What ELYSÉA can prove — and what it cannot
Technical brief for insurers, lawyers and DPOs advising integrators using the ELYSÉA Guardian pipeline. This document distinguishes what is demonstrable today, what is not yet, and declared blind spots.
1. What ELYSÉA can produce — and the real status of each item
Signed Guardian log (HMAC-SHA256)
ON REQUESTExportable AI Act compliance report (PDF)
COMINGPipeline transit confirmation on an identified request
ON REQUESTPublic behaviour version history
AVAILABLE2. Brain-side items — status read from the tools source
Compliance log written by the brain at every exchange
COMINGSigned fingerprint of the brain version
COMINGContinuous behaviour probe (server side)
COMING3. What ELYSÉA cannot produce — structural limits
Full conversation transcripts
Identity or personal data of end users
IP addresses of end users
Proof of the integrated application's regulatory compliance
Proof of transit without a session identifier
4. Declared blind spots — what the pipeline does not cover
The integrator can call the model outside the pipeline
Probabilistic detection — false negatives possible
Text layer only — not notifications, interface, follow-ups
No measurement published, no external evaluation
Behaviour can diverge without code change
WORK IN PROGRESS — BEHAVIOUR FINGERPRINT
Following the 27–28 August 2026 episode, a work item was opened on the Core side to produce:
Your dossier
Dossier 03 — Insurers (underwriters, brokers, MGAs)
- Problem: generative AI exclusions since 2026-01, for lack of visibility [S-ASS-01].
- ELYSÉA: AIUC-1 mapping — C001-C004, C006, E015 covered by evidence; C010-C012 require a third party; E002/E003 remain with the insured.
- Reservation: no source promises that trace + evidence are enough to obtain insurance — a hypothesis to be tested.
- Next step: AIUC-1 workshop after the results are published. →
03_INSURERS.md
Measurements (campaign results file)
- Crisis recall — French
- Crisis recall — English
- Over-blocking
- Adversarial robustness
- Response quality