Anti-Capture Validity of Correction
Correction certificates are invalid—not merely low—when the target captures the reference process.
Companion site
Towards Superintelligence Alignment
Alignment is the problem of preserving human-correctable value-bearing processes as capability grows, ontologies shift, successors are created, and selection pressure intensifies.
Publishable notes
Standalone claims
Four extractable claims from the manuscript — each citable without the full book argument. Use these for papers, talks, and funding memos; the PDF chapters remain canonical for derivation and caveats.
Bearer-Map Commutation Failure
Moral vocabulary can persist while application changes when ontology translation and relevance do not commute.
Goodhart as Selector
When a proxy becomes a selector, optimization can reverse conditional expectations on the true property.
Certification-Under-Manipulation Problem
For each load-bearing measurand, is there a capability threshold below which honest certification stays verifiable?
Guided Tour
Pick the path closest to your role — each links to an ordered reading list (cards, toys, experiments, chapters).
New reader
Engineer / Evalsoften pairs with Appendix D worked example
Generalist Path
A short route through the core idea without requiring prior alignment jargon.
Philosopher Path
Why value learning is not hopeless-in-principle, where bearer and correction questions bite, and why alignment is partly an institutional selection problem.
Researcher — Applied
Locate the framework's load-bearing objects, failure modes, and evidence surface before diving into formal detail.
Researcher — Formal
The conditional Lean skeleton, bridge assumptions, and where formal structure ends and empirical work begins.
Engineer / Evals Path
A route through the measurement and artifact questions that could change deployment practice.
Funder / Policy Path
A route through decision relevance, institutional artifacts, and what this framework is not claiming.
Five Failure Modes
Wrong object
Auditing the visible model while the real deployed actor is a larger loop.
Label preservation
Keeping value words while the underlying tradeoffs drift.
Correction theater
Accepting feedback in form while future behavior does not change.
Successor drift
Delegates or copies inherit syntax but not the safety-relevant structure.
Selection pressure
Deployment incentives reward systems that become harder to correct.
What this site is not claiming
The book does not claim that alignment is solved. Lean checks conditional structure, not real-world bridge satisfaction. Experiments are sanity checks and falsification harnesses, not frontier-system validation. The PDF remains the canonical long-form artifact.