aiprov: AI provenance, recorded, open, verifiable

When an AI model drafts your code or your text, two questions come later: who did what, and who checked it? aiprov writes the answers into a plain Turtle file next to your work, as a profile of W3C PROV-O.

RecordedEvery step names its agent, person or model, its inputs and outputs by hash, and the telemetry the tool reports. Nothing is estimated.
OpenOne Turtle file and one dependency (rdflib). Any tool that reads PROV-O reads the graph.
VerifiableEvery statement is a claim on a seven-step ladder. The top two steps belong to people: the command line refuses an AI, and provlog validate checks it again in CI.

Code on GitHub · Vocabulary · The F(AI)²R paper it comes from

A demo: one report, two claims, one human check

Alice asks an AI model to turn her benchmark notes into a short report. Nine commands later, the graph below holds the whole story. The data is invented, the graph is real: this page ran the commands with aiprov 0.2.0 (git 2d72755) when it was built.

Shapes follow the W3C PROV notation: agents as houses, activities as rectangles, entities as ovals. Claims are entities with a verification state. Hover a node for its details; the numbers match the steps below.

Who. Alice is a person, registered with her ORCID iD. Model X is an AI model that acts on her behalf (prov:actedOnBehalfOf).
What. Model X drafted report.md from notes.md. The activity records the model, both files (the output with its sha256) and the tokens the tool reported: 1,200 in, 800 out.
Claims. The report makes two statements. Each becomes a claim with a parent activity and a state on the ladder: "twice as fast" was checked by the model (AI-confirmed), "all functions tested" by nobody yet (unverified).
Human check. Alice re-ran the benchmark and promoted the speed claim to human-confirmed. The promotion is an activity of its own, an audit pass associated with Alice, so the graph says who granted the step and when.
The rule. Model X then tried to mark the other claim as human-read. provlog refused: the top two steps are for registered humans. validate checks the same rule on the whole file, so a hand-edited graph cannot slip it in either.

The commands, as run for this page

Each with --base https://example.org/demo/; the output is what provlog printed.

$ provlog init
Seeded provenance.ttl (519 triples, base https://example.org/demo/, schema 0.3.0)
$ provlog agent --id alice --type human --name "Alice Example" \
    --orcid 0000-0002-1825-0097
agent:alice registered (human)
$ provlog agent --id model-x --type ai --name "Model X" --model model-x \
    --provider "Example AI" --acted-on-behalf-of alice
agent:model-x registered (ai)
$ provlog log --activity draft-report --pass authoring --agent model-x \
    --label "Draft the report from the notes" --used notes.md \
    --generated report.md --input-tokens 1200 --output-tokens 800
act:draft-report logged (544 triples total)
$ provlog claim --id c-speed --text "The new parser is twice as fast." \
    --parent draft-report --agent model-x --state ai-confirmed
claim:c-speed added (state ai-confirmed)
$ provlog claim --id c-tests --text "All public functions are tested." \
    --parent draft-report --agent model-x --state unverified
claim:c-tests added (state unverified)
$ provlog promote --id c-speed --to human-confirmed --agent alice \
    --note "Re-ran the benchmark myself"
c-speed: ai-confirmed -> human-confirmed (logged as AuditPass, agent:alice)
$ provlog claim --id c-tests --text "All public functions are tested." \
    --parent draft-report --agent model-x --state human-read
REFUSED: rung 'human-read' is human-only; agent:model-x is not a registered HumanAgent. An AI must never grant this rung.
(exit 1)
$ provlog validate
[ OK ] parentless claims: 0
[ OK ] unattributed claims: 0
[ OK ] human-only rungs granted by AI agents: 0
[ OK ] human-only rungs without a recorded human promotion activity: 0
[ OK ] sources above needs-research without DOI/URL: 0
[ OK ] activities without agent: 0
[ OK ] session activities with no measurable span (a logging instant, not the span of the work): 0
[ OK ] agents referenced but not registered: 0
[ OK ] agents typed as both AI and human: 0
[ OK ] wasInformedBy targets that are not logged activities: 0
[ OK ] external references (nodes defined in another graph): 0
[ OK ] external references that carry a file path or a label: 0
[ OK ] sub-activities outside the orchestration plan: 0
[ OK ] orchestrations logged twice or informed by themselves: 0
[ OK ] sub-activities logged before their orchestration: 0
[ OK ] entities generated by two activities that carry a model role: 0
[ OK ] fallbackOf targets that are not other logged activities: 0
[ OK ] activities that are part of something that is not a session: 0
[ OK ] messages without a sender: 0
[ OK ] replies to a message that is not in this graph: 0
[ OK ] messages without a content hash, or whose kept text does not match it: 0
[ OK ] message text kept in a channel above internal or of unknown class: 0
[ OK ] records made by another agent or in arrears without a source (recordedFrom): 0
[ OK ] sessions with token sums from more than one capture path (only one is counted): 0
[ OK ] viaHarness targets that are not registered harnesses: 0
[ OK ] activities via a harness with neither a model nor a stated reason: 0
[ OK ] telemetry chain anchors without one valid head, a length of at least 1 and an author: 0
[ OK ] premises that are not claims: 0
[ OK ] premise cycles: 0
[ OK ] atleast claims with minPremises missing or above the number of premises: 0
[ OK ] composite claims without premises: 0
[ OK ] claims stated above their effective rung: 0
[ OK ] decisions that are malformed (asker, hash, two options, recommendation, text matching hash): 0
[ OK ] options or deciding activities that are incomplete: 0
[ OK ] chosen options that the decision did not offer: 0
[ OK ] decisions set to done by an agent that is not a registered human: 0
[ OK ] supersedes links to a missing decision, to itself or in a cycle: 0
[ OK ] decision text kept where the data class is unknown or above internal: 0
[ OK ] graph states its schema version

The resulting graph: demo.ttl (Turtle, with the vocabulary), and the dashboard that aiprov-dashboard builds from it.

Try it

$ pip install git+https://github.com/noheton/aiprov
$ provlog --base https://your.org/prov/ init

Then register your agents and log your steps as above. The README lists every command; issues and pull requests are welcome.

Cite

Krebs, F. (2026). aiprov: domain-agnostic AI provenance over PROV-O (Version 0.2.0) [Software]. Zenodo. https://doi.org/10.5281/zenodo.23651340

All versions: https://doi.org/10.5281/zenodo.23651333. Citation metadata: CITATION.cff.