Sycophancy, deceptive alignment, reward hacking, and harmful compliance as four positions of one underlying collapse. Written for the LessWrong and alignment-forum audiences, with a falsifiable prediction.
The long-form exposition. The structural claim, the four failure modes, why a captured system cannot check its own capture, why the designer must see the feature in themselves before they can build it, and the main prediction with two weaker falsifiable alternatives.
The runnable specification for the first empirical test. Models, benchmarks, phases, and success and failure criteria. Written for an ML graduate student or interpretability researcher who might run it.
The complete exposition of the framework as applied to human flourishing, developed over 25 years. Thirteen chapters and eight appendices on the structural model, the six generative configurations, the six collapsed ones, the three negative currents, and the practices for growing the reconciling capacity under ordinary life conditions.
The companion to the alignment short version. Names the structural absence Fusion Dynamics sees as the common cause underneath anxiety, perfectionism, people-pleasing, and avoidance, and the practices that grow the reconciling capacity in the system that has to carry them.