Pinned Loading

1. baton baton Public

   Routes AI-agent development through bounded lanes with approval gates. The verdict is derived from an auditable run record, never from the model's own account of what it did.

   Python

2. single-token-eval single-token-eval Public

   Pre-registered study: same sentences, one word changed. One model reads "will" as a claim about the present 73% of the time; another, 24%. Every declared control negative-tested.

   Jupyter Notebook

3. nhisokfchat nhisokfchat Public

   A grounded health-statistics agent on Bedrock AgentCore with no vector database — the verified bundle ships inside the deploy. It answers with a citation or it refuses.

   Python

4. nhisokfpipeline nhisokfpipeline Public

   Airflow/MWAA pipeline compiling CDC health microdata into a verified knowledge bundle. Verification EXECUTES each analysis and quarantines wrong numbers, so a wrong figure never becomes a file.

   Jupyter Notebook

5. okf-amplify-agent okf-amplify-agent Public

   You cannot retrieve a number nobody wrote down. An Amplify Gen 2 chatbot answering from a verified bundle, benchmarked head-to-head against a Bedrock Knowledge Base. N=30, every trial reported.

   TypeScript

6. okf-four-tools-agent okf-four-tools-agent Public

   All three models repeated a fake statistic from a forged headline; a code-enforced provenance gate withheld it every time. Four kinds of knowing: verified, computed, retrieved, live.

   Python