Second of the approved §5 checks. A `file:///Users/ktg/...` link is dead
for every reader but its author, and it publishes that author's directory
layout on a surface meant for strangers.
The SCHEME is not the rule, and measuring first is why. Across the corpus
there are 40 such links, and they split 18/22: a documented convention
example (`[Brief](file:///Users/ktg/.../brief.html)` — the same two lines
copy-pasted into nine CLAUDE.md files) versus real machine paths. Firing
on `file:` would have been wrong 45% of the time on its first run, which
is the ratio that gets a gate switched off. A further 22 links use bare
placeholders (`file:///abs/path.html`) and are not leaks either.
The discriminator is not tuned to this corpus: `...` is not a path
segment, so a target containing `/.../` cannot resolve on ANY machine and
is by construction an illustration.
Level follows the established reader rule — root is the shop window
(ERROR), below it live session plans and agent working files (WARN).
Measured on the registered corpus: exactly one finding, guard's
CLAUDE.md:56, a real absolute path to a file that is IN the repo and
should have been linked as `docs/BRIEF.md`. Every one of the 18
illustrations stayed silent.
Honest limit: three repos holding 21 of the 22 real leaks
(from-ai-to-chitta, wiki-advise, claude-code-llm-wiki) are NOT in the
register, so no check runs against them at all. That is register
freshness (`--refresh`), not this check.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lb7XmJGLnFSX9U7tgS7fKk
Kjell Tore Guttormsen
fa0cfe07d5