Galaxy S21 + Gemma 4 E2B Q3_K_M + Qwen3.5-0.8B Q4_K_M: same Amazon task, two different valid carts via structured browser perception
retained text
This continues the small-model browser experiments from the previous posts. Two standard, off-the-shelf models ran locally on the same Samsung Galaxy S21: * Gemma 4 E2B, Q3\_K\_M as the reasoner * Qwen3.5-0.8B, Q4\_K\_M as the executor No fine-tuning, no task-specific training, no adapters. A persistent ledger carried task state, constraints, verification and recovery. Browser perception and actions came through E2LLM/SiFR, developed by us, inside the user’s existing Firefox session. Task: prepare an Instax Mini 12 kit on Amazon under $200. **Run 1:** bundle route, 2 cart entries, **$190.98**. **Run 2:** separate products, budget violation on a $179.95 film choice, recovery to a cheaper alternative, 4 cart entries, **$143.86**. Both runs ended `READY_FOR_HUMAN_REVIEW`. No checkout or purchase. The attached video shows both full runs at **x16**. Original session: about **51 minutes**. **The ledger owns the task. Each model owns only the next bounded decision.** All evidence is public on GitHub in e2llm/edge-browser-agent, folder amazon-dual-slm-ledger: full raw log, exact script, launcher and SHA-256 hashes. Thoughts, objections, failure modes and alternative designs are welcome.
document fetched 2026-10-15 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5o4xn/galaxy_s21_gemma_4_e2b_q3_k_m_qwen3508b_q4_k_m