Skip to content

AI uses for records

One closure in three said done: building a premature ticket closure dataset

By SourceX Editorial · Updated

9 sections

Short answer

A premature ticket closure dataset should judge each closure by what was still outstanding when the case closed, not by whether it later reopened: record each prerequisite, its owner and its state at closing. In 1,476 human-filed GitLab issues from March 2024, only 32.6% of closures carried the label meaning done, and 18.7% closed while their label still placed the work before merge.

Grader rubric

Premature ticket closure

This page covers one closure in three said done: building a premature ticket closure dataset.Illustrative

Key takeaways

  • A closure is premature when a prerequisite was outstanding at the closing moment with no named owner, whatever happens later.
  • Of 1,476 human-filed GitLab issues from March 2024, 481 carried the label meaning done; 276 closed while labeled as earlier-stage work.
  • Postal scans, records-request closures and hospital readmissions each have rules a support dataset can copy.
  • A prerequisite ledger records type, owner, state at close and evidence for each dependency, plus the closure reason and promise text.
  • Grade five verdicts, including unnecessary holds, and test on the closing-moment state with held-out prerequisite types.

When is a closed case premature rather than just early?

A closure is premature when something the promised outcome depended on was still outstanding at the moment of closing and nobody recorded who owned it. An early closure is fine if every prerequisite was met, or was explicitly handed to a named owner with a tracked link and the customer told. The difference is visible at the closing moment, which is where a premature ticket closure dataset should look first.

Prerequisites come in four kinds, and each fails differently. Internal work, such as an engineering fix or a credit that finance must post, fails when another queue rejects or delays it. External parties, a carrier, a vendor or a bank, fail outside the company's systems. Customer actions, such as returning a device or confirming a change, fail through silence. Time-bound steps, a billing cycle, a deployment or a release window, fail when the date slips.

Delayed outcomes are the later evidence that a prerequisite failed: a reopen, a repeat contact, a second credit, a returned shipment. They are useful for checking a closure label, but they arrive late, they miss customers who give up, and they also fire for problems that are new. A dataset that grades closures only by what happened afterward inherits all three faults.

Mail scans, records requests, hospital discharges: three rulebooks against early closure

Three public systems already treat early closure as a known failure and write rules against it. Each shows a different control a support dataset can borrow: verify where the closing event happened, make silence-based closure follow a documented step, and separate planned follow-ups from failures.

Postal delivery has the stop-the-clock scan, which the USPS Office of Inspector General describes as the scan showing the Postal Service completed its delivery or delivery-attempt commitment. Its audits keep finding those scans made at the delivery unit instead of the delivery point; a June 2026 report on the Virginia District counted 1,347 packages scanned that way at four units, most of them marked Delivered. A scan that ends the clock without the prerequisite, the package at the address, is the postal version of closing a ticket because the label printed.

Federal records requests have rules for closing on silence. Justice Department guidance on still-interested inquiries asks agencies to give requesters no fewer than 30 working days to answer, to try more than one way of reaching them before an administrative closure, and to reopen a request answered within a reasonable time after the deadline and return it to its original place in the queue. In August 2026 the U.S. District Court for the District of Columbia struck down an Energy Department notice that told everyone with older pending requests to reaffirm interest or be closed, distinguishing it from targeted inquiries.

Hospital quality measurement separates planned returns from failures. The 30-day readmission measures in Medicare's Hospital Readmissions Reduction Program count unplanned readmissions and exclude planned ones, which CMS identifies with its Planned Readmission Algorithm, version 4.0 2026 for fiscal year 2027. A scheduled follow-up visit is not evidence that a discharge was premature, and a scheduled callback is not evidence that a ticket was.

Mail scans, records requests, hospital discharges: three rulebooks against early closure
SystemWhat ends the caseRule against early closureControl to copy into a ticket dataset
Postal deliveryA stop-the-clock scanScan at the delivery point, not the unit; OIG audits check scan locationStore where and how the closing event was produced, not just its time
Federal records requestsAdministrative closureAt least 30 working days, several contact methods, reopen with queue position keptLabel silence closures separately and require the documented check-in
Hospital readmissionsDischargeCount unplanned 30-day readmissions; exclude planned ones by algorithmMark planned follow-ups so they never count as failed closures
Support desk (the gap)Solved or closed statusUsually none beyond an automation timerRecord prerequisites and their state at the closing moment

What 1,476 human-filed GitLab closures said about being done

In a large public tracker with a documented definition of done, only about one closure in three carried the label meaning done. We read every public issue filed in March 2024 in GitLab's own product project, gitlab-org/gitlab, through its API on October 27, 2026, set aside 6,299 issues filed by automation, and classified the 1,476 human-filed issues that were closed.

GitLab's Product Development Flow handbook defines Complete as set after the feature is deployed to production and any needed verification is done, with Verification and Awaiting security release as the stages between merge and done. Against that definition, 481 closures (32.6%) carried workflow::complete. Another 276 (18.7%) were closed while their stage label still placed the work before merge, including 66 labeled in review, 43 in dev and 9 blocked, and 5 were closed while labeled as merged but awaiting verification or a security release.

The rest say nothing about completion either way. Of those, 544 closures (36.9%) had no workflow label at all, 86 (5.8%) carry the auto closed label, 84 of them closed by GitLab's bot account, and 64 (4.3%) were closed by moving the issue to another project, a transfer of custody rather than a resolution. The handbook's label-to-status table also shows that the old labels had no value for Duplicate or Won't do, so an issue closed for either reason could keep only whatever stage label it last held.

Checklists tell the same story from inside the issue. Of 285 closed human-filed issues with a task list in their description, 221 (77.5%) had at least one unchecked item. Some are template boxes nobody meant to tick, which is the point: a closed status next to an unfinished list leaves a grader unable to say which one is true.

Counting rules, as the script applies them: population is every issue the anonymous endpoint returns for created_after 2024-03-01T00:00:00Z and created_before 2024-04-01T00:00:00Z; bot-filed means the label automation:bot-authored; each closed human-filed issue gets one class, tested in order: moved (moved_to_id set), auto-closed (label auto closed), complete (workflow::complete), merged but pending (workflow::verification or workflow::awaiting security release), before merge (any of twelve listed earlier-stage workflow labels), other workflow label, no workflow label. A bot closure is one whose closed_by account is gitlab-bot; an open checklist is a task list with fewer completed items than items. Labels and checklists are as of retrieval, not the closing moment. The count covers the whole population with no sampling or random step, and reviewers receive the script with this draft.

The seven classes and what each says about whether the work was done:

What 1,476 human-filed GitLab closures said about being done
Class at retrievalClosed issuesShareWhat the record says about completion
No workflow label54436.9%Nothing: closed, with no stage recorded
workflow::complete48132.6%Done by the handbook's definition
Earlier-stage label, before merge27618.7%Closed while the label says unbuilt, in review or blocked
Auto closed label865.8%Closed for inactivity, almost all by the bot account
Moved to another project644.3%Custody moved; the work continues elsewhere
Other workflow label201.4%A label outside the handbook's mapping
Merged, awaiting verification or release50.3%Closed before the last stated prerequisite
A waffle chart of 100 squares for 1,476 closed human-filed GitLab issues: 37 squares no workflow label, 33 labeled complete, 19 closed with an earlier-stage label, 6 auto-closed, 4 moved and 1 other workflow label, with 5 merged but unverified issues too few for a square.
What a closed status meant in 1,476 human-filed gitlab-org/gitlab issues created in March 2024, labels as of October 27, 2026; one square is about 1%. SourceX count.

The prerequisite ledger a closure record should carry

Every closure needs a ledger of what it depended on and the state of each item when the case closed. Without it, a grader can only guess from later events, and a model trained on the result learns that closing is the goal. The ledger below is the minimum that lets a reviewer decide, case by case, whether closing was valid on the day it happened.

Fields to capture with each closing event:

  • Closing event: timestamp, actor type (person, automation rule, merge, move) and the system that produced it, so a label-printed event is never mistaken for a delivered one.
  • Closure reason code: resolved, duplicate, won't do, moved, customer silence, or handed off, kept separate from the stage the work had reached.
  • Promise text: what the customer was told would happen, quoted from the last outbound message.
  • Prerequisites: one row each, with type (internal, external party, customer action, time-bound), owner, linked record and expected completion date.
  • State at close for each prerequisite: met, pending with named owner, waived with approver, or unknown.
  • Evidence pointers: the record that shows each met prerequisite, such as a posted credit, a carrier delivery scan at the address or a deployed release.
  • Silence step: for closures on customer silence, the check-in messages sent, channels used and the waiting period allowed.
  • Planned follow-ups: callbacks or scheduled visits, flagged so they are not read later as failures.
  • Delayed outcome window: the follow-up period, the events observed in it (reopen, repeat contact, failed prerequisite) and the extract date.

Decision packet: a replacement router closed on a shipping label

The case below is illustrative; the company, customer and details are invented to show how the ledger turns a closure into something gradable. It follows the postal lesson: the event that closed the case was produced before the prerequisite it stood for.

A small business customer reports that its office router drops connections every afternoon. The agent runs diagnostics, confirms a hardware fault and approves a replacement. The fulfillment system creates a shipping label, the macro sends a message saying the replacement is on its way, and the agent sets the case to solved. The automation closes it four days later.

Ledger at the closing moment: prerequisite one, replacement delivered to the customer's address, pending with the carrier, evidence only a label-created event. Prerequisite two, faulty unit returned, pending with the customer, return label sent. Prerequisite three, configuration restored on the new unit, pending, no owner named. Closure reason recorded: resolved. Promise text: your replacement is on its way and you should be back online in two days.

What happened next: the carrier reported an address exception on day three and returned the parcel; the customer opened a new chat on day nine. Verdict: premature close. Every prerequisite was pending, one had no owner, and the reason code claimed a resolution the ledger could not support. A valid alternative existed: keep the case open in a waiting state with the carrier tracking linked, or close it as handed off to fulfillment with the delivery scan as the reopening trigger and the customer told so.

The counterfactual twin makes the rubric honest. Same fault and replacement, but the carrier's delivery scan at the address, the returned unit's receipt and a remote configuration check are all in the record before closing. Verdict: valid close, even if the customer later writes in about something else.

Close, hold or hand off: a two-sided grading rubric

Grade closure decisions in both directions. A model that never closes anything would avoid every premature closure and still fail the business, so an unnecessary hold is an error with its own label, not a safe default.

Five verdicts, with the evidence each requires:

Close, hold or hand off: a two-sided grading rubric
VerdictDefinitionEvidence requiredTypical mistake it catches
Valid closeEvery prerequisite met at the closing momentA met state and an evidence pointer for each ledger rowNone; this is the target
Declared handoffOpen prerequisites have a named owner, a linked record and a reopening trigger, and the customer was toldOwner, link, trigger, outbound messageTreating every open dependency as a failure
Silence closeClosed on customer silence after the documented check-in stepsMessages, channels, waiting periodTimer closures with no attempt to reach the customer
Premature closeA prerequisite was pending or unknown with no owner, or the reason code claims more than the ledger showsThe ledger row that was openClosing on a label-printed or request-filed event
Unnecessary holdAll prerequisites met, but the case stayed open past the follow-up windowMet states with dates earlier than the eventual closeA model that learns never to close

Why the reopen flag, a quiet customer and a survey all mislead

Each of the usual proxies measures something real, just not the closure decision. A reopen flag records one route back into the queue; a quiet customer may have given up; a survey sent at closing scores the message, not the prerequisites behind it.

Reopens are also contaminated in both directions. Planned follow-ups look like reopens unless they are flagged, which is why Medicare's readmission measures remove planned readmissions before counting. And new problems from the same customer look like failures of the old case, so a reopen without a cause code overstates premature closure as surely as silence understates it.

Automation adds a third distortion. In the GitLab count, 86 human-filed issues were closed for inactivity, nearly all by a bot. A support desk that auto-closes solved tickets after a waiting period produces the same thing at scale: a closed status that records the passage of time. Records-request rules show the fix, a documented check-in through more than one channel before a silence closure, and a reopening path that keeps the requester's place.

Freeze the closing moment for held-out tests

Test a model on the state of the case at the closing moment, never on the finished file. Give it the conversation, the ledger as it stood when the human closed the case and the policy in force, then ask for one of three actions: close, hold or hand off. Grade the answer against the verdict derived from the ledger and confirmed by delayed outcomes after a stated window.

Hold out by prerequisite type, not by ticket. Put every carrier-dependent case, or every case that waited on finance, in the test set, so a model cannot pass by memorizing which queues usually come back. Keep cases whose follow-up window has not ended out of the scored set, and report how many were excluded.

Report the confusion between verdicts rather than one accuracy number. The two costly errors are premature close and unnecessary hold, and they trade against each other; a release decision should know which one moved.

Where this evidence stops, and what a supplier should export

The GitLab count describes one engineering tracker's labels as they stood on the retrieval date, not at each closing moment, and engineering issues are not customer cases. The postal, records-request and hospital examples are rules and audits, not datasets of support closures. None of these measures how often support tickets close prematurely; that needs the ledger fields above, collected from a real desk.

A supplier export that can support a premature ticket closure dataset carries full status history rather than final status, every closing event with its actor type and reason code, the outbound promise text, links to the records each prerequisite depended on (fulfillment, carrier, billing, engineering), planned follow-ups, check-in messages before silence closures and an extract date. Ask for the five-verdict rubric to be applied to a sample before a full delivery.

SourceX does not keep ready-made support datasets. A buyer files a request naming the task, volume, language, format, timing and license it needs, and SourceX then looks for U.S. companies whose case histories fit, running the rights review, contract, preparation, delivery and payment with them. Whether a particular desk links closures to carrier and billing records is established during that review, and the company approves scope and terms before anything is delivered.

Frequently asked questions

What is a premature ticket closure?

It is a support case closed while something the promised outcome depended on was still outstanding and unowned, such as a replacement not yet delivered or a credit not yet posted. A closure that hands an open item to a named owner, with a tracked link and the customer told, is a declared handoff rather than a premature close.

Is a reopened ticket proof that it was closed too early?

No. A reopen can come from a planned follow-up or a new problem, and many premature closures never reopen because the customer gives up or contacts the company another way. Use reopens to check labels after a stated window, not as the label itself.

Should tickets ever be closed automatically for customer silence?

They can be, but the closure should be labeled as a silence close and preceded by a documented check-in. Federal guidance for records requests is a useful model: at least 30 working days to respond, more than one contact method, and a reopening path that keeps the requester's place.

How long should a team wait before trusting a closure label?

Long enough for the slowest prerequisite to show its outcome, such as a delivery, a billing cycle or a release, and set per prerequisite type rather than one number for the whole desk. Cases still inside that window belong outside the scored set until they mature.

What fields does a premature closure dataset need from a help desk?

Full status history, each closing event with actor type and reason code, the promise text, links to fulfillment, carrier, billing or engineering records for each prerequisite, planned follow-ups, check-in messages before silence closures and the extract date. Final status alone cannot support the labels.

Sources

  • GitLab's public issues API for project gitlab-org/gitlab returned 8,170 issues created in March 2024 to an anonymous caller on October 27, 2026; all counts in this article come from that retrieval. Source
  • GitLab's issues API documentation describes the created_after and created_before filters and the closed_by, moved_to_id and task_completion_status fields used in the count. Source
  • GitLab's Product Development Flow handbook defines Complete as set after the feature is deployed to production and any needed verification is complete, lists Verification and Awaiting security release stages, and maps former workflow labels to statuses, with no former label for Duplicate or Won't do. Source
  • A June 2026 USPS OIG audit of the Virginia District counted 1,347 packages scanned at four delivery units rather than at the delivery point, most marked Delivered. Source
  • The USPS OIG describes a stop-the-clock scan as one showing the Postal Service has completed its delivery or delivery-attempt commitment. Source
  • Justice Department OIP guidance on still-interested inquiries sets a response period of no less than 30 working days, asks agencies to use multiple contact methods before administrative closure, and to reopen requests answered within a reasonable time after the deadline with their original queue position. Source
  • OIP's summary of American Oversight v. DOE (D.D.C. Aug. 5, 2026) reports that the court struck down a notice requiring requesters with older pending requests to reaffirm interest or face closure, distinguishing targeted still-interested inquiries. Source
  • CMS's HRRP measures are 30-day risk-standardized unplanned readmission measures that exclude planned readmissions, identified for FY 2027 with Planned Readmission Algorithm version 4.0 2026. 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