Guide
Dataset Provenance Explained
Short answer
Dataset provenance is the record of where training data came from and how it got from capture to delivery: who a contributor was, what site and jurisdiction it was captured in, what consent basis applies, and what license governs use. Firsthand tracks this per episode in meta.json and summarizes it in a per-batch data card. This is not legal advice.
Last reviewed:
What does dataset provenance actually cover?
Provenance is the chain of facts that lets a buyer answer "where did this data come from, and can we use it" without contacting the collector. It covers the capture site and jurisdiction, the consent basis for each contributor, and the license terms attached to the delivered footage. It is broader than consent alone — consent is one link in the provenance chain, alongside licensing and per-episode identifiers.
What fields does a per-episode provenance record carry?
- episode_id (meta.json field)
- Unique identifier tying every artefact for one episode together
- skill_id / site_id (meta.json fields)
- What task and location the episode was captured for
- participant_id (meta.json field)
- Links the episode to its signed release and consent artefact, without exposing identity
- duration_ns (meta.json field)
- Episode length, used for validated-hour accounting
- license (meta.json field)
- The license terms attached to that specific episode
- reject_flags (meta.json field)
- Records why an episode was excluded, if it was
- Consent artefact IDs (batch data card)
- Listed per episode at delivery, per the consent chain
- Jurisdictions of capture (batch data card)
- Where the underlying footage was recorded
Provenance-relevant fields already documented in the episode schema and data card
These are the same episode-level and batch-level fields already documented on the episode schema page and on /provenance — this table does not add any field beyond what each already states.
How is provenance different from a consent chain?
A consent chain is the sequence of checkpoints — release signed, site permission on file, de-identification applied, human review, delivery, post-delivery — that establishes a contributor agreed to their footage being used. Provenance is broader: it is the full record of an episode’s origin and handling, of which the consent chain is one part, alongside the site, jurisdiction, and license fields listed above. See consent requirements for AI training data for the consent-chain walkthrough on its own.
How is provenance different from a license?
A license is the grant of rights to use the data — what a buyer is permitted to do with it. Provenance is the evidence trail behind that grant: who the data came from, where it was captured, and under what consent basis, which is what lets a license be warranted rather than just asserted. Firsthand ships a perpetual, buyer-owned commercial license alongside the consent artefacts and data card that document its provenance.
Why does provenance matter for a commercial training pipeline?
- A buyer training a commercial model needs to know the license permits that use, which depends on tracing the license back to a documented consent basis.
- Regulators and enterprise customers increasingly ask for the origin of training data, not just a license statement.
- A reject_flags field and per-episode identifiers make it possible to explain why specific footage is or is not in a delivered batch, rather than treating a batch as an unauditable block.
FAQ
Dataset Provenance Explained, answered.
01What is dataset provenance?
The record of where training data came from and how it moved from capture to delivery — capture site, jurisdiction, consent basis, and license terms, tracked per episode and summarized in a batch data card.
02Is provenance the same thing as a consent chain?
No. A consent chain is the specific sequence of checkpoints that establishes valid consent. Provenance is the broader record that includes consent along with site, jurisdiction, and license fields.
03What fields does Firsthand track per episode?
episode_id, skill_id, site_id, participant_id, duration_ns, license, and reject_flags in meta.json, plus consent artefact IDs and jurisdictions of capture in the batch data card.
04Does a provenance record replace the license agreement?
No. The executed license agreement controls. The provenance record is the evidence trail that supports and can be checked against what the license grants.
05Is this page legal advice?
No. This is a plain-language explanation of provenance concepts and the fields Firsthand already documents. Confirm license and compliance questions with your own counsel.
Check it against the sample pack.
40 episodes across 4 environments, delivered in the exact schema these guides describe.