What the engineering result allowed us to decide next

In July 2024, a rehearsal of a policy-document service used 6,400 synthetic references drawn from an approved map of product editions. The service returned a document for every reference, which made the initial availability result look satisfactory. Inspection found that 184 references pointed to the newest wording rather than the edition assigned to the sample contract. A successful retrieval had concealed a different correspondence problem.

The application team preferred a single current document store because it simplified maintenance. Product colleagues needed the historical edition relationship to remain visible. I supported an edition-aware reference service, with the approved mapping supplied by the responsible product owners. Engineering would preserve and test that relationship; we would not decide which wording legally governed a real customer's circumstances. That distinction mattered because the technical team could demonstrate a match without possessing the authority to determine coverage.

The nine-week correction cost JPY86 million. We retained the required historical editions, revised the reference logic and repeated the same 6,400 synthetic cases. All then matched the approved map. The 184 corrections were part of that defined rehearsal, not a discovered population of affected customers. We did not infer financial loss, compensation or legal error from a test designed to examine document association.

The chosen approach increased storage and maintenance obligations. It required an owner for future edition changes and a method for preserving the mapping when products were withdrawn from new sale. I accepted those obligations because a simpler current-only store could not answer the historical question the service was intended to support.

The outcome was a reproducible technical reference account and a clearer boundary between retrieving a document and interpreting it. Product and legal colleagues retained their decisions. Looking back, I would make historical-edition examples part of the initial acceptance discussion, before the presence of a returned file is allowed to stand in for evidence that the relevant file was returned.

In January 2025, actuarial colleagues were waiting 9.6 hours for a defined processing replay on an agreed 1.2 million-record snapshot. The delay restricted when they could examine results during their working day. A proposal to distribute the processing offered a shorter window, but it would increase recurring compute expenditure and introduce more coordination between tasks. The decision was not simply whether the newer arrangement was technically faster.

I asked the actuarial owner to fix the input snapshot, assumptions and output-comparison criteria before we compared the two arrangements. Platform colleagues wanted to demonstrate maximum throughput using a smaller convenient sample. Finance wanted the change funded through a claimed reduction in staff effort. I rejected both shortcuts. A different sample would not answer the stated processing question, and we had not established a staffing consequence from reduced machine waiting.

The ten-week engineering programme cost JPY120 million. Repeated runs of the agreed snapshot and assumptions completed in 4.1 hours under the revised arrangement, against the 9.6-hour baseline conditions. The responsible actuarial team accepted the output comparison within its specified tolerances. The new arrangement added JPY3.2 million monthly compute cost. We recorded the repeated-run conditions and that recurring charge beside the duration result.

The business gained a different opportunity to examine outputs during the day. We did not claim that every future portfolio would finish in 4.1 hours, that the estimates became more accurate, or that the shorter interval removed the need for actuarial judgement. Larger inputs, different assumptions and competing workloads could change the result.

I supported the additional expenditure because the receiving team had explained a useful working sequence, not because a speed ratio looked attractive in a presentation. I would ask that team to describe its decision calendar at the start of the next proposal. That would let us compare the value of several possible completion windows before engineering effort becomes organised around the shortest technically achievable run.

In June 2025, our portfolio review treated an older product engine as though its replacement would soon remove the need for specialist maintenance. The successor still had acceptance work ahead, while the existing engine continued to support a defined book of policies. Reducing its funding immediately would have made the investment presentation simpler without making the operating obligation disappear. I asked for the transition to be costed as work in its own right.

The replacement sponsor wanted to move the remaining specialists onto new development. Maintenance leaders argued for keeping the existing arrangement indefinitely, concerned that an exit date would become a reason to withdraw support too early. Neither position gave the portfolio a satisfactory decision. I proposed an eighteen-month bridge with named service ownership, a defined maintenance remit and staged reviews of the successor's acceptance evidence.

The approved funding was JPY240 million for the bridge period. It retained a small specialist team and the environment needed to maintain the older engine while the replacement progressed. We deferred two discretionary analytics features to release the necessary engineering capacity. That was a visible opportunity cost, rather than a claim that the bridge could be absorbed without affecting other commitments.

By the March 2026 review, the older service remained supported and the successor had completed two of its three planned acceptance stages. The remaining stage was still open. I reported that position without claiming retirement savings or treating the bridge approval as proof that the replacement would meet its revised date. Any further funding decision would need the current evidence and a fresh account of the continuing obligation.

The value of the decision was an honest, funded transition with explicit alternatives. It did not turn temporary spending into an entitlement to permanent preservation. I would put the bridge option into the original replacement business case next time, alongside the optimistic transition route, so that the organisation can examine the cost of delay before committed service funding depends on a single assumed completion date.

In October 2025, we rehearsed a surge in first-notification intake using synthetic messages and an agreed traffic pattern. The service could acknowledge ordinary traffic, but a burst of 1,800 submissions per minute caused acknowledgement latency to climb and duplicate retry messages to enter the processing queue. The exercise concerned technical intake. It did not represent an actual catastrophe or the insurer's ability to assess and settle the resulting claims.

Infrastructure colleagues proposed increasing the front-end capacity alone. Claims operations were concerned that rapid acknowledgements could be presented as completed claim handling while work accumulated behind them. I supported a queueing and duplicate-recognition change with separate measures for acknowledgement and downstream processing. The decision required more operating visibility and a defined way to show pending work, rather than a single green status for the entire service.

Eight weeks of work cost JPY94 million. In the repeated thirty-minute rehearsal at 1,800 submissions per minute, the 95th-percentile acknowledgement time fell from eighteen to six seconds. Of 54,000 synthetic submissions, 2,160 were deliberate retries. The revised service associated those retries with their original intake records, leaving 51,840 distinct records for downstream work. We retained the distinction between submissions and distinct records in the report.

The queue still took seventy-four minutes after the burst ended to finish the defined downstream technical processing. That elapsed time was not a claim-assessment or settlement measure. I included it because faster acknowledgement alone could obscure an important continuing workload. Operating colleagues retained responsibility for the staffing and decisions needed beyond the technical intake boundary.

The result improved the tested handling of a specified burst and made the remaining work easier to see. It did not establish readiness for every catastrophe scenario or a service promise to policyholders. In a future rehearsal I would agree the language of the operating status report before choosing the technical benchmark, so that a useful acknowledgement result cannot be mistaken for completion of the customer's wider claim.

In February 2026, a product-platform change could support a revised commercial-property configuration before the insurer had completed its product approval process. Engineering had demonstrated the configuration in a controlled environment. The commercial programme wanted a single launch milestone, while product and operating owners still needed to settle wording, pricing and service arrangements. The risk was that a technical release label would become an implied operating authorisation.

I recommended separating installation of the capability from activation for live business. The commercial sponsor worried that this would weaken momentum and create two dates to explain. Engineering managers preferred a single release package because it reduced coordination. I accepted the additional sequencing work and asked for explicit activation authority, with the capability inaccessible to live product use until the responsible owners had completed their decisions.

The six-week programme cost JPY72 million, including the separate activation controls and rehearsal. The platform capability was installed on 18 March 2026. Product approval was recorded on 8 April after the remaining owners completed their review, and the controlled live activation followed on 13 April. The interval was deliberate. We did not describe the March installation as the date on which the insurer began offering the revised product.

During the separation period, the operating team used the prepared environment to rehearse its service steps without accepting live business through the new configuration. This exposed an unclear referral instruction, which the authorised owner corrected before activation. I do not attribute sales growth or underwriting performance to that correction; the observed benefit was a clearer instruction at the point of introduction.

The decision preserved useful engineering progress while keeping the commercial authority visible. It also required careful communication so that readiness in one workstream did not become pressure to approve another. Next time I would define capability installation and product activation as separate milestones from the programme's first charter, instead of introducing that distinction after teams have already organised their reporting around a single launch date.