APPLICATION A3 / DOE GENESIS · SPOTTER-AI

Provenance and integrity for AI-driven science.

SPOTTER-AI is a DOE Genesis Mission project at Argonne National Laboratory. Argonne National Laboratory leads SPOTTER-AI, with Oak Ridge National Laboratory, Illinois Tech, and Hewlett Packard Enterprise as partners. It addresses cybersecurity for AI-driven science through data provenance and integrity verification. The project began in August 2026. The IOWarp team builds the agents that investigate what the provenance shows.

Partners: Argonne National Laboratory · Oak Ridge National Laboratory · Illinois Tech · Hewlett Packard Enterprise
Status: Project began August 2026
Stack layers: Agents and tools · Applications
Components: CLIO Agent · Agent marketplace · CLIO Kit
On the web: Argonne National Laboratory ↗ · Oak Ridge National Laboratory ↗ · Illinois Institute of Technology ↗ · Hewlett Packard Enterprise ↗ · U.S. Department of Energy Genesis Mission ↗ · SPOTTER-AI on GitHub ↗ · Flowcept ↗ · CMF ↗
Record: What the work has produced · Sources

01 / The scientific question

When a scientific AI workflow produces a wrong answer, can we tell whether it was tampered with, and trace the cause to its source?

AI models, datasets, and workflows are increasingly shared across facilities. A poisoned passage in a retrieval corpus or a manipulated tool call can quietly change what a model concludes, compromising the logic of discovery rather than a single result.

SPOTTER-AI captures detailed provenance of workflows and of the models inside them, learns how much influence each input should have on an outcome, and treats departures from that expectation as a signal worth investigating.

WHERE CONTEXT BREAKSProvenance helps only if someone can question it. The records span workflow events, artifact lineage, and signals from inside the model, across systems never designed to be read together. An investigating agent needs bounded, read-only access to all of it, and its own actions must meet the same standard of evidence.

02 / Workflow

The workflow, stage by stage.

FIG A3.1 · DOE Genesis · SPOTTER-AI workflow4 stages

Capture
Workflow and model provenance, including token-importance signals from inference, is recorded by provenance systems such as Flowcept.
Partner systems
Expose
A read-only provenance server gives an investigating agent tools for correlation, artifact lineage, and timelines across providers.
Agent marketplace · MCP
Watch
A standing watcher agent observes a session and raises alerts. While it is armed, destructive tool calls wait for its clearance, and they are refused if no watcher is live.
CLIO Agent · Spotter mode
Attribute
Agentic forensic analysis works backward from an anomaly to its likely root cause and reports it to the people responsible.
CLIO Agent

Stages as the project sources describe them. What each stage has produced so far is listed in the record below.

03 / The record

What the work has produced.

CLIO Agent shipped a Spotter approval mode in August 2026: a watcher agent observes the session, and destructive tool calls fail closed unless the watcher clears them.[1]
A unified, read-only Spotter provenance MCP server was implemented in the CLIO Agent marketplace and qualified live against the Flowcept and CMF provenance providers.[2][3]
CLIO Agent can stamp its records of model calls with the response identifier that joins them to token-importance signals captured from vLLM inference, so an investigation can join agent activity to evidence from inside the model.[4][5]
CONDITIONS · Off by default; enabled through the kvnorm option with Flowcept
A SPOTTER provenance investigator pack is published in the CLIO Agent marketplace.[6]

04 / Software in the loop