Exchange
Connects scientific expertise and institutional source material to suitable programmes.
Scientific data infrastructure for frontier AI
Train AI on how science is done.
Path Invariant captures how leading scientists navigate discovery—so scientific agents can learn to revise hypotheses, recover from misleading results, and choose what to do next.
The missing record
The result is published. The reasoning path is lost.
Publications preserve conclusions, not the full process that produced them. Trajectories retain the decisions, revisions, failed approaches, tool use, and evidence scientific agents need to learn how expert research unfolds.
To learn the scientific process, models need the path to discovery—not merely its published output.
Published output
- Final result
- Curated narrative
- Limited process context
- Citation trail
- Unclear model difficulty
- Implicit permissions
Complete trajectory
- Scientific task and initial files
- Negative intermediate results
- Tools, databases, and decisions
- Verifiable answer and explanation
- Baselines and human difficulty
- Rights and provenance summary
Trajectory
The scientific process, captured as training data.
Each Trajectory records an authentic research path: what the scientist knew, which action they took, what happened, and how that evidence changed the next decision.
- 01Scientific taskQUESTION
- 02Initial filesSOURCE
- 03Relevant tools and databasesMETHOD
- 04Objectively verifiable answerOUTPUT
- 05Explanation of how the answer is knownEVIDENCE
- 06Baseline performance from current modelsBASELINE
- 07Estimated human difficultyDIFFICULTY
- 08Provenance and rights summaryRIGHTS
Research-grade by design
Teach models to adjust their path.
Long source histories expose the feedback loop at the heart of science: form a hypothesis, act, observe, revise, and continue. Expert reference trajectories turn that process into a learnable and measurable target.
- 01
Months or years of source history where available
- 02
Negative and misleading intermediate results
- 03
Expert reference trajectories
- 04
Objective scoring where possible
- 05
Blinded holdouts
- 06
Full rights and provenance report
- 07
Baseline performance from current models
- 08
Separate licences for evaluation and training
Path Invariant system
From expert research history to model capability.
Path Invariant works with universities and expert contributors to turn long-term scientific work into rights-documented training environments and evaluations.
Foundry
Converts source histories into documented research environments.
Trajectory
Represents the task, process, evidence, and result of scientific work.
Atlas
Organises available collections and their technical metadata.
Gauge
Measures model performance against expert work and blinded holdouts.
ProtocolPublished data standard
Defines representation, provenance, and quality requirements across the system.
Provenance by construction
Rights are part of the data.
Every delivery is designed to include a provenance and rights report, with distinct licensing for evaluation and training. Blinded holdouts and objective scoring are included where the source material permits them.
- 01Source history
- 02Contributor permissions
- 03Transformation record
- 04Evaluation licence
- 05Training licence
- 06Holdout policy
Work with Path Invariant
Build the scientific record AI needs.
For frontier AI laboratories
Develop a data programme.
Build datasets and evaluations around specific scientific capabilities, grounded in expert trajectories and verifiable outcomes.
Discuss a data programmeFor universities and experts
Preserve valuable scientific work.
Explore responsible ways to document and license scientific work for model research while preserving its provenance.
Become an institutional partnerBegin a conversation
Teach scientific agents how experts find the way forward.
Tell us the scientific capability you want models to learn, the evaluation question you need to answer, or the research history you want to preserve.