Pre-termination defensibility review for German employers · Verfahrenscheck by ARAM Algorithm

Know if the Kündigung holds — before it goes out.

Verfahrenscheck runs the termination file against the eight statutory gates German courts test it on. Each gate passes, fails, or flags a gap — silence is never a pass. The gap is the finding.

§102 BetrVG §626 Abs. 2 BGB §1 KSchG SGB IX AGG

One live matter, de-identified locally, documented read in 72 hours. Deterministic engine on your device — AI is used at design time only, never on a client matter.

I’m in HR I’m in-house counsel I’m the employment lawyer

Procedural completeness review (Tier 1) · not Rechtsberatung

The output is not “here is the answer.” The output is here is what is defensible — a per-finding map of evidence, gate, and BAG warrant. Inspectable end to end.

  • 50BAG decisions parsed
    200 planned
  • 8main gates
  • 53sub-gates
  • 0LLM calls at runtime

Eight gates. One fail-closed standard.

Each gate is anchored to a statutory provision or established authority. Where authority cannot be verified, the gate flags the limitation rather than asserting certainty. Same methodology, every matter, every manager, every site.

  • G0Special Protection Screen · §17 MuSchG · §168 SGB IX
  • G1Works Council Hearing · §102 BetrVG
  • G2Social Justification · §1 KSchG · §626 BGB
  • G3Ultima Ratio · Verhältnismäßigkeit
  • G4Consistency · Equal-treatment principle
  • G5Co-Determination Tech · §87(1) Nr. 6 BetrVG
  • G6Model Transparency · GDPR Art. 22 · AI Act
  • G7Anti-Discrimination · §1, §22 AGG

The challenge is not generating recommendations. The challenge is defending decisions.

The engine that reviews a matter is deterministic infrastructure, not a language model. Replay equivalence — same record, same verdict, today and in five years — is the property the audit trail rests on.

Deterministic runtime
No probabilistic step
The same record always produces the same reconstruction. Fail-closed gates, no narrative, no scoring softening the verdict.
AI offline only
No runtime LLM
AI is used at design time only — to interpret BAG case law into a tested, hash-locked ruleset. AI never runs on a client matter.
Local · on-device
Windows desktop, fully offline
The deterministic engine runs on the user's own hardware. No cloud inference. Raw documents never leave the device.
Human-in-the-loop
Indeterminates route to a human
The system identifies what requires professional judgement and routes it to a human; the indeterminate is never resolved automatically.

What we check. What we don't.

We test whether a dismissal will hold up procedurally — not how a judge will feel about it. Every boundary below is deliberate, and every gap is named.

What we check
Germany · 8 checkpoints · 4 dismissal types
26 procedural checks across every German dismissal route — conduct, health, redundancy, and extraordinary. Covers the core employment, co-determination, anti-discrimination, disability, data-protection, and parental-leave statutes. Cross-border cases are out of scope.
What we don't model
Reputational cost · two niche clauses · judge temperament
We don't estimate reputational cost. We don't score retaliation claims (§612a) or business-transfer protections (§613a), and we don't predict how a specific chamber will rule. Back-pay and severance estimates are on our roadmap. Every gap is listed openly — nothing is quietly skipped.
What we flag, not score
The 3-week filing deadline · good-faith defences
The 3-week deadline to challenge a dismissal (§4 KSchG), and good-faith / fairness arguments (§242, §138 BGB), need human judgement. We surface them as notes for your lawyer — we don't pass or fail them on our own.
Not legal advice
We stop where your lawyer starts
This is a procedural completeness review, not a legal opinion. When you need binding advice, a named specialist employment lawyer (Fachanwalt) takes it from there — on a separate engagement.

Run one file through the gates.

Bring one live matter. Generate the de-identified intake locally. Receive a documented defensibility read, typically within 72 hours.

Prefer to read first? Review a sample read.