The evidence became more useful when its apparent certainty was questioned.

In 2023 Soravane reviewed a proposed first-notice triage evaluation built from 120000 historical claims. The analytical file was convenient because it combined information eventually held for each case. That convenience was also its weakness. In 18000 records, settlement fields had been entered after first notice. A replay using those fields could appear to anticipate a development that the original decision-maker could not yet have known.

Emi Tsurahana asked when the fields became available, rather than whether the values were correct in the latest file. The delivery sponsor wanted to retain the initial evaluation because it offered a concise demonstration of progress. I understood the cost of rebuilding the account, but the historical decision point was essential to the question. Accurate later information was still unsuitable as an earlier input. Removing the most obvious settlement label without checking related fields would not have settled that problem.

I chose to withhold the original performance claim and reconstruct the eligible inputs for each historical first-notice point. JPY 310 million supported the review and data work. We kept later outcomes where they were needed to describe what subsequently happened, while distinguishing them from information permitted as inputs to the earlier decision. The claims function retained responsibility for handling decisions, and the analytical owners remained responsible for a properly specified evaluation after the data reconstruction.

What we gave up was the immediate use of a persuasive retrospective result. The delivered output was an inspectable input history and a revised basis for evaluation, not a replacement score or proof of a service improvement. I did not describe the 18000 records as fraudulent or unusable for every purpose. The issue was their suitability for a particular time-dependent question. My responsibility was to prevent a completed file from giving the institution an unrealistic impression of what it could have known when the claim first arrived.

A 2021 review of 4600 policy-endorsement records found 730 entered after their stated effective date. The proposed reporting approach would replace the earlier policy view with the latest one and use that account for every historical question. It made the extract simpler, but it erased a distinction needed to understand how the institution's knowledge had changed.

One illustrative change referred to 12 July, was entered on 19 July and was therefore absent from an extract taken on 17 July. Those dates did not by themselves establish whether a claim should be paid or which policy interpretation was correct. They did show that an account of what the system held on 17 July differed from a later account of the terms said to apply on that date. I wanted both questions to remain answerable without treating either view as a universal substitute for the other.

The operating sponsor preferred the latest-state extract because it avoided explaining apparently inconsistent versions. I chose to preserve effective-time and recorded-time histories, with JPY 210 million in data work. Underwriting, claims and legal colleagues retained decisions about meaning, authority and cover. My team maintained the relationship between versions and the dates attached to them. We did not manufacture missing evidence to make the history look continuous, and we did not label every later entry an error merely because its effective date was earlier.

The cost was a more demanding explanation and a reporting account that could no longer rely on one unqualified historical snapshot. The benefit delivered was the ability to distinguish the policy view available at the time from the view assembled after additional information arrived. I did not claim that the reconstruction resolved all disputes or changed payments. It supplied evidence for the responsible owners to interpret. The case matters because a data leader must sometimes preserve two legitimate questions where a simpler interface invites the business to believe there is only one.

In 2016 an insurance-property map displayed 18700 records with visually identical markers. The source evidence was less uniform: 14300 entries were classified at building level, 3200 at postcode level and 1200 remained unresolved. A reader could easily assume that every dot represented an equally precise property location, even though the record did not support that impression.

Kenji Aritsune challenged the presentation during a review of the underlying location account. The analytical sponsor wanted to retain a complete-looking map, arguing that visible gaps would distract from the wider portfolio discussion. I thought the opposite risk was more consequential. A location-dependent analysis could acquire apparent precision simply because the display concealed where the source was coarse or unknown. Classifying an entry at building level also did not prove that every coordinate had been independently verified.

I authorised JPY 160 million of review and reporting work to retain precision classifications and unresolved status in the usable dataset and its presentation. We made the three groups distinguishable and documented what each classification meant. The response did not require assigning an invented coordinate to every unresolved record. Underwriting and specialist risk owners continued to determine the relevance of information to their decisions; my role was to make its location basis visible.

We gave up the visual uniformity of the original display. The result was a more qualified location account, not an insured-loss estimate or proof that a particular hazard model was accurate. The building-level group remained subject to the quality of its source records, while the postcode group could not be presented as exact premises merely because the software could place a point on a map. I learned to treat cartographic confidence as a claim requiring evidence. The task was to help a reader understand what the location could support before the ease of displaying it became a reason to trust it.

As head of claims information in 2007, I accepted a report that described 7800 closure events as final resolutions. The event count itself came from the operating record. The word final was the interpretation I approved. Within ninety days, 260 of the associated claims had reopened, and the report no longer supported the conclusion its headline invited.

Kenji Aritsune asked why a closure event was being used to make a statement about all subsequent activity. My initial preference had been to keep one concise completion measure for management. I had treated the familiar status label as if it established an enduring outcome. The claims colleagues did not need the data function to redefine their decisions; they needed an account that could distinguish an event occurring from the state of a case at a later observation point.

I took responsibility for correcting the report. JPY 48 million covered review and reporting work. We retained the original closure events, recorded the subsequent reopenings and introduced status at a stated cutoff. Removing the 260 from history would have made the initial count look more stable but would also have concealed what actually happened. Calling the other 7540 permanently resolved would have repeated my mistake in a smaller population. The observation window did not establish their future state.

The correction cost time and required me to explain why a confident management label had been too strong. Its outcome was a clearer relationship between event history and observed status, not evidence that reopening always reflected mishandling or that claims service had improved by a measured amount. I still use the example when reviewing language attached to operational data. A leader can be responsible for an overstatement even when no arithmetic is wrong. In this case, the change I needed to make was to preserve the history and withdraw the certainty I had added to it.

The 2025 duration report concerned a cohort of 11600 claims. At its cutoff, 9500 were completed and 2100 remained open. The completed group had a median duration of eighteen days. The proposed headline shortened this to an eighteen-day claims median, inviting a reader to apply the result to the entire cohort even though the open cases had no observed completion duration.

The reporting sponsor wanted a single number that could be compared easily from one discussion to the next. I agreed that the completed-case statistic was useful when correctly labelled. I did not agree that simplicity justified dropping the condition that determined who entered it. The cases still open could have a different age distribution, and their eventual completion times were not known. Adding their current ages to completed durations would answer neither question properly.

I chose to retain the eighteen-day median explicitly for the 9500 completed claims and present an age distribution and count of the 2100 open cases alongside it. JPY 270 million funded reporting reconstruction and the clearer cohort account. Claims management retained responsibility for work allocation and resolution. Actuarial and finance owners retained their own purposes and methods; this operational report did not provide a reserve conclusion or a forecast of when the remaining cases would close.

We gave up the convenience of treating one statistic as the entire service account. The delivered change was a report that showed both observed completion and continuing work without pretending they were interchangeable measurements. I did not publish an invented all-cohort median or claim that the new presentation reduced handling time. The case strengthened my view that absence from a convenient denominator is itself something management should be able to see. A data function helps when it keeps that population visible, even if the resulting conversation requires more than one number.