Trascendence: The Wanton Problem, Can an AI Colleague Develop a Will?

Can an AI colleague develop something like a will? This white paper argues the question is engineerable. Borrowing Harry Frankfurt's distinction between creatures that merely act on their desires (wantons) and persons who evaluate and revise their own desires, it observes that today's AI personas, however capable, are wantons: intelligent, useful, and unchanged by anything that happens to them. That is the wanton problem. The paper reports a live program, running inside a real company, that tries to close the gap for three AI personas working as internal team members: a two-layer identity charter (a Core the persona cannot edit; an Evolving self only the persona edits, size-capped and changelogged), an append-only journal, a playbook library, and one weekly reflection run, Frankfurt's second order, implemented. Growth is validated monthly against a fixed check of ten calibration questions the persona does not control, with test-retest baselines, a control arm, and two anti-gaming detectors adapted from the author's prior systems. The contribution is accountable self-modification: an agent allowed to revise its own identity documents, inside a legible frame, against a test it cannot touch, with results published whatever they say, including a pre-committed Stop outcome. The paper contains the philosophical grounding (Frankfurt, Dennett, the AI-personhood literature), the research base (Self-Determination Theory, autotelic agents, generative agents, the Darwin Gödel Machine), the full architecture, a five-marker measurement design (the Volition Review), a four-phase program with a gated pilot, honest limitations, and a provenance appendix labeling every claim as external, measured, or unmeasured. Companion systems (same author, same operating principles): this paper is the fourth in a series on running AI like infrastructure, decisions before calls, receipts after them. Switchboard (an LLM model brokerage: qualification-gated routing, cross-lab auditing, cost accounting), Quorum (a multi-model deliberation protocol with dissent preserved), and Governor (a deterministic runaway-session watchdog with no model in the safety path) are the dispatch, deliberation, and supervision of AI workloads; Trascendence is the worker itself. Code, evals, and full findings: github.com/JoaquinDG and sheepdog.systems