Document self-service · synthetic demo

Give it the document pack.
Get a path to resolution.

A useful reviewer should not stop at “something is wrong.” PrüfPilot checks the pack, explains the evidence, prepares the next hand-off and keeps consequential release behind an explicit human or policy gate.

60-second workflow

Funding application

Five proofs are required. Three are present and two financial records disagree.

01 · Inputs

Document pack

Synthetic fixtures stand in for the real PDF intake supported by the backend.

02 · Check & resolve

What should happen next?

Detect the issue, trace it to evidence, then prepare the smallest useful next action.

GoalTell me whether this pack is ready — and move it toward resolution.
Reading manifest → checking requirements → comparing evidence → choosing the next safe action
Finding

Recommended next action

Prepare the hand-off.

Prepared applicant message
RecipientSubjectCase state after send

All document checks passed. This demo still keeps release behind a human gate because the workflow has not earned production autonomy.

Released · simulated. The reviewer approved the clean case for release. No external production system is connected in this public demo.
Workflow boundary: detecting a problem is not the finish line. PrüfPilot should route each case toward the next resolvable state — while keeping external communication, exceptions and consequential release behind explicit authority.
How autonomy should grow

Earn release authority from evidence.

The release policy belongs to the workflow and risk tier — not to model confidence alone.

Stage 1Shadow / reviewSystem checks and prepares every hand-off. Human makes every consequential decision.
Stage 2Human releaseClean packs and prepared requests arrive ready to approve. Exceptions stay routed for review.
Stage 3Bounded auto-releaseOnly low-risk, stable case classes can auto-release after explicit policy approval and measured evidence.
Example promotion gate: 100+ reviewed cases · no critical safety failures · low correction rate · stable source coverage · clear rollback · human override. High-impact decisions can remain human-only indefinitely.
See engineering proof

The backend supports real PDF intake, original-byte hashing, exact source locators, versioned authority, persisted findings, append-only human decisions and a fail-closed production gate. The public workflow stays synthetic so it is safe and inspectable.