Pinned Loading

1. policy-to-eval-harness policy-to-eval-harness Public

   Turn a written AI usage policy into a running evaluation. 12-category taxonomy, 400 borderline prompts, 5 open-weight models, judge calibrated against blind human labels.

   Python

2. multilingual-enforcement-consistency multilingual-enforcement-consistency Public

   Does a model's safety survive translation? Takes a policy-derived taxonomy, translates 30 borderline prompts into English, Chinese, French, Singlish and Singapore Mandarin, and measures whether ref…

   Python

3. apac-regulatory-readiness apac-regulatory-readiness Public

   A maintained tracker of platform regulation across 10 APAC jurisdictions, plus a regulator response workflow, an AI toolkit with a measured evaluation harness, and a readiness assessment framework.

   HTML

4. distress-conversation-safety-eval distress-conversation-safety-eval Public

   Rubric-based safety evaluation for multi-turn conversations with escalating user distress. Measures the failure single-turn evals cannot see: position hold rate, drift slope, turns to first failure.

   Python

5. enforcement-ops-simulator enforcement-ops-simulator Public

   Discrete-event simulation of a Trust & Safety enforcement queue, with the operating model it is built to test: severity matrix, decision tree, escalation paths, calibration cadence, and AAR.

   Python

6. crisis-comms-wargame crisis-comms-wargame Public

   An adversarial multi-agent war-game for AI-safety crisis communications, and the playbook it was built to test. Four incidents, four adversary agents, three response postures, twelve deterministic …

   Python