Data Schematic: add visual cues denoting agent/machine/human produced/curated/verified data

[jqnatividad]

Have a subtle, tasteful visual hint (emoji? color hints? hover text?) for users to easily identify:

• deterministic, computed data
• deterministic and human-verified/approved data
• AI-generated data
• AI-generated, but human-verified/approved data
• AI-generated, and adversarially reviewed by another AI data
• Human-generated data

cc @avdata99

[jqnatividad]

Investigation notes: current state & suggested shape

What exists today

Provenance in the Data Schematic is document-level only, never per-field:

• x-qsv.generated_by — describegpt bakes a full attribution block into the schema (tool+version, command line, model, API URL, language, timestamp). viz renders it in three places: the dictionary drawer footer, the page footer, and the tour's meta panel (Model: line).
• WARNING lines — describegpt appends per-prompt-type warnings ("Label and Description generated by an LLM… Verify before using!"). That's the closest thing to an "AI-generated" cue, and it's a blunt block of text, not a per-datum hint.
• x-qsv.tour.generated_by — same convention on the tour envelope.

Mapping the six categories onto current reality

Category	Today
1. Deterministic computed	Everything from the stats cache: chart geometry, types, cardinality, min/max, the self-inferred dictionary fallback. Implicitly computed — no marker.
2. Deterministic + human-verified	No mechanism.
3. AI-generated	describegpt's title / description / x-qsv.tags / content_type / tour narration. Only detectable by the presence of generated_by.
4. AI + human-verified	No mechanism. A human editing an LLM dictionary leaves no trace (viz deliberately tolerates hand-edited sidecars and re-validates on read).
5. AI + adversarial AI review	No mechanism. Closest is --two-pass DictionaryRefine, but that's same-model cross-field refinement, not adversarial review — and it re-emits as PromptType::Dictionary, erasing even that distinction.
6. Human-generated	A hand-written dictionary — detectable only by the absence of a describegpt attribution block.

Categories 2, 4, 5 need a new schema contract; 1, 3, 6 are inferable today from what's already stored.

Suggested shape (two phases)

Phase 1 — no schema change, inference only. viz can already classify coarsely: dictionary text with generated_by present ⇒ AI (✨-class badge); absent ⇒ human/unknown (👤); stats-derived facts and self-inferred labels ⇒ computed (⚙️). Render as a small badge + hover tooltip (title / aria-label) in the dictionary drawer rows, KPI tile corners, panel subtitles, and tour popovers, with a legend next to the existing attribution footer. This covers 3 of the 6 categories and establishes the visual language.

Phase 2 — explicit x-qsv.provenance contract, per-property (and on the tour envelope), loosely aligned with W3C PROV vocabulary for portal interop (relevant to the CKAN/DCAT angle):

"x-qsv": { "provenance": { "source": "ai", "verified": "human",
           "model": "gemma-3", "verified_at": "2026-08-27T…" } }

• describegpt stamps source: "ai" on what it authors
• a future --review-model (different model, adversarial prompt) could stamp verified: "ai"
• human attestation via a hand edit or a small qsv schema helper

This unlocks categories 2/4/5.

Implementation considerations

• Every badge label/tooltip needs curated i18n catalog keys in all supported locales.
• Any tour/HTML template change requires regenerating the gallery pages.
• KPI tile headers are contested real estate — viz: KPI tile doubles the aggregation prefix when the dictionary title already carries it ("Total Total Revenue") #4499 touches the same spot; coordinate the tile-corner badge with that fix.
• The version-1 x-qsv.tour envelope pattern (Rust builds the envelope, LLM supplies prose only) is the right precedent for provenance — keep the machine-readable part out of the LLM's hands.

🤖 Generated with Claude Code

[jqnatividad]

Also check DTPR for AI

https://www.dtpr.ai/en