computable constructs and routed to agents.
Theory → Computable Construct
Pedagogical formalization
Four functional layers organize a century of learning theory into computable constructs that route to specific components of the Hyperagent Swarm.
Teleology
Why and for whom — sets goals, learner stance, ethical frameArchitecture
How to design learning — pathways, evidence, ZPDMechanism
How learning actually happens — task, feedback, retentionAffect & Identity
Who the learner is becoming — agency, kinship, dignityOne Coaching Interaction
Runtime pedagogical flow
When Saki coaches Alastair through a Financial Reporting quest, each step is governed by a specific, formalized learning theory — not improvisation.
- STEP 01
Session opens
Alastair sets her SNBH sliders
Knowles — andragogy
Self-directed adults control their own learning conditions
- STEP 02
Goal confirmed
Financial Reporting gap identified
McTighe — backward design
Define mastery evidence first, then design the path
- STEP 03
Difficulty calibrated
Bloom's Apply level — not Remember, not Create
Vygotsky — ZPD targeting
Stay in the zone between can-do and can't-yet
- STEP 04
Task presented
NHA variance report — a real artifact
Merrill — task-centered instruction
Real task, activate prior knowledge, then apply
- STEP 05
Alastair builds artifact
Constructs the variance report; Saki watches
Bjork — desirable difficulty
Productive struggle builds durable competency
- STEP 06
Friction injected
Saki withholds the answer; asks a sharper question
Generative Friction Constraint
Prevent cognitive surrender — protect the learner's thinking
- STEP 07
Artifact reviewed
Competency claim written to PCR
Mislevy — evidence-centered design
The artifact itself is the credential evidence
- STEP 08
Identity update
Alastair sees herself as someone who can do this
Bandura — self-efficacy
Mastery experience is the strongest source of agency
Mathematical foundations
Core equations
The Generative Friction Constraint and the Charity Reward Function — what mathematically blocks cognitive surrender.
Mathematical foundations
Core equations
The Generative Friction Constraint and the Charity Reward Function — what mathematically blocks cognitive surrender.
Generative Friction Constraint
α = empathy multiplier (behavioral telemetry)
β = penalty for cognitive surrender
Fprod = productive friction level
Ssurr = surrender signal intensity
Charity Reward Function
Λ = Rawlsian multiplier (scales with acute distress)
Ω = Ezekiel Penalty (blocks "solving" over coaching)
ΔH = delta toward Hózhó (holistic balance)
s⃗ = SNBH state vector
Sovereignty & structure
The PCR & the dual-entity model
Learner-owned credentials on open standards; a 501(c)(3) + for-profit pairing that keeps the moat in relationships, not data.
Sovereignty & structure
The PCR & the dual-entity model
Learner-owned credentials on open standards; a 501(c)(3) + for-profit pairing that keeps the moat in relationships, not data.
Data Sovereignty
Personal Capability Record
What it is
Learner-owned · Portable · Immutable · Sovereign · Not institution-controlled
Open standards
Open Badges v3 · CLR · 1EdTech · Tribal data sovereignty
Corporate Structure
Dual Entity Model
HumaneFrame Institute
501(c)(3) · Nation relationships · Federal grants · Employs Weaver network
HumaneFrame Inc.
For-profit · Platform builder · Venture-ready · Licenses to nonprofit
Cite this spec
An open spec, freely adoptable.
The Sovereign AI Architecture Spec is published under Creative Commons Attribution 4.0. Peer Nations, institutions, and researchers are welcome to adopt, adapt, and extend it — with attribution.
Plain citation
HumaneFrame. (2026). Sovereign AI Architecture Spec (v1.0). Aspire Ability Foundation. https://humaneframe.org/architecture
BibTeX
@techreport{humaneframe2026spec,
title = {Sovereign AI Architecture Spec},
author = {HumaneFrame},
year = {2026},
number = {v1.0},
institution = {Aspire Ability Foundation},
url = {https://humaneframe.org/architecture},
note = {Licensed CC BY-SA 4.0}
}