Leadership and readiness
Dataset sample selection bias diligence: redraw datasets, then records
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Dataset sample selection bias diligence starts from one fact: the company chose the data room sample. Curation enters at three levels, which dataset, which records and what was fixed afterward, and each makes quality look better and preparation look cheaper. Have the company commit full lists, draw datasets at random and then records within each, stratified by supplier and period, and underwrite on the drawn estimate.
Checklist
Dataset sample selection
Key takeaways
- A zero-defect sample of 50 records still allows a defect rate near 6 percent, before asking how the 50 were chosen.
- Curation works at three levels: the dataset shown, the records picked and the edits made after picking.
- A flattering sample understates preparation cost as well as quality, so it inflates margin in the model.
- Draw datasets at random from the full catalog first, then records within each, on a seed the investor controls.
- Underwrite on the drawn estimate when the curated figure falls outside its interval and nobody can explain why.
How far can a data room sample overstate dataset quality?
Far enough to change the investment case, and a flawless sample proves less than it appears to. If a company shows 50 records and none has a defect, the most an investor can say at 95 percent confidence is that the true defect rate is below about 6 percent, and that assumes the 50 were drawn at random. The bound comes from solving (1 - p) to the 50th power = 0.05; the shortcut is three divided by the sample size.
Clinicians met this trap long ago. Hanley and Lippman-Hand's 1983 paper in JAMA, If Nothing Goes Wrong, Is Everything All Right?, deals with studies that report a zero numerator, where none of the events of concern occurred, and explains why inferences from such data are problematic. A data room sample with no visible defects is a zero numerator chosen by the party being evaluated.
Machine learning research shows how much familiar material can flatter results. When the authors of GSM1k wrote new grade-school math problems matched in style and difficulty to the GSM8k benchmark (arXiv 2405.00332), some models showed accuracy drops of up to 8 percent on the fresh problems, while frontier models showed little or no drop. Benchmark overfitting is not the same as a curated sample, but the lesson transfers: material the evaluated party has seen, shaped or selected is weaker evidence than a fresh draw, and how much weaker depends on who built it.
Three layers of curation before a record reaches the data room
Curation enters at three levels: which dataset is shown, which records inside it, and what was done to those records after they were picked. Buyer-side audits usually test the middle layer. An investor valuing a company has to test all three, because dataset choice and post-selection edits can flatter every quality number without any single record looking odd.
Some of this is ordinary preparation. Croissant's Responsible AI vocabulary, maintained by MLCommons, has a dataPreprocessingProtocol property for describing the steps needed to bring collected data to a state a model can process, and it gives filtering out incomplete entries as an example. A completeness filter declared that way is a documented choice. The same filter applied quietly to a sample, but not to the population the company would deliver, is curation.
The table lists the mechanisms to expect, the layer each works at, the quality signal it flatters and the draw that exposes it.
| Curation mechanism | Layer | What it flatters | Draw that exposes it |
|---|---|---|---|
| Showing the best supplier or system | Dataset | Every quality metric at once | Random choice of datasets from the full catalog list |
| Showing only the newest records | Dataset | Field completeness and format consistency | Strata by year, including years before system migrations |
| Keeping resolved or accepted cases | Record | Outcome labels and case completeness | Strata by outcome status, including open, abandoned and reopened cases |
| Dropping records with empty fields | Record | Fill rates and apparent cleanliness | Simple random draw from a committed list of all eligible records |
| Hand-correcting the picked records | Edit | Label accuracy and redaction quality | The same record IDs compared in the raw extract |
| Reusing a sample a buyer already accepted | Edit | Apparent acceptance likelihood | A fresh draw on a seed the investor generates |
| Reporting drawn IDs as unavailable | Record | Every metric, quietly | Count of drawn IDs not produced, treated as failures until produced |
Why a flattering sample understates cost, not just quality
Because preparation cost scales with the defects a sample hides. An investor's model of a data company usually turns quality into hours: records to repair, labels to correct, redactions to redo and deliveries that fail acceptance and must be reworked. A curated sample feeds that model a lower defect rate, so it understates cost and overstates margin in one move.
Illustrative arithmetic, with invented numbers: if the data room sample implies almost no defects and a fair draw finds 10 percent, a 100,000-record delivery needs about 10,000 repairs that the plan never included. Hand-corrected samples are worse for the model, because the corrections are exactly the labor that will have to be repeated across the whole population, and the sample has made that labor invisible.
So ask where the sample came from before asking what it shows. A company that can say how each sample record was selected, from which list, by whom and with what changes afterward has handed the investor an estimate. One that cannot has handed over a brochure.
The two-stage redraw: pick datasets at random, then records within each
Have the company commit two lists before anything is drawn, then draw from both on a seed the investor generates: first which datasets to inspect, then which records inside each. Two stages matter because a company with a catalog can pass a record-level redraw on its best dataset and still have weak supply everywhere else.
The first list is every dataset the company counts as supply, with supplier, source system, period and status. The second is, for each dataset drawn, a list of every eligible record identifier, hashed so it can be committed without exposing content. Stratify both draws on whatever could plausibly change quality: supplier, source system, year and outcome status. Oversample small strata that matter, then weight them back to their share of the population when estimating overall rates.
Treat records that cannot be produced as findings. If a drawn identifier comes back unavailable, deleted or held for privacy, count it as a failure until the company produces it or shows that it is excluded from the deliverable population too. Otherwise the redraw quietly becomes a second curated sample.
Then project rather than describe. The PCAOB's audit sampling standard, AS 2315, treats misstatements found in a sample as estimates for the population the sample came from, and says projected misstatement results for all sampling applications should be considered in the aggregate. For a data company, project each stratum's defect rate onto that stratum's records, add them up, and set the total beside the quality claims in the deck.
Reading curated, random and stratified estimates side by side
Put three estimates of the same metric next to each other, with intervals, and read the gap. If the curated estimate falls inside the random draw's interval, there is no evidence of curation on that metric. If it falls outside and the company can point to documented selection, such as deliberately oversampled rare cases, reweight and move on. If it falls outside and nobody can explain why, underwrite on the drawn numbers and treat the company's quality claims as unverified.
Methods note: the curation table, the two-stage redraw, the decision rules and the request list are editorial frameworks built for this article on 2026-10-27. The comparison values are illustrative and describe no company. Single-draw intervals use the Wilson method, the weighted stratified interval a normal approximation, and the zero-defect bound the exact binomial calculation. Factual anchors are AS 2315, Hanley and Lippman-Hand (1983), GSM1k, Datasheets for Datasets, Croissant's Responsible AI vocabulary, Article 10 of the EU AI Act and the Data Provenance Initiative audit. Nothing here measures SourceX or any real data room.
The illustrative comparison in the table below covers a catalog with three suppliers holding half, 30 percent and 20 percent of the records. The curated sample shows no defects in 50 records, so its 95 percent upper bound is about 6 percent. A simple random draw of 400 records finds 40 defective, 10.0 percent with an interval of 7.4 to 13.3 percent, so the curated bound sits below the random draw's lower limit, the signature of selection. The stratified draw, weighted back to supplier shares, estimates 10.6 percent and shows where the problem sits: the smallest supplier runs at 20 percent, which needs repair work or a price before those records count as supply. The figure plots the same intervals.
| Estimate (illustrative) | Records drawn | Defective | Defect rate and 95% interval |
|---|---|---|---|
| Curated data room sample | 50 | 0 | 0%, upper bound about 6% |
| Simple random draw | 400 | 40 | 10.0% (7.4% to 13.3%) |
| Stratified draw, weighted to supplier shares | 400 | 47 | 10.6% (7.7% to 13.5%) |
| Stratum: supplier A, half of records | 150 | 9 | 6.0% (3.2% to 11.0%) |
| Stratum: supplier B, 30% of records | 150 | 18 | 12.0% (7.7% to 18.2%) |
| Stratum: supplier C, 20% of records | 100 | 20 | 20.0% (13.3% to 28.9%) |
When is a curated sample the right evidence?
When the question is format, rarity or privacy rather than quality. A hand-picked handful is the quickest way to check field names, file structure and the de-identification approach, and a deliberately chosen set of rare cases can show that the long tail exists at all. Neither tells an investor how often problems occur.
Documented selection is not a red flag; undocumented selection is. Datasheets for Datasets, the documentation framework proposed by Gebru and colleagues (arXiv 1803.09010), asks creators whether a dataset is a sample of a larger set and whether it is representative, and its later version adds: "If the dataset is a sample from a larger set, what was the sampling strategy (e.g., deterministic, probabilistic with specific sampling probabilities)?" A company that answers those questions for its data room sample has turned a curated sample into evidence with known limits.
One case changes the population rather than the method. If the product is itself a curated subset, for example only resolved cases with complete outcome labels, the fair comparison is a random draw from that curated product population, not from the raw archive. The diligence questions then become how much of the archive survives the curation and whether buyers pay for the curated version.
The sample provenance request list
Ask for the provenance of the sample before its contents, and expect each answer as a record rather than a description. The list below doubles as the evidence request for the redraw.
At a glance
The sample provenance request list
- Selection record: who chose each sample record, on what date and by what rule.
- Frame: the full list of datasets and, for each sampled dataset, the dated, hashed list of eligible record identifiers the sample came from.
- Edits: every change made to sample records after selection, from label fixes to re-redaction, with the raw versions kept.
- Prior use: whether the sample was shown to or accepted by a buyer before, and in what form.
- Exclusions: records removed from the frame for privacy, legal or quality reasons, with counts by reason.
- Metric definitions: how each quality number in the deck was computed, on which records and by whom.
- Reviewer independence: who scored the sample, and whether they work for the company.
- Unavailable records: the count of drawn identifiers the company could not produce, with reasons.
- Documentation: a datasheet-style answer on whether the sample is representative and how that was checked.
Where a clean redraw can still mislead an investor
A redraw fixes selection; it does not fix everything that can go wrong with sampled evidence. Keep these limits next to the result.
- The frame can be filtered before it is committed. If awkward records leave the list before it is produced, a perfect random draw inherits the filter; compare frame counts with counts pulled from the source systems.
- The grader can be the company. If company staff score both samples, their judgment travels with them; use an independent reviewer, at least for disagreements.
- Rare, severe defects escape random draws. One unredacted file or one record covered by a restrictive contract can sit outside any sample; pair sampling with full scans for the highest-severity defects.
- Quality drifts. A redraw from older records says little about records created after a system change; stratify by period and repeat the draw at each refresh.
- Buyer standards rise. Article 10 of the EU AI Act requires training, validation and testing data for high-risk AI systems to be relevant, representative enough, largely free of errors and complete for the intended purpose; buyers building systems in that scope will test harder than an investor's sample does.
- Documentation can be wrong. The Data Provenance Initiative found license omission above 70 percent on widely used dataset hosting sites, so treat any labels that travel with a sample as claims to test, not facts.
Sample questions for any data company, SourceX included
Use the same request list for every company. SourceX is one application of the question: it says it keeps no stock of datasets for AI buyers and sources each one from U.S. businesses when a request arrives, so a sample would be drawn against a specific request rather than taken from a shelf. The question for SourceX, as for anyone, is whether each sample arrives with a frame, a seed and a record of edits.
SourceX has not published sample, defect or acceptance results; its company and commercial metrics are shared with investors directly. Investor materials are provided on request at sourcex.si/investors, as part of a Rule 506(c) capital raise limited to accredited investors. Bring the provenance list and ask for one redraw you control.
Frequently asked questions
What is selection bias in a dataset sample?
It means the sample's records were chosen in a way that makes them systematically different from the population they are meant to represent. In a data room the usual cause is that the company picked the datasets, records or periods that look best, or fixed records after picking them. The remedy is a draw from a committed list on a seed the reviewer controls.
How can an investor tell whether a data room sample was curated?
Compare it, metric by metric, with a fresh random draw from a list the company committed before the draw. If the curated estimate falls outside the random draw's interval and nobody can point to documented selection, treat it as curated. Missing records, edits after selection and samples reused from buyer pilots are the other signs.
Is a sample with zero defects good evidence of dataset quality?
Weaker than it looks. With no defects in 50 records, the defect rate can still be close to 6 percent at 95 percent confidence, and that assumes a random draw. A zero-defect sample the company picked says little about the population until someone redraws it.
Why draw datasets before records?
Because a company with many datasets can pass a record-level check on its best one. Choosing which datasets to inspect at random from the full list, then sampling records inside each, tests the catalog an investor is actually valuing rather than its showcase.
Does SourceX publish sample quality or defect rates?
No. Sample, defect and acceptance results from SourceX are unpublished; investors receive company and commercial metrics directly. Accredited investors can request materials at sourcex.si/investors, and any offering happens solely through offering documents, to verified accredited investors.
Sources
- Hanley and Lippman-Hand, If Nothing Goes Wrong, Is Everything All Right? Interpreting Zero Numerators (JAMA, 1983), addresses studies in which none of the events of concern occurred and why inferences from them are problematic. Source
- GSM1k, a new set of grade-school math problems matched to GSM8k, found accuracy drops of up to 8 percent for some models, while frontier models showed little to no overfitting. Source
- Croissant's Responsible AI vocabulary defines rai:dataPreprocessingProtocol as the steps needed to bring collected data to a state a model can process, for example filtering out incomplete entries. Source
- The PCAOB's audit sampling standard, AS 2315, has auditors project sample misstatements to the population sampled and consider projected misstatement results for all sampling applications in the aggregate. Source
- Datasheets for Datasets asks whether a dataset is a sample of a larger set, whether it is representative, and what sampling strategy was used. Source
- Article 10 of the EU AI Act requires training, validation and testing data sets for high-risk AI systems to be relevant, representative enough, largely free of errors and complete for their intended purpose. Source
- The Data Provenance Initiative's audit of more than 1,800 text datasets reported license omission rates above 70 percent and error rates above 50 percent on widely used dataset hosting sites. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.