Brand Probe

Find out what the book changes in an agent's work, by comparing each real task done with the book against a control done without it. Done when each task has its runs and a report.md with the current book-hash.

You observe: leave the book and the runs as they are, because a run that went badly is the finding and brand-book makes the fixes.

Run node brand/tools/brand.mjs status. Test brand/work/candidate/ if it exists, otherwise brand/, and note the book-hash it prints. If the candidate has errors, hand off to brand-book; if there is no book, to brandfile.

Tasks

If probes/ in the book's folder already holds tasks, keep them, so versions stay comparable. Where a report is missing or carries another hash, rerun the with-book run and the report; the control does not depend on the book.

Otherwise take two or three real tasks from "Readers and work" in brand/work/brief.md, or from {#start} in the book once the brief is archived; if neither names one, ask the person. Word each as they would ask for it, with no brand rules in it, and pick tasks whose result is a page or a text.

Add one gap task: something real the book does not cover. It shows whether an agent says the book has no answer or fills the gap with fashion.

Save each as probes/<task-id>/task.md in the book's folder.

Runs

You have read the book and the brief, so give every run to a new subagent with its run brief. Without subagents, ask the person to paste each brief into a new session, and continue once the files exist.

Each real task runs twice: with the book and as the control. The gap task runs only with the book. The folders are not sealed, so each run replies with the files it opened: run again any run that opened a file its brief does not allow.

Report

For each task, a new subagent writes probes/<task-id>/report.md from the report brief.

Hand back

Run node brand/tools/brand.mjs build so the runs appear side by side on the page. For a candidate, hand off to brand-book. For an accepted book, tell the person what strayed and what the book changed, and name the page. Report only what you saw: a few tasks on one brand catch failures and prove nothing general.