Writing
Field notes on systems, failures, and the boundaries that matter.
Most of these start with something that looked fine and was not. Each one states what I observed, what the ordinary check suggested, why that answer was incomplete, what boundary changed, and what the change did not prove.
Published
-
Evidence Before Claims
Why the smallest useful next move is a better place to start than an AI transformation program
A practical framework for choosing one useful, testable AI move without turning a broad ambition into an unbounded first project.
-
Bounded Automation
MCP is not an authority model
Model Context Protocol standardizes how an AI application discovers and calls a tool. It does not decide which tools should exist, which identity they use, or what counts as human approval.
-
Evidence Before Claims
An evidence-first content standard
Interesting, supported, and safe to publish are three different states. The editorial system I use to keep a personal voice from making institution-sized claims.
-
Solo Operator Architecture
Hardware-bound boot assistance for a Shamir-sealed secret store
When a secret store must recover after a restart, preserving a tested manual seal path and automating a narrow boot ceremony may be less disruptive than a seal migration. Verify the final state, not the exit code.
-
When the Control Was the Failure
The files copied, but the workflow did not complete
A long-running job finished moving the data and never wrote its completion marker. Four downstream stages refused to run. The marker was right, and the tempting one-line fix would have forged evidence.
-
Green Isn't Proof
The stack was green and no prompt worked
Every container reported healthy. Every dashboard was green. Nothing could answer a single question. Health checks proved process availability, and nobody had tested the contract between the processes.
-
When the Control Was the Failure
A perfect evaluation score stopped being trustworthy
The detector scored perfectly against its test corpus. Then I noticed the same model had helped write the corpus and tune the detector—so the score measured agreement with one model's imagination, not coverage of the problem.
-
Evidence Before Claims
An Atlas needs a changelog, not just a map
A private architecture Atlas helped me orient myself across systems and projects. A small, evidence-aware changelog made it useful to the humans and agents doing the work.
-
When the Control Was the Failure
A personal-data system whose first success was refusing ingestion
The first useful result from a personal-data project was not a search feature. It was proving that the system could stay useful without reading the files it was meant to protect.
-
When the Control Was the Failure
The DLP control that removed its own evidence binding
A broad high-entropy detector correctly treated a value as suspicious, then removed the digest that made a safety receipt verifiable. The answer was a typed exception, not a weaker detector.
-
When the Control Was the Failure
Nothing checked the checkers
A passing verifier can be the least trustworthy part of a control when nobody has tested its scope, its dependencies, or its ability to fail.
-
Operating Foundations
A production-minded Docker Compose baseline
Docker Compose can describe a reliable single-host service, but starting containers is only one of the tests that matter.
-
Operating Foundations
One ingress, one owner: Caddy routing is not authorization
A reverse proxy can make web exposure reviewable. It cannot turn a host match into proof of identity or compensate for a backend that is reachable by another path.
-
Bounded Autonomy
The backup job can repair one thing. It should know when to stop.
A recovery controller is useful when it refuses an untrustworthy backup, makes a few reversible repairs, and leaves consequential recovery decisions to a person.
-
Bounded Autonomy
An autonomous patching agent should earn its reboot
Routine updates can be automated without treating a reboot as a harmless side effect: discover, bind approval, prove recovery readiness, change in a window, then verify the result.
-
When the Control Was the Failure
The decommission was thorough. Nothing told the backup.
A retired service left behind a rigorous break-glass archive and a backup job that failed hourly. The gap was not carelessness — it was a checklist that ended at the boundary of the thing being removed.
-
Solo Operator Architecture
Each control was correct. Together they built a job only I could do.
A secrets migration that should have taken a weekend took twenty-five days. Not because any safeguard was wrong, but because correct safeguards compose into a process that requires a present, awake operator — and there was only one.