Specification

How a Provenance Label works

A label is a short declaration: who made this, when, and how much of it came from a human versus an AI. That's it.

v1.2 Published 2026-02-17 Updated 2026-06-01 CC BY 4.0

Disclosure, not judgment.

A Provenance Label makes no claim about quality. It simply answers: what happened here, and who is accountable? Like a nutrition label — informative, not prescriptive.

What a label looks like

Two formats. Same information. Use whichever fits your context.

Long — for articles, posts, documentation

─────────────────────────────────────
PROVENANCE LABEL v1.2
─────────────────────────────────────
Author:  Shelton Davis
Date:    2026-02-17
Human:   65%
AI:      35%
Tools:   Claude Sonnet 4.6
Note:    Concept and direction human. Draft and structure AI. Edited for voice.
─────────────────────────────────────

Short — for bylines, footers, social posts

PL v1.2 | Shelton Davis | 2026-02-17 | Human 65% · AI 35% · Claude Sonnet 4.6

The fields

A label has four required fields and several optional ones. Required fields make it a valid label. Optional fields add context.

Required

FieldWhat it means
authorThe human accountable for the content. Not the AI. Not the tool. You.
dateDate of completion or publication. Format: YYYY-MM-DD.
humanYour percentage of contribution. Must be an integer 0–100.
aiAI's percentage. Must sum to 100 with human.

Optional

FieldWhat it means
toolsWhich AI tools you used. Comma-separated if multiple.
noteOne plain sentence describing how you worked together.
shared-byIf you're republishing someone else's work, your name goes here.
updatedDate of most recent substantive revision.
refSources referenced during creation. Repeatable.

How to calculate percentages

Aim for honest, not perfect. No one is auditing your math. The point is a good-faith estimate of how the work actually came together.

Things to consider: Who initiated the content? Who made the decisions? Who edited, cut, and shaped the final output? What percentage of the final text, code, or image came directly from each contributor?

If you used an AI throughout a session, ask it to calculate at the end — it has context. You review, adjust, and approve. The final number is yours.


Where to put it

Wherever makes sense for your medium. Some common placements:

MediumWhereFormat
Blog postFooter or end of postLong or short
Social postInline or as a noteShort
GitHub repoREADME or .plgen fileLong or raw file
Academic paperAcknowledgments or appendixLong
ArtworkArtist statement or captionLong

The Registry

Submitted labels receive a permanent short ID — for example, PL-000312 — that can be embedded in your content. Anyone can look it up to verify the original, timestamped disclosure.

Self-generated labels are free and unregistered. Registered labels get a permanent ID and public record. The difference is verifiability, not validity.

Become a member →

Building with Provenance Label?

Full API reference, integration guides, and the open registry endpoint.

Registry →

Label Confidence

Registered labels display a computed confidence score — PLGen-C — alongside the label itself. It isn't a judgment on whether your work is good, or even whether the label is true. It's a measure of how much of the disclosure can currently be independently verified versus taken on your word.

The score is computed from a fixed set of observable signals — not estimated or judged by an AI. Each signal is checked independently and shown with a symbol:

SymbolMeaning
Signal fully satisfied.
Partially satisfied, or not yet independently confirmable.
Not satisfied, or the information needed isn't present.

A △ doesn't necessarily mean something's wrong. Some signals depend on session data that not every AI platform exposes yet — it means that piece can't be independently confirmed right now, not that the disclosure itself is false.

What's measured

SignalWhat it checks
Session initializationWhether the disclosure session was set up before the work began, versus reconstructed afterward.
Checksum integrityWhether the label carries a token confirming it wasn't edited after the AI generated it.
Token/time plausibilityWhether the pace of AI output over the session is consistent with a human directing the work.
Membership depthHow many labels this member has registered previously.
Work URL resolutionWhether the linked work actually exists and is reachable.
Paste event ratioHow much of the human-provided input was pasted in versus typed directly.
Session gap detectionWhether there were unusually long pauses during the session.
CorroborationIndependent confirmation from an institution or other verified source, beyond self-report.

The exact scoring weights and thresholds are intentionally not published. Publishing them would let the score be gamed rather than earned — what's shown here is what's measured, not how heavily each part counts toward the final number.


Changelog

VersionDateChanges
v1.2 2026-06-01 Non-persuadable split estimation — AI may not adjust contribution percentages based on user preference, only new factual information. ChatGPT Custom GPT deployed.
v1.1 2026-03-06 Added work_title, human_role, ai_role fields to long format and JSON schema. Introduced plgen command set: init, generate, register, status, validate, formats, help, cancel.
v1.0 2026-02-17 Initial release. Core label format, required and optional fields, short and long output formats.