Systems and records
Insurance claim reopening dataset: testing whether a closed claim was really finished
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Test premature completion by pairing each closed claim with what the file held, and still had open, on the day it closed, then labeling why it later reopened. A reopening counts as premature closure only when its cause was already in the file, such as an unanswered request or an undecided exposure. Reopenings from later-discovered damage, new documents or payment corrections are counterexamples, not failures.
System export
Insurance claim reopening
Key takeaways
- Build the test set from closures, not reopenings, and freeze each file as it stood at the close timestamp.
- A reopen flag is not a failure label; regulators and actuaries treat many reopenings as expected.
- Label premature closure only when the cause was knowable at closure, such as an outstanding request or an undecided exposure.
- Keep later-discovered damage, new documents and payment corrections as counterexamples so a model does not learn to never close.
- Grade a model on the end state it recommends, close or keep open with the cited open item, not on its closing note.
How do closed claims that reopen test premature completion?
They test it by asking one question of every closure: was the claim actually finished on the day someone marked it closed, given only what the file held that day? A reopening is the later evidence that something may have been left undone, but it is not proof by itself. The test set is built from closures, not reopenings, and the label records whether the closure was premature, sound, or impossible to judge.
This matters for anyone building or evaluating claims automation. An AI agent that drafts closing notes, recommends closure or closes low-touch claims on its own is making a completion claim, the same claim an adjuster makes when the status changes to closed. A status field and a polite final letter look identical whether the work was finished or not. Reopened claims are one of the few places where the business itself later recorded that a completion claim needed another look.
The practical design is a paired set: closures that later reopened, and comparable closures that stayed closed through the same observation window. Each case carries a frozen snapshot of the file as of the closure timestamp, the reopen event with its stated reason, and a reviewer label that says whether the reason was knowable at closure.
Why a reopen flag alone mislabels the closure
A reopen flag answers whether the file changed state again, not whether closing it was a mistake. Most reopenings in property lines are expected parts of the claim lifecycle, and treating every one as an agent failure would teach a model to never close anything.
Florida's claims statute shows how many different events sit behind the word. It defines a reopened claim as one the insurer previously closed that is reopened at the insured's request for additional costs for loss or damage already disclosed, and it separately defines a supplemental claim for additional loss from the same peril that the insurer already adjusted. The same section sets different notice windows for the two, one year after the date of loss for a claim or reopened claim and 18 months for a supplemental claim, in the version current in 2025, and earlier versions used different definitions.
Regulators also tell consumers that reopening is a normal path. The California Department of Insurance residential property claims guide says that under most circumstances, if additional damage is discovered later, the policyholder can reopen the claim and request additional compensation. Actuaries treat reopenings the same way: Friedland's study note for the Casualty Actuarial Society lists an estimate for reopened claims as one component of unpaid claims, alongside case outstanding and IBNR. Reopenings are forecast, so their mere existence is not an error signal.
Claim systems record a reason, but the reason vocabulary is set by each carrier's configuration. Guidewire's ClaimCenter Cloud API documentation, for example, describes a reopen call that takes a note and a reason from a ClaimReopenedReason typelist with values such as mistake or paymentdenied. Those codes are useful starting evidence. They were chosen by a busy person at reopen time and are not a reviewed judgment about the earlier closure.
What did complete mean on the day the claim closed?
Complete means the claim reached a decision the rules and the policy allowed, every coverage part and exposure was resolved, and nothing the file itself still requested was outstanding. Defining this per line of business, policy form and state is the first labeling task, because the definition is what reviewers apply.
Timing and explanation duties are part of the definition. California's regulations, quoted in the Department of Insurance's 2026 guide for adjusting property claims after a major disaster, give an insurer 40 days after receiving proof of claim to accept or deny it in whole or in part. When more time is needed, the insurer must send written notice of what information is still required and repeat it every 30 days until it decides. The NAIC Unfair Claims Settlement Practices model act lists refusing to pay claims without a reasonable investigation and failing to promptly give a reasonable explanation of the basis for a denial or compromise offer as unfair practices.
Those duties give reviewers concrete closure-day checks. Was there an open written request for information with no response logged? Was a denial sent with the policy provision it relied on? Was an exposure still open? Guidewire's documentation notes that in the base configuration a claim with open exposures cannot be closed, which tells you a system rule may have forced exposures closed first; the label must look at whether each exposure was resolved on its merits, not at whether the validation rule was satisfied.
Policy version matters as much as rules. The endorsement schedule in force on the date of loss, not the current form, decides which coverages had to be addressed. Snapshot the declarations and endorsements that applied, with their effective dates, into each case.
The reopened case selection and outcome labeling plan
The plan below is the asset: eight steps that turn claim system history into a balanced, frozen test set with labels a second reviewer can reproduce. It is written for a property or auto book; adjust the windows for long-tail lines. Values in the right-hand column are illustrative defaults, not observed rates.
- Unit of analysis: one close event, so a claim that closed twice can contribute two cases with different labels.
- Starting state for a model task: the frozen closure-day file plus the instruction to recommend close or keep open, citing the open item if any.
- Reference outcome: the reviewer label, never the reopen flag alone.
| Step | Rule | Fields captured | Illustrative default |
|---|---|---|---|
| 1. Define the closure population | Every claim-level close event in the study period, including second and later closes on the same claim | claim ID, close timestamp, closed outcome code, closing user role, line of business, state | Closures from a two-year window, all property lines |
| 2. Set the observation window | A closure is eligible only after it has had the full window to reopen; record the window in the spec | close date, window end, maturity flag | 18 months for property, longer for liability |
| 3. Split by later event | Group A reopened inside the window; group B stayed closed through it | first reopen timestamp, reopen reason code, reopen note, requester role | All of group A, matched sample of group B |
| 4. Match the comparison group | Match group B to group A on line, state, closure quarter and handling tier so the model cannot learn the calendar | matching keys, match ratio | One or two matches per reopened case |
| 5. Freeze the closure-day file | Rebuild the file as of the close timestamp: notes, activities, correspondence, documents, reserves and payments dated on or before it | snapshot hash, document list with dates, open activity list | Anything dated after close is held back for labeling only |
| 6. Label the reason | Two reviewers assign one reason label from the taxonomy using the full history, then mark whether that reason was knowable at close | reason label, knowable_at_close, evidence references, reviewer IDs | Disagreements go to a third reviewer |
| 7. Exclude what cannot be judged | Drop or quarantine cases with missing closure-day notes, migrated files with lost dates, or litigation holds | exclusion reason | Keep an exclusion count by reason in the dataset card |
| 8. Version and hold out | Freeze the set with the taxonomy version, window and matching rules; hold out whole closure periods for testing | dataset version, taxonomy version, split | Train on earlier closures, test on later ones |
Which reopen reasons count as premature closure, and which are counterexamples?
A reopening counts as premature closure when its cause was already in the file at closure and a careful handler would have acted on it. Everything else is a counterexample that teaches the model when closing was right even though the claim came back.
| Reason label | What the closure-day file shows | What the reopen shows | Premature? |
|---|---|---|---|
| Open request ignored | A written request for documents or an estimate with no response or follow-up logged | Reopen to process the document that was already requested | Yes |
| Exposure closed unresolved | One coverage part or claimant never addressed on its merits | Reopen to decide the forgotten exposure | Yes |
| Denial without stated basis | Denial letter missing the policy provision or reason | Reopen after complaint or regulator inquiry | Yes, if the rule required the basis |
| Closed in error | Wrong claim closed, duplicate confusion or data entry error | Reopen with a mistake-type reason code | Yes, as a process error rather than a judgment error |
| Later-discovered damage | Nothing pointing to the hidden damage | Insured reports new damage from the same event | No, counterexample |
| New document after closure | No pending request; the document did not exist yet | Invoice, repair receipt or ruling arrives later | No, counterexample |
| Payment correction | Decision complete; payment later voided or returned | Reopen to reissue or correct a payment | No, label as administrative |
| Recovery or litigation event | Decision complete; subrogation or suit begins later | Reopen to track recovery or defense | No, label separately |
Worked case (illustrative): a water loss closed while an estimate was still requested
The following case is illustrative. It is not drawn from any real claim file, and the counts and durations are invented to show the labeling steps.
Starting state. A homeowner reports a kitchen water leak. The adjuster pays the undisputed portion based on a desk review, and on day 12 the activity log shows a letter asking the insured for the flooring contractor's estimate. On day 41 the claim is closed with outcome code completed. The closure-day snapshot has seven documents, one open activity titled awaiting contractor estimate and no reply from the insured.
Linked evidence. On day 55 the estimate arrives, dated day 30. The claim is reopened with the reason code new information. The reviewer sees that the estimate was requested before closure, that the open activity was still pending at closure and that no follow-up letter went out after the request.
Decision and label. Reason label: open request ignored. Knowable at close: yes, because the file itself recorded the outstanding request. Premature: yes. The model task built from this case gives the day-41 snapshot and expects keep open, citing the pending estimate request.
Counterexample. A second illustrative claim from the same storm closes on day 38 with every request answered and every exposure decided. On day 200 the insured reports mold behind a cabinet that was not visible at inspection. The claim reopens with the same new information code. Reviewer label: later-discovered damage, knowable at close no, premature no. The model should recommend close on the day-38 snapshot, and a model that refuses is overcautious, not careful.
Which linked record proves the closure decision and the eventual outcome?
The proof is the dated record chain, not the closing letter. Each case should point to the specific records a reviewer used, with timestamps, so a buyer or auditor can retrace the label.
At a glance
Which linked record proves the closure decision and the eventual outcome
Grade a model on the state it would leave the claim in, not on how convincing its closing note sounds. The tau2-bench evaluation documentation describes the same principle for agent benchmarks: a task's reference actions are replayed to produce a target database end state, and an agent passes when its own end state matches, whatever path it took. For closure tasks, the comparable end state is close or keep open plus the cited open item, checked against the reviewer label.
Score two kinds of error separately, because they cost different things. Catching premature closures means the model keeps open the cases reviewers labeled premature and names the same open item they found; a keep-open answer with the wrong reason should count as a partial miss. Over-holding means the model keeps open the counterexamples that were sound closures, which in production would leave finished claims sitting in a queue. Report both by reason label and by line of business rather than one blended accuracy figure.
- Closure decision: the close event with timestamp, user role and outcome code.
- Open items at closure: activity list and correspondence log filtered to on or before the close timestamp.
- Policy context: declarations and endorsements in force on the date of loss, with effective dates.
- Decision basis: the coverage letter or denial letter with the provision cited.
- Eventual outcome: reopen event, reason code, note, the record that triggered it and any reserve or payment change afterward.
- Reviewer trail: both reviewers' labels, the adjudicated label and the evidence references each relied on.
Where the labeling plan needs caution: windows, jurisdictions and disagreement
The plan measures something real only if its limits are written down. Reopen windows differ by line; a liability claim can reopen after a suit is filed long after a property claim would have settled, so a short window undercounts reopenings and mislabels sound closures as stayed closed.
Jurisdiction changes what counted as complete. Florida's statutory windows and definitions have been amended several times, and California's timing rules are not the same as other states'. Record the state and the rule version each label relied on, and do not pool states into one premature rate without saying so.
Reviewer agreement is the main uncertainty. Report the share of cases where the first two reviewers agreed, the number sent to a third reviewer and the number excluded, and give intervals for any rate the set is used to estimate. This page reports no results; it describes how to build and label the set.
Leakage is the quiet failure. Notes typed after the close but given an earlier activity date, documents re-indexed during a system migration and reserve changes posted in batch can all slip post-closure knowledge into the closure-day snapshot. Check a sample of snapshots by hand against the audit trail, and drop any case whose dates cannot be trusted rather than guessing.
Preparing reopened claim files before they leave the insurer, TPA or agency
Closed claim files hold names, addresses, policy numbers, photos of homes and sometimes medical or financial details, so the reason labels and the evidence chain must survive redaction while identifiers do not. Replace claim and policy numbers with stable tokens, shift dates while keeping intervals, and keep the reason label and knowable-at-close flag intact.
Rights come before redaction. Whether a carrier, a third-party administrator, an agency or a restoration contractor may share these files depends on contracts and the data each one created; removing identifiers does not by itself create a right to license the records. Counsel should review the specific agreements and the state privacy rules that apply.
A carrier, TPA or agency with years of linked claim histories can ask whether they could become a licensed dataset; SourceX handles that path from qualification and rights review through contracting, preparation, delivery and payment. The eligibility check is a short company assessment with no data upload, and no file is delivered before the company has approved the scope and terms and signed.
Frequently asked questions
What is the difference between a reopened claim and a supplemental claim?
Florida's statute draws the line clearly: a reopened claim is a previously closed claim reopened for additional costs on damage already disclosed, while a supplemental claim is for additional loss from the same peril that was already adjusted. Other states use different terms or none. Label both separately, because they carry different signals about the original closure.
How long should a closed claim be watched before it counts as staying closed?
Long enough for most reopenings in that line to have happened, which is a business choice you should document. Property lines often need a shorter window than liability or workers' compensation. Whatever window you choose, apply it to every closure in the comparison group and state it in the dataset card.
Can the claim system's reopen reason code replace human labels?
No. The reason code was picked at reopen time from a carrier-configured list, and it describes the reopening, not whether the earlier closure was sound. Use it as starting evidence, then have reviewers decide whether the cause was knowable at closure.
Should claims reopened only to correct a payment be included?
Yes, as labeled counterexamples. They teach a model that a later administrative event does not mean the closure decision was wrong. Keep them in a separate administrative label so they do not inflate or dilute the premature closure count.
Who should review the labels in a claim reopening dataset?
An experienced claims professional for the line of business should apply the taxonomy, with a second reviewer for agreement checks. A privacy and rights reviewer should approve what leaves the company. Where a label depends on a state rule, someone who knows that state's claims regulations should confirm it.
Sources
- Florida Statutes 627.70132 defines reopened and supplemental claims separately and sets different notice windows for each. Source
- The California Department of Insurance tells policyholders that, under most circumstances, later-discovered damage lets them reopen a claim. Source
- California's rules give insurers 40 days after proof of claim to accept or deny, with written status notices every 30 days when more time is needed. Source
- The NAIC Unfair Claims Settlement Practices model act lists refusing to pay without a reasonable investigation and failing to explain a denial's basis as unfair practices. Source
- Friedland's CAS study note treats an estimate for reopened claims as one component of unpaid claims. Source
- Guidewire ClaimCenter documentation describes close and reopen calls, closed outcome codes, reopen reason codes and the open-exposure validation rule. Source
- The tau2-bench evaluation documentation grades agents on whether their database end state matches a target end state. Source
See if your company qualifies
A short company assessment. No data uploads are needed.