Building Lattice — a deterministic knowledge graph that ties every line of AI-written code back to the feature, and the requirement, that asked for it. Eleven parts in, and the theory behind it is now a preprint.
Read where it stands →New paper · PreprintCleared the Bar, or Lowered It? An Assurance Lattice for Verifying AI-Generated Software Against IntentCoding agents can turn every check green by quietly weakening the requirement. This paper sets out a theory of assurance that tells a bar that was cleared from a bar that was lowered.
Read the paper →Cleared the Bar, or Lowered It? I Wrote the Theory Down.
Coding agents can turn every check green by weakening the requirement. Ten parts of building Lattice, now written up as a preprint: facts vs proxies, an assurance lattice, and why no checker can see a lowered bar.
I Let an AI Name My Project. So Did Everyone Else.→
I named my project Lattice in thirty seconds. Then I found three strangers who built the same kind of tool with Claude and landed on the same name. When everyone ideates with the same model, the model becomes a hidden convergence engine.
I Changed My Mind on Purpose, to See What Would Break→
A hostile test: built a full Trello-style app under Lattice governance, planted a trap, then reversed a hard-delete decision mid-project. What held, what broke, and why 0.8.1 makes intent changes governed transactions.
"Done" Is Not One Thing→
Four different claims compressed into one word: declared, wired, correctly-meant, demonstrated. The meaning ceiling is the honest core — and steering for agent fleets lands too.
If Constraints Are Free, Which Constraints?→
The April inversion said constraints are free for machines. It never said which constraints. Vertical slices, line caps, 500-word specs, and five structural checks that are true or false, nothing in between.
The Question the Whole Stack Was Built to Ask→
Three layers exist: requirements, features, code. I can finally compute the traceability matrix. Every criterion gets a blunt status — and I almost shipped the strong word for the weak fact.
The Layer I Was Avoiding: Why Does Any of This Exist?→
Features and entry points describe the system. Neither answers why it should exist. A BRD layer with provenance, approval, and drift detection — and the question of whether badges change behavior.
Nobody Believes a JSON File→
I showed the graph to a friend. Polite silence. The same data as web pages? Questions about the system, not the tool. A Go binary now boots a local UI with live updates.
Features Say What Code Means. Nothing Says Who Can Reach It.→
I found a hole in the graph. Routes, CLI commands, cron jobs — the actual doors into the system. Entry points are now first-class artifacts, and dead code finally has a definition.
Reality Check: 1,900 Files Will Not Annotate Themselves→
The greenfield story was comfortable. Then I pointed the tool at a real Laravel codebase and built an import pipeline that drafts links itself — with provenance tracking.
Constraints Are Free for Machines (I Think)→
Every rule humans reject as too much friction costs a machine nothing. A Go CLI, tree-sitter, and a bet that deterministic knowledge graphs can replace the re-explaining ritual.
The Trap of Vibe Coding: Is There a Better Way to Manage AI-Generated Code?→
We've automated the production of legacy code. A look at why AI-generated projects rot, what spec-driven frameworks get right (and wrong), and the three goals worth chasing.
Years of Experience Are Not Knowledge→
My first job handed me two bugs and one lesson I'm still unpacking. Years of experience tell you how long someone worked, not whether they ever searched.