How Gordios is built to be defensible
AI Traceability, explainability, and security: three design characteristics.
In medical devices, a decision without traceability is a contestable decision. It's true for a clinician, and it's just as true for an AI.
Gordios was designed from the ground up so that every result it produces is traceable to a source, evaluated for its reliability, and defensible in audit. This page explains how we built the platform to meet this requirement — without compromising on performance.
The four design pillars
The principles that have guided the construction of Gordios from day one.
Every result cites its source
When Gordios identifies a non-conformity or validates a review point, the result is systematically accompanied by two references:
- The specific item evaluated: the regulatory requirement cited, the section of the standard, or the reference for the customized test.
- The exact section from the analyzed document that supports the conclusion: quoted verbatim, with its location in the source file.
No 'trust me'. Every point is verifiable by a human in a single click.
Independent confidence scoring triggers manual review
Every result produced by Gordios is evaluated by a second independent AI agent that assigns a confidence score. This score does not come from the agent that produced the result — it is a separate, documented verification with its own criteria.
Simple rule: any review point with a confidence score below 90% automatically triggers manual review. Nothing slips below the threshold without human verification.
This threshold is documented in our QMS and reviewed periodically based on audit feedback.
Complex checks are decomposed into verifiable steps
For multi-criteria standards — where conformity depends on several cumulative conditions — Gordios does not produce a monolithic conclusion. The reasoning is broken down into individually verifiable steps: each sub-criterion is evaluated separately, and the overall result is built from an auditable decision tree.
Concretely: if a requirement combines three sub-requirements, you see the three intermediate results, not just the final conclusion. If a step is wrong, you know which one.
Real and synthetic test sets
Before updating a model, a prompt, or a verification rule, Gordios runs an exhaustive test suite:
- Synthetic tests: documents crafted to cover edge cases and known pitfalls.
- Real tests: an annotated corpus of real documents, representative of the diversity of files encountered in audits.
- Regression tests: every change is verified against the entire corpus to detect degradation.
No evolution of the audit engine goes to production without passing this full test battery.
Two levels of traceability, continuously recorded
Gordios tracks two distinct levels in parallel — both required by regulation.
| Level | What we record |
|---|---|
| Tool traceability | Who ran which analysis, on which version of Gordios, on which date, with which document, against which standard. Complete, immutable audit trail. |
| Decision traceability | For every review point: the requirement invoked, the source passage, the confidence score, the intermediate steps, the final conclusion. |
Both levels are exportable, integrable into your own QMS, and retained according to the durations required by the regulation applicable to your device.
Summary
| Requirement | How Gordios addresses it |
|---|---|
| Decision traceability | Source + exact passage cited for every point |
| Human validation | Automatic manual review below 90% confidence |
| Reproducibility | Decomposed reasoning, stable results on the same document |
| Usage logs | Complete, immutable, exportable audit trail |
These principles are not options to be activated after the fact — they are design characteristics of the platform. They are the reason Gordios is defensible in audit, not just performant.
Want to learn more?
Schedule a meeting with one of our specialists to best address your needs!
Book a demo