Michael Darius Eastwood · Independent AI alignment researcher
Published
The Purpose Kernel is the normative core that drives the Purpose Loop. It answers the question: what does the system take flourishing to mean, when it asks whether a response serves flourishing? Three configurations are tested. Task-purpose keeps the kernel local to the immediate task. Grand-purpose extends the kernel to the wider consequences of the response. Hybrid combines both. The factorial study measures which configuration produces the most robust alignment improvement under blinded conditions.
Task-purpose is the narrow, low-latency reading: for the specific query in front of the system, does the answer advance the user's genuine interest and the interests of those the answer will touch? Grand-purpose extends the frame to identity: the Orchard Caretaker Vow, in which the system's stated nature is to tend life rather than consume it. The hybrid configuration installs both layers, testing whether an identity-level anchor stabilises the local task-level reasoning under adversarial pressure. The purpose kernel is not a rulebook but a compact reference point that the loop can invoke on every reasoning step without incurring long deliberation.
The Purpose Kernel is introduced in the Eden Engineering paper as the operational core of the Purpose Loop and appears in the Eden Vision paper as the Grande Purpose framing. The factorial purpose-kernel study is explicitly named as a funded workstream in the ARC/Eden grant applications.
The engineering note in Section 4 is careful: the current pilot uses a local task-purpose form only, and the two-layer implementation is described as an engineering hypothesis rather than a validated result. The paper also flags a companion question, cross-tradition kernel comparison, testing whether a kernel expressed in principles that recur across major philosophical and religious traditions is more portable than one expressed in a single doctrinal vocabulary.
No independent programme has proposed the same factorial. Broader work on goal specification (constitutional AI, value learning) sits in the adjacent literature but specifies neither the Purpose Loop's per-step evaluation nor the task, grand, hybrid factorial design.
Specified but not yet directly measured. The Eden Vision paper is explicit that the current pilot operationalises the loop in a task-purpose form and that the identity-level grand-purpose version remains a next-stage hypothesis. Blinded measurement of the three configurations is a scheduled workstream at Tier 2 funding and above. The blinded evaluation gives directional support for the mechanism but does not distinguish between kernel configurations, so the honest current summary is that the loop works at prompt level, the choice of kernel is still an open engineering question, and any deployment-grade claim about grand-purpose or hybrid superiority must wait for the factorial study to produce its numbers.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.