KL Exec Gateway

A small execution layer that lets you see and govern what an LLM actually does.

Most LLM calls are a black box: you send a prompt, something happens in between, and a result appears.

You don't see the steps, the decisions, or the filters.

KL Exec Gateway makes the entire process visible.

It runs each request through a deterministic pipeline:

LLM → Policy → Sanitization → Formatting → Trace

Every step recorded.
Local. Reproducible. Inspectable.

---

Try it in 10 seconds

pip install kl-exec-gateway
kl-gateway-web
# On first launch, the Browser will prompt you to enter your OpenAI API key.

Opens automatically at http://localhost:8787

You immediately see:

• Live pipeline execution with animation
• what you sent to the model
• what the model returned
• which rules were applied
• what was removed or transformed
• the final output
• the full step-by-step trace

Alternative: CLI mode

kl-gateway --key "sk-..."

Interactive terminal chat with full trace logging.

---

Examples

1. Allowed

A normal request flows through the entire pipeline:

hello

Result:

LLM → policy → sanitize → format → done

2. Denied after LLM (length limit)

tell me a love story

This usually produces a long answer that exceeds the default 500-character length policy:

LLM → policy (DENY_LENGTH) → done

3. Denied before LLM (forbidden pattern)

If a request contains a forbidden pattern (configured in the policy engine), it is blocked before the model is even called:

my secret code

Result:

policy (DENY_PATTERN) → done

The LLM step is skipped entirely.

---

Use cases

• enforce policies on LLM output
• remove or mask sensitive data
• analyse model behaviour
• build safe internal tools
• reproduce responses
• explain decisions to auditors or teams

Simple building blocks. All deterministic (except the LLM call).

---

Configuration

Default limits: 500 characters, basic pattern blocking.

To adjust:

• Edit policy.config.json (policy limits)
• Edit pipeline.config.json (pipeline steps, logging, trace)
• Use templates: configs/production.config.json, configs/compliance-gdpr.config.json

---

Documentation

• docs/ARCHITECTURE.md – how the pipeline works
• docs/USAGE.md – examples and recipes
• docs/THEORY_ALIGNMENT.md – how it maps to KL Execution Theory

---

About the KL Execution Model

This project demonstrates how deterministic execution, policy evaluation and step-level traces can be composed into a practical LLM gateway.

It uses KL Kernel Logic as its underlying execution substrate.

For details on how the gateway maps to the execution axioms, see docs/THEORY_ALIGNMENT.md.

---

License

MIT