Know how it was made.
Build on what you can explain.
Customer-owned lineage connects scientific results to their origins—and gives AI training and evaluation a reproducible foundation.
Our applications can capture scientific intent.
When we own the scientific application and its domain APIs, we can record the operation as it happens: which exact data was used, which calculation or interpretation was applied, with which parameters, and what result was produced.
A smoothing run, coordinate transformation or manual interpretation is meaningful scientific history. Recording clicks alone cannot explain it. External connectors can only preserve the history their host exposes; partial history must be labeled.
A result you can trace.
A dataset you can explain.
Explore a fictional curve workflow. Each step carries the context needed to understand the next one.
- 01 / SOURCE
Original curve
Asset A · revision 17Keep the original evidence and its scientific meaning.
Inspect source context
Exact source revision, depth axis, curve units, missing-value mask, source rights and coordinate uncertainty. Preserve the original file where permitted.
- 02 / OPERATION
Recorded recipe
Run R · version 08Capture what the scientist asked the application to do.
Inspect the derivation
Input A17 → a five-sample smoothing operation → output B04. Record algorithm version, window size, null handling, software identity, actor, outcome and quality checks.
- 03 / RESULT
Derived curve
Asset B · revision 04Keep the result connected to its exact inputs and review.
Inspect result history
The result links back to A17 and R08. Later edits create new revisions. Review decisions, interpretations and label origins remain distinct records with their own permissions.
- 04 / AI DATASET
Frozen selection
Dataset D · revision 02Give training and evaluation a reproducible reference.
Inspect the dataset manifest
Record selected revisions and hashes, transforms, label versions, permitted use, split assignments and preprocessing. A later change to the live curve does not silently change this dataset.
This explanation runs entirely in the page. It is not a live lineage viewer or an implemented AI dataset builder.
From a live project
to a reproducible dataset.
Our planned dataset workflow preserves source evidence, governs publication through supported OSDU services and ArrayBridge contracts, then derives a view for a specific task. Tables, large arrays and documents need different representations; there is no universal AI-ready format.
An immutable manifest records exact revisions, content hashes, transforms, code versions, features, label origins and permitted use. Live assets, frozen training datasets, model artifacts and inference outputs have distinct identities and lifecycles.
Better evidence for training.
More meaningful evaluation.
Lineage helps a team inspect why a sample was included, how a label was created and which transformations affect the result. A review rejection can carry useful context without automatically becoming a trustworthy training label.
Split training, validation and test sets by well, field or time as the task requires. Fit preprocessing on training data only. Record null, unit and coordinate handling so hidden leakage and inconsistent inputs can be investigated.
Traceability supports review and reproducibility; it does not by itself prove label quality, model accuracy or suitability for a decision.
Your compute.
Your model choices.
The intended workflow runs in an approved customer environment with scoped access and customer choice of model provider. Retrieval, inference and training are separate uses with separately understood rights.
Data access is not training permission. No cross-customer training without explicit agreement. Customers retain permitted dataset and model artifacts subject to the applicable licenses. Agreeing to diagnostic sharing never grants training consent.
Take the derivation history
with the result.
Essential lineage belongs with customer project data. It must remain readable and exportable even if you stop using our clients. Advanced impact graphs and review tools can add value above that baseline.
What data ownership means ↗