The oldest idea in Groundwork.

1939: the convoy problem

British naval intelligence in the Second World War had a version of the exact problem you have. Reports poured in from everywhere: trusted agents, intercepted signals, neutral captains, rumor mills in port cities. Some of it was gold. Some of it was bait. A convoy routed on the wrong report died. So they refused to treat information as one undifferentiated pile, and built a grading system instead.

It became known as the Admiralty Code, and NATO still uses its descendants today. Every source gets a reliability grade, A through F. Every piece of information gets a credibility grade, 1 through 6. A "B2" is a usually-reliable source reporting something probably true. An "F6" is an unknown source reporting something that cannot be judged. The report does not just say what was seen. It says how far to lean on it.

The idea kept getting reinvented

Because it is correct, the graded trust ladder keeps showing up wherever information volume outruns human review.

• Wikipedia runs a five-level ladder for its sources: generally reliable, no consensus, generally unreliable, deprecated, and blacklisted. The bottom level is mechanically enforced. You cannot cite a blacklisted source even if you try.
• Enterprise data teams mark their datasets endorsed, warned, or deprecated, and grade data bronze, silver, or gold as it earns validation. Analysts downstream know at a glance what a dashboard is standing on.
• Search systems have ranked "authoritative pages" above the rest for decades, so the intranet answer comes from the policy manual, not the lunch menu.

Intelligence agencies, encyclopedias, and data engineers all landed on the same shape: grade the trust, not just the content, and make the grade something the machine can act on.

The gap nobody filled

Here is the strange part. All of that machinery exists for spies, editors, and engineers, and none of it was ever packaged for the person who needs it most right now: an owner handing their business's files to an AI.

When you connect an AI to your documents without trust grades, you have rebuilt the 1939 problem exactly. The signed lease and the abandoned draft sit in the same folder, weighted the same. The policy you enforce and the idea you jotted at midnight read identically to the model. The AI is not careless. It is ungraded. It will route your convoy on whatever report it finds first.

What Groundwork does with it

Authority Classification, Document 04 in Order 2, is the owner's version of the ladder. Five levels, plain language, written into a plain text file the AI loads with everything else:

• Level 0, Core Doctrine: charter, principles, non-negotiables. Obeyed without asking.
• Level 1, Institutional Standards: approved policies and ways of working. Trusted unless doctrine overrides.
• Level 2, Validated Knowledge: executed contracts, cleaned data, reviewed procedures. The working library.
• Level 3, Working Material: drafts, notes, transcripts. Useful, never quoted as final.
• Level 4, Untrusted / Raw: clippings and unverified material. Context only, never truth.

The grades change how the system behaves, which is the whole point. Higher levels override lower ones. Level 3 gets flagged as not final whenever it is used. Level 4 is never quoted as fact. A convoy captain from 1939 would recognize the design at a glance.

If you have not built your ladder yet, it is one document with a template waiting in Order 2, and the scorecard will tell you how much the gap is costing you.