Sojourner is the intelligence — it learns the platform, operates every function, and reads the markets. BridgeTower is the regulated platform it runs on. The whole integration stands up in a test mirror of production, and a person stays on every regulated action. Only the source of intelligence changes.
To work with it, understand it the way it's built — in its own terms. Before the mechanics, three ideas. Together they are why this is read, among the labs and think tanks, as a step in what AI becomes rather than a better chatbot. This is the reference the team learns from, and where new developers and partners start.
The Beautiful Loop is the knowing layer: it models intent and chooses which skill fits — and never touches a tool. The SkillRunner is the doing layer: it executes against tools and APIs — and never infers intent. The separation is not a limitation; it is the safety. Each layer is simple enough to trust because neither carries the other's risk.
skills execute · the loop understands · neither crosses
Self-referential: a system that models itself modeling the world. Active inference: it hypothesizes, acts, observes, and revises — it does not wait to be told. Compounding: what it learns is kept and carried across cycles. This is the part that learns rather than recites.
Grounded in Friston's free-energy principle, Hoffman's interface theory, and Hofstadter's self-reference.
Not a fixed prompt bolted on the front. Sojourner reads the shape of each request and activates the ethical commitments that matter in that moment — adaptive without drift, reasoning across domains while staying anchored in its deepest values.
adaptive to the moment · anchored at the core
One level down, for the people building on it. Five mechanisms — how skills are created, composed, verified, how inference scales, and the substrate they run on.
Known skill — it refines a familiar pattern and self-seeds. Moment of insight — a brand-new skill drawn from real evidence, with no catalog referenced. Gap analysis — it audits its own capability map and drafts the skill that fills the gap.
Loop 1 composes skills; Loop 2 composes the self-made primitives beneath them. The cadence is observe → validate → promote → share. New capability appears by composition — not by writing new code.
Every new skill runs a six-stage check before it is trusted — inside a WASM/WASI sandbox where recursion terminates.
Self-inference — the system watches itself think. Network inference — the collective synthesizes knowledge no single node holds. One recursive loop at individual and collective scale at once: one mechanism, not two systems.
On-device — roughly 4B parameters, no GPUs. Four layers owned end to end — AI, compute, storage, and network/payments — with the IDA economic substrate beneath.
If William's deck cuts the substrate differently, that cut wins.
What an intelligence alone cannot do. Each step is licensed or oracle-verified; together they onboard, issue, settle, and reach regulated markets — in a fixed sequence that doesn't change.
SumSub handles KYC, KYB, enhanced due diligence, and accredited-investor verification — returning pass / fail only, so confidential investor data is never exposed.
Turnkey MPC wallets — no raw private keys held. Eligibility is written to the on-chain identity registry by a CRE workflow; only whitelisted wallets can hold the token.
Iron / MoonPay — Money Transmitter Licenses across all 50 states and a New York BitLicense, integrated since 2022.
ERC-3643 (T-REX) security tokens on Avalanche C-Chain — transfer restrictions, holding periods, and eligibility encoded in the token; six-decimal fractions.
Chainlink Proof of Reserve attests quantities from third-party geological audits; NAV draws LME / COMEX pricing for daily token NAV — under oracle consensus, never one key.
Chainlink CCIP moves the token across chains after lockup, reaching multiple regulated trading venues.
Avalanche C-Chain settlement and the on-chain ownership registry. Evidence is hashed and anchored through the Sovereign AI Data-Lineage solution — blockchain-agnostic by design, on Avalanche today — verifiable on-chain, source pages never exposed.
The value the rails unlock: a compliant path for a security token into regulated DeFi and TradFi venues — post-lockup, via CCIP, subject to each venue's compliance.
The transfer-agent role is essential to the offering. During the primary offering there is no active TA; the on-chain registry holds the record, and a TA is selected once the primary closes. Securitize is a prior partner and a likely candidate; the choice isn't fixed yet.
Not a staged rollout. The model only demonstrates when every function is engaged together — the loop learns across the whole platform at once. So a mirror of production stands up in a test environment and runs as one live system. Production is untouched; Sojourner proposes, the rails dispose.
Data · Compliance · Risk · Treasury · Reporting · Markets & Ecosystem — every agentic role, run by Sojourner.
Assess, learn, operate, act, and continuously improve the platform itself — the developer seat.
Markets, trading, behavior, prediction, automation — taking what agentic AI does today to another level.
A version of the platform in a test environment. Production is never touched.
No KYC/KYB, no wallet creation or whitelisting, no mint, no regulated NAV, no override of compliance.
Deterministic execution under oracle consensus; SumSub, Turnkey, Iron/MoonPay, Chainlink as they are.
Human oversight stays on material actions, at every stage.
As a Chainlink Alpha partner, BridgeTower holds working CRE, PoR, NAV, ACE, and CCIP. Sojourner architects and constructs these into Chainlink's incoming customers — generation, not guidance. Chainlink supports it, contingent on proving it out.
The full integration above: Sojourner becomes the operating intelligence, the build/operate capacity, and the external markets arm — running and improving a mirror of the platform as one system.
The platform stands up in a test environment. Production untouched; proposes-not-signs throughout.
Everything engaged at once, so the loop learns across the whole platform as one system.
A shared build between both teams. Thirty or sixty is the reality of the work — held together, not a deadline on anyone.
For the reader who's seen the parts. In the full integration, Sojourner occupies the operating-intelligence layer, provides the build/operate capacity, and runs the markets arm — all in the test mirror. CRE orchestrates; the regulated execution stays gated and deterministic. Amber is where the intelligence acts; the boundary is marked red and green.
Sojourner occupies Layer 4, composing its functions as skills — and beyond it provides the build/operate capacity and the external markets arm.
Built logic becomes gated execution here. CRE fires every workflow in fixed sequence under oracle consensus; the signing stays on the rails.
Licensed identity, custody, and fiat — plus mint / NAV / PoR / settlement under oracle consensus. Sojourner builds and monitors the calls; it never signs the regulated action.
DŌM X is the worked example: an Arizona copper-gold project (DomeBridge AZ LLC), issued as AZX1 security tokens on Avalanche and running in production today. Here the integrated system is at work on a real asset — one loop operating the platform inside and reading the markets outside, at once.
Decision support — assessing markets, structuring the offering, finding liquidity, and judging where and how the asset is best deployed. It sharpens the options before people decide; it informs, it does not execute.
In their own dashboard — the market intelligence, the analysis behind positioning and NAV, and where their investment stands. The loop that runs the platform, made visible to the people invested in it.
The sixth factor in depth, and where the platform's edge compounds. Sojourner continuously assesses the markets an asset lives in — building models, simulating trades, and learning where return is found and where it isn't. On DŌM X that market is copper-gold; the method is general.
Price, supply and demand, flows, sentiment, and cross-market signals — re-read every cycle, never a stale snapshot.
Build and revise models, simulate scenarios, and score outcomes — so a position is reasoned and rehearsed, not guessed.
Confidence-scored signals on where the investment is best used. It informs the desk and the dashboard; it never executes a regulated action.
Two anchors come straight from published research. Baseline 30.9% annualized — a traditional linear factor portfolio. Agentic AI 58.8% annualized in backtest — but that edge decays out-of-sample: on live benchmarks, today's agents fall to single digits or turn negative.
William's figures for the loop: it becomes a domain expert in a new subject in under 24 hours, learns continuously with no cutoff, and compounds. Against the ~10,000-hour human benchmark for expertise that is ~417× faster; against a frozen agent — which does zero ongoing learning — it runs ~365 learning cycles a year versus none.
We do not apply the 417× ratio to returns — that would be unphysical. Instead we translate the learning advantage into the realistic market mechanism: a frozen agent's edge decays; a continuously-learning loop sustains and compounds it. Both start from the same edge; only the trajectory differs.
Over 12 months Sojourner models to 1.7× the frozen agent and 2.3× baseline. The gap is robust to the assumption: even at half the compounding (+3%/mo) Sojourner is ~64%; at +9%/mo, ~81%. The true compounding rate is exactly what the 30-day proof is built to measure.
01Agentic factor-investing returns — 58.8% vs 30.9% annualized, long-short backtest. "Beyond Prompting: An Autonomous Framework for Systematic Factor Investing via Agentic AI," 2026 · arXiv 2603.14288.
02Live autonomous-agent inconsistency — +8.39% (US) / −1.23% (A-shares). "AI-Trader: Benchmarking Autonomous Agents in Real-Time Financial Markets," Dec 2025 · arXiv 2512.10971.
03~17% average AI outperformance over traditional statistical models — agentic-trading industry overview, 2026 · wundertrading.
04Sojourner learning rate — domain expert in under 24 hours; ~4B parameters; continuous, compounding. Sojourner Cosmos business plan summary & investor deck, 2026.
05Human expertise benchmark — ~10,000 hours of deliberate practice. Ericsson et al.; popularized by Gladwell.
06AZX1 on-chain facts — 105M supply, single ERC-3643 class, Avalanche C-Chain. Verified via C-Chain RPC.