Link models, add discovery tags

Add Space thumbnail

Add Probes H + I (large-alpha stress test and geometric report) + final synthesis

feat: /geometry endpoint — per-layer direction norms, residual projection, cross-layer cosine matrix

Add Tab 4 layer-scan/ablation results + base-prior interlude

feat: /steer_layer endpoint — layer scan + ablation + multi-pair difference-of-means

Add Tab 3 activation-patching results

feat: add /steer endpoint — real Qwen forward pass with L20 activation patching

docs: attractor characterisation (compliance-specific basin)

docs: refusal-axis deep dive (R1/R2/R3)

docs: add interpretability probe results (tab1 + tab2)

ui: bump max_new_tokens slider to 512 (default 128) — fits ZeroGPU 180s budget

zerogpu: load on CPU at import, .to(cuda) inside @spaces.GPU (first call)

zerogpu: slim requirements (trust base image torch/transformers/gradio/accelerate)

zerogpu: python 3.10.13 (3.11 not supported by spaces CUDA emulation), restore module-level .to(cuda)

zerogpu: python 3.10.13 (3.11 not supported by spaces CUDA emulation), restore module-level .to(cuda)

zerogpu: prefetch shards at import (CPU), load+to(cuda) inside @spaces.GPU

zerogpu: import spaces before torch/gradio to install fake-cuda patches

zerogpu: eager-load backbone at import (.to(cuda)) and shorten GPU slot to 120s

pin python 3.11 (gradio 4.44/pydub needs audioop, gone in py3.13)

initial: round-trip + arithmetic demo on ZeroGPU

initial commit