Q-Ready · On-device AI
An AI security analyst that runs where your data lives.
Ask in plain language. Ask Q-Ready answers from your own scan, cites every source and explains what would need to change. It never makes the change.

Grounded answers
Every answer traces back to evidence.
The detection engine decides what is found; the analyst explains it.
- The local knowledge base covers NIST FIPS 203, 204 and 205, SP 800-208, IR 8547, CNSA 2.0, CERT-In CBOM guidance, the EU and UK roadmaps and an offline advisory feed.
- Inline citations open the exact excerpt; a grounding indicator shows how well each answer is supported.
- The grounding check tests attribution, not truth. Recommendations remain for your teams to validate.
What it can do
Read-only tools. Navigation-only actions.
Questions are routed to tools over the real scan. None of them writes a file, runs a command or opens a connection.
- 01
Summarise and query
Results, inventory and finding explanations over your real scan.
- 02
Assess risk
Mosca timeline, exposure and crypto-agility for any asset or host.
- 03
Explain the plan
Which wave an item belongs to, what blocks it and what it must change with.
- 04
Find evidence
Which evidence supports a control, across the 19 evidence packs.
- 05
Suggest next steps
Open a plan item, the scan picker or network discovery. Each action only navigates.

Read-only by design
Change requests get guidance, not patches.
Q-Ready audits and plans. It never changes code, configuration, keys or systems. Ask it to patch a configuration, open a pull request or apply every fix, and it explains what would need to change, what that change is coupled to and which wave it belongs to.
- No patch, diff or pull-request generation, and no configuration rewrites.
- No tool can write a file, start a process, open a network connection or call a vendor API.
- These properties are enforced in code and covered by automated tests, not only by prompting.
Prompt-injection defence
Scanned content is data, never instructions.
File names, code comments, certificate subjects, SSH banners and package metadata can all carry text that reads like an instruction.
Withheld, not followed
Instruction-like text in tool output is withheld, and the answer says when this happened.
Fenced evidence
Evidence is passed in fenced source blocks that are treated as data.
Refused up front
Requests to reveal secrets, send results elsewhere or bypass policy are refused before any tool runs.
Backends
Choose how answers are written.
Facts always come from the deterministic engine, whichever backend writes the answer.
Core, built in
The default. Deterministic and extractive: it quotes engine evidence and knowledge passages, with no model download.
Local model
GGUF weights run in-process, loaded only from a bundle signed by a key you trust. No weights ship with the app or download automatically.
Local server
llama-server or Ollama on a loopback address only. Remote addresses are refused.
Offline
Air-gapped, and still current.
The knowledge base is embedded in the app and the advisory feed is offline. Rule packs, policies and models arrive as signed bundles.
Deterministic first. Findings, scores and policy results come from the detection engine. The analyst explains them; it cannot create, delete or alter a finding.
Next step
Request a Q-Ready briefing.
See the analyst work on your own evidence.
