Every system Meridian builds runs on the Cortex — a domain-agnostic cognitive architecture for autonomous operation.
The Cortex — Meridian Labs
The Cortex is not a model. It's not a prompt chain. It's a five-layer cognitive loop that runs continuously — perceiving state, reasoning about priorities, executing within safety constraints, verifying its own outcomes, and modifying its own capabilities based on operational experience.
Every system Meridian builds runs on the Cortex. The architecture is domain-agnostic — it doesn't know about commerce or trading or research. It knows how to observe a complex environment, detect what matters, act decisively, confirm the results, and get better at all of the above.
The Cortex is the intellectual property. The systems are the proof.
Aggregates signals from APIs, databases, market feeds, system health monitors, and web intelligence into a unified state tensor. Implements data freshness gating — if a source returns errors or stale data, downstream reasoning is automatically constrained. No hallucinated inputs.
Three-stage reasoning pipeline. First: deterministic rule engine scans for critical violations — drawdown breaches, missing deliverables, system failures — zero latency, zero ambiguity. Second: pattern matcher identifies opportunities from market intelligence. Third: priority ranker scores every item by severity × urgency × estimated impact. Top items get resources. Everything else queues.
Actions are classified into two zones. Autonomous: protective measures execute immediately without human approval — disabling a dangerous position, pulling a broken product, restarting a failed process. Escalated: anything touching external reputation, capital deployment, or irreversible changes requires human authorization via structured decision prompts. The boundary between zones is defined in a guardrails file that the system cannot modify.
Every action is logged with predicted outcome and actual outcome. A verification layer confirms whether the action achieved its objective — not whether it executed, but whether it worked. Failed verifications trigger one automatic retry with modified approach, then escalate. Successful and failed action records are formatted as training pairs for the local model, creating a continuous improvement loop from operational experience.
Three tiers of self-improvement. Tier 1: autonomous parameter tuning — the system adjusts its own thresholds, timing, and search strategies based on observed performance. Tier 2: autonomous rule creation — when the system detects a recurring pattern, it codifies a new detection rule. Tier 3: structural proposals — changes to core architecture are drafted as diffs with rationale and submitted for human review. The system can expand its own capabilities but cannot weaken its own safety constraints.
