Chat: use the 16-bit v0.5 build (4-bit rounding breaks v0.5); knowledge base on chips without 16-bit

Chat: treat a few stray words as a garbled answer and use the knowledge base

Chat uses fine-tuned model v0.5 (0.5B, about 300 MB, knows the visitor's device)

Chat: stop the AI on looping text or when too slow, answer from the knowledge base

Chat: show the detected device and tell the AI about it; download progress inside the message box

Security assistant uses the fine-tuned Devseis model v0.4 (1.5B)

Flag repeated-text answers as not grounded in chat

Rename to Arx: links point to spaces/Devseis/Arx

Rename to Arx (display name) with a new fortress icon; knowledge base: Who is Arx

Chat: instant answers from the knowledge base, AI download in the background with consent, auto-start when stored, rewrite with AI

Desktop app 0.4.0 (Ask the assistant about this report), audit-aware chat search, v0.4 pipeline, converter for 1.5B

Security assistant chat as the front page (Nox-style, WebLLM on device, grounded in the Devseis knowledge base, Run audit from chat); phone check moved to phone.html

Use the fine-tuned auditor model v0.3

Download page for non-technical users: Mac / Windows / Linux tabs, step-by-step with pictures of each message, FAQ; Windows and Linux builds

Mac app 0.3.0: download page, desktop-aware front page, model version from the Space, user settings folder, icon, smoke test

Use fine-tuned model v0.1 (WebLLM, constrained JSON decoding, q0f16 on desktop / q4 on phones); MLC converter; v0.3 pipeline

v0.3: phone check (iPhone, Android) as the Space front page, catalog 0.3 (74 checks), shared vulnerability rules, real-CVE training data

Offline vulnerability matching: inventory in collectors, signed OSV/NVD/KEV/EPSS bundle builder, matcher in the app, tests

Collectors 0.2: user account type, USB device access, pending updates, outdated apps, software inventory; controls document

Desktop app (Electron + WebLLM), collector progress output, offline LoRA training script