Dashboard

System & AI Settings

OpenRouter AI model priority, local isolated storage, and academic privacy diagnostics.

Active Execution Mode
Deterministic Demo Fixtures Active

Using verified deterministic fixtures for reproducible zero-token testing.

Demo Mode

AI Model Hierarchy & OpenRouter Benchmark Report

ExamLoop runs with a multi-tier fallback architecture prioritizing NVIDIA Nemotron models:

1st Priority: NVIDIA Ultra 550B
nvidia/nemotron-3-ultra-550b-a55b:free

550B parameter powerhouse. Tested across extraction, generation & rubric grading with 100% valid JSON adherence. Zero cost.

2nd Priority: NVIDIA Lightning
nvidia/nemotron-3.5-lightning:free

Fast secondary free tier fallback for quick subjective evaluation and question generation.

3rd Priority: Cohere Fallback
cohere/north-mini-code:free

Benchmark latency: 14.5s. Resilient free backup in case of upstream provider load.

Storage & Retrieval Diagnostics

Retrieval Modelexical_bm25

Strict subject-scoped lexical BM25 ranking (vector-ready)

Database Storage

SQLAlchemy with JSON support and migrations

Local File Storage

Subject-isolated safe storage adapter

Academic Privacy & Security Contract

  • Server-side ownership verification: Retrieval queries are strictly restricted to the authenticated student and subject.
  • Untrusted reference treatment: Uploaded notes and student answers cannot override examiner grading instructions or inject prompts.
  • Private document isolation: Uploaded materials are stored in isolated subject directories and excluded upon deletion.
  • No training on notes: Study materials are queried ephemerally for test generation and rubric grading only.