For AI organisations

Tell us what your models cannot currently learn.

bnd originates data that does not exist as a product yet. It is built to take a defined requirement, map which workflows and organisations already produce that signal, and work toward a qualified, permissioned package. The first transaction is being executed now; the honest ledger is on Current State.

Define a Buyer Requirement
01 / The requirement

A requirement, not a dataset wishlist.

The requirement format is deliberately concrete. It gives bnd enough to search reality without asking you to expose research.

Missing capabilityWhat the model cannot currently do or be evaluated on
DomainThe field, industry, or context the signal must come from
Data structureTrajectories, corrections, judgements, outcomes, conversations
Provenance requirementWhat must be known about the source
Quality thresholdWhat separates acceptable from unusable
VolumeHow much is needed for the purpose
Permitted rightsTraining, evaluation, safety research, internal use
Acceptance criteriaHow you will judge the delivered package
BudgetA range is enough to qualify the search
02 / What happens next

What happens to a requirement.

  • bnd interprets the requirement and maps where the signal is naturally generated.
  • Likely sources enter confidential discovery. Raw archives are not browsed; descriptions are produced.
  • You evaluate a proposed package against your acceptance criteria before any commitment.
  • Rights, delivery, provenance, and payment are coordinated through one counterparty.
  • Accepted and rejected packages both improve the maps, so the next search is faster.

Buyers disclose enforceable commercial terms: purpose category, identity terms, rights, retention, and payment. Model architecture, unpublished research, and proprietary evaluation methods stay confidential.

Inside the railillustrative sequence, synthetic values
  1. Stream selection
  2. Metadata
  3. Provenance
  4. Matching
  5. Routing
candidate_streams3 identified
domain_fitsupport_operations
sequence_typefailure_resolution
outcome_capturepresent
source_confirmedowner_controlled
rights_positionstanding_permissions
requirement_fithigh
acceptance_criteriaoutcome_evidence_required
package_statequalified
deliverylicensed_output_only
03 / Market evidence

Organisations already pay heavily for human data.

Two public reference points for what is paid, and one for what unconsented data now costs:

$14.3B

Meta's investment for a 49% stake in Scale AI, a company that pays people to produce training data.

Forbes, June 2025
$60M/yr

Google's reported licence for Reddit's user posts.

CBS News, February 2024
~$3,000

Per covered book in Anthropic's proposed copyright settlement: the cost side of unconsented data.

NPR, September 2025

Figures as publicly reported by the named outlets. The full source list is on the About page.

Define a Buyer Requirement