Recorded demo · runs in your browser
Futury Build Days 2026 · ITSG challenge

§pecs

From specification to certified test, in one session.

Every December a new ~400-page specification lands, and ITSG rewrites 3,000–4,000 test cases by hand. Our pipeline reads the spec, finds what changed, drafts the rules for an examiner to approve and turns them into test cases. Every expected value traces back to a sentence in the spec.

▶ Start demo Explore the app

5 minutes: four slides, then 17 steps through the app. → next · ← back · Esc ends.

QR code to the live app on our server
130.83.85.181:34080 The live app with the real AI, on our hackathon server

Try it

5 minutes

Guided demo

The pitch as we present it: four slides, then 17 steps through the app with recorded data. → and ← step through.

 Recorded data

Explore the app

Click through every screen yourself: switch between the ITSG examiner and a vendor (SAP, Lexware) at the top.

 PDF

Pitch deck

The five slides from the pitch, with speaker notes.

Data · Model · Logic

Data

Real ITSG documents

  • The government's DTA EEL Verfahrensbeschreibung v13.4 PDF: all 220 sections, full text.
  • Last year's v13.3 from the archive, for the diff.
  • ITSG's reference case (prior-illness request) and its real TEEL0-Test1-DBVO file, which our expected output matches byte for byte.
Model

Kolibri-1 on hessian.AI

  • Aleph Alpha's Kolibri-1 through the hessian.AI OpenAI-compatible gateway, with tool calls.
  • It gets one § section, the sections it links to and the old version. Never the whole archive.
  • It extracts parameters with verbatim German quotes and writes the formula. It never decides an expected value on its own.
Logic

Deterministic guards

  • G1Every quote appears verbatim in the spec
  • G2The formula parses in a small, safe language
  • G3Our oracle's result equals the AI's own value
  • G4Every legal number appears in its quote (sechs Wochen = 42 days)
  • G5The formula was actually run

Failed guards go back to the AI. An examiner approves every rule. The value a developer receives always comes from tested code.

The pipeline

  1. 01ImportDrop the government's new spec PDF.
  2. 02Index & highlight§ index from the PDF's own headings and links.
  3. 03Detect changesChange log + section diff against last year.
  4. 04Understand & approveAI drafts each rule; the examiner approves.
  5. 05Generate test casesIn plain words; tested code computes the values.
  6. 06Hand overChecklist and AI recap for each vendor's developer.
  7. 07CompareVendor file vs. expected, every difference in words.
  8. 08Report & certifySealed report, certificate, next audit scope.

spec § → AI reads (quotes + formula) → 5 guards → examiner approves → tested function → expected value → vendor check