td_lang/examples/demo_td_loop.td

# demo_td_loop.td — The complete TD self-improvement pipeline
# This is what td_loop runs: merge, then iterate to get smarter

gate { must_pass = [canary, perplexity, thinking_mode] }
budget { max_gpu_hours = 50  max_cost = 200.00 }

data_contract {
    required_fields = [prompt, response]
    min_samples = 50
    max_perplexity = 50.0
}

reward_contract {
    verifiers = [code_compiles, math_correct]
    min_reward = 0.3
}

# Step 1: Load base model
load "Qwen/Qwen3-VL-8B-Instruct" as base

# Step 2: Fuse all donor models in one shot
fuse ["deepseek-ai/DeepSeek-R1", "MiMo-7B", "meta-llama/Llama-3.1-8B", "tiiuae/Falcon-H1R-7B"] into base

# Step 3: Heal the merge damage
heal base lora_r 32 epochs 2
snapshot base -> snapshots/

# Step 4: Self-improvement loop (the core of TD)
repeat 5 {
    diagnose base -> weaknesses.json
    synth base from base filter cherry_llm -> training_data.jsonl
    train base on "training_data.jsonl" using grpo steps 64 lr 5e-5
    eval base -> eval_results.json

    if eval_passed base {
        commit base
        snapshot base -> snapshots/
    } else {
        reset base to "snapshots/"
    }
}

# Step 5: Final report
report -> final_economics.json