Concepts

What is Lexicor

Lexicor is a governed transformation layer for multi-agent systems. It represents meaning canonically first; transport codecs and language renderers are secondary. This page defines the model and the vocabulary the rest of the docs assume.

The problem

Agent pipelines serialize intent, context, and intermediate state through verbose natural language. That drives compounding costs at scale:

• High token cost on every retransmission across agent-to-agent handoffs.
• Repeated ambiguity resolution — each receiving agent re-parses intent from prose.
• Brittle handoffs where prompt bloat and format drift cause silent failures.
• No fidelity guarantees when compression quietly drops constraints or relationships.

The cost lever is internal context: handoffs and repeated state across multi-step workflows. Savings concentrate on hops 2 and beyond, where a compact packet replaces re-sending the full directive.

The pipeline

A directive flows through a fixed sequence of boundaries. Each stage is observable and each transition is explicit:

• Ingress — adapters turn external inputs (narrow English, tool JSON) into an IntentGraph at the edge.
• Normalize — deterministic ordering, defaults, and field alignment so two graphs that mean the same thing compare reliably.
• Encode — the line codec turns a graph into compact lines using internal symbol ids and a versioned symbol table.
• Packet — codec output is wrapped in a typed, versioned transport envelope with metadata.
• Decode & render — a receiver decodes the packet back to a graph and renders it to natural language, tool-shaped JSON, or a protocol payload.

Core concepts

IntentGraph

The canonical, language-neutral JSON graph of entities, predicates, and constraints, validated against JSON Schema. It is not raw embeddings or free text — it is a discrete structure that tools, audit, and equivalence checks can reason about. A graph_version (semver) identifies the IR schema used to produce it.

Transport packets

A typed, versioned envelope wrapping codec output plus metadata — codec version, IR version, symbol-table version, payload mode, and fidelity class. Packets are embeddable payloads: they ride inside whatever carrier protocol you already use.

• Fidelity class — the guarantee level for a packet (for example semantic_lossless or semantic_compact). Consumers check this before acting on payload content.
• Payload mode — the transport intent of a packet (for example semantic_transport or execution_ready).

Symbol tables

A versioned map from internal ids to renderable symbols (for example core-sym-1). Encoders and decoders must agree on the table version. Unknown or incompatible tables fail closed — receivers must not guess.

Fail-closed

On a version or policy mismatch, Lexicor rejects with an explicit error instead of silently decoding or coercing. An unknown codec or symbol-table version is a hard error at decode time, not a best-effort guess. This is what makes handoffs auditable rather than merely lossy.

Semantic equivalence

Correctness is checked by scoring and diffing graphs — entities, predicates, and constraints — after a transform, not by string-comparing prose or JSON blobs. The round-trip normalize → encode → decode → equivalence path can be held to a repeatable bar so you can prove meaning survived the trip.

When it fits

Lexicor is built for teams that own a multi-hop pipeline:

• Platforms and orchestration layers standardizing handoff formats across models and tools.
• Vendors embedding a governance-aware semantic layer between agents and execution.

It is not the right fit for single-shot chat wrappers with no multi-hop state, or for teams that only need a fixed API schema with no ambiguity handling or equivalence testing.

Next: run the Quickstart, or jump to the API reference.