Data lineage
Record the origin, version, transformation, inclusion rules, and limitations of data used in an experiment or evaluation.
Research area 02
Reproducible AI research infrastructure
We study the technical records and workflows needed to connect data, configurations, system versions, evaluations, outputs, and conclusions.
Research problem
Code is only one part of reproducibility.
An AI result depends on more than a repository. Data versions, preprocessing, prompts, tools, model endpoints, environment settings, evaluation procedures, reviewer decisions, and interpretation can all change the outcome.
InfluxWave focuses on infrastructure that preserves these relationships without making the research workflow unusably heavy. The goal is a record another person can inspect, repeat, and challenge.
Infrastructure layers
Preserve the chain behind a result.
Traceability should make important results easier to understand, not merely create more logs.
Experiment provenance
Connect configuration, code, model, prompt, tool, environment, and operator choices to each run.
Result traceability
Link outputs and aggregate findings back to examples, evidence, reviewer judgments, and the system state that produced them.
Research questions
Questions that guide this area.
- What is the minimum record needed to reproduce or audit an AI evaluation?
- How should provenance be represented when an application uses multiple models, tools, and retrieval sources?
- Which parts of a research workflow should be automated, and which require deliberate human annotation?
- How can teams compare results across changing environments without losing context?
- How should sensitive data and private evaluation examples be referenced without exposing them?
Evidence boundary
This page defines a research direction. It does not claim a released provenance platform or reproducibility tool.
Connected work
Infrastructure makes evaluation evidence durable.
A trustworthy evaluation requires the ability to explain where a result came from. Applied systems require the same evidence when behavior changes after deployment.