The Provenance Inspector
The crate holds the delivered dataset — it glows gold because its hash verified. Run each provenance check to peel back a layer and reveal what integrity checks can't see. Drag to orbit.
drag to rotate
Integrity: PASS
0 of 4 context checks run — trust unearned
Data vs. context: two different verification problems
Data verification (solved-ish)
- Checksums / hashes — bytes unchanged in transit
- Schema validation — right columns and types
- Signatures — sender is who they claim
All of these can pass on a dataset that is biased, stale, mislabeled, or illegal to use.
Context verification (the open problem)
- Provenance records — machine-readable lineage (W3C PROV, C2PA for media)
- Datasheets for datasets — Gebru et al.'s standard questions: why collected, by whom, what's missing
- Attestation — a party stakes reputation (or collateral) on the metadata being true
Why this bites harder with AI agents
- No human smell test. A human buyer might notice all reviews sound alike; an agent checking checksums won't. Context fraud is invisible to syntactic validation.
- Garbage compounds. Bad sentiment data → wrong trading signal → real money lost. In agent-to-agent supply chains, one unverified context claim propagates downstream at machine speed.
- The 2024–25 lesson from media: C2PA content credentials emerged because "is this image real?" turned out to mean "where has this image been?" — a provenance question, not a pixels question. Datasets are hitting the same wall.
- Practical checklist today: demand a datasheet, sample 100 random rows and read them, verify timestamps against the claimed window, and check the license of the original source, not the reseller. Twenty minutes of context checking beats any hash.