Skip to content
Dataset construction

Four tiers.

Every engagement produces the same four artefacts. The tier defines scope and price, not rigour.

The four artefacts

01

Versioned dataset

Lance-native format with fragment-level copy-on-write versioning and schema-enforced annotations against a defined taxonomy. For training, evaluation, and downstream pipelines.

02

Merkle integrity proof

A cryptographic tree over dataset fragments, with proof generation for any individual sample. Verify the bytes have not changed.

03

Provenance bundle

Four layers: SHA-256 content hashes, Merkle root, C2PA content credentials carried as a sidecar manifest referencing the dataset by hash, SLSA v1.2 attestation. Verified independently with external reference tools.

04

EU AI Act evidence pack

Articles 10, 11 and 12 mapped artefacts with coverage classification: fully populated, requires human input, or not applicable. Articles 9, 14 and 15 (risk management, human oversight, and system accuracy) stay with the provider deploying the system.

Each deliverable ships with verification recipes: commands that run against your copy of the dataset using tools Static Signal does not maintain. Acceptance requires you to run at least one successfully.

Tier 1

Standard

100–500
samples

Single defined taxonomy — yours, or a Static Signal template — with modality-appropriate validation and train, validation, and test splits.

Tier 2

Production

500–2,000
samples

Adds a taxonomy iteration cycle, custom schema authoring, stratified splits with a verification report, and branch-aware versioning.

Tier 3

Enterprise

2,000–10,000
samples

Adds hierarchical taxonomies, a reference set curated with your domain experts, and a workshop day.

Tier 4

Bespoke

Scoped
volume, set in discovery

Greenfield taxonomy design, a pilot dataset to validate it, then full production construction. Begins with a paid discovery engagement.

Modality coverage

Images

In production use

Six-tier validation: structural checks, subject gate, data quality, perceptual quality, categorical assessment, consistency. Four of the six gate; advisory findings are recorded as advisory.

Text · tabular · multimodal

Available on engagement

Validation stacks are scoped per engagement; the provenance and compliance artefacts are identical to image deliverables.

Out of scope

  • Software licensingthe instrument is the production line, not the product
  • Continuous SaaS hostingcustody is bounded, not permanent
  • Model training and deploymentwe hand over the dataset; what you train on it stays with you, or becomes an applied engagement
  • Annotation labour at scaleyou or your domain experts provide labels; we verify, structure, and sign them
  • Generic consultingaudit and regulator support is a published add-on; open-ended advisory is not something we sell

Applied systems — capture rig to operator interface — are scoped separately.

Applied work

Negotiate scope and timeline. We do not negotiate evidence quality or the price floor.

Start an engagement