- In short
- Choosing code execution over prose generation for a numeric result is both a verification technique and a format decision: it determines what kind of output the reader receives. A prose-generated number is a fluent estimate of an uncomputed calculation; a code-executed number is the actual re-runnable result of a computation over the real data, and can also return supporting artifacts such as charts. When a number must be trustworthy, the format choice is code execution.
The format choice that decides trust in a number
For numbers, the format choice is a trust choice. The Claude Certified Associate - Foundations (CCAO-F) exam frames the decision between code execution and prose generation as a format decision, because it determines what kind of output the reader actually receives: a fluent estimate, or a computed, traceable result. When a figure must be trustworthy, this is the choice that makes it so, and it belongs in task statement 2.6 as a format decision, not only in the fact-checking task statement as a verification technique.
The two paths produce numbers that look identical on the page and differ entirely in what stands behind them. A prose number is the model's guess at a calculation it never ran; a code-executed number is the result of the calculation actually running over the data. For a consequential figure, that difference is the whole point of the format choice.
- Code execution vs prose generation
- A format decision for numeric results that also serves as verification. A prose-generated number is a fluent estimate of a calculation the model did not perform step by step. A code-executed number is the actual, re-runnable result of a computation over the real data, and can return supporting artifacts such as charts. When a number must be trustworthy, the format choice is code execution, because it determines what kind of output the reader receives.
A prose number is a guess in the shape of an answer
When Claude writes a figure in prose, it produces the most plausible-looking number rather than the result of an actual computation. It did not sum the column or run the model step by step; it estimated. That estimate might be right, but it is not traceable to the data and cannot be re-run, so a wrong figure has nothing to catch it and propagates into everything built on it. This is the same risk as a fabricated specific: a precise-looking number with no underlying computation to back it. For a low-stakes gut-check the prose path may be fine; for a reported figure it is the wrong format.
A code-executed number is a computed result
Code execution runs the real calculation over the actual data and returns the computed result, along with, where useful, supporting artifacts such as charts and processed files. The number is anchored to the rows and operations that produced it, so the reader receives something they can read, re-run, and check rather than a sentence they must take on faith. Choosing this format changes what lands in front of the reader: not a claim about a number, but a computation whose output is that number. That is why it is the reliable format for figures that matter.
Format and verification in one choice
The reason this sits in the format task statement is that the choice determines the output's kind, not just its correctness. Selecting code execution is selecting a different deliverable, a traceable computation with an exposed method, over a prose assertion. It is simultaneously a verification technique, as covered in code execution for numeric verification, and a format decision, because the reader receives a fundamentally different artifact. The one caution carried over from the verification view: the guarantee is traceability, not automatic correctness, since Claude wrote the code and it can still contain a bug.
What the CCAO-F exam trips candidates on
The first trap is choosing a prose or table format for a number that will feed a downstream financial or regulatory decision. A consequential figure delivered as an unverifiable estimate is the wrong format for its stakes. The credited answer chooses code execution so the number is computed and traceable when the decision it feeds matters.
The second trap is assuming that any number presented in a clean table is more reliable than one presented in a sentence. A tidy table can still contain a prose estimate; the presentation does not compute the number. The exam rewards judging reliability by how the figure was produced, code execution over real data, not by how neatly it is displayed.
Worked example
A quarterly revenue total that will go into a regulatory filing needs to be produced from an uploaded spreadsheet. Claude offers it two ways: as a sentence ('total revenue was about $4.7 million') and as the same figure inside a clean, well-formatted table. Which is the right format, and why is the table a trap?
Neither the sentence nor the table, as offered, is the right format, because both are prose-generated numbers, and this figure is headed for a regulatory filing where it must be trustworthy. The sentence is obviously an estimate: "about $4.7 million" is the model's plausible guess at summing a revenue column it did not actually add. The table is the subtler trap. Presenting the same guessed number inside a tidy, well-formatted table changes only its appearance, not its origin. The number was still estimated, not computed, so the table's neatness lends unearned reliability to a figure that has none of the underlying computation a regulatory total requires. Assuming the tabled number is more reliable because it looks cleaner is exactly the second trap.
The right format is code execution. Have Claude write and run code over the actual spreadsheet, returning the computed total, ranked components if needed, and optionally a supporting chart. Now the reader receives a different kind of artifact: a traceable computation whose output is the total, anchored to the rows that produced it and re-runnable to confirm. For a figure feeding a regulatory decision, that is the format the stakes demand, and choosing the prose sentence or the prose-in-a-table over it is the first trap.
One caution remains: code execution guarantees traceability, not automatic correctness, because Claude wrote the code and it could contain a bug. So the computed total still gets its code reviewed and its result sanity-checked, and because a regulatory filing is a do-not-ship-without-review category, the computed figure is the reliable input to that human review, not a replacement for it.
Common misreadings to avoid
Misconception
A number in a clean, well-formatted table is reliable.
What's actually true
Misconception
Prose is a fine format for any number as long as it looks reasonable.
What's actually true
How this shows up on the exam
Domain 2 questions on this knowledge point present a consequential number and ask which output path or format is most trustworthy. The dependable answer chooses code execution for figures that must be trusted, treats the computed result as traceable but still to be checked, and does not mistake a clean table for a reliable number.
This knowledge point is the numeric case of output format as a reliability decision and the format-facing view of code execution for numeric verification. It guards against the guessed precision of spotting fabricated specifics and combines with input handling in choosing format and curation strategy for a given task.
A quarterly revenue total headed for a regulatory filing can be produced as a prose sentence or as the same figure inside a clean table. Which format is trustworthy?
People also ask
When should you use code execution instead of prose for a number?
Why is a prose number less reliable?
Is a number in a clean table more reliable?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.