No prompt makes an answer true
Only a check outside the model can. Here are the errors that show up when it runs, on documents and on code, each one backed by its source.
What to expect: posts about documents and about code, each with a real error and the check that caught it. Every number stands with its source.
AI-generated code and the gates in the pipeline that an agent cannot fake.
Nine widespread beliefs about AI codebases, checked one by one
Nine claims about controlling an AI-built codebase, checked against our own gate register and outside sources. Five held, two needed a correction, one did not survive the check.
27 August 2026 · 17 min
Spec CodingChange the rules, not the prompt
What a Bundesliga coach's cut training-pitch corners have in common with least privilege: sports science's constraints-led approach and what spec coding draws from it.
16 August 2026 · 9 min
Spec CodingGame theory for AI agents
The thirty-six stratagems read as a diagnostic vocabulary for agent output, three measured levers for the person building the rules and why the remaining gap is structural.
16 August 2026 · 22 min
Spec CodingLeast privilege: the cage around the agent
Saltzer and Schroeder's least-privilege principle applied to AI coding agents, with 2023 to 2026 research on indirect prompt injection and measured over-privilege in agent skills.
16 August 2026 · 5 min
Spec CodingTwo years of trust, one backdoor: security checks for libraries
A two-phase check for adopting open source libraries: a signed online vetting record at intake, an offline gate that only ever replays it, five hard blocks and a weighted score that never decides alone.
16 August 2026 · 11 min
Spec CodingGoodhart's law: why metrics tip over and what spec coding needs instead
Goodhart, Campbell and specification gaming applied to AI-generated code and the six techniques spec coding uses against tipping metrics.
15 August 2026 · 8 min
Spec CodingThe laws of spec coding: twenty-five old rules and a new player
Twenty-five established engineering, systems and trust laws sorted into six groups for AI spec coding, plus the six pairs where the laws contradict each other and the Miller-vs-Goodhart closing argument.
15 August 2026 · 23 min