zlm - Zero Language Model

A test utility that simulates LLM output without burning tokens.

Purpose

Generate plausible-looking gibberish text with configurable delays to test Duso's concurrency model and worker swarms at scale without making actual LLM API calls.

Perfect for testing:

• Large-scale worker spawning (tested up to 100,000 concurrent workers)
• Datastore coordination patterns
• Swarm orchestration logic
• Latency-sensitive synchronization

Usage

zlm = require("zlm")

// Generate mock output with default params (200 tokens, 1.5s delay)
text = zlm.prompt()

// Custom parameters
text = zlm.prompt(tokens=500, delay=2.5)

Parameters

• tokens (default: 200) - Approximate number of tokens to generate (varies by ±20%)
• delay (default: 1.5) - Simulated processing time in seconds (varies by ±20%)
• id - Not currently used (reserved for future use)

How It Works

• Generates text using weighted character distribution that creates word-salad gibberish
• Adds natural variation to both token count and delay time
• Uses sleep() to simulate real processing latency
• Characters are weighted to produce roughly natural-looking output (common letters more frequent)

Example: Worker Swarm

zlm = require("zlm")
store = datastore("workers")

ctx = context()

if ctx == nil then
  // Main: spawn workers
  store.set("completed", 0)

  for i = 1, 1000 do
    spawn("swarm_worker.du")
  end

  // Wait for all to finish
  store.wait("completed", 1000)
  print("All workers done!")
end

// Worker code runs when ctx != nil
text = zlm.prompt(tokens=300, delay=1.0)
store.increment("completed", 1)
exit({text = text})

Notes

• Generates different gibberish each run (uses random)
• Character distribution roughly mimics English (more common letters weighted higher)
• Useful for load testing without API rate limits or costs