Broken Clock Doctrine

Broken Clock Doctrine

Axiom

Even a broken clock is right twice a day.

Doctrine

This doctrine states that reliable signal can be extracted from unreliable or high-noise sources when approached with structure rather than trust. Within Convivial Systems Theory, this doctrine explains how systems preserve epistemic agency in high-noise environments by extracting utility without outsourcing judgment to unreliable sources. Low signal density is not useless data; irregular correctness is not random. Patterns repeat even in chaos, and unstable systems still produce moments of alignment if the observer builds tools strong enough to detect them. The task is not to demand consistent accuracy from noisy sources but to construct a framework that identifies, isolates, and amplifies the meaningful flashes when they occur. Structure turns sporadic truth into usable information.

Form

Look for repeatable patterns even in chaos.

Neural Network Mapping

(Broken clocks inside machine learning computational systems)

Neural nets have broken clocks too.
Early in training, most activations are noise and most layers misfire; the system is technically wrong nearly all the time.
But even inside this high-noise environment, useful gradients still appear.

If the architecture maintains a viable path for signal flow:

• sporadic correct activations become anchors
• gradient descent extracts direction from rare correctness
• noise becomes navigable rather than disabling

A neural net does not need to be right consistently.
It needs to be right detectably.

Broken Clock Doctrine in ML terms:

• You don’t need every timestep to be right.
• You need a way to notice the few that are.
• And you need an update rule that learns from those.

This is why architectures with residual pathways, skip connections, and stabilized gradient flow outperform those that drown in noise:
they preserve the ability to catch fleeting correctness and convert it into learning.

Related reading

The Broken Clock (essay)