Sit in enough operating reviews and you notice that arguments rarely turn on logic. They turn on two people holding different beliefs about the same sentence — one hearing a measurement, the other hearing a guess — and neither noticing that this is the disagreement.
The fix is unglamorous. Give every claim in your company one of three labels, and require the label to travel with the claim wherever it goes.
What you’ll learn
- The three labels and the test that separates them
- Why unlabelled numbers degrade as they travel
- How to introduce the vocabulary without adding process
- What changes once software can read the labels too
The three labels
Fact. Something measured, with a source and a timestamp. Cash in the account this morning. Invoices issued last month. Headcount on the payroll run. A fact can still be wrong — the source can be broken — but it is wrong in a way you can go and check.
Assumption. Something believed, chosen, or inherited, which nobody has measured. Churn will stay roughly where it has been. The two open roles will be filled this quarter. The integration work is about two weeks. Assumptions are not defects. Every plan rests on them. The defect is an assumption wearing a fact's clothes.
Prediction. A statement about a future state, made by a person or a model, with a stated range and a stated basis. Not "revenue will be 400" but "revenue lands between 340 and 430 under these assumptions, on this model version".
The test that separates them is one question: what would I have to do to check this? If the answer is "look it up", it is a fact. If the answer is "wait", it is a prediction. If the answer is "nothing — we just decided it", it is an assumption.
Numbers degrade as they travel
A figure starts its life properly caveated. An analyst says "roughly 30, assuming the two pending deals close". It goes into a deck as "~30". The deck goes into a summary as "30". The summary is quoted in a planning session as the basis for a hiring decision.
Nothing dishonest happened at any step. Each step just dropped the cheapest thing to drop, which is the qualification. By the end the assumption has become load-bearing, and the people leaning on it have no idea they are leaning.
Labels stop the erosion because they are attached to the claim rather than to the sentence around it. An assumption that stays labelled as an assumption for four hops is still arguable at the fourth hop. That is the entire mechanism.
Tip: The cheapest version of this discipline costs nothing and starts on Monday. In every document that carries numbers, prefix each one with F, A, or P. People will argue about the labels. That argument is the value.
Why this is a precondition for AI, not a nicety
A company state model that carries sources, freshness, and confidence on every field is this discipline made structural. The label stops being a convention people maintain and becomes a property of the data.
Once it is a property, three things become possible.
Software can decline. A system that knows a field is an assumption with low confidence can refuse to act on it, ask, or widen its output range instead of producing a confident answer built on sand.
Predictions can be graded. A prediction that was recorded with its range, its basis, and its model version can be compared against what actually happened. Predictions carry ranges and a model version, and they are preserved unchanged when reality arrives — which is what makes grading meaningful rather than retrospective narrative.
Ranges can stay honest. Ranges stay honest-wide until calibration evidence justifies narrowing them, and a prediction that cannot be made honestly is an abstention rather than a guess. That policy is only implementable if the system can tell the three kinds of statement apart in the first place. Ranges beat certainty makes the case for why the wide band is the useful one.
Introducing it without adding process
The failure mode here is turning epistemics into a ceremony. Three moves keep it light.
Label at the source, not in review. The person who produces a number labels it. Nobody audits labels in a meeting. If a label is wrong, whoever notices changes it and says why.
Make assumptions nameable. Assumptions that recur — pipeline conversion, delivery velocity, hiring lead time — should be named once and referenced, not restated. Then when one changes, everything resting on it can be found.
Attach a review trigger, not a review meeting. Each significant assumption gets a condition that would make you revisit it, rather than a calendar slot. "Revisit if two consecutive months come in below the band" beats "revisit quarterly", because reality does not respect quarters.
What it feels like when it lands
The visible change is that meetings get shorter and slightly more uncomfortable. Shorter because a large share of every operating discussion is people establishing what kind of claim they are looking at, and the labels do that in advance. More uncomfortable because the count of load-bearing assumptions in a normal company is higher than anyone expects, and seeing them listed is a specific experience.
The second change is slower and matters more. When assumptions are named, they can be tested. When predictions are recorded with ranges, they can be scored. A company that does both starts accumulating a record of how well it understands itself, which is the only durable input to deciding better.
If you want the mechanism, company state is where the labels live as data, and the Business World Model is what consumes them to model futures. The vocabulary comes first, though. It works on a whiteboard, and it works before you buy anything.
Brainis Team
Notes on the company loop — company state, decisions, governed autonomy and verified work — from the people building Brainis and running on it.