Standards & Reproducibility

Standards & reproducibility

Keep the evidence connected.

Pelagia is working toward a durable, interoperable imaging lifecycle: portable source-image archives, reviewable training datasets, and independently verifiable model products. The goal is to preserve the connection between an observation, the evidence used to interpret it, the model that contributed to that interpretation, and the data products shared downstream.

Three connected, independent contracts.

Imaging archives, training datasets, and model releases serve different scientific purposes. Pelagia keeps them distinct, then connects them through stable identities, provenance, and integrity checks. That makes each product easier to inspect, move, reuse, and preserve.

Image interchange

A portable preservation package for retained images and acquisition context.

Training archive

A reviewable record of source assets, annotations, labels, and dataset lineage.

Model artifact

A sealed, inspectable release that records how a model was prepared and evaluated.

Scientific Image Interchange

Pelagia's Scientific Image Interchange format is a portable archival and exchange package for retained scientific imagery. A package is readable outside Pelagia and combines exact retained image bytes with the context needed to understand them: a manifest, scientific metadata, append-only processing history, and checksums.

What the format preserves

Original encoded frame payloads without routine decode-and-reencode steps.
Frame, stream, acquisition-segment, and source-file provenance.
Timestamp evidence and explicitly recorded gaps, failures, duplicates, or removals.
Package and shard integrity through inventories and checksums.

deployment_042/
├── data/                 # SQLite image shards
├── manifest.json         # inventory and identities
├── metadata.toml         # scientific description
├── history.jsonl         # append-only provenance
├── checksums.sha256      # integrity record
└── preview/              # non-authoritative derivatives

Previews are useful for access and review, but authoritative image payloads remain separate. A completed package can be structurally or fully validated without treating a successful check as permission to discard source data.

Read the Scientific Image Interchange specification on GitHub

Training-set archives

Oracle Builder training datasets are durable SQLite archives designed to preserve the difference between source material, human scientific judgment, and derived model evidence. A dataset has stable identity and revision lineage; annotations and reviews are retained as history rather than silently overwritten.

Human labels remain evidence

Model predictions and embeddings are derived workspace material, not ground truth. They require a human annotation or review step before they become training labels.

Freeze a release to train

Working datasets remain editable. A training release is frozen, fingerprinted, and preserved as a distinct checkpoint of the evidence used for a model.

Keep split decisions with the run

Training, validation, and test assignments belong to a model-run protocol, allowing the same frozen dataset to support transparent, repeatable experiments.

Read the Oracle Builder dataset contract on GitHub

Model artifacts, not just model files

A trained network alone does not explain how it should be used. Oracle Builder treats a model release as a durable scientific artifact with a defined input and output contract, a record of the frozen dataset it used, its split protocol, resolved configuration, evaluation products, a model card, and an integrity inventory.

A sealed model release includes

Stable artifact and run identities, plus dataset fingerprint and schema version.
Portable preprocessing, architecture, and output semantics.
Item-level train, validation, and test split assignments.
Metrics, logs, load-test evidence, model card, and checksummed inventory.

model-run/
├── artifact.json
├── MODEL_CARD.md
├── config/               # source and resolved settings
├── protocol/splits.json
├── provenance/
├── metrics/ and evaluation/
└── model/                # tested model formats

Read the Oracle Builder model-run artifact specification on GitHub

Oracle Builder and Pelagia

Oracle Builder is Pelagia's companion environment for curating training data, training and evaluating image models, and preparing models for operational use. It keeps the machine-learning environment separate from the field-processing system: Oracle Builder owns model preprocessing, inference behavior, and immutable model provenance, while Pelagia retains operational processing, scientific review, measurements, and observation records.

PelagiaProcess, review, measure, and preserve observations.

Oracle BuilderCurate, train, evaluate, seal, and serve model products.

Versioned contracts under active development.

These are Pelagia and Oracle Builder technical contracts, published for inspection and iteration. They are intended to make scientific imaging work more portable and auditable—not to claim that every imaging program must use one application, one database, or one model framework.

Explore Pelagia source and documentation · Explore Oracle Builder source and documentation