Assuming one reading and running with it is faster, and it also enlarges the cost of being wrong

What it is

The user says “book a table.” The system books two people, tonight at seven, at the nearby place, because that is the most common reading. The dialogue saved a turn. When it is wrong, the cost moves from “ask one more thing” to change-the-booking, apologise, maybe pay a cancellation fee. Execute-by-assumption wins latency and loses the cleanup after a misunderstanding has been written into the world. It is fast because it skipped joint attention. It is expensive because the skipped step was the chance to see the wrong road.

This is not a debate about whether to ask, or how often. It pins the trade: take interpretation out of the dialogue, and the cost of being wrong moves from language into the world.

Why it happens

Ambiguity is normal in language. People cheaply close it with the next turn. If a system treats the first reading as already grounded and calls a tool, repair has to walk the tool’s inverse. Inverses are often asymmetric: a booking can be cancelled; a sent mail, an outbound payment, a third party already notified, either has no inverse or is embarrassing.

The default reading also carries a frequency bias. The most common reading is best for the crowd and can be worst for this user (a table for eight booked as two). The more “worldly” the system, the more willing it is to skip the ask, and the more likely it is to drop a high-frequency script onto a low-frequency situation. What the user sees is decisiveness, not an assumption.

Studying it

Pick tasks with tool consequences (booking, sending, changing permissions) and instructions with two legal readings. Conditions: ask then act / act on the majority reading. Dependent variables: time to finish, rate of wrong execution, recovery cost after a wrong act (steps, whether irreversible), whether the user realised an assumption was made before the error landed.

Recovery cost must be measured on its own. Reporting only “task success” will score execute-by-assumption as a win — success mixes in lucky guesses. Split guessed-right from asked-and-known, or the enlarged tail of cost stays hidden.

Where it stops holding

Read-only generation that can be swapped at any time (another slogan line) has almost no world-layer cost; execute-by-assumption is a fair default. If the same user just succeeded under the same reading, with a short gap, reusing it is close to joint attention, not a blind assumption. In high-stakes domains, even a reading that looks unique should be spoken before acting; looking unique is not being unique.

Applying it

• For any act that touches a world boundary (outbound send, payment, changing someone else’s data), do not send an unconfirmed reading into a tool. Spread the reading that will be executed first.
• For pure generation that can be swapped in one click, allow the majority reading to run first, but mark the assumption next to the result in one human sentence so a wrong road is visible.
• Do not treat decisiveness as a demo of skill. Seconds saved by skipping a turn do not cover one wrong booking.
• Check: run the same line “book a table” against two account profiles (dinner for two / dinner for eight). Under execute-by-assumption, at least one profile books wrong; log the steps to change and cancel. After spreading the reading first, both profiles should see headcount before an order exists. An order with no visible headcount is an assumption written into the world.

Related

• Same group: L2.15.2 A follow-up is worth asking only when the ambiguity would lead to clearly different outcomes · L2.15.3 Too many follow-ups turn the exchange into a questionnaire and cancel the efficiency natural language was for · L2.15.4 Leading with a result and stating the assumption is a third path between asking and running · L2.15.5 Weigh the cost of clarifying against the cost of undo; if undo is cheap, do not ask
• Nearby: L4.07 Pre-action Confirmation · L4.11 Pre-action Confirmation and Irreversible Operations · L2.01 Openness of Natural-Language Commands and Its Cost
• Search terms: execute-by-assumption · grounding before tool use · cost of misinterpretation