My Notes

Working notebook

Short observations from building and studying AI systems. These are provisional by design: compact enough to revisit, concrete enough to test.

MODEL BEHAVIOR

Context is a budget, not a drawer

Adding relevant material can still lower performance when it obscures the decisive evidence or instruction.

EVALUATION

Failures deserve names

An error taxonomy creates a bridge between a bad example and a design change.

SYSTEMS

The harness owns the loop

Models can propose actions; software must own permissions, state transitions, budgets, and completion checks.

Recent notes

2026-08-22 · Optimize for diagnosis

The first question after a regression should be answerable from the trace: which input, evidence, configuration, tool result, or validator changed the decision? If the answer needs a code-reading expedition, the system is under-instrumented.

2026-08-18 · Build the smallest meaningful grader

A rough rubric attached to real cases is more useful than a sophisticated metric detached from the actual task. Measurement should grow from observed errors.

2026-08-11 · Interfaces reduce semantic entropy

Typed boundaries do more than prevent parser errors. They force the system designer to decide what information matters, what is optional, and how uncertainty is represented.

Note format

Each note records an observation, a hypothesis, the evidence, and the next test. A note graduates into the knowledge map only after it survives repeated use.