AI Glossary / Generative AI and language / Beginner

One right answer

Definition

One right answer, often abbreviated ORA in generative-AI evaluation discussions, describes a prompt with a single correct answer under its stated assumptions. A specified arithmetic problem or a question asking for one field from a supplied record can have this property. The term concerns correctness of the requested result. It does not necessarily require one exact wording, and it does not guarantee that the model will produce the answer correctly.

What it means

An ORA task gives evaluators a relatively clear target. If a prompt asks for the sum of 12 and 8, the numerical result is 20. “20,” “twenty,” and an appropriate sentence containing that result can express the same answer, unless the prompt requires a specific format.

The assumptions must be complete enough to make the answer unique. “What is the best delivery option?” is not an ORA question without criteria, constraints, and available options. “Which listed option has the lowest price?” may be, provided there is no tie and the prices are defined.

This is a task-description label, not a special model type. The same assistant can face both ORA questions and open-ended tasks with multiple acceptable responses.

How it works

Specify the question, information source, assumptions, and required output form. Establish the correct reference result through a reliable process, such as a verified calculation or an authoritative supplied record.

Evaluate the model's response against that result. For structured outputs, parse the requested field and apply a clearly defined comparison. For free text, allow legitimate equivalent expressions while checking that the response does not contradict itself or add unsupported claims.

Separate answer correctness from other requirements. A response may contain the right number but use the wrong units or violate a required format. Decide in advance how such cases are scored. If ambiguity, missing information, or multiple equally valid results are discovered, revise the task or its evaluation instead of forcing one arbitrary answer to appear uniquely correct.

Three practical examples

1. Record extraction: an assistant identifies the order number from a supplied confirmation containing exactly one clearly labeled order number.

2. Arithmetic: a model calculates a specified total from a small list of quantities, with units and rounding rules explicitly stated.

3. Closed comparison: an application selects the earliest date among a provided set when one date is unambiguously earliest.

Think of it this way

Think of checking a key against a lock with one required pattern. The result has a definite criterion, even if the key can be described in several ways. The analogy is limited because language prompts can contain hidden ambiguity, and the reference answer can itself be wrong. A clear-looking question still needs careful specification and a trustworthy checking process.

Worked example

Consider a fictional prompt: “A carton contains 12 packs, and each pack contains 8 labels. How many labels are in the carton? Return an integer only.” The verified calculation is 12 × 8 = 96.

Response A is “96”. It has the correct value and satisfies the requested format. Response B is “There are ninety-six labels.” Its meaning is correct, but it fails the integer-only output requirement. Response C is “96 packs.” It introduces the wrong unit. Response D is “108.” It is numerically incorrect.

An evaluation could therefore record answer correctness and format compliance separately. If the product requires machine-readable integers, a combined acceptance rule might require both. That rule should be declared, not improvised after results are seen. The example shows that one correct underlying result does not remove the need to define equivalence, units, and output constraints.

Illustrative case study

Fictional case: Ashford Orders tests an assistant that extracts shipment dates. Its initial dataset contains both an order date and a dispatch date, but some prompts ask only for “the date.” Reviewers disagree because the question is underspecified.

The team rewrites prompts to identify the required field and checks the reference labels. It then evaluates exact date values after a documented normalization step. The case illustrates that a reliable ORA benchmark depends on clear questions and sound references, not merely on storing one answer per row. It does not report a real company's results.

Common misconceptions and mistakes

One reference answer does not prove that only one answer is valid. A dataset can contain incomplete references or ambiguous prompts. Treat unexpected valid answers as potential evaluation issues rather than automatic model failures.

Do not confuse one correct result with one exact string. Equivalent number formats, dates, or wording may need normalization, while a strict application format may legitimately require exact compliance. Those are different decisions.

ORA also does not mean easy. A complex proof or calculation can have a determinate result while remaining difficult. Conversely, a short subjective request can allow many answers. Finally, a correct final value does not establish that every explanation accompanying it is correct or supported.

Business use

ORA tasks support clear acceptance checks for extraction, calculations, structured transformations, and fact retrieval from defined sources. They can provide useful automated regression tests when the target and comparison rules are stable.

Maintain the provenance of reference answers and review edge cases such as missing fields, ties, and conflicting records. Report content accuracy and formatting failures in a way that helps diagnose the system. When facts can change, define the relevant date or source version so an older reference does not silently become an incorrect standard.

People also ask

Can several strings express one right answer?

Yes. Equivalent wording or formatting may express the same result, unless the task specifically restricts the output form.

Does an ORA prompt always allow automatic evaluation?

Often it does, but reliable parsing, normalization, and reference construction may still require care or human review.

What if two listed options tie for the requested best value?

Then the prompt may need a tie-breaking rule or allow both. Without that, treating one option as uniquely correct is unjustified.

Check your understanding

A benchmark asks for the lowest price among three products, and two products both cost 10 units. Its reference names only one product. Is the task clearly ORA as written?

Answer: Not if it asks for the product and provides no tie rule. Both cheapest products are valid. Clarify the question or permit the complete set of correct answers.

Related terms

• No one right answer
• Exact match
• Golden response
• Data Quality
• reference answer
• structured output
• evaluation rubric
• factual accuracy

Learn next

Write three seemingly simple questions with unique reference answers. Check each for hidden assumptions, units, ties, and equivalent representations. Revise any question whose supposed uniqueness depends on information the prompt does not provide.

More in Generative AI and language

• Multimodal Transformer - A multimodal Transformer is a model that uses Transformer-based processing to work with information from more than one modality, such as text and images, or...
• Natural language generation - Natural language generation, or NLG, is the production of human-language text by a computer system.
• Negative sampling - Negative sampling is a training technique that contrasts observed positive examples with a selected set of negative examples instead of comparing against...
• Oversampling - Oversampling increases the representation of selected examples or classes during training, commonly to address a class-imbalanced dataset.
• Parameter-efficient fine-tuning - Parameter-efficient fine-tuning, or PEFT, adapts a pretrained model while training substantially fewer parameters than full fine-tuning.
• Personal access token - A personal access token, or PAT, is a secret credential that software can present to a service when making authenticated requests.

Sources

Google, Machine Learning Glossary, one right answer: https://developers.google.com/machine-learning/glossary#one-right-answer

The primary glossary supplies the ORA label and its contrast with NORA. Calculations, examples, fictional case, and exercise are original teaching material. The discussion distinguishes unique semantic results from exact-string and format requirements.

Previous: One-hot vector Next: One-shot learning