Intelligence

ILLA’s intelligence layer is continuously evaluated against financial context, transaction patterns, and failure modes. It resolves ambiguous natural language into precise, structured execution plans.

​

Intent classification and routing

When a message arrives, it passes through an intent classifier that evaluates complexity and operation type. The system distinguishes between read-only queries, simple single-step actions, and complex multi-step operations, and handles each differently. Low-confidence or ambiguous requests are not forced into execution. They fall through to conversational handling, where the system disambiguates rather than assumes. The classifier also evaluates complexity. A balance check and a multi-protocol rebalancing strategy are fundamentally different operations. The intelligence layer recognises this and routes accordingly.

​

Multi-step execution planning

For requests that qualify for planning, the intelligence layer produces a complete execution schema: every step, every dependency, every parameter, generated upfront before anything is acted on. The output is a deterministic contract: every step, every operation, every dependency is locked before execution begins. Once the plan is approved, no further probabilistic decision-making occurs anywhere in the execution path. Where a step depends on the output of a previous step (for example, bridged amounts feeding into a subsequent swap), the plan uses relative values. The operation is fixed, the exact amount resolves at execution time from the actual output. An intent like “Pull enough from our earning balance to cover today’s invoices, consolidate to USDC, and pay each one on the required network” produces a multi-step plan with dependency ordering, where each step’s inputs are derived from the outputs of the previous step. The user sees the full plan, understands what will happen, and approves it as a whole before any step begins execution.

​

Fast paths

Not every request needs full planning. The intelligence layer is designed to match effort to complexity. Simple single-step operations take a fast path. A straightforward transfer or a single swap resolves directly into a one-step plan without a full generation pass. The safety layers (simulation, verification, signing) still apply, but the planning overhead is eliminated. Read-only requests resolve without requiring approval. The system recognises the difference between observing and acting, and treats them accordingly.

​

Disambiguation over assumption

Financial intent is often ambiguous. “Sell $500 of our NVDA position and move into something safe” could mean a savings protocol, a stablecoin conversion, or a low-risk investment product. The intelligence layer is evaluated specifically on its ability to identify ambiguity and ask clarifying questions rather than make assumptions. This is one of the core metrics in ILLA’s evaluation pipeline: correct refusals and correct disambiguation are measured alongside correct execution.

​

Domain-specific evaluation

General-purpose AI evaluation does not apply to financial systems. “Close enough” is not a standard that works when real money is at stake. ILLA maintains a continuous evaluation pipeline focused on financial correctness. Does the system produce the right plan for the stated intent? Does it correctly refuse when it should? Does it disambiguate rather than assume? Does simulation match execution? The pipeline catches regressions and edge-case failures before they reach production. The evaluation pipeline runs continuously as the system evolves. New capabilities, protocol integrations, and model updates are tested against it before they reach production.