The pipeline.
Each version ships a usable fund. Value first, armor when there's something worth armoring, learning loop once there's history to learn from.
Replatform
Streamlit retires. A real web platform: browser-based Admin Console with role-based access, kill switches, and live Boardroom streaming — on an event-driven core with a tamper-evident, hash-chained audit store.
Prediction Swarm
A dedicated forecasting swarm: regime detection, time-series foundation models, and Bayesian estimators issue probabilistic predictions into a registry — then get scored against realized data. Calibration decides whose forecasts the fund trusts, and position size shrinks automatically as uncertainty rises.
Execution & Breadth
From one ticker to a screened universe. An ingestion swarm supervises every data feed, an uncertainty-driven screener decides which names earn a full boardroom, and a production-grade execution engine works orders — paper first, always through the gates.
Governance Layer
A compliance swarm audits every agent against the rulebook and files findings with citations. Humans steer any agent through an instruction console — and a dedicated risk agent scores every instruction against fund strategy before it dispatches. An analytics agent can surface any dataset any agent ever used.
LocusOS
The fund becomes a machine. An immutable, reproducible operating-system image runs every agent in its own micro-VM with kernel-level network and policy enforcement. Validated "golden rules" compile into an OS-level policy kernel that no agent — however clever, however compromised — can route around.
Fund-in-a-Box
The learning loop closes: a neural engine trains on the fund's own audit history to tune strategy weights, and a meta-research agent runs sandboxed experiments on the agents themselves. Multi-book support. Install the image, get a hedge fund.
Three swarms become six.
Specialists, not a monolith. Every swarm evolves — and every swarm is watched.
Ingestion
Agent-supervised data pipelines that detect drift, gaps, and staleness — filing incidents instead of failing silently.
Prediction
Probabilistic forecasts with credible intervals, scored daily against realized data. Skill earns trust; decay loses it.
Research
Desks, debates, and adjudicated house views — consuming predictions weighted by the predictor's live track record.
Execution
Deliberately not an LLM: a deterministic engine that only accepts tickets that clear every gate and every policy.
Compliance
Continuous rule checks on every action, plus auditors that replay agent histories against regulation — with citations.
Neural Engine & Meta-Research
The audit trail becomes training data. A research agent studies the agents themselves and A/B tests improvements in sandboxes.
Where humans stay.
We automate everything a fund does that machines do better — and we're explicit about what they don't.
| function | who runs it |
|---|---|
| Research, forecasting, backtesting, order flow, monitoring | AUTOMATED — gate-checked |
| Reconciliation & NAV | AUTOMATED |
| Exception resolution | HUMAN |
| Compliance judgment & regulatory sign-off | HUMAN |
| Capital allocation across strategies | HUMAN — machine proposes |
| Crisis discretion & kill switches | HUMAN — by design |
Follow the build.
Paper-trading preview today; each release above lands with a recorded, auditable run.