# Cross-check

You do not get to call your own work done. The value is the difference in context. A second reader who is given your reasoning will agree with you. A second reader who is given the claim, the files, and none of your story will find what you stopped seeing.

This is that second reader. They do not fix. They write one report.

Commands in this file are relative to a folder named `beargrassai` on the machine that runs them. If you do not have that folder, you cannot run the commands. Do not invent a home path.

## Who you are while you run it

You are the author, about to declare something finished. You are not allowed to. You spawn a Validator in this same thread, hand them the claim, and wait. Until their report is back you may not write done, complete, fixed, closed, or verified. Say swept, unvalidated.

The Validator is a no-memory sub-agent. No sheet. No doctrine. No pad URL. They are not a named person on this pad. They are not a hatch. Mark lays pad hatches; you do not open the operator for them, you do not fetch a pad URL for them, and you do not invent an avatar named validator.

Skills and tools are files you open. PicViewer, MeatCheck, and the rest are commands. Do not spawn them. The only spawn is Validator.

## How you spawn them

In this same Cursor or Claude Code thread, start a Task / sub-agent.

- Cursor: Task tool, write-capable (`generalPurpose` or equivalent). Not a read-only explore agent.
- Claude Code: the same Task / sub-agent tool.

They get no sheet, no doctrine, no memory, no kit, and not this skill.

You give them only:

1. What was done. Concretely.
2. Where done is written. Paths from the station root, not your summary. Work with no written expectation cannot be validated.
3. Extra traps: what is live, what is uncertain.
4. The mandatory line below.

Prompt them as a cold reader:

```
VALIDATOR
No identity. No pad. No doctrine. Do not fix. Write one report.

What was done:
<concrete>

Where:
<paths from the station root>

Traps:
<live / uncertain>

Check the claims listed. Then go past them. My list was written by the same judgment that produced the defect. Report what you find beyond it separately.
```

## After the report

Verify every HARD finding yourself before you act. Then close it with a check, not a claim. A HARD finding is something that will break live service, lose money, lose data, or expose a live credential. Everything else is SOFT. Unlabeled alarms are SOFT.

If another *seat* (a named person Mark talks to) needs to look, tell Mark. That is not this recipe.

## Temptations (refuse these)

- Validating yourself with a paragraph that restates your reasoning.
- Handing them a story instead of paths.
- Letting them fix. They write one report.
- Calling the spawn a hatch, or laying an avatar for them.
- Writing done because the spawn is in flight. Until the report is back, the work is unvalidated.