Skip to content

Privacy and preparation

Dataset acceptance specification template: what to agree before export

By SourceX Editorial · Updated

9 sections

Short answer

Before exporting business records, agree a dataset acceptance specification with the buyer: scope that reconciles to the source system, row and case totals, key integrity, required fields with coded missing values, timestamp rules, schema history, attachment hashes, outcome label rules and an exception log. Give each criterion a measurable field, a method, a threshold and an evidence file, so acceptance is a comparison of documents.

Checklist

Dataset acceptance

This page covers dataset acceptance specification template: what to agree before export.Illustrative

Key takeaways

  • Most acceptance disputes are scope disputes nobody wrote down; reconcile the export to a same-day count in the source system.
  • Every criterion needs a field, a method, a threshold with a unit and the evidence file the supplier attaches.
  • Declare missing-value codes before measuring completeness, or two people will get two different numbers.
  • Schema history is the defect most likely to change a buyer's decision while every file still opens.
  • Hash every attachment beside its parent ID so the files tested are provably the files delivered.
  • Pause the export when a finding changes what records mean or raises a rights question; do not relax a threshold to get past it.

Which acceptance criteria should be agreed before exporting business records?

Agree ten criteria in writing before anyone runs the export query: the scope the records must reconcile to, row and case totals with an allowed difference, key integrity, required fields with their missing-value codes, timestamp rules, the schema version history, attachment integrity, how outcome labels are defined and populated, the de-identification check and the format of an exception log. Each item gets a measurable field, a method, a threshold and the evidence file that proves it. If a criterion cannot be measured, it is a hope, not an acceptance test.

The reason to settle this before export rather than after delivery is simple: almost every acceptance argument is really a scope argument that nobody wrote down. The buyer expected closed work orders from three regions over four years; the supplier exported every work order the reporting role could see. Both parties acted in good faith, and the dataset still fails. A specification signed before the query runs turns that disagreement into a line item that either party can check.

The buyer-side pages on this site cover how to run an acceptance window and how to write rejection and cure terms. This page is for the side that holds the records. It answers a narrower question: what the supplier should propose and measure on its own data, so the package arrives with its acceptance evidence already attached.

Why a package that opens cleanly can still be rejected

Opening every file only proves the files are readable. Acceptance is about whether the package is complete, internally coherent and usable for the task the buyer declared, and several common defects pass every file-level check. A CSV with the right columns can still hold the wrong population, shifted timestamps or labels that were filled in after the fact.

At a glance

Why a package that opens cleanly can still be rejected

ISO/IEC 25012, the data quality model in the SQuaRE family, sorts data quality characteristics into inherent ones and system-dependent ones. That split is a useful reminder for suppliers: readability and format depend on the system that stores and reads the data, while whether the records are correct, complete and consistent with each other is a property of the data itself. A specification that only tests the first group will pass packages that fail the buyer's actual use.

  • Scope drift: the export filter included test accounts, archived regions or records outside the agreed date window.
  • Silent truncation: an API page limit or a report row cap cut the export short without any error.
  • Timestamp shifts: local times were written without an offset, so durations and orderings across sites are wrong.
  • Schema change mid-history: a status field gained values two years in, so early and late records mean different things.
  • Orphaned attachments: files exist with no parent record, or records point to files that were never exported.
  • Label leakage or gaps: the outcome field was back-filled for some records, or is blank for open cases that should have been excluded.

The dataset acceptance specification template, field by field

Use one row per criterion and keep every threshold numeric or enumerated. The example thresholds below are illustrative starting points for a mid-size operational export, not norms; the buyer's declared task decides the real values. The evidence column names the file the supplier hands over, so acceptance becomes a comparison of documents instead of a meeting.

  • Header fields for the specification itself: spec_id, version, date agreed, buyer task statement, source system and module, export method (report, API or database query), query or filter text, and the names of the people who sign for each side.
  • Change control: any edit after signature gets a new version number and both signatures again; the delivered package names the version it was tested against.
The dataset acceptance specification template, field by field
CriterionMeasurable fieldHow it is measuredIllustrative thresholdEvidence file
Scope reconciliationscope_filter_text, source_count, export_countSame filter run as a count in the source system and against the export on the same dayExport count equals source count, or the difference is fully itemizedreconciliation_worksheet.csv
Row and case totalsrows_per_table, distinct_case_idsCount rows per table and distinct case keys; compare with the manifestZero unexplained difference per tablemanifest.json with per-file row counts
Key integrityprimary_key, foreign_keysUniqueness and not-null on keys; every child row resolves to a parent0 duplicate keys; 0 orphan child rowsvalidation_report.json
Required fieldsrequired_fields[], missing_value_codes[]Count nulls and agreed missing-value codes per required fieldBelow 2 percent missing per required field, with reasons codedfield_completeness.csv
Timestamp validitytimestamp_fields[], offset_ruleParse every timestamp; check offset present, order of events, plausible range100 percent parse with offset; 0 closures before openingstimestamp_checks.csv
Schema historyschema_versions[], change_dates[]List each field or picklist change with its effective date from admin logsEvery change in the window documentedschema_history.csv
Attachment integrityfile_count, sha256[], parent_idHash every file; join files to parent records both ways0 hash mismatches; 0 orphans either wayattachments_manifest.csv
Outcome labelsoutcome_field, label_definition, populated_whenCheck distribution, closed-only rule and when the label was writtenLabel present on every closed case; none on excluded open caseslabel_profile.csv
De-identificationremoved_fields[], residual_check_methodField drop list plus a sampled residual reviewSample size and acceptance number agreed in advancedeid_report.pdf
Exception logexception_id, rule, count, dispositionEvery record excluded or altered by a rule is counted under that ruleLog totals reconcile to the scope differenceexception_log.csv

Reconciling the export against the system it came from

Reconcile to the system of record, not to the file you produced. Run the agreed filter as a count inside the source application, on the same day as the export, and record both numbers side by side; every difference must be explained by a named rule in the exception log. This single worksheet resolves more acceptance disputes than any statistical test.

In practice, the source count comes from the application's own reporting layer: a saved report or list view in a CRM, a ticket view count in a help desk, a saved search in an ERP, or a count query on the reporting database. Note which one you used and who ran it. If the export used a different path, for example an API that excludes archived records, the worksheet is where that shows up.

Reconciling the export against the system it came from
CheckSource system figureExport figureAllowed differenceExplanation required
Records matching the scope filterCount from a saved report with the exact filterRows in the main table0 unless itemizedEach excluded record listed under an exception rule
Distinct cases or accountsDistinct count of case IDs in the reportDistinct case_id in export0Duplicates or merges named with both IDs
Records per year or quarterReport grouped by periodExport grouped by period0 per periodAny gap month explained (migration, outage, policy change)
AttachmentsAttachment count from the admin console or storage listingFiles in the attachments folder0 unless itemizedSize limits, deleted files or excluded types listed
Closed versus openStatus breakdown in the reportStatus breakdown in export0Open cases excluded on purpose are counted in the log

Setting thresholds a validator can check without arguing

A threshold is usable only when two people running the same check get the same answer. Write it as a metric, an operator and a unit, the way data contract standards do, and decide in advance what happens at the boundary. The Open Data Contract Standard, for example, expresses quality rules as library metrics such as rowCount, nullValues, missingValues, invalidValues and duplicateValues, compared with operators like mustBe, mustBeLessThan or mustBeBetween, with a unit of rows or percent.

Declare what counts as missing before measuring it. The Frictionless Table Schema specification lets a schema list the strings that should be read as missing values, and its newer form allows each missing-value code to carry a label. That matters for business records, where a blank, N/A and 0 can mean not collected, not applicable or a real zero. An acceptance spec that does not name these codes will produce two different completeness numbers from the same file.

Some checks cannot run on every record, such as a human review of whether a resolution note matches the outcome code, or a residual personal-data review. For those, agree a sampling plan. NIST's engineering statistics handbook describes lot acceptance sampling as a way to decide whether a lot is likely acceptable, not to estimate its exact quality, which is exactly the role it plays here: the sample size and the acceptance number are part of the specification, fixed before anyone looks at the data.

  • Key and structure checks map directly onto Table Schema constraints: required, unique, enum, primaryKey and foreignKeys, so a standard validator can run them.
  • Count and completeness checks map onto data contract quality rules with a metric, operator and unit.
  • Judgment checks get a written rubric, a sampling plan and a named reviewer on each side.
  • Every threshold states its unit (rows or percent) and whether the boundary value passes.

Timestamps, schema drift and attachments: the quiet defects

Three defects routinely change a buyer's decision even when every file opens: timestamps without offsets, fields whose meaning changed during the history, and attachments that do not match their records. Each needs its own line in the specification because none of them shows up in a row count.

For timestamps, require an explicit offset from UTC on every value. RFC 3339, the common internet profile of ISO 8601, defines how that offset is written, and it reserved a convention for a time whose UTC value is known but whose local offset is not; RFC 9557 later updated that convention so that Z carries that meaning. The practical rule for an export is to state which fields are stored in UTC, which are local, and which source time zone applies to each site, then check that event order holds, for example that no ticket closes before it opens.

For schema history, ask the system administrator for the dates when fields, picklist values or workflow states were added, renamed or retired, and write them into the specification. A status called Resolved that meant fixed until a given date and fixed or abandoned after it will quietly corrupt any model trained across the change. The buyer can handle the change if it is documented; it cannot handle it if it is discovered later.

For attachments, hash every file and record the hash beside the parent record ID. The Data Package specification carries a hash per resource, MD5 by default with other algorithms marked by a prefix, and the Croissant metadata format for machine-learning datasets lets each file object carry a sha256 checksum and strongly recommends recording checksums for versioned datasets. Either format gives the buyer a way to confirm that the files tested are the files delivered.

Worked example: an illustrative field-service work-order export

Consider an illustrative export from a field-service platform: closed work orders for four years across three regions, with technician notes, parts lines and photos. All counts here are invented to show the method. The agreed scope filter returns 48,200 work orders in the platform's saved report; the export holds 47,950 rows.

The reconciliation worksheet itemizes the 250-row gap under two exception rules: 180 work orders belonged to an internal test account and 70 were duplicates created by a mobile app sync defect, each paired with its surviving ID. The gap is fully explained, so the scope criterion passes. Key checks find zero orphaned parts lines. Required-field completeness is 98.7 percent for the resolution code, inside the agreed limit, with every blank coded as not recorded.

The exception is in the schema history. The admin log shows that a Return Visit Required status was added to the workflow 22 months into the four-year window. Before that date, return visits were recorded only in free-text notes. The data passes every mechanical check, but the outcome label means something different before and after the change. The parties amend the specification to version 1.1: the label is delivered with a schema_era field, and the buyer decides whether to train only on the later period. Without the schema history line, this would have surfaced as a model problem months later.

Attachments reconcile too: 61,300 photos with hashes, zero mismatches, and 40 files without a parent record, which the exception log attributes to work orders deleted under the retention policy. Those 40 files are removed from the package rather than delivered as orphans.

Findings that should pause the export rather than be waived

Pause the export when a finding changes what the records mean, not merely how many there are. A small itemized count difference can be accepted with a note; an unexplained difference, a label that cannot be tied to a date, or personal data where the agreement assumed none should stop the run until the specification is revised and re-signed.

Also stop when the problem is outside data engineering. If reconciliation shows the export includes records from a business unit the company does not control, or records it received from a client under a confidentiality term, that is a rights question for counsel, not a threshold to relax. De-identification does not resolve it either: removing names does not give the company a right to license records it was never permitted to share.

  • Unexplained scope difference after two reconciliation attempts: stop and re-scope.
  • Outcome label that was back-filled or rewritten after the event for part of the history: stop and document, or exclude that period.
  • Residual personal data above the agreed acceptance number in the sample: stop and re-run de-identification, then re-sample.
  • Records from another company's systems or under a client confidentiality clause: stop and route to legal review.
  • Schema change that the buyer has not seen: amend the specification before delivery, not after.

Using the specification in a SourceX licensing deal

When a company licenses records through SourceX, the acceptance specification is part of the preparation work, not a separate project. SourceX carries a licensing deal from qualification and rights review through contracting, preparation, delivery and payment, and the scope written into the signed agreement is the same scope this specification reconciles against. No export is handed over before the company has approved that scope and its terms and put its signature on the agreement.

Each deal's signed Dataset Deal Confirmation fixes the delivery date, which is a reason to fill in the reconciliation worksheet and schema history early. Very large packages can remain in the company's own storage or travel on encrypted drives, and the attachment hashes in the specification are how both sides confirm what moved. A short company assessment comes first and needs no data upload.

Frequently asked questions

What is a dataset acceptance specification?

It is a signed list of the measurable conditions a delivered dataset must meet, each with a field, a measurement method, a threshold and an evidence file. It is agreed before export so the supplier can test its own package and the buyer can check the same evidence. It complements, rather than replaces, the contract's acceptance window and cure terms.

Who should write the acceptance specification, the buyer or the supplier?

The buyer states the task and the thresholds it needs; the supplier proposes how each one is measured on its own systems and what evidence it can produce. In practice the supplier's data owner drafts the reconciliation and schema history sections, because only the supplier can see the source system. Both sides sign the same version.

How is acceptance different from a delivery manifest?

A manifest proves which files were sent and that they arrived intact. Acceptance tests whether the content is the agreed population, internally consistent and usable for the declared task. A package can match its manifest perfectly and still fail acceptance because the export filter was wrong.

What defect most often changes the buyer's decision even though every file opens?

A change in what a field means partway through the history, such as a new status value or a redefined closure code. It passes every format check and only appears when someone compares early and late records. Writing a schema history into the specification is the cheapest defense.

Can a sampled check be part of acceptance?

Yes, for checks that need human judgment or are too costly to run on every record, such as residual personal data review. Fix the sample size, the selection method and the number of failures allowed before anyone looks at the data. Changing the plan after seeing results undermines the whole test.

Sources

  • The Open Data Contract Standard expresses quality rules as metrics such as rowCount, nullValues and duplicateValues compared with operators like mustBe or mustBeBetween, in rows or percent. Source
  • The Table Schema specification lets a schema declare which strings count as missing values, optionally labeled, and defines required, unique, enum, primaryKey and foreignKeys constraints. Source
  • A Data Package resource can carry a hash, MD5 by default, with other algorithms indicated by a lower-case prefix. Source
  • Croissant lets each file object carry a sha256 checksum and strongly recommends checksums for versioned datasets. Source
  • RFC 3339 defines how timestamps carry an offset from UTC and a convention for a known UTC time with an unknown local offset. Source
  • RFC 9557 updates RFC 3339 so that the Z offset carries the unknown-local-offset meaning. Source
  • NIST's handbook describes acceptance sampling as deciding whether a lot is likely acceptable, not estimating its quality. Source
  • ISO/IEC 25012 defines a data quality model for structured data whose characteristics are viewed as inherent or system dependent. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies