Your documents make claims. We check every one.

The trust problem

You already know why you don’t trust this category.

You cannot ship documents on vibes.

How it works

A relay of agents, supervised by code.

Every check runs the same five stages. Model stages propose; code stages measure character offsets, validate schemas, and check cited evidence verbatim against what the model was actually shown.

1. 1 parsingcode

   Sandboxed text extraction; every character position tracked.

2. 2 extractingmodel

   Every checkable claim pulled as a verbatim quote.

3. 3 verifyingmodel ×N

   Each claim judged against retrieved reference excerpts, in parallel.

4. 4 reviewingcodemodel

   Deterministic checks, evidence validation, severity grading.

5. 5 renderingcode

   One self-contained report file, fully auditable.

Models propose — code disposes.

The trust architecture

Enforced in code. Not requested in prompts.

1. §1
   Evidence must be verbatim.

   A conflict’s cited sentence is checked character-for-character against the excerpts the verifier was shown. No match — the finding is downgraded, automatically. Fabricated evidence cannot reach a report.

2. §2
   Unverifiable is not verified.

   Claims no reference document addresses are counted and displayed on every report. Silence about unknowns is not permitted.

3. §3
   Injection findings outrank the model.

   Instructions embedded in a document are detected by code and appended after the model review — a document cannot talk the pipeline out of reporting them.

4. §4
   Malformed output fails loudly.

   Every model response is validated against a strict schema. Failures are retried, then surfaced as warnings on the report itself. Nothing passes silently.

The report

The deliverable is a case file, not a chat answer.

Self-hosted — in plain English

Your server. Your keys. Your documents stay home.

Confyro is software you install on a computer your company controls — not a website you upload contracts to. There is no cloud copy, no account with us, and no way for us to see your documents. The questions that decide it, answered plainly:

• Q.

  Where do my documents go?

  A.Nowhere. They are checked and stored on your own machine. You can have uploads deleted automatically after every check; the reports are kept.

• Q.

  Can the vendor read them?

  A.No. The software’s only contact with us is a one-time license check — it sends your license key and an anonymous machine ID, and gets back a signed license. No document content is ever transmitted, and there is no usage tracking. We could not see your documents if we wanted to.

• Q.

  What does the AI see?

  A.It reads your documents through an account your company owns and pays for directly — under your provider agreement, not ours. If policy requires, the AI itself can run on your own hardware, so your document content stays entirely in-house.

• Q.

  Can it change or send anything?

  A.No. It cannot edit, forward, or file your documents. It only marks problems — with the exact sentence that proves each one — and leaves every decision to a person.

The part for your IT team

One container behind your reverse proxy. Documents are parsed, indexed, checked, and stored on your machine; the only outbound traffic is to the model API you configure — or none at all with local models.

# 1. authenticate against the vendor registry (token from your kit)
echo "<PULL-TOKEN>" | docker login ghcr.io -u <user> --password-stdin
# 2. configure: image tag, model API key, admin password
cp config.example.yaml config.yaml && cp .env.example .env
# 3. pull and start
docker compose pull && docker compose up -d
# 4. verify
curl http://127.0.0.1:8000/healthz   # -> {"status":"ok"}

• Bound to localhost by default; expose via your own TLS proxy — a two-line Caddy or nginx recipe ships in the kit.
• Everything lives in ./data/ next to the compose file. That one directory is your backup, your migration, and your audit scope.
• Bring Anthropic, OpenAI-compatible, Azure, or Bedrock keys — or point at a local vLLM/Ollama server so document content never leaves your network.
• Optional retention rule purges uploaded documents after checking; reports are kept. No telemetry, no usage tracking.
• Licensing uses a one-time online activation (license key + machine ID out, a signed license back) — never any document content. Air-gapped deployments activate from a signed license file instead, with no network at all.

The security review is short because there is nothing to review. It runs on your server.

Pricing

Flat, per server. Set like a rate card, because it is one.

Model spend runs on your own keys — about 1¢ per dense page. Word, Excel, PowerPoint, and PDF are all supported: a slide deck checks in about two minutes for cents, while a pricing spreadsheet costs more than its page count suggests — nearly every cell is a checkable value. These figures are modeled, and the product ships the harness that measures them on your documents; we quote the measured numbers, not these.