Demo App

The CRP demo app is a fully interactive showcase of all 9 dispatch strategies, with metrics comparison, provider auto-detection, and live benchmarking.

Quick Start

# Run with auto-detected provider
python examples/demo_app/demo.py

# Or specify a provider
python examples/demo_app/demo.py --provider ollama --model qwen3-4b

What It Does

The demo app provides four modes:

Mode 1: Quick Demo

Runs a single dispatch with your chosen provider and displays the result with quality metrics (tier, facts extracted, envelope saturation, windows used).

Mode 2: Strategy Comparison

Runs ALL 9 dispatch strategies side-by-side on the same task and produces a comparison table:

Strategy	Time	Words	Quality	Facts	Windows
dispatch	2.1s	592	B	12	1
dispatch_with_continuation	18.4s	6,993	A	89	9
dispatch_reflexive	3.2s	645	A	15	2
dispatch_progressive	5.1s	1,203	A	28	3
...	...	...	...	...	...

This is the clearest way to see what CRP's continuation engine does — same model, same task, 12x more output with continuation.

Mode 3: Interactive

Opens a prompt where you type your own tasks and see results. Good for exploring how different prompts produce different quality tiers.

Mode 4: Full Benchmark

Runs a comprehensive benchmark across multiple tasks and produces detailed statistics.

The 9 Dispatch Strategies

The demo exercises every strategy CRP offers:

#	Strategy	What It Tests
1	dispatch()	Basic single-window generation
2	dispatch_with_tools()	Tool-augmented generation
3	dispatch_reflexive()	Self-critique → re-generation
4	dispatch_progressive()	Multi-pass progressive deepening
5	dispatch_stream()	Streaming token-by-token output
6	dispatch_agentic()	Multi-step agent with tool use
7	dispatch_stream_augmented()	Streaming with real-time extraction
8	dispatch() with continuation	Full continuation engine (multiple windows)
9	dispatch_with_continuation()	Explicit continuation dispatch

Provider Auto-Detection

The demo automatically detects available providers:

1. Ollama — checks http://localhost:11434
2. LM Studio — checks http://localhost:1234
3. OpenAI — checks OPENAI_API_KEY env var
4. Anthropic — checks ANTHROPIC_API_KEY env var
5. Mock — always available as fallback

If multiple providers are available, you choose from a menu.

Reading the Output

Quality Tiers

Tier	Meaning
S	Exceptional — all facts grounded, no hallucinations
A	Strong — minor gaps only
B	Acceptable — some parametric claims
C	Marginal — significant gaps
D	Poor — mostly ungrounded

Key Metrics

• Facts extracted: Number of facts pulled from the output by the extraction pipeline
• Envelope saturation: How full the context envelope is (0.0–1.0)
• Windows used: How many continuation windows were needed
• Overhead ratio: CRP protocol overhead vs. useful output

Command-Line Options

python examples/demo_app/demo.py [OPTIONS]

Options:
  --provider TEXT    Provider name (ollama, openai, anthropic, lm-studio)
  --model TEXT       Model name/path
  --base-url TEXT    Custom API base URL
  --api-key TEXT     API key for cloud providers
  --mode TEXT        Demo mode (quick, compare, interactive, benchmark)
  --task TEXT        Custom task description

Example: Comparing With LM Studio

# Start LM Studio with qwen3-4b loaded
# Then run strategy comparison:
python examples/demo_app/demo.py \
  --provider lm-studio \
  --model qwen3-4b \
  --mode compare \
  --task "Write a comprehensive guide to Python async programming"

This shows the dramatic difference between a single dispatch() call (truncated at ~600 words) and dispatch_with_continuation() (complete guide at ~7,000 words across 9 windows).