# Classification Must Be Saved, and Records Must Be Findable A classifier finally saves its result, while a records-management test exposes the gap between applying a label, enforcing retention, and delivering trustworthy search. Language: en Canonical: https://avanticomplex.com/en/blog/classification-must-persist-and-rm-must-search/ Published: 2026-08-20T06:59:00.000Z A clerk who shouts a category down the hall has not classified the file. That is where I start today. An AI demonstration can produce a convincing label and still leave nothing the business can use tomorrow. On August 19, I fix a write failure in VALESKA and test a separate records-management connection. Both matter. Neither earns me the right to call governance search finished. ## First, Make the Result Stay The classifier had three operational blockers: a conflict over access to the graphics processor, a waiting limit shorter than the model needed, and a database rule that rejected the classification method being recorded. The first two hid the third. Empty results made a working model look broken; once those obstacles were cleared, the save attempt exposed the remaining fault. The day's record reports the first successful classification at 396.7 seconds end to end for one document. That is an observation, not a dependable processing rate for a collection. I am not turning it into a delivery promise. More importantly, a saved label proves that the write path works. It does not prove that the label is right for the business. The model's confidence score is not an independently measured accuracy rate. ## A Subject Is Not a Retention Decision There are two different questions here. Subject classification asks what a document is about. Records management, or RM, concerns official records and the rules for keeping, holding, and eventually disposing of them. OpenText Content Server, or OTCS, is the enterprise content management system in this test. Enterprise content management, often shortened to ECM, means managing an organization's documents and related information. APQC, the American Productivity & Quality Center, provides a business-process taxonomy: an organized set of categories for activities such as finance and information management. It can help describe content. It cannot decide the legal retention period for a contract. The same-day research also corrects an earlier accusation: a large pile of documents under one unsuitable code came mainly from a rule-based classifier, not the AI classifier being evaluated. Blaming the wrong component would have sent the repair in the wrong direction. The broader label-ranking benchmark reported the expected category first in 24% of cases and somewhere among twenty candidates in 70%. Those are category-matching results, not business-answer accuracy. They explain why I should not let an uncertain subject label silently exclude documents from search. ## What the Records Test Actually Proved In the training system, one archive document received an existing RM classification, and VALESKA recorded the corresponding state. The test also raised the warning we needed: the document had no linked retention schedule. Being filed under a records category had not established when it should be kept or disposed of. Then we removed that classification and marked the local copy as no longer current. The available test category belonged to finance; the document concerned Guyana petroleum. Leaving it in place would have turned a successful connection test into false records data. Undoing the test was part of doing it properly. Three file plans, the organized structures used to file records, were also prepared and their generated import files checked. That supports a preparation milestone, not a claim that the plans were imported and operating. Retention schedules were deliberately left unassigned. Automatically copying the subject taxonomy into the content system remained unbuilt. ## The Next Test Is Retrieval The research changes how I frame search. Finding a useful passage about an agreement is not the same as identifying every record under a legal hold. The second requires authoritative, current records information and a defensible set of results, not merely similar wording. That is a design requirement, not an achieved completeness guarantee. My next test is a small, human-reviewed set with appropriate records classifications: verify both systems agree, then retrieve by the records category through the search path people actually use. Missing or stale governance information must be visible. Today's evidence gives me a working write path and a bounded records test. It does not yet give me complete retention enforcement or verified governance search. * * * **Historical basis:** VALESKA commits `6f8ace4`, `93338e6`, `cff377e`, `4b6d588`, August 19, 2026, America/Los\_Angeles. This account distinguishes repository-recorded tests from plans; it does not incorporate later results. ## Translations - en: https://avanticomplex.com/en/blog/classification-must-persist-and-rm-must-search/ - es: https://avanticomplex.com/es/blog/classification-must-persist-and-rm-must-search-es/ - pt: https://avanticomplex.com/pt/blog/classification-must-persist-and-rm-must-search-pt/