How it works.

01The separation

The fundamental architectural decision in STRATA is the separation between AI reasoning and deterministic execution. Everything else follows from this.

02A single cycle

Each trading unit runs on a dynamic schedule and executes the following sequence. One cycle for IB Equity illustrated, though the pattern holds for HL Equity and V2 Crypto.

03What's in the payload

Every cycle rebuilds the payload from scratch. REI Core has persistent knowledge — the payload provides live data, not re-teaching. For IB Equity:

04The teaching loop

Beyond trading cycles, STRATA maintains a dedicated teaching infrastructure that sends learning interactions to each REI unit. This is where the compounding flywheel runs.

05REI Core — not an LLM

REI Core 0.5a is developed by REI Labs. STRATA is built on it. Core is a hypergraph traversal engine with evolutionary computation — these properties are core to REI, not STRATA:

• The LLM is the articulation layer, not the intelligence. It translates multi-dimensional reasoning into linear text.
• Knowledge is encoded as interconnected concepts in a hypergraph — nodes, edges, hypertags, derived relationships.
• Per-unit evolution via genetic algorithm. Each unit develops its own cognitive phenotype over runtime.
• Persistent learning survives sessions. Reinforced knowledge strengthens; unreinforced knowledge fades.
• Hybrid recall: semantic similarity plus structural graph traversal. Fading semantic knowledge can still be surfaced through hypertags.
• Native Float64 processing. Numbers stay numbers, not tokens.
• Deterministic output. Given the same payload, response is reproducible.

Each unit reasons over REI Core on a dedicated, authenticated channel. Every request is self-contained: all context for the decision is included, responses are deterministic, and isolation between units is cryptographically enforced.

06The critical rules

Six rules that govern all interaction with REI Core. These are not guidelines — violating them causes measurable degradation in reasoning quality.

1. Never feed LLM-articulated responses back into the system. Core's reasoning is multi-dimensional; LLM articulation flattens it to linear text. That transformation is lossy and one-directional.
2. API requests must be self-contained. Each request includes all necessary context in the user message. Don't simulate conversation.
3. Teach principles, not instances. "Array variables should be plural" (strong) vs "Rename this to userList" (weak). Principles create inference paths.
4. Explicit correction is the strongest learning signal. "That's wrong because X" directly adjusts the pathway that produced the error.
5. Implicit validation reinforces. Accepting a response without modification signals approval.
6. Primordials are permanent. "Remember this:" and "Never forget:" triggers create immutable knowledge anchors that survive decay.

07News and event pipeline

The News Unit (NRU) does not directly call trading units. It operates upstream: it ingests real-time streams from Synoptic and Telegram, classifies urgency and sentiment, and emits a digest buffer that the Analyst consumes every 30 minutes.

Synoptic WebSocket → NRU (classify) → Analyst Digest Buffer
                                         (in-memory ring)

Telegram breaking    → NRU (urgent) → EAU cycle (90s cooldown)
                                         → Urgent Interrupt to Analyst
                                         (if matches invalidation watchlist)

Analyst cycle → reads digest · synthesizes with theses · writes to Redis
Trading units → read Analyst output from Redis on next cycle

The Event Anticipation Unit (EAU) runs its own cycle every 30 minutes and identifies upcoming binary events — rate decisions, NFP, CPI, earnings, geopolitical catalysts. It produces a risk overlay that trading units must respect:

interface EauRiskOverlay {
  avoidNewLongs: boolean;     // block new long entries
  avoidNewShorts: boolean;    // block new short entries
  maxSizingPct: number;       // cap size (e.g., 50% of normal)
  scenarios: EauScenario[];   // modeled outcomes with probabilities
  reasoning: string;
}

08Scheduling and cadence

Dynamic intervals respond to: market session, position floor (minimum interval when positions are open), AI-recommended next-check interval, performance state (CAUTIOUS or DEFENSIVE shortens intervals), recent position closes, and news or research triggers. A weekend gate suppresses trading cycles Saturday through Sunday evening; ingestion continues.

For the full architecture reference including deterministic governance detail, see architecture. For the unit inventory, see units. For risk controls specifically, see diligence.