Method

What would count as evidence?

How I tell claims, observations and outcomes apart.

Kind

Editorial

Status

Working

Published

17 October 2026

Revised

20 October 2026

Reading time

About 3 min

A persuasive output is a starting point. I want to understand what happened when it was used.

Separate the claim from the observation

I distinguish source research, first-hand observations, interpretations, and hypotheses. An account of one operation does not establish a general rule about organisations.

Count the human work

Setting direction, designing workflows, reviewing outputs, and rescuing failed execution are different contributions. Each belongs in an honest account of how a system operates.

Zero employees does not establish zero human labour. Founder effort, external providers, and supervision must remain visible.

Describe authority precisely

Role names such as finance or marketing do not tell us what a system may decide. Case studies should specify its information, permissions, approval boundaries, and escalation paths.

Use the agent delegation framework to examine duties, decision rights, answerability and operating evidence consistently across roles.

How I place a deployment

The reality checks record the deployment behind each claim, an organisation and a function, and I place each one on two axes: how much authority the organisation has handed to its agents, and how far the proof has got. These are the tests I apply, written down so that two readers working from the same sources land on the same place.

Authority handed to agents, from left to right:

• Prepares. The agents gather, analyse, draft or run scenarios. Nothing they produce takes effect until a person acts on it.
• Acts with approval. A source names a step the agents carry out in a system other people rely on, and names the approval that releases each one.
• Acts alone. A source names a step the agents complete without anyone approving that instance, and the limits that apply to it.

Proof, from bottom to top:

• Announced. The sources describe what the agents are to do: a plan, a target, a platform, or a sentence saying agents are in use with no work named and nobody named who uses it.
• In use. A source describes work the agents do now in that function, and the people who use the result or a worked example from practice.
• Scaled and measured. The use reaches beyond one team or one process step, and a figure for reach or effect is published for a stated period with a stated method.

Three rules settle the awkward cases. I place the agents the claim is about: where an organisation also says agents work elsewhere in the same function, I record that sentence and leave the position where the described work puts it. A figure counts as a measurement whether the organisation measured itself or somebody else did, because who measured it, and against what baseline, belongs on the claim page and not in the position. What the sources do not describe is not disclosed: an axis with nothing to place it on stays empty, and the deployment appears in the record without a dot.

The four corners of the graph name what the combinations amount to: announcing, assisting at scale, delegating in pilots, delegating at scale. A deployment on a middle step sits on a dividing line, between two corners rather than in one. Deployments that share a position share a cell, and the graph spreads them inside it in alphabetical order, which carries no meaning of its own: the position is the cell. The tags that describe the organisation, its sector, its size band and its region, come from the organisation’s own published reporting. The position and the one sentence that supports it come only from the sources on the claim page.

Follow through to outcomes

Each case should record the mandate, operating conditions, actions, human interventions, costs, outcome, and unresolved limitations. Failed attempts and changed conclusions are part of the record.

Keep a revision history

When evidence changes our view, we will explain the change. Research becomes useful when readers can see both the current conclusion and the reasons behind it.