# A Vocabulary Must Be Stewarded Before It Classifies

I am adding a controlled-vocabulary schema, a seed, a steward CLI, and a fail-closed ingest preflight. Labels that cannot be reviewed are not a language.

Language: en
Canonical: https://avanticomplex.com/en/blog/a-vocabulary-must-be-stewarded/

Published: 2026-07-22T18:00:00.000Z

Classification without a vocabulary is a private dialect. This week I add a controlled-vocabulary schema, a seed and shared loader, a vocabulary-aware classifier, and a steward review CLI with quality measurement. I also add a fail-closed pre-ingestion preflight and a classifier pilot harness. Later in the week I restore the full CVocab proof gate. The honest measurement finding goes in the same record. I will not sell a vocabulary I have not measured.

## The wrong default

The usual classifier invents labels as it goes, or it binds to APQC as if that were the only language a client speaks. A ministry, a hospital, a datacenter each have terms of art. If I cannot seed those terms, review them, and refuse ingest when the vocabulary is missing, I am classifying into fog.

A shop that paints its own aisle names every morning cannot send anyone for a part. I want a list a steward can stand on, and a gate that will not ingest until that list exists.

## Schema, seed, steward, gate

Migration 46 is the schema. The seed script and shared loader put terms where the classifier can see them. The classifier can use that vocabulary instead of only a generic process code. The steward CLI is how a person reviews quality instead of hoping the model was polite. The preflight is fail-closed: if the conditions for ingest are not met, ingest does not start. That includes identity defaults in the delivered stack and a declared file-plan registry. Deposit now fails closed when no file-plan registry is declared. Identity defaults in the shipped stack fail closed rather than inventing a principal.

I also commit the governed ontology graph track so staged governance work is not blocked on a missing migration. Customer-facing capability claims get corrected against measured ground truth in the same week. I am writing down what we cannot yet do, in the same breath as the schema.

The proof gate is restored at the end of the week. The handoff records an honest measurement finding, not a victory lap. A vocabulary that cannot survive a proof gate is still a prototype. I am not treating the gold-set work of early August as known today. That measurement has not been run yet.

Fail-closed identity defaults in the delivered stack matter as much as the vocabulary. A missing principal is not an excuse to write as admin. A missing file-plan registry is not an excuse to deposit into nowhere. The ontology graph track unblocks staged work; it does not mean the graph is populated. I correct overclaims in the same week I ship the schema, because the two belong together.

Steward review is a person. Quality measurement is a number. Neither is the classifier talking to itself. I want both before I call this a language.

## The next test

The next test is a stewarded seed against a real corpus, with the preflight refusing a run that lacks a file plan, and a quality number I am willing to put next to the vocabulary. Until then this is a schema, a gate, and a warning, not a language the enterprise already speaks.

* * *

**Historical basis:** VALESKA commits `6c0d373`, `677947a`, `cc72b28`, `8522fdb`, `7e06050`, `bda8dc3`, 16–22 July 2026.


## Translations
- en: https://avanticomplex.com/en/blog/a-vocabulary-must-be-stewarded/
- es: https://avanticomplex.com/es/blog/a-vocabulary-must-be-stewarded-es/
- pt: https://avanticomplex.com/pt/blog/a-vocabulary-must-be-stewarded-pt/