Published ranges.
Per-deliverable, published as ranges. Exact quotes depend on scope.
Ranges are published so you can qualify us before a call. A quote moves within its range on volume tier, taxonomy complexity, iteration cycles, and delivery channel. It does not move on negotiation.
Four tiers
Standard
Single defined taxonomy — yours, or a Static Signal template. Modality-appropriate validation. The four artefacts. Train, validation, and test splits. Delivery to a private Hugging Face Hub repository or as a Lance tarball with SHA-256 sidecar. Invoice on acceptance.
Production
Tier 1, plus one taxonomy iteration cycle, custom schema authoring (Pydantic and COCO, or the modality equivalent), stratified splits with a verification report, branch-aware versioning for experimental subsets, and two delivery rounds. Invoice on acceptance.
Enterprise
Tier 2, plus multi-class hierarchical taxonomies, two iteration cycles, a reference set curated with your domain experts, one workshop day (remote or on site), and a customer-specific verification procedure document. Invoice on acceptance and workshop completion.
Bespoke
Greenfield taxonomy design, a pilot dataset of 100–500 samples to validate it, full production construction, and knowledge-transfer documentation. Begins with a discovery engagement. Invoicing is phased per stage; you can exit between stages without penalty.
Negotiate scope and timeline. We do not negotiate evidence quality or the price floor.
The Tier 1 floor is £5,000, including for the smallest engagements.
Standing arrangements
Tier 4 work begins with a paid discovery engagement. The deliverable is a written engagement specification: timeline, deliverables, and exit criteria. It has standalone value — you may take it to another supplier. The fee is non-refundable and is the only commitment until a full engagement is signed.
A twelve-month standing arrangement: unlimited re-verifications, evidence-pack refreshes, compliance updates as standards land, advisory access of eight hours per quarter, and priority in the engagement queue. Suited to customers whose deliverables face recurring audit.
Add-ons
Re-verification
Fresh verification of an existing deliverable against the current instrument. Live, recorded, signed report.
Compliance export update
Evidence-pack refresh when EU AI Act harmonised standards or guidance change.
Synthetic data augmentation
Governed synthetic samples for class balance or rare-case coverage, provenance-marked for Article 50. In development. Available on engagement once released.
Integrity certification
Near-duplicate detection, contamination and leakage across splits, and label-inconsistency poisoning, on a dataset we did not construct. Same-framing candidates the second matching pass rejects come back as review clusters, not verdicts; poisoning coverage is label inconsistency, not backdoor triggers. Cost scales with distinct-content items. Final pre-engagement checks in progress; available on engagement once they close.
Audit & regulator support
On-site or remote support during your conformity assessment or regulator interaction. Day rate, not hourly.
Custom taxonomy development
Greenfield taxonomy from interviews with your domain experts. Output: schema code, integration, documented rationale.
Reference set curation
Build and version a customer-specific subject reference set for validation against your subject matter.
Included in every engagement, no charge
Bounded custody
Receipt acknowledged with a SHA-256 manifest. EU-resident processing. Single-operator access; no subcontractors without written consent. Raw inputs deleted within 30 days of acceptance. Deletion on request within 7 working days, with signed attestation.
Audit trail
Every custody event — receipt, processing, delivery, deletion — recorded in an append-only log. Full trail delivered at handover.
Custody proofs
Transparency-log inclusion proofs and signed erasure attestations. In development. Included at no charge from availability.
Where the premium goes
Data Provenance Initiative, review of 1,800+ datasets
Independent audits of popular dataset hosts — the Data Provenance Initiative’s review of more than 1,800 datasets — find licence information omitted in over 70% of cases and incorrect in over 50%.
The premium over commodity annotation reflects what ships with the labels: verified content hashes per sample, a Merkle integrity tree, C2PA content credentials, SLSA attestation, the Articles 10, 11 and 12 evidence pack, and recipes to check all of it independently.
What a sample is
A sample is the natural unit of the modality: an image, a document, a row of structured data. Mixed-modality datasets are counted by primary unit.
Terms
Acceptance requires loading the dataset and running one verification recipe successfully. Invoices are payable within 30 days. Tier 4 invoicing is phased per stage, with exit between stages and no penalty.
Prices are in GBP and exclude VAT where applicable; engagements may be contracted in EUR at the prevailing rate on the scope date. An NDA is in place before any data discussion, and a GDPR Article 28 data processing agreement is executed with every engagement.
Ranges are reviewed quarterly and hold for engagements scoped within 90 days of issue. Exact instrument versions are pinned in each signed engagement scope and recorded in every evidence bundle.