Systems and records
Status semantics in a cross-system dataset: when Resolved means two things
By SourceX Editorial · Updated
9 sections
Short answer
To get status semantics right in a cross-system dataset, never join on the word. Fingerprint each status in each system on five properties: what sets it, who may set it, whether it can be undone, what it asserts and which companion field completes it. Map every system to one task pivot with an explicit match type, then have each process owner confirm the map on sampled records before any status becomes a label.
Retrieval
Status semantics cross
Key takeaways
- A shared status word is a hypothesis about shared meaning, not evidence of it; Dynamics 365 files merged cases under its Cancelled state.
- Trigger, authority, reversibility, assertion and companion field fingerprint a status; documentation answers the first and last, owners the rest.
- Record a match type on every mapping and map each system straight to a task pivot, because SKOS keeps closeMatch non-transitive to stop errors compounding.
- Owner verdicts are dated by configuration period: a workflow edit can change what a status means from one quarter to the next.
- Add a third system only when it closes a named gap; irrelevant context measurably degraded model accuracy in a 2023 study.
What does a shared status word tell you across two systems?
On its own, almost nothing. A shared status word tells you that two administrators picked the same label; it does not tell you that the same thing happened in the business. Before a cross-system dataset uses a status as a feature or a label, someone has to show what each system meant by it, for which period, and who confirmed that.
Two common products make the gap concrete. In Microsoft Dynamics 365 Customer Service, a case reaches Resolved only through the Resolve Case action; Microsoft's documentation says the Case Status field cannot simply be set to Resolved, and the resolve dialog asks for a Resolution Type and a Resolution. In Jira, Atlassian's documentation says the Resolution field is set automatically when work moves to Done only in default workflows; a custom workflow needs a post function configured by hand, and Jira's plans date completed work from the Status Category Changed Date rather than from the resolution.
So resolved in one system records that an agent ran a deliberate action and typed an explanation, while in the other it may record a card moved to a green column, with or without a resolution value. Join the two on the word and the dataset quietly adopts the weaker meaning for every row. The rest of this guide is a reproducible way to decide whether two statuses may share a label: fingerprint each one, map it with an explicit match type, have the people who run each process confirm or contest the map, and pack only the evidence the map says a task needs.
Five properties that fingerprint a status before anyone maps it
A status is defined by five properties, and two statuses mean the same thing only when all five agree. The label itself is not one of them.
The trigger is the event that sets it: a dedicated action with a dialog, a picklist edit, an automation, a timer or an inbound message. Authority is who or what may set it, including cascades and integrations. Reversibility covers whether it can be undone, by whom, and whether undoing it leaves a trace. The assertion is what the business holds to be true once it is set, such as work stopped, problem fixed, question answered or record superseded. The companion field is the second field that completes the meaning and that most exports drop.
Vendor documentation answers the first and last properties fairly well, and the middle three badly. Salesforce's Case data model describes the closed flag as controlled by the Status field and not settable directly, so which picklist values count as closed is an org-level choice. HubSpot asks an administrator to mark each new ticket status Open or Closed when the stage is created, for every object except deals; its default Support pipeline holds New, Waiting on contact, Waiting on us and Closed. In Dynamics 365 the Resolved state carries two status reasons, Problem Solved and Information Provided, and the Cancelled state includes a status reason called Merged.
That last detail is the whole problem in miniature. A model trained to read Resolved as fixed learns from cases where an agent only answered a question, and a churn feature built on Cancelled counts duplicates folded into another case. The fingerprint below records what the documentation settles and marks the cells only an owner can answer:
| Property | Dynamics 365 case: Resolved | Jira issue: resolved | Salesforce case: Closed | HubSpot ticket: Closed |
|---|---|---|---|---|
| Trigger | Resolve Case action only; Status cannot be set to Resolved by hand | Resolution set on a transition; automatic only in default workflows | Any Status value the org treats as closed; the flag is not set directly | A move to a status flagged Closed when the stage was created |
| Companion field | Status reason: Problem Solved or Information Provided | Status category Done, which plans use to date completion | Closed flag derived from the Status value | The stage's Open or Closed setting |
| Reversibility | Read-only until the case is reactivated | Owner confirms from the workflow's transitions | Owner confirms | Owner confirms |
| Authority | Agents, plus a parent case resolving its active child cases | Owner confirms who may run the transition | Owner confirms, including automation | Owner confirms, including automation |
| Assertion | Owner confirms; Information Provided is not a fix | Owner confirms, per resolution value | Owner confirms, per closed status value | Owner confirms; the flag says only that work stopped |
Map every system to a task pivot and record the match type
Map each system's status to a small pivot vocabulary written for the task, and record a match type on every row instead of an equals sign. A lookup table that says Resolved maps to Done hides the one fact a reviewer needs: how close the two meanings are.
The W3C's SKOS Reference supplies a ready-made scale. exactMatch signals that two concepts can be used interchangeably across a wide range of applications; closeMatch, that they are similar enough to be interchangeable in some; broadMatch and narrowMatch, that one is wider than the other; relatedMatch, an association without overlap. SKOS deliberately leaves closeMatch non-transitive, so that errors do not compound when mappings are chained across more than two schemes. The practical rule follows: never derive the map for system C by chaining A to B to C. Map each system to the pivot directly, so every row is judged once against one definition.
Write the pivot for the decision the dataset serves, not as a universal workflow model. Process-mining practice offers a precedent: the XES event-log standard's lifecycle extension records which transition each event represents and supplies a standard transactional model, with steps such as schedule, assign and start, while allowing any model a log needs. For a task like deciding whether a customer's problem was fixed, five pivot concepts are enough: work stopped, fix confirmed, answered without a fix, superseded by merge, and withdrawn by the customer. The rows below are illustrative, with verdicts of the kind the owner review produces:
| Source status | Pivot concept | Match type | Owner verdict (illustrative) |
|---|---|---|---|
| Dynamics case: Resolved, Problem Solved | Fix confirmed | closeMatch | Confirmed; agents sometimes resolve before the customer replies |
| Dynamics case: Resolved, Information Provided | Answered without a fix | exactMatch | Confirmed for all periods sampled |
| Dynamics case: Cancelled, Merged | Superseded by merge | exactMatch | Confirmed; the surviving case carries the outcome |
| Jira issue: Done with a resolution value | Fix confirmed | closeMatch | Confirmed only after the resolution post function was added |
| Jira issue: Done, resolution empty | Work stopped | broadMatch | Contested: one team used Done for won't-fix |
| HubSpot ticket: Closed | Work stopped | broadMatch | Confirmed; outcome must come from another field |
How process owners confirm or contest each row
Process owners confirm the map by reading real records, not by approving a spreadsheet. Documentation describes product defaults; only the people who configured and ran each workflow know what a status meant in their company, and when that changed.
Run the review row by row. For each source status, draw a small random sample of records from each period in which the workflow configuration differed, for instance a handful per quarter, and show the owner each record's full transition history with its companion fields. Owners answer for their own system only; the data team owns the pivot definitions and answers for those. Record each verdict with the owner's role rather than a name, the date of the review and the period it covers.
Four verdicts keep the outcome honest. Confirmed means the sample matched the mapped meaning. Confirmed for a period means it held only between configuration dates, such as before and after a workflow edit. Contested means owners disagreed with each other or the records contradicted the documentation. Unknown means nobody who ran the process during that period is available. Only the first two allow a row to become a label. Five questions do most of the work in each session:
- What had happened in the business at the moment this status was set on these records?
- Who could set it then, and did any automation, cascade or integration set it without a person?
- Could it be undone, and would the undoing appear in the history we exported?
- Which field, if any, distinguishes the outcomes hidden under this one status?
- On what dates did the workflow, picklist or automation behind this status change?
An escalated support case, packed with only what the map needs
A task packet should carry the transitions, the map rows that interpret them, and the gaps, and nothing else. The illustrative packet below serves one decision: was this customer's problem fixed, and if so, when?
The case lives in Dynamics 365 and links to an engineering issue in Jira. Read naively, both say resolved and the label writes itself. Read through the map, the case was answered without a fix and the engineering issue only stopped, so the honest label is answered, fix not confirmed. Each line below records why it is in the packet or why it was left out:
| Packet line | Value (illustrative) | Annotation | In or out |
|---|---|---|---|
| Case transition | Active to Resolved, status reason Information Provided, day 9 | Map row: answered without a fix, exactMatch, confirmed | In |
| Resolve dialog fields | Resolution Type and a one-line resolution typed by the agent | Written at the action itself, so its timing is known | In |
| Linked engineering issue | Moved to Done on day 14, resolution empty | Map row: work stopped, broadMatch, contested for this team | In, flagged |
| Customer confirmation | No reply or survey after day 9 | Explicit gap: nothing shows the fix worked | In, as a gap |
| Account's other open cases | Four unrelated cases | No map row needs them; adds linkage without meaning | Out |
| Sprint comments on the issue | Scheduling chatter from other work | Not referenced by any map row | Out |
Which parts of a status's meaning survived the export?
Usually the trigger and the companion field survive an export and the other three properties do not. A completeness map makes that visible before anyone trains on the packet, by stating, property by property, where the meaning came from and what would close each gap.
Read the map below as a ledger of trust. Elements marked as recovered from records can be rechecked by a buyer; elements that rest on an owner's confirmation are only as good as the review behind them and should carry its date. Anything marked missing stays missing in the dataset card rather than being filled by inference:
| Meaning element | Where it came from | State in this packet (illustrative) | What would close the gap |
|---|---|---|---|
| Trigger | Transition history and resolve dialog fields | Recovered from records | Nothing further needed |
| Companion field | Status reason and resolution value | Recovered; resolution empty on the issue | Owner states what empty meant for that team |
| Authority | Owner review and the workflow configuration | Owner-confirmed for the case only | Configuration export for the engineering workflow |
| Reversibility | Owner review | Confirmed for the case; unknown for the issue | Reopen events in the issue history, if retained |
| Assertion | Map row and owner verdict | Confirmed for the case; contested for the issue | A second owner from the same team |
| Outcome for the customer | Not recorded in either system | Missing | A confirmation message or survey linked to the case |
Would joining a third or fourth system settle what Resolved meant?
Rarely, and adding systems has costs that are easy to miss. Every new source brings its own vocabulary to fingerprint and map, and every join is one more key match that can attach the wrong record, importing a status from someone else's case.
Extra context is not neutral for the model either. Shi and colleagues showed in 2023 that adding irrelevant information to grade-school math problems dramatically decreased large language model performance, even though the added sentence did not change the answer. A packet stuffed with every linked record asks the model to sort meaning from noise that the data team could have removed with the map.
A third system earns its place only when it supplies a property the fingerprint is missing. In the packet above, a customer survey response linked to the case would settle the outcome; the account's full ticket history would not. The test is simple to run: before adding a source, name the map row or the gap it closes, and leave it out if there is none.
Pivot label, split, or context only: the call for each shared word
Each shared word ends up as one of three things: a pivot label, a split label, or context that never becomes a label. The match type and the owner verdict decide which, and the decision should be written next to the row so a buyer can audit it.
Confirmed exactMatch rows share the pivot label without qualification. Confirmed closeMatch rows take the pivot label but keep the source status as a feature, and any evaluation reports results per source system so a difference between them stays visible. broadMatch and narrowMatch rows are split with their companion field before labeling, which is why Jira issues done with an empty resolution become work stopped rather than fix confirmed; where the companion field is absent for a period, those rows drop to context for that period. Contested and unknown rows never become labels.
Name the limits next to the decision. Owners remember imperfectly and may defend their own workflow, a sample of a few records per period can miss rare uses, and vendor documentation describes defaults rather than any one company's configuration. The next evidence to collect is a repeat owner review after any workflow change, and a second owner for every contested row.
Putting status semantics into a dataset request
A dataset request should name the status fields the task depends on and ask for the fingerprint, the pivot map with match types and dated owner verdicts, and the completeness map as deliverables. Without them, a buyer cannot tell a confirmed label from a word that merely matched.
SourceX does not keep datasets on a shelf: it sources each one from U.S. businesses after a buyer describes the scope, volume, language, format, timing and licensing terms it needs in the request form. Because preparing licensed records is part of the work SourceX carries out for the supplying company, a semantic map can be written into the request as a preparation deliverable; whether a supplier's records can support it is settled during review, not promised up front.
State an acceptance test the supplier can run. A workable one: draw a blind sample of mapped rows, have the supplier's process owner re-check them against the records, and accept the map if agreement meets a threshold the buyer sets in advance. Rows that fail go back to contested, not into the training set.
Frequently asked questions
What is status semantics in a cross-system dataset?
Status semantics is the documented meaning of each status value: what triggers it, who may set it, whether it can be undone, what it asserts and which companion field completes it. In a cross-system dataset it is the evidence that two statuses with the same label, or different labels, describe the same business event. Without it, a join on the label decides meaning by accident.
Why not map status values with a simple lookup table?
A lookup table records that two values are treated alike but not how alike they are or since when. Recording a match type, such as exactMatch, closeMatch or broadMatch from the W3C SKOS vocabulary, plus the period an owner confirmed, lets a reviewer see which rows can carry a label and which need splitting first.
What is the difference between Status and Status Reason in Dynamics 365?
Status is the state of a record, such as Active, Resolved or Cancelled for a case; Status Reason is a finer value that sits under one state. Microsoft's case table lists Problem Solved and Information Provided under Resolved, and Merged under Cancelled, so an export that keeps only the state loses the difference between a fix, an answer and a duplicate.
Why can a Jira issue be Done but still unresolved?
Done is a status category, while resolved depends on the separate Resolution field. Atlassian's documentation says default workflows set the resolution automatically on the move to Done, but custom workflows need a post function, so issues can reach Done with no resolution value. Treat those issues as work stopped unless the owner confirms otherwise.
Who should sign off a status mapping before it becomes training data?
The owner of each source process signs off the rows for their own system after reviewing sampled records, and the data team signs off the pivot definitions. Each verdict should carry the reviewer's role, the review date and the period it covers, because a workflow edit can change what a status means from one quarter to the next.
Sources
- Microsoft documents that a Dynamics 365 case becomes Resolved only through the Resolve Case action, which collects a Resolution Type and Resolution, and that resolving a parent case resolves its active child cases. Source
- Microsoft's case table reference lists Problem Solved and Information Provided under the Resolved state and Merged under Cancelled, and says resolved and canceled cases are read-only unless reactivated. Source
- Atlassian documents that Jira sets the Resolution field automatically on the move to Done only in default workflows, and that plans define completed work by the Status Category Changed Date. Source
- Salesforce describes the Case closed flag as controlled by the Status field and not settable directly. Source
- HubSpot asks admins to mark each new pipeline stage Open or Closed for objects other than deals, and lists the default Support pipeline statuses. Source
- The W3C SKOS Reference defines exactMatch, closeMatch, broadMatch, narrowMatch and relatedMatch, and leaves closeMatch non-transitive to avoid compound errors across more than two schemes. Source
- The XES event-log standard's lifecycle extension records the transition an event represents and supplies a standard transactional model while allowing any model. Source
- Shi and colleagues (2023) found that adding irrelevant information to math word problems dramatically decreased large language model performance. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.