Architecture · Saolix Eagle

Where the line sits.

One pipeline, with a hard line through the middle. This page shows what runs on each side of the line, how the line is enforced in code rather than by convention, and how the whole thing sits inside your perimeter. The evidence for each statement is in the white paper.

The pipeline

Models score. Rules decide. The log keeps both.

AI ADVISES · PROBABILISTICFraud modelAML modelAnomaly modelRISK SCOREDETERMINISM BOUNDARYRULES DECIDE · DETERMINISTICFixed rulesApprovedHeld for reviewBlockedLedgerreplayable record
Three layers

What runs where.

Above the line

Probabilistic, advisory

  • The match chain: normalisation across alphabets, phonetic and fuzzy comparison, n-gram similarity, fused into one score
  • The local language model: a small open model on your CPU servers, weights pinned by digest, behind one port in the code
  • The copilot: re-ranks the review queue, drafts RAG-grounded explanations, drafts case reports
  • Everything here is logged as advisory. A model signal can raise a result, never lower or close one
The line

Enforced, not assumed

  • A build guard fails the build if any path lets a model output become a verdict
  • A database constraint accepts a decision source of RULE or HUMAN only
  • A runtime assertion checks the source as each decision is written
  • A second build guard keeps the verdict core free of the clock, random numbers, locale and the environment
Below the line

Deterministic, decided

  • The rule engine: pinned thresholds and floors under governed configuration, each version hash-chained with a rationale and an approver
  • The verdict: pass, review or fail, from the rules or from a named person
  • The audit log: every screening sealed into a Merkle log, insert-only, verifiable on demand
  • Byte-for-byte replay proven on every build against pinned test vectors
Inside your perimeter

Four parts, no route out.

An Eagle deployment is small. It runs on ordinary CPU servers in your data centre or your private cloud, and it ships with an air-gapped configuration.

01

Application

The screening service and the analyst interface. Stateless; scales out behind your load balancer.

02

Database

PostgreSQL. Screening data in one schema; the audit log in its own schema, written by a role that can only insert and read, with write-once triggers.

03

Model gateway

Serves the local model to the application. The only path from the product to a model. Absent, AI assist is off and screening runs unchanged.

04

Network

All three on an internal network with no route out, under a sovereign mode that assumes no outbound connection. Lists, rulesets and models come in through your change process.

See the guards fail on purpose.

In a working session we plant a violation and show the build guard refuse it, run the integrity check on a live log, and run the replay conformance test.