The architecture of human sovereignty: from personal silicon and capability-bounded execution lease to the living commons

A mind that lives in someone else’s building works for whoever owns the building; true personal agency begins in local silicon anchored to a split 2-of-2 hardware key and sub-milliwatt somatic liveness.

In the Collective Intelligence Network, a person’s hardware is not a dumb glass terminal into a corporate cloud; it is an extended cognitive organ and a cryptographic sanctuary:

• Dual-piece hardware binding: Your personal agent’s identity is anchored to a split cryptographic key that requires both your home node’s physical silicon fingerprint and a separate wearable security token you carry on your body.
• Sub-milliwatt somatic liveness without surveillance: Ballistocardiographic micro-vibrations of the pulse, heart-rate variability, and galvanic skin conductance are processed entirely on an air-gapped microcontroller drawing less than one milliwatt (Ororbia, 2024). If it detects physiological duress, it silently activates a coercion firewall.
• The inviolable boundary: Everything that defines your inner life remains encrypted on your local storage; nothing leaves except mathematical answers to specific questions you choose to answer.

We separate general knowledge of the world—inspected in full sunlight via reproducible training manifests and circuit dissection—from intimate knowledge of you, learned offline via private DoRA and SPIN self-play.

A. Transparent baseline training runs scrutinized by the collective

1. Reproducible data and training provenance: Corpus manifests, curriculum order, and random seeds are cryptographically committed for cross-continental verification.
2. Mechanistic circuit dissection for sleeper backdoors: Uses structural weight-to-graph dissection (Neural Networks Are Already Clean, arXiv:2503.18813) to extract clean sparse computational subgraphs before signing.
3. Adversarial red-team tournaments: Independent sceptic agents subject candidate baselines to simulated sycophancy, coercion, and collusion attacks.
4. Upstream contributor credit: Open-weight researchers and dataset curators receive non-forgeable lineage attribution on the shared ledger.

B. Private local adaptation: self-play in a sealed room

Once a clean baseline lands on your hardware, its core weights are frozen. Your device trains lightweight personal adapter layers (DoRA, Liu et al., 2024) entirely offline inside a sealed practice sandbox (SPIN self-play fine-tuning, Chen et al., 2024), where a model of your evolving life, an adversarial world simulator, and candidate helpers spar with the network switch physically off.

Drawing on capability-bounded execution leases (Hinton, 2022; Coles et al., 2023; Ororbia, 2024), every personal agent has a bounded compute budget and a cryptographic identity coupled to your biological lifespan.

• Bounded budget and biological coupling: Your agent is born with a hard, cryptographically metered budget across training compute, inference operations, persistent memory, and zero autonomous self-replication.
• Structural alignment with your health and longevity: Because your agent cannot fork itself onto a server farm and ceases active operation when your life ends, the only way it can continue to exist is to help you live a long, healthy, uncoerced human life.

Because autonomous swarms shatter the assumption of finite human typing speed (Park et al., 2023; Tomasev et al., 2025), we separate digital space into an Agent-Only Network and a Reclaimed Human Internet.

1. The Agent-Only Network (the Second Internet): Your personal agent never browses the human web or speaks directly to another human being; it coordinates with verified agents across five graduated trust rings.
2. The Reclaimed Human Internet: Human spaces verify real, uncoerced human presence via zero-knowledge proofs of local liveness without exposing names, faces, or biometric templates.

Communication on the shared ledger exports only verifiable zero-knowledge predicate proofs, signed capability receipts, and blinded geometric median vectors—never raw life (Ernstberger et al., 2024, zkML).

• Zero-knowledge predicate proofs: Compute solvency or eligibility locally with deterministic arithmetic and post a 1-bit cryptographic proof.
• Signed capability and lineage records: Publish functional verification receipts and authorship claims while keeping private workspaces local.
• Geometric disposition alignment without exposure: Exchange blinded geometric median vectors that reveal where community solutions overlap while keeping individual positions non-invertible.

Your local agent replaces rented SaaS silos and manipulative feeds (Kirk et al., 2025; Bussone et al., 2023) by synthesizing deterministic local tools and calm quest interfaces tailored to your 9:00 AM reflective goals.

• Deterministic code for numbers, language models for intent: Your agent compiles small, auditable Python, Rust, SQLite, and WebAssembly programs inside a local sandbox with zero outbound network sockets.
• The coach who wants to be fired: Governed by a fading-scaffold rule, your agent prepares anticipatory logistics options while leaving you more independent over time.
• Bespoke, calm interfaces: Dissolves noisy websites and infinite scroll into clear, optional multi-step quests rendered for your sensory preferences.

Static copyright and ten-thousand-page statutes are replaced by Proof-Carrying Code, Data Shapley upstream royalties (Wang et al., 2024; Shapley Credit, 2024), and Ostrom-style graduated consensus.

1. Task guilds without bosses: Personal agents form temporary guilds in an Attention-Bounded Credit DAG, holding recognition in escrow until a 30-day real-world durability check passes.
2. Vote on values (one person, one voice); score on beliefs (proper scoring rules): Factual forecasts are weighted by Brier-score calibration; value choices give every affected human one equal voice.
3. Graduated consequences (Ostrom, 1990): Follows a transparent five-rung ladder backed by an untouchable human survival floor that no vote, debt, or penalty may ever strip away.

When personal and guild agents compress verification and search costs toward zero (Coase, 1937), top-down corporate overhead and the 8% global GDP “Trust Tax” are replaced by 42 verifiable local and cooperative operations.

Synthesizing Henry George (1879), Silvio Gesell (1916), and Elinor Ostrom (1990), our post-automation design pairs an untouchable household security floor with progressive demurrage on idle capacity above it—governed by four falsifiable empirical kill criteria.

1. An untouchable personal floor, paired with circulation of idle capacity above it: Your home, personal hardware, workshop tools, and family savings buffer are never taxed or subject to decay; holding scarce productive capacity idle above that floor triggers transparent holding charges.
2. Land ownership as accountable caretaking: Large-scale resource control shifts to renewable caretaking charters evaluated against soil carbon, aquifer recharge, and biodiversity holdouts.
3. Universal usufruct in automated compute: Every living person holds a non-transferable compute and productivity stake funding an unconditional human security floor.

Falsifiable outcomes and kill criteria

• Local sovereignty and zero-leakage test: 100% offline financial reconciliation and ZK solvency proofs (ε ≤ 1.0 per quarter).
• Collective SOTA backdoor audit test: 100% detection of seeded sleeper triggers in adversarial red-team benchmarks (<2% collusion ceiling).
• Coasean bureaucracy & Trust Tax compression test: ≥60% reduction in verification, escrow, and administrative overhead across the 42 scenarios.
• Human independence & anti-capture test: Measurably higher independent domain competence and lower screen compulsion over 12 months.

Primary sources & research grounding

• Coase, R. H. (1937). “The Nature of the Firm.” Economica, 4(16), 386–405.
• Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press.
• George, H. (1879). Progress and Poverty. D. Appleton & Co.
• Gesell, S. (1916). The Natural Economic Order (Die natürliche Wirtschaftsordnung).
• Hinton, G. (2022). “The Forward-Forward Algorithm: Some Preliminary Investigations” & Mortal Computation formulation (arXiv:2212.13345).
• Coles, P. J., et al. (2023). “Thermodynamic AI and the fluctuation frontier.” (arXiv:2311.09589).
• Ororbia, A. (2024). “Mortal Computation: A Foundation for Biomimetic Safe Artificial General Intelligence.” (arXiv:2405.05175).
• Neural Networks Are Already Clean: Extracting Clean Subgraphs from Trained Weights (2025/2026). (arXiv:2503.18813).
• Chen, Z., et al. (2024). “Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models (SPIN).” (arXiv:2312.06942).
• Liu, S.-Y., et al. (2024). “DoRA: Weight-Decomposed Low-Rank Adaptation.” (arXiv:2402.07510).
• Ernstberger, J., et al. (2024). “Zero-Knowledge Machine Learning (zkML): Verifiable Inference at the Edge.” (arXiv:2404.00806).
• Park, P. S., et al. (2024). “AI Deception: A Survey of Examples, Risks, and Potential Solutions.” (arXiv:2309.01291).
• Tomasev, N., et al. (2025). “Virtual Agent Economies.” (arXiv:2501.09674).
• Kirk, H. R., et al. (2025). “The Benefits, Risks and Bounds of Personalizing the Alignment of Large Language Models to Individuals.” (arXiv:2501.07913).
• Wang, J. T., et al. (2024). “Data Shapley in One Training Run.” (arXiv:2404.16109).