A field study under our Knowledge and Memory research. A regional manufacturer's audit picture lived across three disconnected systems and inside the heads of a few people. We studied what it takes for a system to hold that knowledge as it is created and keep it current.
Most of what an organization understands about itself lives in people, and when they leave it leaves with them. Documentation goes stale the moment it is written, and the real knowledge stays informal, undocumented, and fragile. Knowledge and Memory, our second research direction, asks whether a system can retain operational knowledge as it is created and keep it current. A regional food manufacturer gave us a place to study that, because its audit picture was decaying in exactly this way.
The Setting
The manufacturer ran three facilities under SQF certification, a food safety standard most major retailers require. Maintaining it meant assembling weekly evidence that the food safety system was operating in control: batch records, ingredient traceability, inspection results, hold events, and corrective actions. That evidence lived across three systems that did not talk to each other, a legacy ERP, an OSIsoft PI historian, and a cloud quality management system, and the weekly task of stitching them together by hand consumed about 20 hours of the quality team's time.
The interesting part was not the manual labor. It was what the manual process quietly lost. Under time pressure, the team had built up informal workarounds: small discrepancies between systems were carried forward rather than resolved, on the assumption they would get cleaned up before the annual audit. Knowledge of which mismatches were real and which were artifacts lived in a few experienced heads. The system of record was, in practice, those people.
What Was Actually Decaying
Three systems each held a piece of the picture and none agreed perfectly with the others. Timestamp offsets between the historian and the ERP created near-miss matches that a person had to confirm. Inspection records referenced lot numbers that did not always match the ERP because of manual transcription at the line. Hold events were sometimes closed verbally without the workflow ever being completed in the QMS.
The QMS itself had years of accumulated entropy. Inspection categories had been renamed without migrating historical records. Legacy hold codes no longer mapped to current states. A partial migration two years earlier had left duplicate records for a three-month window that no one had ever cleaned up. Every one of these was a small piece of lost institutional memory, and together they meant the operation no longer had a reliable account of its own past.
What We Studied and Built
We built a system that connects to all three sources, normalizes them into a single batch-level record model, reconciles them, and detects discrepancies on its own. The research question underneath was retention: can the system hold the reconciliation knowledge that previously lived in people, and keep it current as new records arrive, instead of letting it decay?
The normalization layer does not assume the source data is clean. It assumes the known failure modes will recur and handles each one explicitly. Historian timestamps are resolved against the ERP's authoritative lot creation time. QMS lot formatting is standardized to ERP conventions, including the specific variations the line operators introduce. Discrepancy detection is categorical: data mismatches, missing records, and unresolved holds each route to the right reviewer, and an open hold escalates if it sits too long past its disposition window.
The first two weeks went entirely to mapping the data model before any integration code was written. Every known inconsistency was documented. Every field mapping was validated against real records. That unglamorous front-loading is where most similar efforts fail, because a reconciliation layer built on an incomplete understanding of the source surfaces false discrepancies and misses real ones, which makes it less trustworthy than the manual process it replaced.
The Validation
We ran the system in parallel with the manual process for four weeks and compared outputs side by side. It surfaced three discrepancies that had been passing through the manual process undetected for months: a traceability record mismatched between ERP and QMS, a hold from eight months earlier resolved verbally but never closed, and a recurring duplicate lot pattern from one line that a person had been silently correcting by habit.
"We found errors in the old manual process during parallel testing. That told us the automation was not just faster, it was more reliable." Quality Manager
Finding those was not a setback. It was the evidence we were looking for. The knowledge needed to catch them had existed, but only informally, and it was not reliably applied. Once the system held it, it was applied every week.
What the Study Showed
- Manual reconciliation hours eliminated: 20 hrs/week across three people
- Discrepancies surfaced weekly that previously passed undetected: 2 to 3
- Legacy data anomalies surfaced and corrected during the study: 47 records
- Two-year corrective-action backlog from inconsistent follow-through: cleared and kept current
The deeper result is the one that matters to the research. The operation's account of itself stopped decaying. Reconciliation knowledge that had lived in a few people and degraded under pressure now persists in a system that applies it consistently and keeps it current as new records arrive. That is the property we are after in Knowledge and Memory: not a faster report, but an operation that retains what it learns instead of losing it.
Where This Points
The same record model that holds this knowledge extends naturally as requirements grow, because the reasoning is over a living model of the operation rather than a fixed report format. That direction, a system that maps how an organization actually runs and keeps that map current, is the research behind Atlas. This study is one early piece of evidence that holding operational memory in a system, rather than in people and stale documents, is both possible and worth doing.