Choose evidence that tests the actual claim
Work through arithmetic, citations, JSON types, held-out evaluation and sample percentages using fresh invented examples.
Five matched answers sample checking decisions; they do not certify that every future AI output will be assessed correctly.
Recompute from the supplied inputs
Suppose a fictional table contains only 18 and 24, while a draft reports 45. The supported sum is 42, because 18+24=42. The source values and operation are enough to check that narrow statement. Keep the source intact and correct the draft rather than changing the evidence to fit the answer. This example does not establish anything about a named model’s average error rate; it demonstrates how a deterministic witness can resolve one specific mismatch.
Check existence and support separately
A reference can exist without supporting the claim attached to it. Read the actual source and identify the relevant passage, population, conditions and outcome. If a fictional draft claims a study proves a universal benefit but the source describes a small, specific task, the mismatch is in the strength and scope of the claim. A long bibliography or plausible title is not a substitute for this comparison. Record what the source supports and what remains unverified.
Distinguish parseable JSON from a matching contract
For an invented requirement that count must be an integer, {"count":"6"} contains a string and does not meet that type requirement. {"count":6} supplies an integer value. Both texts can be valid JSON, so parsing alone cannot establish compliance with the stated structure. Check expected fields, types and ranges separately from syntax. This is a small data-validation example rather than an instruction to trust any generated object automatically or a claim about a particular vendor’s structured-output feature.
Keep evaluation data from becoming the answer key
Testing a classifier on its training examples can conceal poor performance on new inputs. Separate relevant held-out examples provide a stronger check of generalisation, but repeatedly tuning to the same final test results can undermine that separation. Distinguish training, validation and final evaluation roles. A fresh invented exercise can label which examples were used for each purpose without training a real model. Relevance to the intended setting still matters; a technically separate but unrepresentative test can leave important conditions unexamined.
Attach the denominator and setting to a percentage
A fictional evaluation with 21 correct cases out of 24 gives 87.5% on those cases. It does not establish that every future task has that success rate. Report the denominator, task selection, conditions and misses alongside the number. A result on a small handpicked set can be useful for finding errors while being insufficient for a broad performance claim. The online item checks this evidence boundary rather than assigning you a percentile or measuring a real model in operation.
Describe the failure precisely enough to reproduce the check
For a new exercise, save the fictional input, generated claim, expected result and the method used to check it. Distinguish arithmetic disagreement, unsupported citation, type mismatch and evaluation-scope overreach. Each needs different evidence. This record can be short and local; it does not require publishing an answer sheet or uploading private material. The useful question is whether another reader could follow the same supplied facts to understand why the correction is supported.
Use five choices as revision targets
The section’s correct count describes matching responses to five stated checks. It does not establish domain expertise, universal factual accuracy or a professional credential. All counts receive every explanation and all four full guides. Change the inputs and explain a fresh correction before retaking if you want a more informative exercise. Remembering these exact keys can improve a repeated count without showing how you would evaluate an unfamiliar claim. The website executes no AI evaluation service or private document analysis.