We don't use AI. That's the point.
Every page is filed by a signed barcode, not a content guess. No LLM. No OCR. No training on your documents. The page either has a verifying signature (filed) or it doesn't (flagged UNKNOWN). Loud failures. No silent misclassification.
14 days · no credit card · Windows + Linux today, macOS coming soon
Three things AI doc-filers do wrong.
Not a hit-piece on any one product. The same three failure modes show up across the category, and any auditor on a serious matter will hit at least one of them.
Classifies wrong, and you don't know why
An AI doc-filer reads a page, decides it's an "invoice" with 87% confidence, and drops it in the invoices folder. When it's wrong — and on a long enough timeline it's wrong — there's no audit trail explaining the decision. You only find out when the auditor asks for the SWMS and it's filed under "delivery dockets."
The Matrix can't make that mistake because the destination was decided by a human at print time and signed into the page itself.
Trains on your documents
Most AI doc-filer terms of service grant the vendor a licence to use your documents to "improve the service." For a sole-trader sparkie that means nothing. For a law firm with privileged matter, a clinic with PHI, or a contractor with principal-controlled data, that one clause is the deal-breaker.
The Matrix has no document ingest endpoint. ChunkLand cannot train on what we never receive.
Needs internet
If your fly-camp Wi-Fi is closed off, your site is on a $20/GB satellite link, or your client engagement letter says "no cloud," an AI doc-filer is dead in the water. The model lives on the vendor's servers. Inference happens off-site. No internet, no inference.
The Matrix runs on your machine. Stamping and reading both work air-gapped — first-launch licence check then a 7-day rolling offline grace.
The QR stamp + HMAC chain-of-custody.
Deterministic, byte-for-byte. If you can read the source, you can verify the claim.
The byte layout
When you print a page through The Matrix, a small QR stamp goes in the top-right corner. The payload is 36 bytes total — laid out exactly like this:
MAGIC(1 byte) —0x5C— sort/slash mnemonicVERSION(1 byte) —0x02for v0.2;0x01still readableLICENSE_ID(4 bytes) — uint32 derived from your licence keyDOC_ID(8 bytes) — uint64 hash of the caller-supplied document idPAGE_NUM(2 bytes) — uint16, 1-basedCATEGORY_ID(1 byte) — your library destination, looked up locallyTIMESTAMP_U24(3 bytes) — uint24 minutes from a 2026-01-01 epoch (rolls 2057)
The 20 bytes above are XOR-scrambled with a keystream derived from your licence key. Then 16 bytes of HMAC-SHA256(license_key, scrambled_payload) are appended, truncated to 16 bytes. Verification on read is constant-time via hmac.compare_digest.
That's the whole filing decision. The destination category was chosen by a human at print time, baked into the stamp, signed by your licence key, and routed back into the right folder when the page is scanned. There is no model, no probability, no confidence score.
src/codec.py — encode() / decode()Failure is loud, not silent
If a page is photocopied — fine, the stamp survives at 200 DPI. If it's faxed — fine. If it's coffee-ringed, folded, grease-thumbed, or scanned on a junkyard MFC — fine. If someone retypes the page and tries to pass it off as the original, the HMAC won't verify and the reader raises:
CodecError("HMAC mismatch — wrong license key or forged payload")
The unverified page is parked in a quarantine folder for someone to look at by hand. It is never silently misfiled into the wrong category, because the system never makes a guess.
src/reader.py — group_into_sections()Where AI would actually be useful — and we still won't add it.
There are two places an LLM would genuinely help. We've thought about both. Here's why neither one ships.
1. OCR for legacy paper
If you've got a filing cabinet of pre-Matrix paperwork — invoices from 2018, SWMS from a job two years ago, contracts your previous office printed without a stamp — there's no signed barcode on those pages. An OCR model could read the text, classify the document type, and propose a destination folder. Useful.
We're not adding it. The moment we ship "OCR your legacy paper," our marketing has to start saying "we read your documents," and the security claim that holds the rest of the product up — your scans never traverse the internet, never reach an LLM, never appear on someone else's server — gets a footnote. Audit-grade products don't have footnotes.
If you've got legacy paper, our recommendation is the boring one: stamp it forward from today. The pile from 2018 stays in the cabinet where it was; the new stuff lands in the right folder from the day you install. Two years later, the cabinet's irrelevant.
2. Summarisation
"Show me every isolation cert from this 12-month contract." "Find me the 2023 invoices for that one supplier." Summarising or querying across thousands of stamped pages is a real workflow and an LLM could nail it.
We're not adding it. Same reason. Summarisation requires reading the document text. Reading the document text requires either uploading to a cloud LLM (breaks the no-cloud claim) or shipping a local model (massively expands the binary, requires GPU, opens model-update surface). Neither is a trade we're willing to make against the audit-grade promise.
If you need cross-document search on your local archive, the answer is the operating system you already have — Spotlight on macOS, Windows Search on Windows, recoll or fzf on Linux. They search filenames and PDF text without phoning home. The Matrix files the documents into the right folders so those tools have something predictable to search across.
"No AI" is a product feature.
Half the doc-management category right now is racing to add "AI-powered" to the headline. We're going the other way, on purpose. When a regulator, an auditor, an opposing solicitor, or a procurement reviewer asks "does the vendor's tool ever see the document content?" — the answer is no, and the answer never changes between releases. We can't change it without rewriting the architecture and we have no plans to rewrite the architecture.
That's the trust signal. Not a marketing claim. A choice that constrains every future release of the product.
Try it on a real day's paperwork.
14-day trial, no credit card. Stamp a real morning of work, scan it back, watch every page land in the right folder. No model. No guess. No cloud upload.