noirebox/noirebox

⬛ NoireBox
**The black box for AI agents — tamper-evident journal, guardrails, verifiable proof.**
**Your AI agent writes meeting notes that bind your clients.
Eighteen months from now — who can prove what it *exactly* produced, and why?**
NoireBox seals every decision in a SHA-256 hash-chained + Ed25519-signed
journal, flags poisoned transcripts before they reach the agent, and exports
an attestation anyone can verify offline — **without trusting you**.
> **Works with any agent, in French and English.** The journal, the API, the
> MCP tools and the verifier are domain- and language-agnostic; each language
> gets a ~250 KB detector trained from a reproducible dataset (seed 42) — a new market =
> one dataset generator + make train. **Already have a guardrail? Keep it** —
> it's just another event source. 🇪🇺 Built for the
> European market (GDPR & AI Act) — engineered in France.
Quickstart · Real LLM scene ·
Specs · Threat model ·
The authority ladder ·
ADRs
---
Built for the European market (GDPR & AI Act)
European AI vendors face a compliance reality their US competitors don't:
**proving** — not promising — what their systems did. NoireBox is engineered
around exactly that obligation. And the clock is real: GPAI incident
tracking is **live since August 2025** (art. 55(1)(c)), transparency
obligations applied in August 2026, and high-risk logging lands
**2 December 2027** (AI Act as amended by the Digital Omnibus,
Reg. (EU) 2026/1744) — with fines up to €35 M / 7 % of turnover.
| Requirement | How NoireBox answers |
|---|---|
| **GDPR art. 5(2)** — accountability: the controller must *demonstrate* compliance | Tamper-evident journal of what the AI produced, when, on which input |
| **GDPR art. 15/20** — data subject rights (access, portability) | Signed export of everything related to a meeting — verifiable *by the subject's own auditor* |
| **EU AI Act art. 12 + 19/26(6)** — automatic event logging, retained ≥ 6 months | Every agent decision sealed at runtime; log integrity is cryptographic, not a promise |
| **EU AI Act art. 55(1)(c)** (GPAI systemic risk — live) | Serious incidents kept track of, documented and reported — sealed as first-class events (ai_incident) |
| **AI Act Annexe IV §2(f)** — document the logging characteristics | noirebox audit-pack: one folder for the auditor — export + verifier report + generated logging description |
| **eIDAS art. 41** — a qualified timestamp carries a legal presumption | Anchor profiles for qualified eIDAS TSAs (Universign, Certigna…), public TSAs, and Bitcoin via OpenTimestamps — see ADR 008/009 |
| **DPIA / DPO workflows** | Attestation exportable for the DPO; incident taxonomy feeding risk documentation |
| **Sovereignty** | Self-hosted, no telemetry, no cloud dependency, Ed25519 keys stay on your infrastructure — deploys anywhere (including EU-only clouds) |
> **Honest scope** (we're a building block, not a certification): NoireBox
> proves the *integrity* of your AI's record. Data minimization of payloads
> remains the deployer's responsibility. Key custody now ships in the box —
> the key can live in a dedicated process (0.15.0) or on a PKCS#11 token
> (0.16.0); the pinned *hardware* root is still the deployer's call — see
> the threat model. A tool that oversells compliance
> is a liability; one that states its perimeter is auditable.
Why
- **GDPR**: a platform claiming "GDPR-compliant" must be able to demonstrate
what its AI produced, on which data, and when.
- **EU AI Act (art. 12)**: risk systems must keep *automatic event logs*.
Today, virtually every LLM stack logs for *debugging* — not for *proof*.
- The market gap: observability tools (Langfuse, LangSmith…) are built for
devs, remain modifiable after the fact, and produce nothing presentable
to an auditor, a DPO or a client.
Honest positioning (state of the art)
| Tool | What it does | What it doesn't |
|---|---|---|
| Langfuse / LangSmith / Helicone | LLM observability, debugging, traces | No tamper-evident proof, nothing exportable for an audit |
| Garak / PyRIT / promptfoo | **offline** model-level red-teaming | No runtime guardrail, no proof journal |
| halo-record / gate-oc-audit | hash-chained journals for *coding* agents | No business domain, no third-party attestation, no French |