Every architecture is a hypothesis until it runs on live capital. This post is about what happened when ours did.
The setup
Lifecycle 01 is ETF creation and redemption, starting with a Bitcoin ETF. The mechanism is old and well understood. When an ETF trades above its net asset value, an authorized participant delivers the underlying basket, receives shares at NAV, and sells them at the premium. When it trades below, the participant buys shares at the discount, redeems them, and receives the basket. The gap between price and NAV is the yield, and it appears daily across roughly seventeen thousand ETFs holding about twenty-two trillion dollars.
We ran the full six-gate lifecycle. Signal agents read market data, NAV, flows, and borrow. Prediction agents estimated gap direction, size, and duration. Decision agents proposed sized, routed create-or-redeem actions. The council checked policy, caps, and hours. The execution layer routed each action to the venue and counterparty the routing layer scored best. The settle gate reconciled against the broker and wrote to the ledger.
The result
Ten consecutive live sessions, every one above the floor we set of ten basis points. Six sessions were in premium regimes, four in discount regimes. Every fill was independently confirmed by the broker before it was recorded.
I want to be precise about what this does and does not show. It is a proof of concept over a limited period. It is not a performance guarantee and we do not present it as one. What it does show is that the architecture holds under real market conditions in two different regimes, that the council rejects what policy says it should, that execution is idempotent under real network conditions, and that the ledger reconciles against an independent counterparty every time.
Where the cost goes
The economics of a traditional desk break down into four components. Cost of carry and counterparty risk together are more than eighty percent of the stack. Technology and execution are the next largest. Management and strategy are the remainder.
The swarm eliminates the first two. Agents hold no inventory between sessions and do not wait on a counterparty to settle before acting again, so both cost of carry and counterparty risk go to zero. Technology and execution costs fall by about two-thirds because the execution layer is shared infrastructure rather than per-desk build. Management and strategy fall by about half because the council does the supervisory work. That is where the figure of ninety percent less cost and risk than a human desk comes from.
What we learned about regimes
The most useful research finding from the ten sessions was about regime handling. Premium and discount regimes are not symmetric. Borrow cost matters in one and not the other. Venue depth behaves differently. The prediction agents that were trained on both regimes performed differently from those trained on one, and the post-trained model's exception handling was tested most severely at regime transitions. Those transitions are now the richest part of our training set.
Why we started this small
Matching a fund's price to what it holds is deliberately the smallest, lowest-risk lifecycle we could find. Position caps bound the downside. Counterparties confirm every action. The rules are public. It is the right place to prove the architecture because a failure teaches us something without hurting anyone. Trust is earned here first. Then it compounds into the lifecycles that come next.