Link models, add discovery tags

Add Space thumbnail

Upload app.py with huggingface_hub

ui: default to Completion+code (matches cached trace); clear textbox on switch to Chat, restore seed on switch back if empty

feat(ui): raise max_new_tokens cap from 200 to 512

feat(ui): chat mode shim (User:/Assistant: wrapper) so visitors can type natural questions; raw completion mode still available

docs(av-claim): plain-language rewrite of AV trust caveat for first-time readers

docs(av-claim): fix mis-attributed TruthfulQA reference and lead with best-of-64 rho_norm=0.92 (greedy 0.26 was both the wrong dataset and the worst-case number)

fix(click-to-pin): move srtPin/srtPinClear to gr.Blocks(js=...) so the symbols are defined on window before any onclick fires

demo: bake initial cached trace (quicksort prompt, 107 tok, 30 regime flips)

demo: regime badges, agg curve, entropy, click-to-pin, baseline toggle, prompt revisions, cached-trace loader

demo: regime badges, agg curve, entropy, click-to-pin, baseline toggle, prompt revisions, cached-trace loader

demo: regime badges, agg curve, entropy, click-to-pin, baseline toggle, prompt revisions, cached-trace loader

demo: regime badges, agg curve, entropy, click-to-pin, baseline toggle, prompt revisions, cached-trace loader

demo: regime badges, agg curve, entropy, click-to-pin, baseline toggle, prompt revisions, cached-trace loader

demo: regime badges, agg curve, entropy, click-to-pin, baseline toggle, prompt revisions, cached-trace loader

ux: diversify example prompts beyond Q/A framing

feat: per-layer divergence sparklines (L7/L14/L21)

feat: per-layer divergence sparklines (L7/L14/L21)

feat: per-layer divergence sparklines (L7/L14/L21)

fix: pin transformers==4.53.3 (Qwen2 RoPE broke in 4.55+)

fix: pin starlette<0.40 and simplify launch (requirements.txt)

fix: pin starlette<0.40 and simplify launch (app.py)

initial deploy: SRT introspect (Gradio + ZeroGPU shim)

initial commit