Skip to content
OSA-Midnight Oil

Bundled Models

ROS ships with plain-language model profiles so regular operators do not need to run terminal commands, copy model IDs, or understand provider endpoints. The model cards appear inside the Model Workspace, and ROS handles the local preparation flow from the app.

Included model profiles

The first bundled profiles are:

  • DeepNimSec / Security Model v1 (DNS-v1) — defensive review for project memory, command captures, evidence, risk summaries, and safer next steps.
  • Citizen-AI / Project Blueprint — defensive training, bias recognition, verification coaching, lab scenario planning, and after-action summaries.

These are presented as normal product features, not as developer infrastructure. A user should be able to choose a model card, click Prepare model, and let ROS handle the rest.

Prepare model

From the Model Workspace:

  1. Open Model Workspace.
  2. Choose Security Model v1 or Citizen-AI.
  3. Click Prepare model.
  4. Wait for the status to become Ready.
  5. Use the action chips, such as Review risk or Summarize evidence.

If a model cannot be prepared, ROS should say so in plain language and keep manual capture available. The error details belong in Developer setup, not in the primary workflow.

What gets bundled

The ROS app bundle includes:

  • model cards
  • prompt contracts
  • safety boundaries
  • deterministic project-memory context builders
  • local runtime adapters
  • one-click preparation controls

Large model weights may be staged, downloaded, or installed by the app depending on platform and release channel. The important user promise is that operators do not need terminal commands to use the model workflow.

Developer setup

Developer setup exists for AI builders and advanced operators who want to inspect or override the local runtime:

  • provider/runtime
  • local endpoint
  • technical model name
  • raw runtime status
  • request context preview
  • install or preparation logs

This drawer is optional. It is not the normal user path.

Local-first model rule

Model output is saved as project memory, never executed as a command. DNS-v1 and Citizen-AI can review, explain, summarize, and suggest safer next steps, but they do not run shell commands on behalf of the user.