Guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
How to Verify an Artifact Hash in an Evidence Room
Short answer
To verify an artifact hash, take the artifact's sha256 value from the room manifest — recorded when the artifact was sealed at intake — and use the room's verification for that artifact: the platform recomputes the stored file's hash and compares the two. A match proves the stored file is byte-identical to the file that entered the room. Verify on a sampling rhythm — a few random artifacts monthly, a fresh sample whenever the auditor or a dispute asks — and treat any mismatch as a stop-and-investigate event, not a rounding error.
What is hash verification?
Hash verification is the act of checking an artifact against its seal on AI Audit Exchange: the room exposes a per-artifact verification that compares the stored file's computed SHA-256 with the sha256 recorded in the manifest at intake. Verification is available to the room's owner and to a granted auditor, and the manifest itself can be exported so the seals can be carried with engagement records outside the room.
The comparison at the heart of it
Strip the ceremony away and verification is one comparison. On one side: a 64-character fingerprint computed at the moment the artifact entered the room, preserved in the manifest beside the filename, source, size and import date. On the other: the fingerprint of the file as it exists now, recomputed on demand. If the two strings are equal, the files are the same file — SHA-256 leaves no practical room for two different files to share a fingerprint. If they differ, something differs. There is no third outcome and no expertise barrier: the skill being exercised is not cryptography but the habit of actually running the comparison, regularly, before anyone forces you to.
Sampling: how much verification is enough
Verify samples, not universes. A monthly rhythm that works: choose three artifacts at random from the room — genuinely at random, across sources and months — and verify each. Log the outcome in your review note beside the completeness count and the stale-flag count. Add event-driven samples: at rehearsal before the auditor starts, at the auditor's own sampling during the engagement (theirs will be larger and weighted toward significance), and whenever a specific artifact is questioned at any point in the record's life. Three random checks a month compounds into something powerful across a year: a standing, dated demonstration that the seals were alive and being tested, not decorative.
Choosing samples the auditor's way
When the stakes rise, sample like an examiner. Weight toward consequence: the artifacts a finding would rest on, the largest files, the oldest seals, the direct uploads (whose provenance rests on the auditee's word plus the seal), and any artifact that has a near-twin — an original and its correction both filed. Include one artifact from each source product and each quarter of the period. A sample constructed that way, verifying clean, supports a sentence a random sample cannot: not merely 'some files match' but 'the files that matter, from everywhere in the room, match'. Both sides in an engagement understand the difference instantly.
Reading a match correctly
A match proves sameness since intake — nothing more, and nothing less. It does not prove the artifact was accurate when produced, complete for its purpose, or current for this period; those are coverage and freshness questions with their own checks. Do not let a clean verification round inflate into a general claim about evidence quality, because auditors will deflate it in one question and discount your other claims alongside. State it precisely — 'sampled artifacts are byte-identical to their intake seals' — and let precision do the persuading. Precision about small verifiable things is how a team earns belief about large unverifiable ones.
Reading a mismatch correctly
A mismatch is a stop event. Do not re-upload, overwrite or tidy anything. First, identify which side moved: verify the stored room artifact against its manifest seal — the room's record is the reference. In practice, mismatches almost always indict the outside copy: a file edited after download, a regenerated export mistaken for the original, a namesake from a different period. The room's sealed record resolving the question is the system working. If — the serious case — a stored artifact ever failed against its own seal, preserve everything, place the room's circumstances before the platform and your auditor, and treat the record as contested until explained. The response discipline matters more than the event: calm, documented, nothing deleted.
Making verification a visible habit
Verification that nobody sees still protects integrity, but verification that is seen also builds the engagement's climate. Verify in the auditor's presence at the engagement's opening sample. Record verification rounds in the room's thread where the auditor can read the rhythm — three a month, every month, all matching. By close-out, integrity has stopped being a claim in your brief and become a demonstrated property of how the room was run. Successors inherit the same visibility: a new auditor granted the room can run their own samples in their first session and confirm the habit was real. Controls that can be seen working are worth twice their silent equivalents.
Five actions to take
- Verify three genuinely random artifacts monthly; record the results beside the completeness and freshness counts.
- At rehearsal and at the auditor's opening, sample the auditor's way: consequential, old, uploaded and twinned artifacts included.
- State a match precisely — byte-identical since intake — never as proof of correctness or completeness.
- Treat any mismatch as stop-and-investigate: check the room's sealed record first, change nothing, document everything.
- Verify visibly: in the auditor's presence at sampling, and note rounds in the room's thread.
Worked example
A fictional marine-services firm, Pelican Bay Services, runs a CPPA room of 54 artifacts. Its monthly rhythm verifies 3 random artifacts per month: 36 checks across the year, 36 matches, each round noted in the room's thread in one line. At the engagement, the auditor samples harder — 14 artifacts weighted to the decision records and the 4 direct uploads — and verifies 14 of 14.
In month nine, one routine check produces the year's only drama: a file on an analyst's laptop, believed to be a room artifact, mismatches its manifest seal. The room resolves it in minutes — the laptop file was a regenerated export from after a configuration fix, never imported; the room's sealed original verifies clean. The incident is logged in the thread as what it was: a copy problem, caught by the habit, with the record untouched. That is a verification programme paying for itself in a single afternoon.
Frequently asked questions
Where do I find an artifact's sealed hash?
In the room manifest: each entry lists the artifact's filename, source, format, sha256, size, import date and retention note. Export the manifest to work from it outside the room.
How long does verification take?
Seconds per artifact — the cost is the habit, not the computation. A three-artifact monthly sample, including noting the result, takes minutes.
Can the auditor verify hashes themselves?
Yes. Verification is available within the auditor's scoped access to the room, and examiners will typically run their own weighted samples during the engagement.
What if a file I downloaded no longer matches?
Then the downloaded copy differs from the sealed record — it was edited, regenerated or is a different file with a similar name. Work from the room's stored artifact; its seal is the reference.
Does verification check the evidence is correct?
No. It checks sameness — that the file is byte-identical to what was sealed at intake. Correctness, coverage and freshness have their own disciplines: the audit, validation and the completeness review.
Keep reading
All AI Audit Exchange guides · Auditor directory · About AI Audit Exchange · Security · Pricing · AI Audit Exchange home · Previous: Immutable Evidence Storage: Why It Matters for Audits · Next: AI Auditor Vetting: How the Owner-Run Workflow Works
Auditor Credentials: Claims vs Verified Facts, Explained · How Matching Works Between Auditees and Auditors · Multi-Client Evidence Rooms: An Auditor's Practice Guide