Skip to content

Systems and records

Export context loss in a dataset: diff the screen, not the schema

By SourceX Editorial · Updated

9 sections

Short answer

Export context loss is the gap between what a decision-maker's screen displayed and what the exported file holds. Find it with a screen-to-file diff: list every element the decider's view showed at decision time, mark each as carried, altered or dropped in the export, then restore omissions one at a time. Rebuild only the omissions that change a reviewer's explanation of the decision; log the rest as gaps.

System export

Export context loss dataset

This page covers export context loss in a dataset: diff the screen, not the schema.Illustrative

Key takeaways

  • Exports lose most of what a screen computes or fetches at display time: formulas, roll-ups, rule badges, related panels and the conversation.
  • Audit against the decider's own view, by role and decision date, not against the table schema.
  • An omission is critical only if restoring it changes the reviewer's stated reason or decision.
  • Rebuild critical elements as of the decision time with their method stated, and log everything else by loss type.
  • Stop adding context when the explanation stops changing; extra attributes add noise and re-identification risk.

What does an export drop that the decision-maker's screen showed?

Mostly it drops what the screen computed or fetched at the moment someone looked: formula values, roll-up totals, badges and timers produced by rules, panels pulled from related records, and the conversation behind a case. Stored fields usually survive the trip. The evidence that explains why someone acted on those fields often does not.

Vendor documentation states these gaps plainly. Salesforce's Data Export FAQ says formula and roll-up summary fields are always excluded from exports. Zendesk's help center says a view exported to CSV carries only the ticket details displayed in that view, without comments, and that its account-level CSV export leaves out ticket comments and descriptions; in its JSON export, tickets larger than 1 MB have their comments removed. ServiceNow's documentation says a CSV or Excel list export contains only the columns displayed in the current list view unless an all-fields parameter is set.

Spreadsheets then trim what is left. Microsoft's Excel specifications cap a single cell at 32,767 characters and 253 line feeds, so a long case history pasted into one column can arrive cut short. Power Apps exports a model-driven view to a static Excel worksheet at up to 100,000 rows at a time, using the view's columns. None of these limits is a defect. Each is a documented design choice that turns into export context loss only when a dataset treats the file as the full record of a decision.

Why schema-complete exports still fail the explanation test

Because people decide from screens, not schemas. The usual completeness check compares an export with the object list or table schema and passes when every stored field is present. That check cannot see a tier badge computed from contract totals, a banner that appears because a linked incident is open, or a counter of prior credits drawn from a related list. Those elements exist only while the page renders.

Exports also fail in the opposite direction. View-scoped exports, such as the Zendesk and ServiceNow behaviors above, are complete for whichever view someone configured, which is rarely the view the decision-maker used. Support leads, finance approvers and account managers often work from different layouts of the same record, so the reference for a context audit has to be the decider's own view, for their role, as it stood on the decision date.

That changes the question from whether the export captured the data to whether a reviewer holding only the file would explain the decision the way the decider would have. The second question can be tested, and it keeps the audit small: elements that never reached the decider's screen stay out of scope unless the explanation turns out to need them.

Running the screen-to-file diff for one decision type

Run the diff per decision type and per role, on a small sample of real decisions, before anyone approves a bulk export. Each step leaves a written artifact a buyer can inspect later.

The sequence for one decision type:

  • Name the decision and its decider, for example service credits above the standard limit, approved by a support director. One decision type per diff.
  • Build the screen inventory: every element the decider's view displayed, grouped as stored fields, computed values, related-record panels, conversation and files, and system banners or timers. Note whether the inventory comes from the current layout, layout change history or a recorded walkthrough.
  • Inventory the export actually delivered: every column or key, with its source field, format and time zone.
  • Diff element by element: mark each displayed element as carried, altered (present but changed) or dropped, and name its loss type from the map in the next section.
  • Run the restore test on a sample: a reviewer reads the export only, then omitted elements come back one at a time while the reviewer restates the reason and the decision.
  • Reconstruct, as of the decision time, only the elements that changed the explanation, each with its source and method. Everything else goes into a gap log with its loss type.
  • Freeze the result as a specification: packet layout, gap log, and the count of sampled decisions whose explanation changed, with the sample size beside it.

Six ways a displayed element changes on its way into the file

A displayed element can be dropped outright, swapped for a key, replaced by a later value, cut short, filtered out along with its row, or stored in another object altogether. The type decides the repair, so classify before fixing anything.

Two repairs are never allowed: filling a missing past value with today's value, and writing a missing reason from what someone remembers. Both produce a tidy record that teaches a model a story the evidence does not hold.

The context completeness map below classifies each altered or dropped element and names the repair that keeps the packet honest:

Six ways a displayed element changes on its way into the file
Loss typeHow it happensHow the diff exposes itRepair allowed in the packet
Computed value droppedA formula, roll-up or rule output is calculated at display time, or excluded from the exportAn element on the screen inventory has no matching columnRecompute as of the decision time from stored inputs and the rule version then in force, or log a gap
Label swapped for an identifierLookups export as record IDs while names, roles and picklist labels stay in other tablesThe column holds keys where the screen showed wordsMap IDs to role tiers or labels valid on the decision date; leave personal names out unless the task needs them
Present value instead of value thenThe export reads current field values, not the values on the decision dateField history or an audit trail shows a later changeRoll back from history; if history has expired, withhold the current value and log the field
Text cut, split or flattenedCell limits, line-feed limits or markup stripping shorten long fieldsLengths cluster at a limit, or threads end mid-sentenceRe-export in a format without the limit; never patch text from memory
Rows outside the view or capView filters, permissions or row caps decide which records leaveRecord counts differ from a count run inside the source systemRe-run the export by date range until counts reconcile
Related record held elsewhereComments, emails, attachments and linked records live in separate objects or exportsThe screen showed a thread or panel; the file has only the case rowJoin by the link the screen displayed, at the version that existed at decision time, with personal data minimized

The restore test: which omissions change the explanation?

An omission is critical when restoring it changes the reason a reviewer gives for the decision, or changes the decision the reviewer would make. Everything else can stay out of the packet and go into the gap log, however prominent it looked on screen.

Run it as a restore test rather than a removal test. Start from what the buyer would actually receive, the export alone, and ask a reviewer who knows the policy domain to write the decision's reason in one sentence and say whether they would have decided the same way. Then restore one omitted element at a time, in the order of the screen inventory, and ask again. Elements that move the answer are critical; elements that leave it unchanged are not.

Starting from the export tests the delivered artifact rather than an idealized one, and it exposes a failure a removal test misses: an export that produces a confident wrong reason. If the export-only reviewer writes that a credit was approved outside policy while the decider was applying a documented exception, the file is not merely incomplete. It is misleading, and a model trained on it learns that approvals ignore policy.

Report the result with its denominator: how many sampled decisions changed explanation after restoration, and which elements did the changing. With samples in the dozens, report counts rather than percentages with decimals, and repeat the test per role, because the critical set for a finance approver rarely matches a support lead's.

Worked diff (illustrative): a service credit above the policy limit

In this illustrative case the export alone makes a policy-compliant credit look like an exception. A support director approved a service credit above the standard limit. A reviewer given only the export row wrote: approved above limit, no reason recorded. After the restore test the reason read: outage compensation for a Strategic account, within the annual credit cap, at the amount the customer asked for.

Four of the seven displayed elements changed the explanation and three did not, including the SLA timer, the most visually prominent item on the screen. The case, account, amounts and policy values are illustrative and describe no real company.

The diff for that one approval, element by element:

Worked diff (illustrative): a service credit above the policy limit
Element on the approver's screenWhere the screen got itWhat the export row holdsLoss typeChanges the explanation?
Account tier badge: StrategicFormula over a roll-up of active contract valueNothingComputed value droppedYes: Strategic accounts carry a higher credit limit
Banner: linked major incident openRelated incident recordNothingRelated record held elsewhereYes: the credit follows outage compensation, not goodwill
Prior credits in 12 months: 2Count over a related listNothingComputed value droppedYes: shows the annual cap of 3 was not exceeded
Customer's request with invoice attachedCase conversation and fileNothing; the view CSV carries no commentsRelated record held elsewhereYes: sets the amount requested
SLA timer: first reply 9 hours lateComputed from timestamps and the SLA targetCreated and first-reply times, no targetComputed value droppedNo, once the incident link is restored
Approver shown by name and titleLookup to the user recordApprover user IDLabel swapped for an identifierOnly the authority tier matters
Priority: HighStored fieldUrgent, changed after the decisionPresent value instead of value thenNo, but flagged in the gap log
Side-by-side diff of an illustrative service credit approval: of seven elements on the approver's screen, three arrive altered in the export row and four are dropped, and all four dropped elements are marked as changing the explanation.
Screen-to-file diff for one approval. Illustrative: the case and its values are examples, not a real company.

The minimal task packet the diff leaves you with

The packet keeps the export row and adds back only the four critical elements, each rebuilt as of the approval time and annotated with how it was rebuilt. The annotations are data rather than documentation: a reviewer or a model should be able to tell exported values from recomputed ones and from gaps.

At a glance

The minimal task packet the diff leaves you with

Packet contents for the credit approval:

  • Case row as exported: case key, category, created and first-reply times in UTC with the original offset kept, and the credit approved. Annotation: exported, unaltered.
  • Account tier at approval time: Strategic, recomputed from contracts active that day under the tier rule version then in force. Annotation: recomputed, rule version cited.
  • Linked incident: incident key, start and resolution times, and the outage-compensation policy version in force. Annotation: joined through the case-to-incident link the screen displayed.
  • Prior credits in the 12 months before approval: 2, against a cap of 3. Annotation: recounted from credit records dated before the approval only.
  • Customer request: message text with personal details removed and the invoice amount kept. Annotation: thread version as sent before approval.
  • Approver authority: director tier. Annotation: mapped from the user ID through the role table valid on the approval date; name withheld.
  • Gap log: priority at approval unknown because history expired and the current value differs; SLA target version not recovered. Neither changed the explanation in the restore test.
  • Left out on purpose: the account's other tickets, contact roster and full contract history. Annotation: failed the restore test; excluded to limit noise and linkage risk.

Usually not. Joining everything linked to the account gives a model more to read but rarely a better reason, and every added attribute raises the chance that a packet points to a specific business or person. The restore test exists to stop the join list at the elements that change the explanation.

The privacy cost compounds quickly. Rocher, Hendrickx and de Montjoye estimated in Nature Communications in 2019 that 99.98 percent of Americans would be correctly re-identified in any dataset using 15 demographic attributes. Business packets are not demographic tables, but the arithmetic carries over: contract values, locations, incident dates and staff roles combine into the same kind of fingerprint, so an element that does not change the explanation should not travel.

The noise cost is quieter. A long packet buries the few elements that mattered among dozens that did not, and a model can learn to key on whatever appears most often rather than on what was decisive. There is one good reason to go beyond the screen: the explanation still diverges after every displayed element has been restored. That usually means the decider relied on something off-system, such as a call, a message in another tool or a manager's verbal approval. Record it as a named gap with its likely location instead of joining a new system to see whether it helps.

Writing export context loss into a dataset requirement

A buyer can turn the diff into acceptance terms. Name the decision type and the deciding role, ask for the screen inventory and the loss-type classification, require critical elements to be rebuilt as of the decision time with their method stated, and accept the dataset only when a sample of packets passes the restore test with the gap log attached.

Describe the source characteristics the packet depends on rather than a volume target: field history kept long enough to roll values back, links between cases and incidents or orders, and conversation exports that preserve versions. Those are the properties that let a supplier rebuild what a screen showed, and they can be checked before anything leaves the company.

SourceX finds datasets for AI buyers among U.S. companies once a buyer has described the need, instead of selling from a catalog, so a request built around one decision and its restore test shapes what candidate companies are asked to demonstrate from the first conversation. Whether any company's records can meet the requirement is decided in review; nothing about a fit is promised in advance. Submit the requirement, its intended use and the acceptance test through the request form.

Frequently asked questions

What is export context loss in a dataset?

It is the gap between what a decision-maker's screen displayed when they acted and what the exported file contains. Computed values, related-record panels, conversation threads and past field values are the usual losses. It matters when a dataset is meant to teach or test why decisions were made, because the file alone can point to the wrong reason.

Is a full JSON export enough to prevent export context loss?

Not by itself. A full export usually carries more stored fields than a view export, but values calculated at display time can still be missing, and some exports apply their own limits; Zendesk documents that its JSON export removes comments from tickets larger than 1 MB. Run the screen-to-file diff on the export that will actually be delivered.

Can screenshots of the source system replace the screen inventory?

Screenshots help list the elements, but they show the screen as it looks today, with current values and the current layout. Historical decisions need elements as they stood on the decision date, which means field history, rule versions and layout history. Use screenshots or a recorded walkthrough to build the inventory, then rebuild past values from records.

How many decisions should the restore test cover?

Enough per decision type and role that new loss types stop appearing; there is no universal number. Start with a small sample from each role, report the count of changed explanations next to the sample size, and extend the sample when a new critical element shows up late. Results for one role should not be assumed for another.

Does rebuilding context make a dataset riskier to share?

It can. Each restored attribute adds a way to single out a company or a person, which is why the restore test limits the packet to elements that change the explanation. Restore roles instead of names and counts instead of full histories where the explanation allows it, and send the packet through privacy review before release.

Sources

  • Salesforce's Data Export FAQ says formula and roll-up summary fields are always excluded from exports. Source
  • Zendesk's help center says exporting a view to CSV exports only the ticket details displayed in the view, not comments, which require a JSON export or the API. Source
  • Zendesk's account export article says the CSV export does not include deleted tickets, ticket comments or descriptions, and that the JSON export removes comments from tickets larger than 1 MB. Source
  • ServiceNow's documentation says CSV and Excel list exports include only the fields displayed in the current list view, with a parameter to export all fields. Source
  • Microsoft's Power Apps documentation says static Excel worksheet exports follow the view's columns and can export up to 100,000 rows at a time. Source
  • Microsoft's Excel specifications list a limit of 32,767 characters per cell and 253 line feeds per cell. Source
  • Rocher, Hendrickx and de Montjoye (Nature Communications, 2019) estimate that 99.98 percent of Americans would be correctly re-identified in any dataset using 15 demographic attributes. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies