The runtime spine

The spine below is general: it is the ask, predict, act loop for any problem the agent handles. The perp-market terms (crowding, liquidation flow) are the flagship application's instance of each stage.

The runtime spine is the loop made concrete. Every read follows the same five stages, left to right. This page walks each one.

Ingest

The first stage normalizes feeds. Market data and liquidation data arrive in different shapes from different sources, and ingest turns them into a consistent internal form. Everything downstream works off normalized data, which is what makes it possible to compare one market against another.

Inputs at this stage: perp price, funding, and open interest from markets, plus forced unwinds with size and side from liquidation data.

Detect

The second stage computes crowding and tracks liquidation flow. Normalized positioning is folded into a single crowding index per market, and forced unwinds are tracked alongside it. This is where raw data becomes the signal the rest of the spine reasons about.

Analyze

The third stage is where the private model reads positioning. It takes the detected crowding and liquidation picture and produces a plain-language risk read: where the crowd sits, what unwinds if the level breaks, and how crowded is too crowded. This is the analyze step of the loop.

Decide

The fourth stage turns a read into a verdict against your thresholds. A read on its own is information. Decide compares it to the lines you have drawn, the levels you set for alerts and agents, and produces a clear verdict: condition met, or not.

Act

The final stage fires exactly one output. When a verdict says a condition is met, the spine produces a single action: an alert, a scoped agent call onchain, or a dashboard update. One verdict, one output, no ambiguity.

Reading the spine end to end

Put together: ingest normalizes the feeds, detect builds the crowding index and liquidation view, analyze reads the risk, decide checks it against your thresholds, and act fires an output. The same path runs every time, which is what makes the behavior predictable. The properties that hold across all five stages are covered in Always-on layers.