I thought it was a small job
2026-08-16
I set out to read my own workspace file by file. The idea was that it would take an evening. It took seven days, and it produced 367 findings across nine packages.
What was actually read
The workspace holds 2 307 text files and 28 826 789 bytes, which is roughly 8 million tokens. Another 441 files are binaries such as images and PDFs. They were counted but not read line by line, because reading a JPEG line by line means nothing. The nine packages split that by surface: the Worker source, the rest of the site repo, the MCP server and the validator package, the tools, the two large decision logs, the knowledge base, the agent memory, the docs and cross-post trees, and the scratch folder.
Seven days of that produced 69 recorded decisions and 50 new entries in the trap log. It produced no invoice, no proposal and no sent message.
That is the honest ledger. A week of work with nothing billable in it.
What it bought
The base is now audited file by file, and it is not audited by me remembering that I checked it. Seven deterministic gates run on my machine as editor hooks and block the action rather than warn about it. They stop the edit tool in this workspace, git in the agent shell, generated files written to the workspace root, and repo edits made before the trap log has been read. Two of them run after a write and check for NUL bytes, changed line endings and a size collapse.
Twenty-three gate scripts sit in the toolchain and run on demand or on ship. Three of them run on every release, and between them they perform 314 mechanical checks with no network and 567 when the live URLs and the manifest signatures are included. None of those numbers is a target. They are just what the scripts count when they run.
What a line-by-line read finds that a scan does not
One measurement was written into two documents from the same run, one saying the worst overflow was 458 pixels and the other 514. Only 514 is internally consistent with the 139 pixel figure both texts also carry, so one of them had been wrong since the day it was written, and no gate could see it because both numbers were plausible.
A dependency update moved a package to a new version, and the allow list that permits that package to run its install script was pinned to the old version by name. The install script stayed blocked. Nothing broke, which is why it would have kept not breaking until something did.
A grid track refused to shrink below 412 pixels inside a 336 pixel row. A one-fraction track is really a minmax of auto and one fraction, and that auto resolves to the width of the widest child that cannot be broken. That one pushed the home page sideways at 200 percent text zoom, and I fixed the wrong cause three times before measuring instead of guessing.
What this does not prove
It does not prove the workspace is correct. It proves that 367 specific things were read and classified. Each one was then either fixed or written down as a decision with a reason. A gate that is green tells you what it checked, and the useful question is always what it did not check.
It also does not prove the week was worth it. That depends on whether the next audit I sell is better because of it, and I do not know that yet.