# Soap Lab — Agent Trace

## Example: Agentic Recipe Optimiser

**Input:** "300g coconut oil, 400g olive oil — optimise this recipe for me"

**Step 1 — Parse:** coconut oil: 300g, olive oil: 400g

**Step 2 — Score:**
- Cleansing: 27.9 (HIGH — target 12-22)
- INS: 103.5 (LOW — target 136-165)

**Step 3 — Iteration 1:**
- Cleansing too high → transfer 35g coconut → olive
- Recipe: coconut 265g, olive 435g

**Step 4 — Rescore:**
- Cleansing: 24.6 (still high, improved)
- INS: 97.6 (still low)

**Step 5 — AI Explanation:**
- Model: Qwen/Qwen2.5-3B-Instruct via Featherless AI
- Explains changes in plain English to user

## Architecture
- Chemistry engine: deterministic Python (no AI)
- AI layer: explanation only, never calculates
- Model: Qwen2.5-3B-Instruct (Tiny Titan eligible)