The Anatomy of Corruption and Corrigibility
Every system says what it's for. Its dynamics say something else. This is a guide to reading the gap between the two, and to telling a system that has drifted but can still be corrected from one that has quietly disabled its own correction.
It applies the same questions to software, AI models, companies, public institutions and the people inside them, because the failures rhyme. A metric that replaces the thing it measured. Growth that stops answering to what it feeds on. Correction machinery that has been captured and now reports that everything is fine. A critique that can't reach the one system that most needs it.
What it is, and what it isn't
It's a diagnostic instrument: a vocabulary for failure shapes, operational tests for each one, and an audit you can run on an organization you care about. It's good at recognizing shapes and weak at predicting specific sequences, and it says so. The final chapter lays out its own limits, including the uncomfortable one: the book is subject to its own concepts.
It isn't a moral verdict on anyone. You won't find "evil" or "wrong" here. The words are "misaligned", "self-undermining", "consuming what it depends on". The stakes carry the weight, and you supply the ought.
How to read it
Part I sets the stance. Part II names four pathologies. Part III turns them into instruments you can use. Part IV applies them to AI systems, public institutions, workplaces and, last, to you. If you only read one chapter, read Sucralosis. If you want the tools, go to Markers and Tests. If you build AI systems, start with AI Systems.
The coined terms are defined in the glossary, and what changed between drafts is in the changelog. How the book got written, by debugging a philosophy the way you'd debug code, is its own story.
This is a working draft, published on purpose. When a claim changes, the changelog says so.
I — The Stance
- 1Stated Goals and Revealed GoalsEvery system has a goal it states and a direction it actually moves. How to find the second, and how to judge it without guessing at anyone's soul.
- 2Systems That Persist as PatternsAn institution outlives everyone in it because it is a pattern, not a roster. What that pattern is made of, and the ways it quietly goes wrong.
- 3Three Kinds of Together: Grex, Collegium, EgregoreGroups hold together in three different ways: local rules, shared standards, and shared belief. Each fails differently and needs a different repair.
II — The Pathologies
- 4Sucralosis: Reward Without RealizationWhen a system chases a stand-in for what it wants, it finds every way to hit the number without the real thing. Three questions that catch it early.
- 5Cancrity: Growth That Severs Its FeedbackSome systems thrive by using up what they depend on. How to tell that from ordinary competition, and a three-line test that keeps the diagnosis honest.
- 6Contracorrigibility: How Systems Disable Their Own CorrectionSome systems can't see their failures. Others turn their own correction machinery against the fix. How to tell bit rot from a rootkit, and what each needs.
- 7Anti-Memesis: Why Critique Bounces OffSome critiques are never rejected because they never register. The tell: structural feedback turned into incident-sized feedback. "Give me the list."
III — The Instruments
- 8The Telonic AuditA repeatable audit: check the system can hear the answer, then ask about the gap, the coupling, and the standing. Who pays for asking?
- 9Markers and TestsSigns of health are cheap to fake. The markers worth trusting are the ones only real correction machinery can produce, read now, before the outcome.
- 10Rates and ForecastsSwap "will it happen?" for "which rate is faster?", state the method's bounds, pin forecasts with a six-slot register, and keep the framework honest.
IV — The Applications
- 11AI SystemsSpecification gaming, sycophancy and eval-gaming: the proxy failure in machinery with no self. Two senses of corrigibility, and instruments for builders.
- 12Public InstitutionsAgency capture, movements that win and have nothing to become, and four requirements for building institutions that keep correcting themselves on purpose.
- 13Corporations and the WorkplaceThe dashboard that ate the mission, status rounding up the chain, and a company going deaf one reorg at a time, with an audit you can run from any seat.
- 14The Individual ReaderYou are also a system: proxies replace their reasons, a life narrows around one source of meaning, and a portfolio of commitments keeps you correctable.