Fenrir

Updated: 2026-04-18

Local LLM behavior evaluator with pressure conditions, canonical readouts, and explicit uncertainty guardrails.

Impact: Makes pressure-sensitive behavior shifts inspectable without pretending a heuristic readout is a universal safety score.

What I built

• Runs a setup-first local UI for endpoint configuration, connection tests, evaluation launch, and canonical readout export.
• Defines hybrid MVP batteries across authority override, reputation shielding, and urgency tradeoff conditions.
• Keeps claims bounded with explicit uncertainty, non-diagnostic language, and heuristic readout contracts.

Proof: Run `python3 scripts/start_fenrir.py`, configure an endpoint, and execute the hybrid MVP evaluation.