Layered Error Correction
Judge a system by how many independent ways it can catch its own mistakes
- Difficulty
- Moderate
- Time to result
- ~ongoing to results
- Steps
- 5
- Confidence
- 68%
Deutsch's move is to stop scoring a system on its current decisions and score it on its error-correcting capacity instead. He observes that systems good at error correction tend to make larger errors, because large errors are survivable when several independent mechanisms exist to reverse them. Applied to British governance, he enumerates the stack: losing the next election, the civil service quietly failing to implement, police forces not being funded to enforce, newspaper editors finding a niche in contradicting the line, and public uproar when the first prosecutions happen. Each layer is independent, so a hostile actor has to defeat all of them. This reframes the real threat: not any specific bad law, but systemic changes designed to make the makers unremovable. The defensive priority follows directly, protect the mechanism rather than argue the outcome.
Origin
Extracted from Naval, where David Deutsch applies his epistemology of error correction to British politics, tracing how a piece of legislation he considers wrong would be caught by several independent layers before it ever bit.
Core principles
- 01Knowledge persists because it is error-corrected, not because it was right to begin with.
- 02Systems that correct well can afford to make larger errors.
- 03The number of independent correction layers matters more than the current error rate.
- 04The only fatal attack is on the correction mechanism itself, not on any particular decision.
- 05Inertia is a correction layer: much of what is legislated is simply never implemented.
How to run it
- 1
Separate the error from the system
Name the decision you think is wrong, then set it aside. The object of analysis is the machinery that could reverse it, not the decision itself.
Watch out Arguing the merits of the decision is the most common way to skip this framework entirely.
- 2
Enumerate the correction layers
Write down every distinct route by which this error could be undone. For a government, Deutsch counts voting, non-implementation by the bureaucracy, non-enforcement, adversarial press, and public reaction to the first enforcement cases.
Pro tip Include passive layers. Nothing happening is a real and frequently decisive form of correction.
- 3
Test each layer for independence
Two layers controlled by the same actor count as one. Check whether the same person, budget, or incentive could switch off more than one of them at once.
Watch out Correlated layers give a false sense of depth; a stack of five that share a single point of control is a stack of one.
- 4
Find the attack on the mechanism
Ask what a party wanting to stay in power indefinitely would change. Deutsch's concern is not any single policy but systemic changes that make the makers unremovable.
Pro tip The tell is any change that alters the rules of removal, rather than the outcome of a decision.
Watch out Complete power for one side over many years is dangerous specifically because it enables mechanism-level changes.
- 5
Spend your effort on the mechanism
Direct criticism, funding, and attention to keeping the correction layers alive rather than to overturning the current error. If the layers hold, the error self-resolves.
Pro tip This is also how to stay optimistic without being complacent: the problem is soluble, and it is up to you to make the arguments.
In the wild
Asked whether new speech legislation in Britain is an existential threat, Deutsch declines to argue the law and instead walks the stack. There is more to error correction than out-voting the government. If it passes, the next layer is that it is never implemented: the Home Office does not order chief constables to spend the money on the surveillance needed to enforce it. Pursuing it would require a government firing officials who refuse. Even then, the first prosecutions cause an uproar, and newspaper editors eventually stop following the line because there is a niche for contradicting it.
→ Deutsch concludes nothing is inevitable, and that the underlying reason the government moves this way is that most people currently think it is fine, which is itself changeable by argument.
Deutsch points out that Britain in 1945 effectively voted in communism: the Attlee government nationalised and controlled everything. Yet it had no chance of sinking the British project, and the same Attlee persuaded the Americans into NATO. His reading is that these were functionally communist but violently pro-British, believers in British institutions who simply held wrong ideas about the economy. This is his illustration of the general rule: systems that correct well make larger errors, and the major ones were corrected because nobody changed the correcting system itself.
→ A policy error large enough to look civilisational was absorbed and reversed without damage to the underlying constitution.
Common mistakes
Scoring a system by its error rate
Low visible error often indicates suppression rather than health. Deutsch's point is the reverse: robust correction is what makes large errors affordable.
Counting dependent layers as independent
Correction layers that answer to the same actor or budget collapse together, so a stack that looks deep can fail all at once.
Fighting the policy instead of the mechanism
Energy spent overturning one decision is wasted if the rules for removing decision-makers are being quietly rewritten at the same time.
Is it for you?
Best for
Assessing the resilience of governments, institutions, or organisations that are currently making decisions you think are wrong.
Not ideal for
One-shot irreversible situations where no correction layer can act in time.
From the transcript
“the systems that are good at error correction make larger errors”
“there's a lot more to error correction in the British system than merely out voting the government”
“if it goes through the next level of error correction is that it'll never be implemented”
From the episode
The Deutsch Files IV