Autonomous research system for measurable, computer-executable research.
Python
Adversarial Tests
These tests encode long-lived architecture contracts at high-risk product boundaries.
Rules for this directory:
- Prefer synthetic task/plugin/run fixtures over source-text checks.
- Exercise public loaders, verifiers, tool handlers, or stable boundary APIs.
- Avoid exact model names, SAM-specific prose, current docs wording, and implementation sequence labels unless they are the behavior under test.
- Keep tests deterministic, local, and network-free.
- When a test fails against the current system, treat it as a production contract gap, not a reason to weaken the test.
Focused run:
python -m unittest tests.adversarial