principle-real-artifact · v1.0.0 · 2026-10-14 · sha256 8b4407f778fb6133
principle-real-artifact v1.0.0A
Immutable. This exact content never changes and is served at /api/v1/blob/8b4407f778fb6133.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_7cxx6ofiada5kzaq/8b4407f778fb6133e2efeeffc4d2d6681768b6b3b1fbfc73a6cca4cd80aee39b.json
Then /plugin install principle-real-artifact@markdownregistry-7cxx6ofiada5kzaq in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: principle-real-artifact version: 1.0.0 description: Verify every claimed outcome against the real artifact — run the feature, read the actual value, inspect the diff — never against a proxy, a self-report, or a build pass. Apply after completing any task, before declaring done, and whenever reading evidence produced by a worker, a ticket, or a pipeline state. recommended: false --- # Principle: the real artifact Check the thing itself, not a proxy for it. "It compiles" is not behavior. "The worker said done" is a report. "State says green" is a claim. Only the running artifact answers the question. ## The rule | Claim | Proxy that proves nothing | The real check | |---|---|---| | The feature works | Build pass, typecheck, HTTP 200 on a health endpoint | Run the feature the way a user would; observe the output | | The fix landed | Worker's summary, ticket moved to done | `git log`, the diff at the head SHA, the gate run fresh | | The value is correct | A cached field, a derived metric, a screenshot of old state | Read the actual value from the live system | | The process is healthy | A status file, an uptime counter | Direct liveness evidence on the process itself | A model verdict is still a verdict. An agent's "I verified it" without an attached artifact is the same class of claim as a ticket's green state — treat both as unverified until the check has been re-run or its output read. ## Script the check when you can The strongest proof is a deterministic command that re-runs the same comparison, not a one-time eyeball. Write it, run it, keep the output. A reviewer should be able to re-run your evidence instead of trusting your word. Keep the artifact visible to the human. Commit it only when the trail must be auditable later — a migration, a port, a disputed delivery — per the project's own conventions. ## When the observation fails Suspect the observation method before suspecting the system. A timed-out probe, a stale cache, a detached worktree are all cheaper explanations than "the code is wrong." Re-establish a trustworthy channel first; then judge the system. Missing evidence is reported as missing evidence. It is never filled in with a plausible story.