Most AI trading systems are LLM wrappers with prompt libraries. STRATA is built on REI Core 0.5a, a hypergraph-reasoning engine — and wraps it in a deterministic governance layer. The difference is structural, not incremental.
Three assumptions we don't make:
Most AI trading systems sit on top of GPT-4, Claude, or Gemini. They make stateless calls. Each request begins without memory. Prompt engineering becomes the entire craft, and the "model" can be swapped by any competitor with a subscription.
STRATA is built on REI Core 0.5a, developed by REI Labs. Core is a hypergraph traversal engine, not an LLM — the LLM is only the articulation layer, translating Core's multi-dimensional reasoning into linear text. The intelligence lives in the graph. Ecliptica does not build the engine; we build the multi-asset product, the governance layer, and the integration surface around it.
| Capability | Standard LLM-based system | STRATA on REI Core 0.5a |
|---|---|---|
| State | Stateless per call; context stuffed each time | Persistent hypergraph; learning survives sessions |
| Learning | Fine-tuning runs, or none | Primordials, corrections, research teaching — continuous |
| Reasoning | Linear token prediction, stochastic | Hypergraph traversal + genetic algorithm, deterministic |
| Numerics | Tokenized floats, lossy | Native Float64 throughout |
| Specialization | One general model for all tasks | Per-unit evolution; each unit diverges into its domain |
| Moat | Rented model access; swappable by competitors | Accumulated graph state; unique per unit, non-transferable |
Research is commoditized. Every serious quant shop has news feeds, Benzinga, Polygon, and increasingly, open LLM reasoning. What is not commoditized is the infrastructure that decides whether a trade should actually execute, how large it should be, what time of session, after how much chasing, with what stop discipline, and what profit-taking ladder.
STRATA enforces ten separate deterministic governance layers. Each is pure math — no AI calls, no cycle dependencies. Each can run in OFF, ADVISORY, or ENFORCING mode, which lets us paper-test new rules before they gate real capital.
Recommended posture (FULL_DEPLOYMENT, TACTICAL_ONLY, REDUCE, NO_NEW_RISK) is a hard filter before any trade is evaluated.
Max gross, max per-sector, max per-strategy-bucket. Proposed trades are evaluated against the full book, not in isolation.
Velocity, range position, breakout-vs-retest. Extended prices trigger CONVERT_PULLBACK or DELAY_RETEST or REJECT_EXTENDED.
Rungs by asset class, progressive stop tightening. Never relies on AI to protect a win — math handles partial profits.
Session-boundary tightening, shock-move detection, profit giveback, event invalidation. Independent of AI.
After a stop-out, the system enters a cooldown and waits for thesis reconfirmation before re-entering at half size, then full.
If no inputs have materially changed since last cycle, the system suppresses repeat actions. Eliminates churn.
Independent of trading cycles. Shock-move detection, cascade protection, extended-drawdown monitoring, news invalidation.
Additional layers: Exposure Graph for book-level what-if analysis, and Post-Trade Forensics which labels every closed trade as thesis failure, expression failure, entry timing, stop placement, profit harvesting, or premature exit. Those labels feed back to Analyst and trading units for the idea-quality-vs-execution-quality separation.
Static systems decay as market regimes shift. STRATA's edge is designed to compound. Four mechanisms drive this:
Each unit receives domain-specific invariant rules that become immutable inference nodes in its hypergraph. These never decay. They anchor reasoning across all future cycles.
Every detected error generates an explicit correction. Per REI docs, correction directly adjusts the reasoning pathway that produced the mistake. Auto-corrections fire for invariant violations.
Every research document (PDF, article, note) is extracted by Kimi OCR, parsed by REI, taught to the relevant unit as thesis + tickers + themes + catalysts + risks. Multiple independent sources raise conviction on the shared theme.
Closed trades are scored along six failure modes. Units and Analyst see whether the thesis was wrong, the instrument was wrong, the timing was wrong, or the sizing was wrong.
The serious question is: what else is out there? We frame competitors in three buckets:
| Category | Examples | Gap vs STRATA |
|---|---|---|
| LLM wrappers | Most "AI trading" projects launched since 2023 | Stateless, no governance layer, no persistent learning |
| Crypto-only bots | Hummingbot derivatives, algo desks, signal bots | Single-asset, no research pipeline, no cross-asset thesis |
| Institutional quants | Two Sigma, DE Shaw, Renaissance | Closed systems, factor-model based, do not do AI-generated thesis reasoning |
| Retail "AI trading" apps | Various consumer products | No real autonomy, no governance, no institutional posture |
We do not compete with Two Sigma on co-located HFT. We do not compete with retail products. We compete in the narrow band where an institutional allocator wants multi-asset AI-generated thesis trading with institutional-grade governance — and that band is currently empty.
When a fund allocates to STRATA, they are buying four things, in decreasing order of substitutability:
The first three are engineering. The fourth is the moat, and the moat gets deeper every cycle.
For the technical walkthrough of how this is implemented, continue to how it works. For operational rigor, risk controls, and LP diligence detail, continue to diligence.
STRATA is in selective deployment with institutional counterparties. For allocator diligence, partnership inquiries, or licensing discussions, contact Ecliptica directly.