AEST · MMXXVI
Adjudon · Frankfurt · Decision Audit Layer

Decided.  Signed.  Audited.

Making sure your AI agents are what they claim to be — at the moment of decision, in Frankfurt.

Mapped against
BaFinDORA Art. 28(7)EU AI Act Art. 13ISO 42001MaRisk AT 4.3.1
Our wedge

Three things no one else ships in combination.

We checked the alternatives in May 2026. Each capability below exists somewhere in the category. None ship all three. That gap is why Adjudon exists.

01 Cryptographic

SHA-256 hash chain + multi-anchor witness layer

Every decision lands in an append-only chain. The chain itself runs today. The three witness anchors — eIDAS RFC 3161 (D-Trust), OpenTimestamps (Bitcoin), and Trillian Tessera Signed Tree Heads — are wired and ship deterministic-mock proofs until each procurement gate fires (D-Trust contract, OTS calendar endpoints, Tessera Fly.io deployment). A regulator runs curl GET /api/audit/verify against your tenant without an Adjudon login today; the witness layer activates per-anchor as each gate closes.

02 Multi-Clock

Five parallel SLA clocks per incident

GDPR Art. 33, EU AI Act Art. 73, DORA Art. 19, NIS2 Art. 23, CRA Art. 14 — each clock runs in parallel from the same incident, with its own staged checkpoints. No competitor surveyed has shipped this. We built it because one incident triggers five regulators on different SLAs.

03 DACH-Native

BaFin / MaRisk / VAIT vocabulary in the product

Field names, export formats, audit-pack template — all in the German supervisor's language. EU-native vendors don't ship the same vocabulary. Frankfurt eu-central-1, German support hours.

The problem

Three questions you can't answer when the regulator calls.

The model works. The use case is validated. The pilot ships in two weeks and then sits in legal review for six months. Three reasons, every time.

01

Your AI scores its own confidence.

That number is self-reported, uncalibrated, and legally meaningless. When the model is wrong, it's wrong with confidence. No one catches it.

02

There's no gate before the customer.

Risky and safe decisions look the same in your pipeline. Both ship. Observability tells you what already happened — it doesn't decide what gets to happen.

03

The regulator will ask. You can't answer.

Article 13 demands transparency for every decision. “The model decided” is not an explanation. Today, that record doesn't exist.

How it works

From AI decision to compliance record. p50 under 10 ms.

Adjudon doesn't replace your AI agents. It wraps them — three engines: Score, Enforce, Prove. One POST. One HTTP response. One chain entry.

Agent14:23:07Z0x7f3a…
Policy+ 4msPERMIT
Audit+ 6ms0x4c2e…
Witness Seal+ 11msSEALED
Step 01

Send

Your agent POSTs the decision to /api/traces. Two SDK lines. Node.js or Python.

Step 02

Score

The Confidence Engine triangulates from three runtime signals — then calibrates the score per-organisation against your reviewed decisions. Brier, ECE, conformal coverage, drift, bias-stratified. Not the model's self-report.

Step 03

Enforce

The Policy Engine returns one of three deterministic HTTP responses. Pure deterministic Python. A regulator can re-run any rule offline.

Approve201Flag for review202Block403
Step 04

Prove

The decision lands in a SHA-256 hash-chained log. A regulator verifies it in three commands — without an Adjudon login.

BaFin is asking now. The AI Office will ask soon. Either way, the chain answers.

10,000 decisions per month, free, no credit card. Wire one agent — a card-block decision, a triage routing, a credit approval — and watch the chain build in real time. Frankfurt eu-central-1, two SDK lines, ten minutes.

Start free Talk to an engineer