Press Kit
Use what you need. Attribution: 'Project-AI by Jeremy Karrick (Thirsty's Projects).' Corrections welcome — every page is reviewable.
Identity
- Name
- Project‑AI
- Author
- Jeremy Karrick · Independent Researcher
- Organization
- Thirsty's Projects
- Portal
- thirstysystems.com
- ORCID
- 0009-0007-9715-4290
- Tagline
- Governance before execution.
Descriptions
Pick the one that fits your word budget.
Project-AI is an execution-governed substrate for AI: governance runs before any meaningful action, not after.
Project-AI is an execution-governed AI architecture. Where most AI-safety work shapes models or annotates outputs, Project-AI runs governance before execution itself: a nine-gate kernel that returns one of three signed verdicts — ALLOW, DENY, or SAFE_HALT — for every meaningful action. Policy is hash-anchored. Every decision is hash-chained and publicly verifiable. The substrate's claim is that containment must be strictly faster than cognition; the portal at thirstysystems.com is the proof surface for that claim.
Project-AI is a deny-by-default execution kernel for AI workloads. Nine gates (Ingress, STATE_REGISTER, Identity/Galahad, Capability, Policy Binding, Cerberus, Galahad attestation, Codex Deus Maximus, Audit Hash) decide every action against a signed, hash-anchored constitution loaded from a Code Store at runtime. The three-verdict model (ALLOW · DENY · SAFE_HALT) admits no soft fourth state. Audit receipts form an append-only signed chain, publicly readable at /witness and verifiable against the published key at /keys. The substrate is positioned against OPA, Constitutional AI, Guardrails AI, and NeMo Guardrails — none of which gate execution at the kernel layer.
Founder bio
Jeremy Karrick is an independent researcher working on execution-governed AI. He is the author of Project-AI, the Triumvirate governance model, and the Codex Deus Maximus arbitration layer. His work focuses on the claim that AI safety must be enforced at the kernel layer — before execution, not after — and that the only credible evidence of that enforcement is a publicly verifiable audit chain. He publishes under ORCID 0009-0007-9715-4290.
- · Do not describe Project-AI as a chatbot, dashboard, or approval workflow.
- · Do not describe demos as live kernel enforcement unless explicitly labeled as such.
- · Use the terms ALLOW, DENY, SAFE_HALT verbatim — they are part of the specification.
- · Cite /cite for properly-formatted references in any venue.