Hi again! 👋 Last week I asked here whether building software with AI makes me a dev. Thanks to everyone who answered. Today I want to share something I built and ask for your feedback.
Some context first. I’m building Provenia, a tool for documenting how projects and work files evolve over time. I wanted to keep dated records of my files while keeping the originals on my device. The browser calculates a hash of each file, and the server receives that hash and the metadata needed for the record.
That raised a question for me. When a page says “this record is signed and timestamped,” how can someone check what that actually means?
I wanted a way to explore that question with your own files and your own terminal. So I built the Provenia Audit Lab.
## How it works
Pick two files from your computer. Any type, up to 50 MiB each. Use test files with generic names and no personal information, since the names appear in the dossiers.
Your browser calculates the SHA-256 hash of each file. This happens in a Web Worker. The originals stay on your device. The server receives their hashes, names, and sizes.
The Lab generates two dossiers, A and B. Each is a ZIP containing a PDF and a JSON manifest. One passes all the audit checks. The other contains an inconsistency.
Investigate and choose which dossier passes the audit. After you submit your guess, the Lab explains what changed and where to find the evidence.
No account is required. A session lasts up to 15 minutes, and the round is held in memory.
## What you actually check
You can check the downloaded dossiers locally, using your original files and the public trust references provided separately from the ZIPs.
The Lab includes a downloadable local auditor (local-audit.zip, Python 3.12+). It doesn’t import the generator or know the answer. It checks the evidence in each dossier.
Some of the checks include:
File hashes and association. Does each record in the JSON match the name, size, and SHA-256 hash of the corresponding original file, in the correct order? You can also calculate the hashes yourself using
sha256sumon Linux,shasum -a 256on macOS, orGet-FileHashin PowerShell.JSON signature and signing key. Is the Ed25519 signature valid? Does the signing key match the public reference listed separately on the Lab page?
PDF signature and integrity. Is the PDF signature valid? Does the signer match the expected public reference? Does the final signed PDF match the hash declared in the JSON?
Timestamp. Is the RFC 3161 timestamp token from the expected authority, FreeTSA? Does it cover the correct file hash?
Each check ends as VERIFIED, FAILED, or PARTIAL.
A dossier signed by an unpublished key can have a mathematically valid signature and still fail the audit. You also need to check whether that key belongs to the issuer you expected.
The public references matter here. A key supplied inside the ZIP can help you check a signature, but establishing who that key belongs to requires a trusted reference outside the bundle. The Lab provides those references on its page so you can examine that step too.
## What I learned building it
The hint inside the Lab helped me separate three things:
- A hash lets you compare a file’s bytes with a recorded fingerprint.
- A digital signature lets you check whether the signed data verifies under a particular key.
- A timestamp binds a hash to a time certified by a timestamp authority.
Each answers a different question. The timestamp provides temporal evidence for the recorded hash. Authorship and the date when someone first created the content require other evidence.
The other lesson was simple: keep your original. To check whether a file matches a recorded hash later, you need the exact bytes that produced it.
The Lab generates training dossiers. Its document-signing keys and PDF trust root are separate from the official product’s. The timestamps come from the external authority, FreeTSA.
## Try it and tell me what you find
If you identify the inconsistent dossier, which check gave it away?
And if a step is confusing, or you think a check is missing, I’d love to hear about it. As I mentioned in my first post, I’m still learning to read and debug this kind of code on my own. Feedback from people who work with it every day would help me a lot.
Thanks for reading!

Top comments (0)