Silicon observability
Debug and state-capture primitives for complex AI/HPC hardware systems: scan paths, JTAG-style control, functional state capture, and repeatable root-cause workflows.
The Observability Stack
The Observability Stack
Research notes on silicon debug, AI infrastructure, autonomous agents, and complex compute systems.
Core thesis
Complex systems do not become trustworthy just because they perform well. They become trustworthy when their behavior can be inspected, replayed, and debugged.
The Observability Stack is my public research surface for exploring how complex compute systems become inspectable: from scan-based silicon state capture to AI infrastructure economics, autonomous agent evals, and quantum-classical control systems.
Why this exists
A structured body of work, not a newsletter feed.
This publication exists for one question: as compute systems become more capable, distributed, and autonomous, what instrumentation do we need to understand them when they fail?
In silicon, observability means state capture, scan paths, controlled debug flows, and root-cause workflows. In AI agents, it may mean goal traces, tool-call traces, uncertainty traces, replayable execution, and escalation checkpoints. In quantum-classical systems, it may mean low-latency measurement, feedback, and control. Different systems, same pressure: hidden behavior must become inspectable.
Core pillars
Four pillars, separated by maturity.
AI infrastructure economics
Frameworks for reasoning about useful AI output under real constraints: latency, reliability, quality, safety, and cost.
Autonomous agent observability
Research notes and eval ideas for making tool-using AI agents inspectable, replayable, and debuggable.
Quantum-classical systems
Notes on the classical control, calibration, and observability layers that may make future quantum systems operationally useful.
Featured artifacts
Public work you can read, cite, or inspect.
The Cost of Usable Intelligence
A framework for reasoning about useful AI output under constraints like latency, reliability, safety, and quality.
Debuggability for autonomous agents
What AI safety can learn from silicon bring-up: state capture, replay, traces, and root-cause workflows.
Observability is an underbuilt safety primitive for frontier AI systems
As AI agents move from answering to acting, safety will depend on whether we can trace, replay, and debug their behavior.
Scan Chain / TAP Visualizer
Interactive visual model for scan-chain movement, state capture, and debug intuition.
Frameworks
Reusable models for complex-system observability.
Agent Debuggability Stack
A layered model for tracing agent behavior from instruction to goal interpretation, planning, tool use, uncertainty, escalation, outcome, and replay.
Observability maturity model
Five levels of agent observability, from final-output-only systems to replayable execution with failure taxonomy and eval integration.
Silicon debug to agent observability map
A conceptual bridge between state capture, scan/debug paths, signal traces, and agent traces, tool logs, uncertainty, and escalation.
Publication roadmap
Published artifacts first. Future directions second.
- The Cost of Usable Intelligence
- Debuggability for autonomous agents
- Observability is an underbuilt safety primitive for frontier AI systems
- The Agent Debuggability Stack
- A taxonomy of autonomous agent failure modes
- Operational evals for AI agents under ambiguous authority
- Quantum computing needs an accelerated classical control plane
Discussion / citation layer
Share, cite, or build on this work.
Each major piece includes a thesis, citation block, and related artifacts. If you build on a framework, cite the page or link back to the canonical article. For serious feedback or collaboration, email me.