New approach: formal academic with subtle imperfections, no casual overcorrection

Fix: Remove suggested transitions (they're detectable too), let model write naturally

Final prompt: Turnitin-focused, no dashes/shapes, temp 0.6, Gemma user-role format

Fix: Use Gemma native format - NO system role, instructions in user turn, proper control tokens

Revised: Gemma-optimized prompt format, concise prose instructions, strict length matching

Revised prompt: match input length exactly, no over-generation, stronger anti-AI"

Improved system prompt - no dashes, academic style, better detection bypass

Switch to HF Inference API for Gemma 4 27B - instant speed, free tier credits

Switch to Q4_0 - simplest/fastest dequant on CPU, revert from slow Q3_K_M

Switch to Q3_K_M for faster generation speed on CPU

Fix: Q3_K_M for faster generation, n_batch=256, auto-detect threads

Fix TTFT: cache system prompt, use smaller quant Q8_0 for speed, reduce n_ctx

Speed + concurrency improvements: streaming, queue, batch size, n_batch, flash attention"

Remove temperature, fix sweet spot for academic AI removal, tuned for Turnitin evasion"

Fix: add streaming, improve prompt to ban slang/bad grammar, keep it clean and fast"

Revert to original simple form"

Full UI overhaul: all modes, polished theme, streaming, word counts, gauge, sentence highlighting"

Add app.py - Gemma 4 E4B humanizer + RoBERTa detector, API for React"