Skip to content

Provenance — the full vocabulary

This page is generated. Do not edit it by hand — run scripts/atlas-vocab.py.

There are 4 registered values and nothing in use outside them.

Where it is stored: nowhere yet. The vocabulary is registered and nothing writes it — that is work not done, not a fault in the model.

Worth knowing: Written on every classification junction as a column of its own; the vocabulary governs all of them at once.

What this axis answers

Proposed — not yet in the database. Who said so — whether a classification came from the pipeline or from a person. It is written on every classification, and it is what keeps automated work from overwriting human work.

The values

A bold value carries more than its row — tagging rules, a confusion map, or worked examples — and has its own entry under “Value details” below.

ValueStored asIn useWhat it means
AnalyzeranalyzerThe pipeline's AI classification wrote this row (the default actor). The replace-not-accumulate key: analyzer re-runs replace ONLY analyzer rows.
HumanhumanA person asserted this row (admin edit, review save, Label-Studio write-back). Human rows are hard constraints: analyzer replace passes never touch them.
Importimportworking note, not a definition: A one-off script/backfill wrote this row outside the analyzer engine. Named per D1: script IDENTITY never goes in this …
Curated ruleruleA standing deterministic rule wrote this row (e.g. the DN-194 Sora-Showcase content_origin pass). Re-runnable; replace scope = rule rows.

Value details

Analyzer

analyzer · not used yet · kind: governance

The pipeline's AI classification wrote this row (the default actor). The replace-not-accumulate key: analyzer re-runs replace ONLY analyzer rows.

Tag it when — Rows written by the analyze/classify/scene-classify engines, including re-runs.

Never tag it when — Never for human edits, one-off scripts, or curated rules — even if an LLM was involved in building them.

Easily mistaken for:

  • import (a script writing rows outside the engine)
  • rule (a curated deterministic pass)

Worked examples:

  • scenes-classify writing content_segment_categories
  • post_processor discipline harvest

Human

human · not used yet · kind: governance

A person asserted this row (admin edit, review save, Label-Studio write-back). Human rows are hard constraints: analyzer replace passes never touch them.

Tag it when — Admin CRUD edits, review-queue saves, labeling write-backs.

Never tag it when — Not for scripts a human RAN — the actor is who asserted the classification, not who pressed enter.

Easily mistaken for:

  • rule (human-designed but machine-asserted)

Worked examples:

  • Joe correcting a discipline in the Studio
  • P2b Label-Studio write-back rows

Import

import · not used yet · kind: governance

The registry records a working note here rather than a definition: “A one-off script/backfill wrote this row outside the analyzer engine. Named per D1: script IDENTITY never goes in this axis (the openai_campaign_source_audit.py lesson — filename provenance silently exempted 231 rows from every analyzer replace scope).”

Tag it when — Bulk backfills, migration-time repoints, external-data imports.

Never tag it when — Not for engine re-runs (analyzer) or standing curated rules (rule).

Easily mistaken for:

  • analyzer
  • rule

Worked examples:

  • a backfill script seeding disciplines from an external dataset

Curated rule

rule · not used yet · kind: governance

A standing deterministic rule wrote this row (e.g. the DN-194 Sora-Showcase content_origin pass). Re-runnable; replace scope = rule rows.

Tag it when — Deterministic, reviewable, re-runnable rules maintained in the repo.

Never tag it when — Not for ad-hoc one-offs (import) or model inference (analyzer).

Easily mistaken for:

  • import

Worked examples:

  • curated Sora-Showcase origin tagging (DN-194 R1)

Registered, never used

4 of the 4 registered values have never been written to a row. For a young axis that is expected; for an old one it is a question worth asking.

analyzer, human, import, rule

How complete this page is

Of 4 registered values, 3 carry a written definition (75%) and 4 carry tagging rules (100%). A further 1 hold a working note where a definition should be — a spec reference or a rename history, useful to whoever wrote it and no help to a reader asking what the value means. Every number on this page was counted from the live database when the page was generated — none is stored, and none was typed by hand.

Where to look next

Atlas — the BrandTrackers domain model. Source: git markdown, drift-checked against the live DB.