Operator instrumentation. The IC discipline of running your own evals.

In 2018 I ran a gaming behavior audit with no apparatus at all. Just a spreadsheet and the reflex to look at the numbers and name what they showed. The tooling was primitive. The habit that grew from it was not. Seven years later that same reflex runs through Toggl time logs, Jupyter notebooks, and a second brain with a typed ontology. The tool changes. The habit doesn't. Instrument yourself the way you'd instrument a system.

The PM who cannot be debugged from data is not doing their job. The same discipline applies inward. Look at your own patterns. Name them. Externalize them into a substrate that synthesizes them back. Introspection is memory. Memory is lossy. Instrumentation is a closed loop.

The externalization discipline

Instrumentation starts with observation, not apparatus. The first step is labeling: name the pattern before reaching for a tool. "My urge here is competitive, not escapist." Labeling plainly before building anything around it. Unnamed patterns do not get instrumented; they accumulate into noise.

The substrate must do something with what the operator puts in. Paper notes do not synthesize. Tracking data that is never queried does not synthesize. The loop only closes when the substrate returns something actionable. A Toggl report that surfaces how time actually distributes across project types, a Jupyter analysis that identifies where estimates consistently miss, a second brain that surfaces a decision pattern the operator did not know was stable across four years of choices.

The transition sequence matters less than the principle behind it. Toggl replaced a paper log. A Jupyter notebook replaced Toggl. A second brain with a typed ontology replaced the notebook. Each is a hardware upgrade, not a principle revision. The disposition that each upgrade serves is identical: name the pattern, externalize it, let the substrate return it.

Domain-specific over generic

A general-purpose log is low-friction to write and low-value to query. Domain-specific wrappers return targeted synthesis. A GPT structured around a curriculum for a specific subject area returns the kind of Socratic feedback that a general LLM session does not. A notebook running the same analysis weekly against the same data schema surfaces drifts that are invisible in one-off queries.

The corpus of self-instrumentation that compounds into a second brain is not a single system. It is per-domain configurations, each built to close the loop on a specific recurring decision. The gaming retrospective from 2018 was a one-domain configuration. It was also the seed.

The IC economics argument

In the AI era, the IC advantage is not throughput. Agents handle throughput. The IC advantage is judgment: knowing what to build, what to cut, what pattern the data is actually showing. Self-instrumentation is how that judgment gets trained on real evidence rather than intuition.

The audit that makes the discipline real: identify every recurring decision made from memory alone, without data. Each one is a gap. The question is not whether the gap is tolerable. The question is what pattern is being missed by not closing it. At the enterprise level, FarEye's data onboarding compressed from 60 days to 7 days because instrumentation surfaced exactly where the bottleneck lived. The IC who runs their own evals finds the equivalent bottleneck in their own decision-making. And discovers it in hours, not after a quarter has closed.

Where this is going

The second brain is the current form of this discipline. It is not the final form. As the graph deepens and agent traversal improves, the synthesis layer gets better. Not because the habit changed but because the substrate got smarter underneath a constant disposition. The operator who built that habit before the tooling arrived is in the better position. The operator who is waiting for a seamless tool before starting the instrumentation practice is paying a delay tax on every decision in the interim.