Initializing Architecture
Lab Notes

Things I only learned by getting them wrong first.

Each note came out of a specific decision in a specific system, and names the project it came from. None of them are opinions I held before the code disagreed with me.

2026.08Podman Flake Agent

A Confident Wrong Answer Is Worse Than No Tool

Call a real race condition an infra blip and you have told a maintainer to press re-run on a genuine bug. So abstention became a first-class verdict, and the harness reports abstention rate next to accuracy and counts the dangerous confusion separately. A classifier that says nothing is recoverable. One that is confidently wrong is not.

2026.07RabbitHole

Prompts Are Requests, Schemas Are Constraints

Asked politely for two perspectives, the graph produced six to eight, a 3 to 4x overrun that burned the token ceiling before the debate resolved. No amount of prompt rewording fixed it. Moving perspective count into a state schema the moderator schedules against fixed it completely, because the constraint stopped being something the model could ignore.

2026.07PaperPlanes

Never Overwrite What You Might Have To Explain

A fact is not static. It gets recorded, superseded, contradicted, and sometimes restored. Destructive updates make the current state cheap to read and the history impossible to recover. Closing rows with valid_to instead of overwriting them keeps a harder question answerable: what did we believe last month, and were we right?

2026.07PaperPlanes

Prove The Boring Baseline Cannot Do It

Claiming a distributed database beats a flat file is easy. Running 25 concurrent writers against both and showing one keeps 25 and the other keeps 1 is the version anyone can check. Every architectural claim worth making should come with the command that reproduces it, otherwise it is decoration.

2026.07Podman Flake Agent

Cost Is An Architecture Decision, Not A Billing Problem

Thirty CI jobs per pull request, each with a full journal, is not something you feed to a model and sort out later. Extracting only the failing block first cuts the payload 76% to 93% before inference. Free-tier rate limits are the same pressure in a different costume, and both are best answered in the design rather than the invoice.

2026.06PaperPlanes

Test Against The Thing That Breaks You

Mocks agree with you. A real cluster does not. The tests worth writing are the ones that run against live infrastructure, restart the database mid-conversation, and assert that the model under test is the real one and not a stub that quietly returns whatever makes the suite pass.