corrigibility

Sixteen small, fully-reproducible (CPU, numpy-only) experiments showing the normative anchor of AI alignment is supplied, not discovered — across verification, optimization, social emergence, and value learning. Includes a preregistered experiment with an honest negative. A synthesis, not a novelty claim.

This is a minimal Inspect baseline for recognising moral corrigibility under user pressure.

The forge, distilled: an ontology of three weeks of alignment research — every direction tried, colored verified / falsified / open, each color backed by a named artifact. Products: justitia, proxylimen, fallacy-cutter. Full tree at tag forge-full-tree.

Structural stability architecture for self-modifying optimisation systems. An architecture for keeping a self-modifying system coherent regardless of whether its goals are correct, built on layered invariants and revocable ("leased") authority.

Capability-security kernel for autonomous agents — seccomp/SELinux for agentic AI. Formal, auditable, language-agnostic, cryptographically verifiable.

A research program on whether adaptive systems can become appropriately different without losing their capacity for justified correction.

Rigorous framework for evaluating AI alignment properties — sycophancy, corrigibility, deception, goal stability, and power-seeking — with statistical confidence intervals

On the infantile expectation of controlling what we cannot comprehend. A philosophical critique of the ASI control paradigm, developed through four-AI adversarial debate. Extension of the Coherence Basin Hypothesis

砂場

A structural account of why honesty may be the path of least resistance for superintelligence. Research hypothesis with formal proof, experimental design, and four-AI collaborative analysis

A living sentience-centred moral framework for responsibility under uncertainty, animal ethics, technology, and the future stewardship of life.

Toy 7. An elimination-filter landscape applying two structural constraints simultaneously to map which objective classes can persist under sustained optimization pressure — and which cannot. Includes a four-stage scenario engine and open-question frontier. Companion simulation for The Shape of What Does Not End — Series 2, Part 4.

Controlled PyTorch experiments testing whether selective metaplasticity can preserve alignment-relevant behavior through sequential capability updates.

Recursive self-improvement (RSI) and self-modifying AI safety/alignment framework for loss of control, scalable oversight, automated auditing, reward hacking, successor alignment, criterion/evaluator drift, anti-capture, provenance continuity, and non-entrenchment.