Proofdesk.

A working system — running live on this page

The knowledge desk that refuses to guess.

Proofdesk answers your team’s questions from your own contracts, policies, and manuals. Every answer opens the exact passage it came from — and an answer the documents can’t prove is declined, never dressed up.

Built for your business · Integrated into your tools · Owned by you

No subscription, no per-seat pricing — we build it around your documents and systems, hand it over, and it’s yours.

Watch it runSee a recorded session
Get it built — free test runbuilt for your business · $10 test budget on our infrastructure
live now: 342 documents · 4,400 passages · 554 questions weighed, every one recorded
watch a full run

How it works

Ask a question. Get a proven answer — or an honest no.

Five stages, every time. The checking cannot be skipped: if any part of the system has a bad day, the desk declines rather than guesses.

  1. 1

    You ask

    In plain language, like messaging a colleague.

  2. 2

    It finds

    The passages in your documents that bear on the question — with their exact location.

  3. 3

    It drafts

    An answer where every part carries its source.

  4. 4

    A second AI checks

    A different company’s AI verifies every part against the documents. It never grades its own homework.

  5. 5

    Proven — or declined

    Every part proven: the answer ships with its proof. Anything unproven: the desk declines and shows the closest passages.

And it runs itself

Reads

Documents are split into passages that keep their exact place — that’s why a citation can highlight the exact lines.

Records

Every answer and refusal is stored — what was found, what was checked, what it cost, how long it took.

Examines itself

It re-takes a fixed exam of graded questions; every result is stored and traceable.

Never goes dark

If the pipeline is ever down, recorded real runs play instead — you never meet a dead demo.

All of it is running on this page right now — ask it yourself. Open the desk

Watch a real question

Watch the checking happen.

The pipeline you just read, playing back a recorded question — the passages found, the draft, every part checked, the stamp.

As amended, what monthly service retainer does Meridian now charge Brookfield Clinic?

The passages are found, the draft is written, and a second AI checks every part — you are about to watch it happen.

a real recorded run

The part no other tool does

When the documents don’t prove it, it says so.

A real declined run, at full ceremony. No guess, no error screen — the closest passages, and the decision left with you.

What monthly gym-membership allowance does Meridian offer its staff?

The passages are found, the draft is written, and a second AI checks every part — you are about to watch it happen.

a real recorded run

How we know

Every claim here has receipts.

Nothing on this page is asserted by hand. Three things back it up — a fixed exam of graded questions, the recorded trail behind every answer, and a refusal we caused on purpose.

The graded exam

You don’t have to take our word for it: the desk is run against a fixed set of real graded questions — some answerable, some deliberately not — and every result traces to a stored run.

The paper trail

Every answer and refusal is recorded — what was asked, what was found, what was checked, what it cost, how long it took. Open any answer’s proof pack to see the whole receipt.

The kill switch

We broke a key on purpose to see what happens when the checker can’t run — it refused, rather than answer unchecked. That recorded refusal is in the replay bank, clearly labeled.

Scale, stated honestly

A public example to browse. A million-word project used to test the system.

They are separate on purpose. The public example is open to browse. The larger project uses a made-up company only for scale testing. Neither contains client data.

Public example

342 documents · 4,400 passages

The recorded runs on this page use this prepared example.

Recorded scale project

997 documents · 9,798 passages · 1,265,816 words

Authenticated production read-back · 25 August 2026 · generated test company, not a client.

Recorded knowledge state

  • DocumentsReady
  • ConnectionsReady
  • Across documentsReady

These labels describe saved documents and their connections, not answer quality; every answer still has to prove itself from exact passages. The full 73-question comparison between the usual search and the broader across-document search is still pending, so there is no result or better-answer claim yet.

One document was re-read and connection sets for 3 of 997 documents were recalculated. An independent full check still matched all 2,991 nearest-document links.

What we build for you

Your desk is assembled from these blocks.

Proofdesk isn’t a subscription you sign up for. It’s a system we build for your business from proven blocks — the gold-chipped ones are running on this site right now; the rest we add on request, wired to your documents, your rules, and your tools.

Builtrunning live on this site — watch it work

Build on requestadded to your build when your business needs it

The trust core

why this desk cannot lie to you

Built

The honesty check

No confident wrong answer can reach your team: a second, independent AI checks every part of an answer against your documents before you see it.

Checking strictness tuned to your risk — a clinic is not a sales team.

watch it work
Built

Proof on every answer

Verify in one click: every answer opens to the exact place in the exact document it came from.

Your documents, your numbering, your language.

watch it work
Built

The honest “I don’t know”

Unproven answers are withheld — you get the closest passages instead of a guess.

The refusal message and the escalation path are written for your team.

watch it work
Built

The graded exam

You don’t have to take our word for it: the desk is graded on a fixed exam of real questions — answered, proven, declined — and every result is stored.

A private scoreboard on your own questions comes with every engagement.

watch it work

Loading knowledge

getting your documents in

Built

Drop in your documents

Drop files, say in your own words what they are, confirm the plan — then ask.

The setup proposals learn your document types.

watch it work
Built

PDF documents

Contracts and manuals as PDFs — read with page-exact citations that open to the right page.

Office files and web pages on request; scanned paper on request.

watch it work
Build on request

Connected sources

The desk reads straight from SharePoint, Drive, Notion, or Confluence and stays current.

We wire your two or three systems — not a 275-connector zoo.

Built

Re-read updated documents

Add or replace a document: changed words change the answers.

Automatic daily, weekly, or monthly runs are added with a funded operating plan.

watch it work

Where your team asks

the desk meets your team where it already works

Built

The web desk

Ask in plain language; get the proven answer or the honest no.

Your wording, your example questions.

watch it work
Build on request

Ask from Slack or Teams

Questions answered where the team already talks, proof attached.

Your workspace, your channels, your escalation rules.

Build on request

Department desks

An HR desk, a contracts desk, an on-call desk — each seeing only its documents.

Scope, wording, and refusal path per team.

Built

Share an answer

Any proven answer becomes a link a colleague can open — proof included.

Access follows your team boundaries.

watch it work

Staying current

the desk that knows what it knows

Built

Corpus health

Flags what nobody cites, documents that contradict each other, and superseded versions.

Health rules that match your document lifecycle.

watch it work
Build on request

Verified by your expert

Your people stamp documents “verified” with an expiry; the desk prefers them and nags on expiry.

Your verifiers, your expiry policy.

Whose desk it is

control, access, and ownership

Built

Invites and the test budget

Access by invite, spending a real owner-funded budget with a hard cap in code.

Evaluation budgets sized per prospect.

watch it work
Built

The corpus console

Add, replace, or remove documents over time; see what your team asked and what it cost; answer the desk’s follow-up questions — or tell it what to change and confirm. Nothing happens unconfirmed.

Your document lifecycle, your reviewers.

watch it work
Build on request

Per-team access

Different groups see answers from different document sets — nothing leaks.

Mapped to your real teams.

Build on request

Runs on your own machine

The whole desk — including the AI — on a machine in your building. Nothing leaves.

We size and set up the hardware.

Your choice of AI

It runs on the AI you choose.

When we build your desk, you choose what powers each stage — capable AI we run at no per-answer charge, paid frontier AI, or open AI on hardware you own, where your documents never leave the building. One rule holds in every choice: the checker is always a different company’s AI than the writer. Cost is per answer, never per seat — and you never pay inside this app.

Budget

measured

$0.00 per answer

runs on our own model subscription — no per-call charge; the checker is still a different AI than the writer

Part of your build — ask about this

Quality

estimate

about $0.01–0.05 per answer

paid frontier AI checks the answer — a different vendor than the writer, so the checker never grades its own homework

Part of your build — ask about this

Private

on request

runs on hardware you own

the whole desk on a machine in your building — your documents never leave

Part of your build — ask about this

The honest comparison

Why not just use the big tools?

Use them if they fit — here is the one thing they don’t do. And they are subscriptions to someone else’s system; this one is built for you, and it’s yours.

ToolWhat it does wellWhat it won’t do
GleanCompany-wide search and assistant for large enterprises, with strong citations.Its own docs say ungrounded answers still reach users — it monitors quality afterward; it never refuses. Sales-led, ~100-seat economics.
GuruA verified company wiki — humans stamp documents as trusted.Verification is a badge on documents, not a check on answers; its guardrails are off by default and fail open when the checker is down.
OnyxOpen-source chat over your company tools; you can self-host it.Bets everything on good retrieval — no check that the answer is actually proven, and the permission layer is paywalled.
Do it yourselfFree open tools can chat with your documents on your hardware.A raw tool, not a service: no checking layer, no scoreboard, and someone in-house has to build, tune, and maintain it.
ProofdeskAnswers proven against your documents, or honestly declined — every run recorded.It will not show you a guess. That is the point.

The next step

Get it built for your business — with a free test run first.

One door: tell Danylo about your business, get a funded test run, and we build it for you if it fits.

Everything on this page is the live product — not a mock-up, not a slide.

We build and hand over a version of it for your business: wired into your tools, owned by you.

You get $10 of test usage on our infrastructure first — on our prepared example documents, or on your own. The invite code comes from asking Danylo.

Your documents, and your team’s corrections, make it smarter over time — and it refuses to guess. It becomes your system, not a rented tool.

Tell Danylo about your business

What is your business, and what would you want the desk to answer from your own documents? This starts a real conversation — you get a funded test run before anything else, and you never pay in this app.

Proofdesk is a tool: you review and decide. Answers are drawn only from the documents you provide. Not professional advice. Provided as-is.

What the manual way costs

For a 20-person firm, digging answers out of contracts and policies by hand runs to roughly $14,000 a year of interrupted time — before counting the one missed clause that auto-renews a bad rate. A worked example, not a measurement: your numbers replace these.

worked example

For a clinic, the bigger cost is risk: one confidently wrong protocol answer is an unbounded liability. Proven-or-refused — never guessed — is the property that makes AI answers usable in that setting at all. A real buyer with $276,000 of spend history was hiring, from scratch, exactly this checking layer. A worked example, not a measurement.

worked example