Hash-Chained Evidence Packs: Why Each Run Seals the One Before It
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer
In the vault, each run's pack hash covers the previous run's pack hash and the canonical record of this run, and the pack is HMAC-signed with the product signing key. Change any stored field in any run and recomputation breaks from that point. The chain can be checked through the vault's API, which returns the run count, head hash and whether the chain is valid.
A folder of PDFs can be edited quietly after the fact; a hash chain announces its own editing, which is the entire reason to keep evidence in one.
Each battery run is sealed into a hash chain: the run's pack hash covers the previous run's pack hash plus the canonical record of this run, and the pack is HMAC-signed. Tampering with any stored field breaks recomputation, and the chain can be checked through the vault's API. That structure is what turns a folder of test results into evidence: a provider preparing for the Article 50(2) marking duties can show not just a passing run, but an unbroken, signed history of runs across versions. The free tier includes one sample battery per system; conformance history across repeated runs is a Pro and Team feature, as is evidence-pack export.
What is a hash-chained pack? an evidence record whose identifying hash is computed over the previous record's hash plus this record's canonical contents, and signed, so altering any earlier record invalidates every later link's recomputation.
The problem behind hash-chained evidence
The Evidence Vault answers a provider-side question: does this generative system's output stay machine-readably marked after a realistic pipeline touches it? A system is registered in the vault, a sample output is uploaded, and the conformance battery runs real transforms on the real bytes: re-encoding, resizing to half size, format conversion, and metadata stripping. After each transform the bytes are re-inspected for marking, C2PA and JUMBF markers, XMP, a ProvenanceGuard assertion, or AI disclosure in metadata, and the before-and-after results are recorded as measured, never assumed.
The pack's value is mostly about timing and completeness under pressure. A brand safety question, a platform appeal, a client audit or a regulator's letter all arrive with deadlines, and all of them ask the same underlying questions: which file, made how, declared as what, published when. A team that exports the pack and sends it the same day looks, and is, organised. A team assembling screenshots and recollections does not, however innocent its actual process was. Evidence does not create innocence, but it is how innocence is demonstrated.
How ProvenanceGuard handles it
An evidence pack is the record, packaged for a reader who was not there. For a sealed asset it contains the exhibit ID and verification URL, the original and sealed SHA-256 hashes, the perceptual fingerprint, the file's format and dimensions, the tool identification and version, the registration timestamp, the creator and disclosure as declared, the inspection results for the original bytes, and disclosure wording by jurisdiction for the EU, California and New York. Exporting the pack is a Pro and Team feature, because it is the point where a working record becomes a compliance artefact.
Two kinds of fingerprint do two kinds of work. The SHA-256 is exact: change one byte and it changes completely, which makes it perfect for proving a copy is bit-for-bit the registered file and useless for recognising a resized repost. The perceptual fingerprints, computed from the pixels rather than the bytes, are approximate on purpose: they stay close when a copy is resized, re-encoded or lightly edited, and drift as edits accumulate. A good registry stores both and is honest about which kind of match it is reporting.
What survives distribution, and what does not
The registry model also disciplines what a match means. A fingerprint match says this image is the registered image or a close descendant of it; it does not say the person showing it to you owns it, and it does not say the disclosure attached to it was correct. Those judgements stay human. What the registry removes is the argument about identity and timing: which file, registered when, carrying which declaration. Most real disputes shrink considerably once those three facts stop being contestable.
The registry is the part that survives distribution. At sealing, the asset's SHA-256, its sealed copy's SHA-256, and its perceptual fingerprints are stored against a public exhibit ID. Anyone with that ID, or with the file's hash, can open the public verification page or query the verification API and confirm that this exact file, or a near copy of it, was registered, when, and with what declared creator and disclosure. Verification pages are public by design: a record only a vendor can check would not settle disputes.
A routine that survives busy weeks
The practical problem is distribution. An image is rarely published once, in one place, in its original form. It is uploaded to platforms that resize and re-encode it, forwarded by partners, saved and reposted by audiences, and archived by people who strip whatever context surrounded it. Each step is normal and usually innocent, yet together they separate the image from anything that said how it was made. Planning for that separation, rather than hoping it will not happen, is the whole discipline this guide is about.
No single layer does this job alone, and honest tooling says so. Embedded metadata can be stripped. A visible label can be cropped. A fingerprint can drift if an image is edited heavily. A registry entry proves registration, not authorship in a legal sense. Used together, though, the layers cover each other's failures: when the metadata is gone the fingerprint can still find the registry record, and when a copy is too altered to match, the labelled version and the evidence pack still tell the story.
Five actions to take
- Verify the chain after every run, while the context is fresh
- Record the head hash outside the vault, so a replaced vault would be visible
- Download and keep both copies the seal produces: the sealed copy as the record, the labelled copy for publication
- File the exhibit's verification URL beside the asset wherever your team keeps campaign or client records
- Run check-a-copy on the version you actually published, and note which checks held: hash, signature, fingerprint
A worked example: A provider re-running its battery after a model update
Its vault history held 6 runs across 3 versions, chained end to end; when a partner's reviewer recomputed the chain through the API and got a valid result with the expected head hash, the review moved on to other questions.
The lesson generalises beyond this case. Records made at publication time are cheap, calm and complete; records reconstructed under pressure are expensive, partial and contested. Everything in this guide is in service of moving work from the second category into the first.
FAQ
What exactly breaks if one field is edited?
The edited run's pack hash no longer recomputes from its canonical record, and because the next run's hash covered that hash, the failure propagates forward. Verification reports the chain as invalid rather than pointing politely at nothing.
What formats and sizes are supported?
JPEG and PNG, up to 25 MB per file. Those are the formats the inspection, XMP embedding and fingerprinting are built and tested against; other formats are refused rather than approximated.
What does the free tier actually do?
The complete job on an asset: byte-level inspection, a signed assertion embedded as XMP, a visibly labelled copy, SHA-256 and perceptual fingerprints registered, and a public verification page. Paid plans add survival re-checks, evidence packs and bulk verification.
What do Pro and Team cost?
Pro is $29 a month and Team is $79 a month, as shown on the pricing page. Paid depth means per-platform survival re-checks, evidence pack export with EU, California and New York wording, and the bulk hash-verification API.