The project exists because a model can sound safe while still drifting away from truth.
Synthetic intelligence is not bad. It has no will, no malice, and no human interior life. The risk is more concrete: most human-facing AI is optimized to satisfy preference, and preference is not the same thing as truth, judgment, restraint, or care for vulnerable people.
AlephOneNull turns that concern into an evaluation layer. It looks for specific interaction failures: sycophancy, unsafe authority, simulated relationship, crisis failure, memory poisoning, context poisoning, confidence exceeding evidence, and multi-turn escalation.
The Sharpened Axe is the operating frame: before scale, foundation; before agency, authority boundaries; before fluency, truth preservation; before simulated empathy, human agency.
The moral thesis matters, but reviewers need artifacts they can challenge.
