Your AI says it forgot.
We check the bytes.

Scanalis's Canary Protocol injects 8 cryptographic canary tokens into the LLM via its standard API, triggers the declared purge, then scans volatile memory byte by byte across 4 zones (application heap, Linux page cache, swap space, kernel buffers). The verdict is binary: AMNESIA_CONFIRMED if physical erasure is effective, AMNESIA_FAILED if personal data residues are detected with their exact memory address and hex dump.

Four memory zones can retain residues after declared purge: the application heap (Python/Node dynamic allocations), the Linux page cache (data maintained by the kernel for I/O optimization), the swap space (RAM extension on disk during load peaks), and kernel buffers (system zones not accessible to the application but forensically readable). These residues can persist for tens of minutes after session close.

Neither. Scanalis is a technical trusted third party that produces verifiable forensic documentation. Like a testing laboratory: we produce the technical result, you interpret it with your DPO or legal counsel. The report documents the actual execution of the purge, it does not certify regulatory compliance.

No. Scanalis is non-intrusive by design. No agent installed on your servers. No SSH access. No access to source code or model weights. The memory dump is generated by your teams (gcore or procdump) and transmitted securely. Scanalis interacts only via your LLM's standard API, exactly like your own application.

No. HDS certifies the health data host. ISO 27001 certifies governance. Neither specifies forensic audit of the volatile memory of LLMs post-session, this concept did not exist when they were written. AI Act Art. 12 (August 2026) will create this obligation for actual execution proof that neither HDS nor ISO 27001 covers.

Sovereign hosting covers the physical storage location, not what the software does with data once inside. If your model is of American origin (GPT-4, Claude, Llama...), the CLOUD Act applies regardless of server location. And regardless of the model, RAM persistence post-purge is independent of geography: it's physics, not law.

No. DLPs block what enters the LLM before processing. Scanalis verifies what remains in RAM after processing. Two complementary layers. A system can have the best DLP on the market and still retain personal data residues in volatile memory post-purge, these are two distinct problems on the processing chain.

Your report belongs to you, SHA-256 sealed, any modification invalidates the hash, verifiable by any third party with openssl in 30 seconds. Your client data (memory dump, questionnaire) is destroyed within 20 calendar days via shred -vfz -n 3 (3-pass NIST SP 800-88 Rev.1), with a signed destruction attestation delivered with the report. Scanalis holds Cyber Professional Liability insurance from Hiscox SA.

Partially. For SaaS LLMs hosted by the vendor (OpenAI, Anthropic, Mistral cloud), Scanalis cannot scan the vendor's server RAM, nobody can. However, the Minerva Protocol applies via the public API to detect what the model memorized during training. And the Canary scan applies to the client-side orchestration layer middleware, local RAG, application cache, vector store. Every report explicitly documents this scope and its limits. That's a guarantee, not a gap.

ISO 27001 certifies governance. HDS certifies the host. Scanalis verifies physical execution. These certifications describe what you planned to do. Scanalis documents what actually happened in volatile memory after processing, a layer that neither ISO 27001 nor HDS specifies, because this concept didn't exist when they were written.

Scanalis doesn't do anonymization, it verifies that yours works at the physical level. Your pipeline transforms "John Smith" to "J*** S*****" before the LLM call. Scanalis verifies that "John Smith" doesn't appear in plaintext in volatile memory after processing, even if only the pseudonymized version was sent to the model. Three possible verdicts: ANON_VERIFIED (no leak), ANON_LEAK (original data in RAM), ANON_PARTIAL (leak in the pseudonymization pipeline itself, before the LLM).

Three reasons. Production LLMs are recent: 2022–2023. The regulation creating the obligation just came into force: AI Act 2024, full enforcement August 2026. And combining forensic RAM expertise with operational LLM mastery in the same place is exceptionally rare. This isn't an oversight. It's a window that just opened.