Five decisions where the factory challenged the first technical account

Before a September 2024 production-application migration, we rehearsed 420 work orders through the proposed route. A total of 336 agreed with the authorised engineering and operating records. Eighty-four did not: fifty-two involved engineering revisions and thirty-two involved routing. The migration team could technically load either version. That was not the same as having authority to decide which instruction the plant should receive. Some differences reflected outdated records; others required an explanation from engineering or the relevant production owner.

I asked the programme to preserve the conflicting versions and assign each question to the person authorised to resolve it. The alternative was to apply a common conversion rule and correct exceptions after introduction. That would have made the transfer appear simpler while moving an unresolved decision into the plant's working day. Masato Yurinobe supported a phased introduction over three additional weeks. The remediation and parallel work cost JPY430 million, including specialist participation and continued application support.

During that period, engineering owners addressed the fifty-two revision questions and production owners addressed the thirty-two routing questions. My team then applied the agreed records and repeated the relevant rehearsals. All eighty-four conflicts were reconciled for the examined orders. We retained the decision trail, including the reason for a local routing where it remained appropriate. The result was not a claim that every possible future order had been checked or that the migration itself made the factory faster.

The useful operating change was a clearer route for questioning an instruction when the enterprise record and the plant's understanding diverged. We could now distinguish a technical conversion error from a change requiring engineering or production authority. I took the three-week extension and its cost into the programme account rather than subtract them from the story of completion. A shared record earned greater credibility because its difficult differences had been examined, not because the technology team had made them disappear from the report.

I would bring engineering and production owners into the first conversion rehearsal, before the programme treats disagreement between records as a technical defect with a single correct repair.

An enterprise connectivity proposal in February 2025 covered eighty-four machine cells across a defined group of plant areas. The initial presentation treated them as one introduction cohort. Review with plant engineering and equipment suppliers found that sixty-three fitted the agreed maintenance and support route, while twenty-one required further examination of their configurations and operating conditions. The eighty-four cells were the programme scope, not the total number of cells across Kosemori's seven plants.

The choice was whether to retain a common introduction date or separate the cohort according to the evidence available. I recommended proceeding with the sixty-three under the agreed conditions and holding the remaining twenty-one outside that first stage. The technical team could not remove the uncertainty by applying office-system assumptions to operating equipment. Plant engineering retained authority over control and safety matters, and the relevant suppliers needed to contribute the knowledge their equipment required.

We funded JPY260 million for the defined connectivity and segmented support programme. That included the work needed to keep the two cohorts distinguishable, maintain support ownership and prepare the unresolved cells for further review. A narrower initial scope reduced the apparent completeness of the programme, but it gave operations a usable account of what had actually been introduced. I did not present the decision as a safety certification or assign a financial value to an incident we could not show would otherwise have occurred.

The sixty-three cells entered the agreed arrangement through their scheduled maintenance windows. The other twenty-one remained on their existing supported routes pending the required examination. Our closure report kept that outstanding scope explicit and assigned it a continuing owner. The practical lesson was that enterprise coordination can improve without forcing every piece of plant equipment into the same timetable. My responsibility was to provide the resources and coherent account; it was not to replace the plant's operating judgement with a percentage of connected assets.

I would assess cell readiness before publishing the connectivity sequence, giving plant colleagues an earlier opportunity to explain conditions that a common equipment inventory cannot establish.

A planning integration trial in June 2025 used 120 proposed production orders. When examined against the agreed crew, tooling and maintenance conditions, eighty-eight were feasible as proposed and thirty-two were not. The system had produced a plausible sequence using the constraints it knew. The missing or outdated conditions meant that plausibility was not yet an operating commitment. I wanted the trial to expose that gap before the enterprise treated the output as a promise to another function.

We could have simplified the demonstration to the eighty-eight feasible orders and continued extending the planner. Instead, I asked production colleagues to explain the thirty-two exceptions and identify which conditions needed better representation. Some required a different sequence; others needed a tooling or crew decision outside the planner's authority. Masato Yurinobe's team revised those proposals and agreed a staggered release. My function delivered the JPY190 million integration work and the route through which updated conditions could reach the planning process.

We kept the owner of the operating plan visible. The application supplied proposals and supporting material; authorised production colleagues accepted or changed them. This was not a manual workaround hidden behind an automation claim. It was an intentional division of responsibility while we learnt which constraints the common model could represent reliably. We also recorded why an order changed, so the next planning cycle could benefit from the explanation rather than encounter the same exception as if it were new.

At the end of the trial, all 120 proposals had an agreed disposition, including the revised sequence for the thirty-two exceptions. That did not establish that every order subsequently shipped on time or that the investment caused a measured increase in output. Those outcomes would require their own operating evidence. The immediate result was a planning process whose assumptions could be inspected by the people expected to act on it. I considered that a stronger basis for expansion than a faster calculation that left the factory to discover its missing conditions after release.

I would include crew, tooling and maintenance constraints in the first planning demonstration, so the factory can challenge feasibility before a fast calculation is presented as a usable schedule.

We evaluated a component-surface inspection model in November 2025 using a held-out pool of 18,000 labelled images. The model flagged 432 images. Review against the agreed labels identified 360 correctly flagged defects and seventy-two false flags. It also found twenty-eight labelled defect images that the model had missed, leaving 17,540 correctly unflagged images. These were results from a controlled image pool, not an estimate of defects escaping from an unreviewed production process.

The proposal initially emphasised the speed with which the model could screen images. I asked for the complete error account and the intended operating decision before supporting further use. False flags created review work, while misses raised a different question about reliance on the signal. A single overall accuracy number would have concealed much of that distinction because the pool contained many more non-defect images than defect images. We also needed to understand how the selected material related to changing lighting, component families and production conditions.

I supported the JPY110 million pilot as an aid to human inspection, with product disposition remaining under the quality organisation's authority. The alternative of automatic release was not supported by this exercise. We retained the test material and labels, documented the error categories and asked the sponsors to define a further evaluation covering the proposed use. We did not infer a staffing reduction from faster image processing or claim that the system detected every kind of safety-relevant defect. The model's purpose was narrower than that.

The pilot gave the enterprise a concrete basis for deciding what to investigate next. It showed useful signals and limitations in the same account, rather than asking colleagues to choose between enthusiasm for a new tool and rejection of it. My role was to make that account fit the decision. Quality colleagues determined whether and how the signal could assist their work, while my team maintained the technical evaluation and the resources for a bounded next step.

I would agree the consequences of false flags and misses with quality colleagues before selecting the model, ensuring the evaluation answers an operating question rather than rewarding one attractive aggregate.

An energy-information programme in April 2026 sought to relate machine readings to production intervals. We examined forty-eight machines in a defined pilot. Forty supplied readings aligned with the agreed common clock; eight showed sufficient clock drift to make the proposed fifteen-minute attribution unreliable. The meters were producing numbers, but the apparent association between energy and a particular operating interval could be wrong. Ranking machines from that first combined data set would have presented an unexplained timing difference as an operating conclusion.

I recommended correcting the measurement arrangement before publishing comparative performance claims. The alternative was to include the eight with a broad caveat and refine the analysis later. That would have made a qualification easy to miss once the rankings travelled into another presentation. We invested JPY85 million in the defined instrumentation and time-alignment work, with plant colleagues involved in checking which operating intervals the readings actually described. We kept the original observations so that the correction did not erase the explanation for the earlier limitation.

After the work, all forty-eight machines supplied usable aligned readings for the agreed review period. That result supported further analysis, but it did not establish an energy saving. A comparison also needed to account for product mix, utilisation, idle periods and other operating conditions. The technology team could make the timing more dependable without deciding which production practice should change. We therefore presented the corrected measurement capability separately from any proposed operating intervention or later claim about consumption.

The case mattered because the first version of the report looked precise. Fifteen-minute intervals and machine-level figures can make an account seem more authoritative than its underlying alignment warrants. I wanted the enterprise to know which part of the measurement it could now rely on and which questions still belonged with operations. Our immediate outcome was a better-founded evidence base for the pilot cohort, not a company-wide efficiency result. That distinction kept the investment useful without requiring us to invent a benefit before the next operating decision had been made.

I would check clock alignment before selecting the energy comparison period, preventing measurement work from accumulating around observations that do not yet describe equivalent operating intervals.