How communities update principles like a git merge—and why we need you on the test-net

When your personal agent helps you navigate a week of life—blocking late-night outrage traps, negotiating an emergency-room shift schedule, or routing zero-interest credit to a local solar co-op—it observes where existing network principles helped you thrive and where a rule was too rigid or incomplete.

Until recently, AI researchers assumed that a neural network’s internal representations of those lived experiences were a hopelessly tangled mess (“polysemantic superposition”), requiring massive corporate datacenter autoencoders to decode. In October 2026, interpretability researcher Nick Cammarata at Goodfire published a landmark 26,000-word analysis, “Neural Networks Are Already Clean” (covered in MIT Technology Review on 19 October 2026), showing that as modern models mature, their native neurons and subspaces are already overwhelmingly structured along clean, single-purpose concept axes.

For a decentralized network of personal agents, this changes everything:

• On-device concept activation without cloud surveillance: A compact 1B–8B personal model running on your own phone or laptop can directly read its own native concept activations and project them onto a 12-dimensional symbolic principle basis (v_u,k ∈ ℝ¹²) in under 10 milliseconds—with zero raw diaries or biometrics ever leaving your device.
• Direct 1-to-1 mapping from vector shifts to symbolic clauses: Because each basis coordinate d ∈ {1, …, 12} corresponds to a specific parameter in the principle’s symbolic syntax tree (such as night-shift fatigue ceiling, monsoon harvest grace window, demurrage rate, or 5-hop creator royalty), a geometric vector shift translates deterministically into a human-readable git diff!

Suppose N = 1,200 verified humans across five real-world communities—plus a billionaire attacker running 50,000 synthetic bots—submit principle update vectors in the same epoch.

What happens under standard aggregation rules?

• Failure Mode 1 — Sybil & Outlier Hijacking (Naive Arithmetic Mean, 4.9 / 100): The arithmetic mean (1/N) ∑ v_u has a statistical breakdown point of 0%. A single coordinated bot swarm or extreme outlier vector drags the global average wherever the attacker wants—wiping out demurrage on idle wealth and weakening safety quorums.
• Failure Mode 2 — “Muddy-Middle” & Zero-Majority Dilution (Unsharded Global Median, 50.2 / 100): Suppose 220 emergency-room nurses in Shard S1 discover that 12-hour night shifts cause dangerous burnout and activate the clinical fatigue boundary (v₅ = +0.64). Meanwhile, 830 farmers, coders, grid technicians, and retirees in Shards S2–S5 have v₅ = 0.00 simply because they don’t work hospital night shifts. If you dump all 1,050 people into a single unsharded global median, the four non-hospital shards outvote the nurses 4-to-1 with zeros, erasing 82% of the domain-specific safety signal!

To achieve 99.6% fidelity across both shared global rules and specialized community domains while holding communication complexity to O(N + S²), our protocol coheres communities in two sharded stages:

1. Stage 1 — Intra-Shard Smoothed Weiszfeld Geometric Median (50% Byzantine Tolerance): Inside each community/domain shard C_s, after hardware liveness verification (H_human ≥ 0.80) filters out unverified bot farms, the shard computes the L₁ Fermat–Weber geometric median:
   μ_s^{(k, t+1)} = [ ∑_{u ∈ Cs} w_u v_u,k / √(‖v_u,k − μ_s^(k,t)‖₂² + ε²) ] / [ ∑_{u ∈ Cs} w_u / √(‖v_u,k − μ_s^(k,t)‖₂² + ε²) ]
   Between epoch merges, shards track real-time signals in O(p) memory using the streaming Cardot–Cénac–Zitt stochastic geometric median update.
2. Stage 2 — Cross-Shard Orthogonal Subspace Decomposition & Meta-Median Upsert: When cohering shard medians {μ₁, …, μ_S} into the meta-median policy chain ΔM_meta, the protocol measures inter-shard cosine alignment along each symbolic axis:
   • Convergent Global Axes (MAIN_FAST_FORWARD): Where multiple shards move together, their vectors combine via the square-root-weighted Weiszfeld meta-median (W_s = √|C_s|).
   • Orthogonal Community Discoveries (ORTHOGONAL_SUBSPACE_MERGE): When a domain shard activates an axis orthogonal to other shards (|cos(Δμ_a, Δμ_b)| < 0.25), its vector is merged losslessly via direct-sum subspace projection (⊕ P_d^⊥ Δμ_s(d))—just like Git cleanly merging edits to two separate functions in the same file!
   • Antagonistic Context Trade-Offs (CONTEXT_BRANCH_MERGE): When urban and rural shards push a cadence parameter in opposite directions (cos ≤ −0.50), the compiler parameterizes the clause by verified local context rather than forcing one community to overwrite the other.
3. Stage 3 — Floor-Invariant Symbolic Gate (Π_Floor): Every candidate commit is checked against the 19 foundational floor protections. Any vector component that would weaken local data ownership, equal human voice, or the protected livelihood floor is mathematically projected to zero before the SHA-256 commit is signed.