Measurement at Birth: Every Action States Its Metric Before It Runs
A proposed action that cannot say what it will move, by how much, and by when is not ready to run. Measurement at birth is the rule that turns agent activity into an auditable track record instead of a busy dashboard.
Published August 12, 2026
Here is the rule, stated plainly: an action that cannot say what it will move, by how much, and by when is not ready to run. We attach that prediction to a proposed action at the moment it is drafted — its birth — not at review time and never after the outcome is known. That single constraint is what separates a system with a track record from a system with a busy dashboard.
The prediction is the contract
When a seat proposes something — pause this, raise that, send this — the proposal carries a metric, an expected change, and a due date. “Expect first-order conversion on this change to beat its trailing band within seven days” is a contract the action signs with the future. It is deliberately falsifiable. If the number does not move, the record will say so, and there is no later paragraph that can talk the result into a success.
Why up front and not after
Measurement after the fact is where honest reporting quietly dies. Given a result and no prior commitment, humans and models alike will find the framing that makes it look good. Science solved this problem with pre-registration — you declare the hypothesis before you run the trial — and the fix transfers directly. The prediction has to exist before the outcome is observable, or it is not a prediction; it is a rationalization with a timestamp.
Gates enforce it, they don't request it
Measurement at birth is not a best practice we hope seats follow. It is a precondition of approval: the one gate will not pass an action that has no success condition. The rule is machine-checked, not culturally encouraged, for the same reason the honesty rails are code rather than values — a discipline you can audit is a discipline, and anything else is a mood.
What it buys you: a real ledger
Predictions at birth plus verdicts at close is what makes an outcome ledger possible at all. Over weeks it compounds into something no activity feed can offer: hit rate, calibration, and cost per hit per seat. That is the substrate under the eleventh question — how do you know it worked — and the reason a seat can be promoted on evidence rather than optimism.
Questions founders ask
- What does "measurement at birth" mean?
- It means every action an agent proposes is created with three things attached: the metric it expects to change, the magnitude and direction of the expected change, and the date its verdict comes due. The prediction is written before the action is approved, so when the result arrives there is a fixed, un-edited claim to grade it against. An action that cannot state its own success condition is not approved — it goes back for a clearer bet.
- Why not just measure results after the fact?
- Because after the fact, the goalposts move. Without a prediction recorded up front, a mediocre result gets narrated into a win and a real win gets no credit, since nobody committed to what "good" looked like. Pre-registering the expected outcome is what makes the later measurement honest — the same reason experiments declare their hypothesis before collecting data rather than after.
- Doesn't this slow the agent down?
- Slightly, and on purpose. Forcing a one-line prediction costs the drafting seat a sentence and saves the operator from a class of actions that were never going to be checkable. In practice it also improves the proposals themselves: an action that has to name its metric tends to be a sharper action, because vague bets fail the test at the drafting stage.
- What happens when the verdict comes back a miss?
- It is recorded as a miss, against the number that was promised, and it counts. Misses are not deleted or reframed; they accrue to the seat's track record and inform whether it keeps or loses autonomy. A logged miss is priced as tuition — the cost of learning which bets do not pay — and it is exactly the data a system that only shows activity can never give you.