Fetching metadata from the HF Docker repository...
Jun 12 17:35 feat(lora): train-only SFT experiment scaffold (cookbook recipe, ms-swift) Auenchanters +1: Claude Fable 5 Production-shaped SFT targets (filled explanations from gold labels),
swift-format converter, Modal LoRA trainer (frozen ViT, A10G, adapter
persisted to volume). NEVER deployed pre-judging - measurement only.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> 17:31 feat(loop): agent-trace draft (consent-gated), demo kit, robustness v2 Auenchanters +1: Claude Fable 5 Items 8+9+10 scaffolding: trace bundle drafted for owner review (NOT
published), shot-by-shot demo script per field-guide beats, graded
degradations (blur/dim/rotate/perspective/shadow/noise x3 levels) with
outcome-based scoring on Modal (honest refusal passes, wrong-low fails).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> 11:26 docs: non-Latin OCR findings — ja scams fixed to warning, hi documented Auenchanters +1: Claude Fable 5 Prompt-fix probe simulated through full triage: ja gift-card scams 3/3
warning (original-script quotes + multilingual heuristics), ja bills
stay low; Devanagari remains under-read (1.3B limit) -> honest README
limitation. sim_probe_levels.py re-scores probe dumps offline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> 06:24 feat(training): letter forge + labeled dataset + Modal prompt-A/B eval Auenchanters +1: Claude Fable 5 guardian/letterforge.py: 8 parameterized letter factories with gold
labels derived through triage itself; scripts/make_dataset.py renders
images + labels.jsonl (+SFT-ready targets); modal_eval.py scores level/
doc-type accuracy, scam recall, false alarms, amount/deadline hits,
latency and output size for the production vs compact prompt on A10G.
Also: deep-link/easter-egg retry budget 7.5s -> 30s (Space mounts
slower than local; observed live).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Jun 11 13:59 feat(ui): ?lang= deep-link via head script + UI screenshot set Auenchanters +1: Claude Fable 5 launch(js=) is ignored on load in Gradio 6.17.3 (verified), so the
deep-link rides in head=. ?autorun is GUARDIAN_MOCK-gated: it can never
fire a GPU run on the Space. scripts/shoot_ui.py = CDP screenshot driver
(dev-only). docs/ui: en/hi/ja landing+result, mobile, loader captures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> 13:22 feat(ui): one-tap example letters shared with the Modal harness Auenchanters +1: Claude Fable 5 guardian/samples.py is the single source for synthetic letters;
modal_validate.py imports it; gr.Examples(cache_examples=False) so no
GPU runs at build time. Browser-verified example->analyze flow.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Jun 10 17:59 Fix crg.ps1 false failure on PS 5.1 stderr wrapping Auenchanters +1: Claude Fable 5 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> 17:58 Mystery-Mail Guardian: complete app for the Build Small Hackathon Auenchanters +1: Claude Fable 5 Privacy-first scam-and-action triage for confusing mail, fully local:
MiniCPM-V 4.6 (1.3B, central engine) + VoxCPM2 (2B TTS) = 3.3B params
deployed (32B cap); optional MiniCPM4.1-8B GGUF via llama.cpp (full
config, 11.3B). Modular pipeline (extract/triage/explain/speak), safety
layer enforced in code, trilingual accessible UI (en/hi/es), ZeroGPU
request-driven wiring, mock mode, 66 offline tests, GPU isolation checks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> 
