Specify Record Samples Before AI Review

Before asking an AI assistant to review a sample of records, specify what the sample is a sample of. A list of selected rows does not by itself identify the source population, the selection rule or the question the review is meant to answer. Ten well-chosen examples may teach a boundary clearly while providing no basis for estimating how frequently that boundary occurs. A randomly selected set may answer a different question while missing the one difficult case an implementer needs to examine.

This guide separates the sampling frame, selection positions, record identities and interpretation of results. It uses original fictional arrays and manually authored selections, not observed model outputs or client data. No random draw, model evaluation, database query or production import was performed. The downloadable CSV and JSON support narrow consistency checks of those teaching examples. They do not supply an executable sampling utility or certify a statistical study.

Define the frame before defining the selection

A frame is the actual ordered list from which a selection is made in this exercise. Describe its source snapshot, inclusion rule and row meaning. A business population and an available export may differ. For example, an export might contain only records with a supplied status or only a chosen date window. Do not call the export every record unless the available evidence supports that scope. The distinction remains important even when every selected row is read correctly.

Write what one position represents. It might be one transaction, one occurrence of a message or one entity with several observations. If repeated observations exist, selecting rows is not necessarily selecting distinct entities. Keep the row identifier and any entity identifier as separate fields. A later deduplication step changes the selected material and needs its own explicit rule rather than being treated as harmless cleanup.

The Python3.12 random module documentation distinguishes sampling without replacement from choices with replacement. It also describes reproducibility in relation to generator state and notes that algorithm behavior can vary across versions. These are interface concepts, not evidence that a real dataset has been sampled representatively. Our selections below are deliberately written into the fixture rather than produced by a random generator, so no claimed random outcome needs to be reproduced.

Specify positions without confusing them with record labels

Fictional Case A defines an ordered frame containing K1 through K10 and uses zero-based positions one, four and eight. Those positions select K2, K5 and K9. There are three selected positions out of ten frame positions, giving a selection fraction of three tenths. The fraction describes how much of this authored list was selected by position. It does not measure semantic coverage, model accuracy, business completeness or the probability of a future error.

The difference between a zero-based position and a printed record label is worth making visible. Position one is the second element in this particular array; it is not the record named K1. A prompt that says review rows one, four and eight may be ambiguous if a spreadsheet's visible numbering and a program's array positions follow different conventions. State the convention and show at least one selected identity as a check.

Keep the order of the selected positions when order matters to the task. A review might use the source order, the order in which selections were supplied or another declared order. Those are separate choices. Sorting selected identities alphabetically may make a table easier to read while changing a sequence-sensitive exercise. Record any such transformation explicitly rather than allowing a formatting step to alter the intended review material.

For an AI handoff, include the selected identities with the source snapshot identifier. A bare list of positions becomes difficult to interpret if the frame later changes. If the source system supplies stable row identifiers, preserve them along with positions when appropriate. This guide does not establish which identifiers are stable in a reader's actual system; that question needs evidence from the relevant source contract.

Distinguish replacement from repeated identity

Fictional Case B uses the same ten-position frame and an authored selection list of one, one and four. The selected sequence is K2, K2 and K5. There are three selection occurrences but only two distinct selected positions and two distinct labels. This illustrates a replacement-style selection description. It is not an output from Python choices or an actual random process, and it does not establish a sampling probability.

Repeated selections are not automatically mistakes. They can be meaningful under a declared with-replacement design. But if the task expected distinct positions, the repeated position indicates a mismatch between the task and the supplied selection. Do not silently remove the second K2 and claim the original three-selection design has been satisfied. Removing it changes both the occurrence count and the material the assistant receives.

A count labeled sample size therefore needs a definition. It could refer to draw occurrences, distinct frame positions or distinct business entities. Case B has different answers under those definitions. The difference affects what a later summary means even before any model reads the records. Write the intended count alongside the selection rule rather than relying on a familiar term to carry an unstated interpretation.

If a repeated record receives two different AI answers, that would create another evaluation question, but no such answers were collected here. Do not infer repeatability or inconsistency from the authored repetition alone. The fixture establishes which material would be presented under the supplied sequence. A claim about model behavior would require actual runs, their conditions and an appropriate comparison method.

Keep row identity separate from entity identity

Fictional Case C contains three row keys, row-A, row-B and row-C. Their entity labels are E1, E1 and E2. Selecting all three row keys gives three rows and two distinct entity labels. Neither count is wrong; each answers a different question. The example does not claim that two real records belong to the same person or organization. The repeated entity labels are deliberately assigned in a fictional teaching table.

Suppose the intended review concerns whether each row preserves a supplied date. All three rows might be relevant even when two share an entity label. If the intended review concerns one consolidated entity description, the aggregation rule needs to be specified first. It cannot be inferred from the fact that an identifier appears twice. Keep the review unit visible before choosing an appropriate selection unit.

Do not collapse the two E1 rows by keeping whichever one appears first unless that behavior has been explicitly chosen for the task. The rows could represent different observations, states or source fragments. Their shared label alone does not establish which should survive. A selection exercise is not permission to resolve business identity, overwrite data or decide which observation is authoritative.

When reporting a review result, name both the unit checked and the number checked. A statement such as three examples inspected is incomplete if the audience assumes three independent entities. A clearer teaching statement is three supplied rows representing two authored entity labels. That wording preserves the evidence without turning the small fixture into a claim about independence, real identity or general performance.

Make targeted coverage different from population proportions

Fictional Case D defines a frame with twenty Cedar-labeled rows and eighty Maple-labeled rows. A deliberately authored review set selects two rows from each label. The frame shares are one fifth Cedar and four fifths Maple. The selected set's shares are one half and one half. The selection was constructed to include both labels, not to reproduce their population proportions.

This can be useful for teaching two branches of a rule. An implementer may need to see examples from both categories even if one is less frequent in the frame. But the resulting equal-count review set should not be described as a proportionally representative sample. Its category composition was chosen for the exercise. State that purpose explicitly so a later reader does not interpret the balanced set as a discovery about the source population.

Do not pool hypothetical category outcomes and call the result overall model accuracy without an appropriate evaluation design. No outcomes exist in this example. Even if outcomes were later supplied, the selected composition and the intended population would remain relevant to interpretation. This article does not propose a performance estimator, confidence interval, sample-size recommendation or statistical guarantee for a real workflow.

A useful output can report the two compositions side by side before any evaluation narrative. The frame contains one hundred authored rows, and the selected set contains four. Its selection fraction is four hundredths, or one twenty-fifth. Those arithmetic facts establish scope. They do not establish that four rows are enough for a particular decision or that the selected examples cover all important variations within either label.

Preserve the frame order when reproducing a selection

Fictional Case E defines frame version V1 as K1, K2, K3, K4 and frame version V2 as K4, K3, K2, K1. The authored selection positions are zero and two in both versions. V1 yields K1 and K3; V2 yields K4 and K2. Identical positions do not imply identical selected identities when the ordered source list changes. No actual generator was used to produce either pair.

A seed alone cannot identify the complete material reviewed in such a case. A reproducibility note should include the source snapshot, ordered frame, selection method, relevant parameters, implementation context and the actual selected identities when the task calls for them. This is a practical recording suggestion, not a claim that any particular software function will return the same result across every version or environment.

Keep revisions separate rather than replacing the old frame in place while retaining an old result label. If the available export changes, identify the new snapshot and select or reconcile the intended material under the appropriate process. A previous result belongs to its original inputs. Reusing its title on a changed frame can make an evaluation look repeated when it actually reviewed different records.

The same issue can arise before sampling if a query has no defined ordering. Two exports with the same membership may arrive in different orders. This guide has not executed such a query or identified an ordering defect in any system. The fictional reversal simply makes the relationship between positions and identities easy to inspect without relying on a complex database example.

Choose boundary examples for the rule they exercise

Fictional Case F describes an authored rule that accepts a numeric value only when it is at least ten and less than twenty. The supplied candidate values are nine, ten, nineteen and twenty. Under that rule, ten and nineteen are accepted; nine and twenty are rejected. These four values deliberately exercise the two endpoints and their nearby alternatives. They were not drawn randomly and do not estimate how often those values occur.

A targeted boundary set can reveal a misunderstanding that a small random set might never include. Its purpose is to distinguish candidate implementations or interpretations of a rule. But passing that set does not establish coverage of every input type, missing state, numerical representation or application behavior. The teaching rule uses exact supplied integers and does not claim a complete validation specification for a real process.

Keep a record of why each targeted example exists. Nine tests the lower outside case, ten the included lower endpoint, nineteen an interior accepted value and twenty the excluded upper endpoint. That explanation helps a reviewer tell whether a future revision still exercises the intended distinctions. Replacing an endpoint example with another interior value can make the list look similar while removing an important test.

Do not blend targeted examples into a randomly selected evaluation set without explaining the composition. A combined review can be useful when its components are reported separately. The important point is that the examples answer different questions. One component may teach the declared boundary; another may describe selected records from a frame. Neither role should be hidden inside a single unexplained sample-size number.

Separate selection evidence from review evidence

A selected-record list establishes which inputs were intended for review. It does not establish that a model processed every selected input, returned a complete response or preserved every relevant field. Keep those later stages separate. If a workflow eventually performs a model run, its evidence should identify the supplied inputs, run conditions, outputs and checks rather than using the selection list as a substitute for execution records.

Likewise, a review note written by a person or assistant should say what was checked. Finding that selected identifiers match the fixture does not establish the correctness of every statement inside each record. Finding that an output contains all identifiers does not establish semantic accuracy. The present downloads are checked for authored position, identity, count, fraction and rule-membership consistency only.

Avoid percentages with an implied performance meaning unless the numerator and denominator actually describe that performance question. Three selected positions divided by ten frame positions is a selection fraction. It is not thirty percent accuracy, thirty percent semantic coverage or a confidence level. A concise label beside the number is more useful than a broad disclaimer hidden at the end of a report.

Write an AI task that preserves the selection contract

An original teaching instruction could ask an assistant to read only the supplied fictional selected records, preserve the selection order, report repeated occurrences separately from distinct identities and leave population-level conclusions unresolved. It could also require the source frame version in the answer. This defines a reviewable request without claiming that any model will follow it reliably or that the request authorizes a destination update.

If the assistant proposes a different selection, treat that as a proposal. Ask which rule changed and which question the replacement is meant to answer. A shorter set may omit a required category; a larger set may still omit a critical boundary. Compare the proposed material with the declared purpose before accepting a change. The number of examples alone cannot settle whether the task has been preserved.

For a handoff, keep the frame metadata, selected identities and expected teaching distinctions together. A colleague should not have to infer the zero-based convention from one filename or the replacement policy from an old chat message. The related dataset-design guide helps distinguish authored expectations from observed evaluation results. The related missing-state and filter guides address separate input and predicate questions; none turns this fixture into a validated sampling study.

Check the small fixture before making a larger claim

The independent checks reproduce the selected labels in Cases A, B and E, the row-versus-entity counts in C, the composition fractions in D and the integer boundary membership in F. They operate on the exact downloaded teaching files as well as the local source files. No client records, random outcomes or model outputs are supplied. The checks therefore support the consistency of these original examples and no broader performance conclusion.

Before sharing a summary, read each numerical label aloud with its population and unit. Three positions out of ten frame positions is clear. Three records correct may imply an evaluation that did not happen. Four examples covering two category labels is clear. Representative of one hundred records would add an unsupported inference. Precise wording lets an implementer use the examples for learning while keeping real evaluation decisions attached to their own evidence.

A useful revision note states whether the change affected the frame, selection or review criterion. These are different layers. If the frame grows from ten records to twelve, an earlier three-tenths fraction no longer describes the revised list. If only the question changes, the selected identities may stay fixed while their adequacy for the new task remains unresolved. Keep that distinction visible when comparing versions rather than treating every changed file as the same kind of update. The fixture counts are intentionally small enough to inspect directly, so a reviewer can follow the dependency without relying on a hidden sampling service.

Frequently asked questions

Does selecting three of ten records establish thirty percent coverage?

It establishes a three-tenths selection fraction for the declared ten-position frame. It does not establish semantic coverage, completeness or model accuracy. Case A supplies manually chosen positions and identities, not observed evaluation outcomes. State what the fraction counts before deciding whether the word coverage is appropriate for a different real task.

Can a selected record appear more than once?

That depends on the declared selection design. Case B deliberately repeats a position, producing three occurrences and two distinct labels. It is an authored replacement-style example, not a random draw. If distinct positions were required, silently removing the repetition would change the supplied selection rather than prove the original requirement was met.

Are three rows necessarily three distinct entities?

No. Case C supplies three row keys with entity labels E1, E1 and E2. It therefore has three rows and two authored entity labels. The example does not resolve real identity or independence. Define whether the review concerns rows, observations or consolidated entities before choosing how to select and count material.

Is an equally balanced category set proportionally representative?

Not in Case D. Its frame has twenty Cedar and eighty Maple rows, while the deliberately selected four-row set contains two of each. Those compositions differ. The balanced set can teach both categories, but it does not reproduce the frame shares or establish an overall performance estimate. No model outcomes are provided.

Is a seed enough to reproduce the exact reviewed material?

Record more than a seed. Case E uses identical authored positions on reversed frames and obtains different identities. Preserve the ordered input snapshot, method, parameters, relevant implementation context and selected identities as appropriate. The example makes no cross-version software guarantee and does not execute a random generator.

Do the boundary examples constitute a complete validation suite?

No. Case F checks only four supplied integers under one authored interval rule. It does not test every representation, missing state, application behavior or model response. All six cases are fictional teaching material. No production workflow, professional statistical review, security certification, endorsement or operating approval is claimed, and no executable utility is supplied.