Outcry, a four-layer activist AI, runs on older phones without network calls by stacking quantization, domain adaptation, steering vectors, and a learned prefix

Outcry combines a quantized open-weights base model, a QLoRA adapter trained on activist literature and scrubbed user conversations, contrastive activation steering (with user-adjustable axes for radicalism, persona, and identity), and a soft-prompt wellbeing layer — all fitting in ~3 GB of RAM on Apple MLX.

The stack layers operate independently: the base model is quantized aggressively; the LoRA adapter specializes it for movement work without fusing into base weights; steering vectors inject at a single mid-stack residual layer and compose additively; and the soft-prompt prefix (eight virtual tokens trained against the AI-Wellbeing objective) ships in the next iOS release and is already in final testing.

The system prompt is precomputed as a key/value cache at build time and shipped in the app bundle, so the user's first message becomes the first token the model actually computes attention over, eliminating the need to process the system prompt at conversation start.