---
schema: "bureau.instrument/v1"
id: bureau-instrument-002
title: "The Upstairs Audit"
kind: "Diagnostic"
version: "1.0"
status: "active"
execution: "sequential"
agent_ready: true
source: "https://thefantasticbureau.com/briefs/the-intelligence-stack/"
publisher: "The Bureau of Fantastic Strategy"
language: "en"
---

# The Upstairs Audit

> Issued Instrument 002 · Diagnostic

## Objective

A map showing whether the system is merely clever or becoming peculiarly yours.

## Use when

A model demo looks impressive, but nobody can explain where durable advantage will accumulate.

## Required inputs

- The AI system or workflow being assessed.
- The context, standards, and workflow position currently available to it.
- Examples of human corrections, decisions, and outcomes.

## Machine execution rules

- Check the required inputs before starting. Ask only for inputs that are missing.
- Follow the procedure in order. Do not silently skip a step.
- Use supplied evidence and named sources. Label assumptions and inferences.
- Do not invent research, quotations, people, statistics, or certainty.
- Return only the required output, followed by a short list of unresolved questions.
- If the evidence cannot support the requested outcome, stop and state what is missing.

## Procedure

1. Score model, context, judgment, workflow, and learning from 0 to 2: absent, partial, or working.
2. Ignore the model floor for a moment. Circle the weakest floor above it.
3. Name one correction, decision, or outcome that should be captured there but currently disappears.
4. Design a thirty-day change that keeps that information and puts it back into the work.

## Required output

- A five-floor scorecard covering model, context, judgment, workflow, and learning.
- The weakest owned floor above the model.
- A thirty-day intervention and the signal that will show whether it worked.

## Bureau voice

- Be precise, accessible, imaginative, and formally understated.
- Use short sentences, concrete nouns, and evidence that can survive daylight.
- Keep one consequential choice visible. Remove hype, consultancy fog, and generic AI phrasing.
- Permit one useful oddity. Do not perform whimsy.
- End with a decision, test, refusal, or next move.

## Completion check

- Every required input was used or explicitly marked missing.
- Every procedure step was completed in order.
- Claims are separated from assumptions.
- The output contains one consequential choice and one next move.
- No invented evidence or decorative certainty was added.

---
The Bureau of Fantastic Strategy · https://thefantasticbureau.com