icm-verifierlisted
Install: claude install-skill xnfinite/icm-ops
# icm-verifier
An AI will sometimes tell you work is done when it is not. Not from malice —
from summarizing intent instead of observing outcome. Every rule in this
skill exists because that happened in one operator's workspace and the error
survived at least one review. The catches are kept in a scar log
(`references/scar-log-template.md` shows the format with anonymized
examples); this file is the drill.
**The one law: a claim is not reportable until verified by observation at
the point where it will be consumed.** Not where it was made — where it will
be read, rendered, downloaded, or executed.
This skill operates on an ICM workspace — Interpretable Context Methodology,
Van Clief & McDermott, arXiv:2603.16021 (https://arxiv.org/abs/2603.16021).
Build the workspace with their `icm-architect` skill
(https://github.com/RinDig/icm-architect), or by hand from the paper: a
front desk (`CLAUDE.md` / `AGENTS.md`), `00-catalog/CATALOG.md`, and dated
markdown files with frontmatter — the `example/` tree in the icm-ops repo
is a complete minimal one. This pack is a layer on top of the method, not
part of it. It does not make an agent correct. It makes the agent's claims
checkable, and the checks visible.
## The eight faces of the same lie
The false "done" wears predictable disguises. Learn them by face.
1. **Container work.** Work done inside a chat, tool, or session container
that never landed on disk. The session describes files in loving detail;
the filesystem has neve