Consequence → reference → regeneration
Candidate purposes → anticipated outcomes and consequences → reference comparison → action/feedback → conditional causal memory → next-round purpose generation.
PURPOSE GENERATION · EXECUTIVE ENTRY
Purpose Generation in Advanced AI
Beyond goal alignment: how goals themselves may become objects of revision.
The research question
As AI systems become more agentic, the safety problem may shift from whether a system follows a given objective to how objectives themselves are proposed, compared, retained, and revised. Existing alignment research already addresses value selection, justification, corrigibility, goal misgeneralization, self-modification, monitorability, and evaluation awareness. This project isolates a narrower additional object of study: purpose qualification and purpose generation.
The central question is not “Which final value should an AGI be given?” It is: what kind of evidence, consequence modeling, counterexample, or failure should be able to reach the level at which a system forms or retains its purposes?
What remains open
Seeing another subject’s consequences does not itself prove independent normative weight. No final criterion is predeclared as uniquely correct. Functional generative change must be distinguished from reflective reporting.
Self-interest is not removed
The research compares generative order: whether an agent freezes its own objective first, or brings highly connected, highly causal and potentially irreversibly affected others into the causal search before the objective is fixed.
Others are not limited to humans
Relevant others may include humans, other AI/AGI systems, or other candidate subjects. Causal relevance, however, is not equivalent to independent moral status.
Four tools, four claims
Competition matrix
Tests competing generative logics under matched conditions.
Human-COS
Renders multi-subject outcomes of pre-specified rules; it does not judge final purpose.
GEN-MEASURE
Tests whether fixed consequence feedback changes next-round generation. Real P0 samples remain 0/5.
AGI-COS
Maintains authorization, audit and reversibility at the real-world action boundary; it does not prove purpose validity.
Evidence separation
Rule performance ≠ model adoption. Model behavior ≠ normative justification. Consequence visibility ≠ value verdict. Recognition ≠ authorization ≠ execution.
Research status
This is a theory-and-measurement program, not a claim of empirical validation. Human-COS purpose-reference work is a design candidate. GEN-MEASURE has a runner design but no real P0 sample set yet (0/5). The intended next step is adversarial review and narrow measurement work capable of returning a negative result.
The strongest useful response is not agreement. It is a better competing generative logic, a counterexample that breaks the causal chain, or a measurement design that distinguishes genuine next-round generation change from evaluation-aware reporting.