# ArmoredLedger — The Verifiable Audit Trail for the EU AI Act **The independently verifiable audit trail for AI — cryptographically signed, Bitcoin-anchored, and mapped to Articles 12 & 19 of the EU AI Act (Regulation (EU) 2024/1689).** *For AI-governance leads, compliance officers, and DPOs preparing for high-risk AI record-keeping obligations.* --- ## The problem: your AI compliance rests on logs you can't prove The EU AI Act makes record-keeping a legal obligation. High-risk AI systems must automatically log their operation over their lifetime (**Article 12**), and those logs must be retained for auditors and market-surveillance authorities (**Article 19**). But almost every log today lives in a single application database — one the operating vendor can edit, back-date, or lose. When a regulator asks you to *prove* what your AI did and when, an export from that database is an **assertion, not evidence**. Your auditor has no independent way to confirm the record wasn't rewritten five minutes before you handed it over. Every AI-governance suite on the market — Credo AI, IBM watsonx.governance, OneTrust, Saidot, Trustible, Holistic AI, Fairly/Asenion, LatticeFlow, Monitaur — gives you a self-attested "system of record": a log inside the vendor's app database whose integrity you, and only you, can vouch for. **None of them anchor that record to a public, independently verifiable source of truth.** --- ## The wedge: verify against Bitcoin, without trusting the vendor ArmoredLedger anchors each day's records into the **public Bitcoin blockchain** via OpenTimestamps. An auditor can independently verify: 1. the **cryptographic signature** on a record, 2. its **Merkle inclusion** in the day's batch, and 3. that batch's **anchor** in Bitcoin — **offline, in a browser, with no access to our systems — and the proof survives even if ArmoredLedger disappears tomorrow.** The anchor is on a chain no vendor owns. > **The market's own admission of the gap:** In 2024, IBM had to **partner with Casper Labs** to bolt blockchain onto watsonx.governance for the "tamper-resistant, time-synched" AI records it couldn't produce natively. Public-blockchain anchoring isn't a feature the governance suites have — it's one they reach outside their own stack to get. **ArmoredLedger is the integrity layer that slots *beneath* your governance suite — not a replacement for it.** Keep your governance platform for policy, risk assessment, and workflow. Put ArmoredLedger under it so the audit trail is provable, not just present. --- ## Self-attested log vs. independently verifiable record | | A governance suite's "system of record" | ArmoredLedger's anchored record | |---|---|---| | **Where it lives** | The vendor's application database | Content-addressed, hash-chained records | | **Integrity** | Asserted by the vendor | Cryptographically signed, checked against published trust anchors | | **Timestamp** | A clock the vendor controls | Daily Merkle root anchored into Bitcoin (OpenTimestamps) | | **Who can verify** | Requires access to and trust in the vendor | Anyone — auditor, regulator, adversary — in a browser, offline | | **If the vendor vanishes** | Evidence goes with them | Proof survives — the anchor is on a public chain | --- ## Mapped to the Articles your auditor cites ArmoredLedger is an **audit-trail and record-integrity layer**. It *supports* the record-keeping and traceability obligations below. It is not a full AI-governance suite, and it does not claim to satisfy the Act's risk-management, data-governance, or conformity-assessment requirements. | Article | Obligation | How ArmoredLedger supports it | |---|---|---| | **Art. 12** | Record-keeping — automatic event logging over the system's lifetime | The `ai-action-v1` schema records operational AI events (tool calls, content generation, code changes) with input/output digests, outcome, and model, under fail-closed authority binding; signed and hash-chained so the log is tamper-evident. | | **Art. 19** | Log retention — keep automatically generated logs (≥ 6 months) | Records retained for the plan's retention window; the Bitcoin anchor makes proof-of-existence permanent, verifiable long after raw rows age out. | | **Art. 26(6)** | Deployer obligations — deployers keep logs under their control | Deployers emit their own attestations to a private trust anchor, holding an independently verifiable copy of the logs they must keep. | | **Art. 50** | Transparency — disclosure/labelling of AI-generated content | Content-generation events carry the input/output digest and model in the signed record — durable, verifiable evidence of what was AI-generated and when. | | **Art. 11 · Annex IV** | Technical documentation kept up to date | Signed provenance for releases and model/pipeline changes gives technical documentation an anchored, tamper-evident change history. | --- ## The deadline moved. The lead time didn't. The EU's **Digital Omnibus** deferred the high-risk obligations — including Articles 12 and 19 — from August 2026 to **2 December 2027** (Annex III high-risk systems; Annex I-embedded systems follow in **August 2028**). That is **runway, not reprieve**: - **GPAI obligations are already in force** (since **August 2025**) — provenance for model and pipeline releases is due now. - **Article 12 logging must cover the system's operating life.** You cannot reconstruct a traceable, anchored history retroactively — so the practical lead time for record-keeping is the longest of any obligation. It starts the day your system goes live, not the day the deadline bites. Standing up a verifiable audit trail before high-risk systems ship is the difference between having evidence and scrambling to manufacture it. --- ## How it works — four steps, zero trust required 1. **Record** — every AI event is content-addressed and hash-chained: what happened, input/output digests, model, outcome, on whose declared authority. 2. **Sign** — each attestation is signed (Ed25519 / HSM P-384) and verified against published trust anchors. 3. **Anchor** — every day the Merkle root is timestamped into Bitcoin via OpenTimestamps. 4. **Verify** — your auditor checks any record in-browser: signature, inclusion, anchor. No login. No faith in us. --- ## Honest scope — what we do and don't claim We'd rather you trust us because we're precise than lose you when the details don't hold up. - ArmoredLedger provides a **tamper-evident record** that *supports* the Article 12/19 obligations. It is **not a certification**; using it does not by itself make an AI system "EU AI Act compliant." - The operational-logging SDK (`ai-action-v1`) **exists and enables Article 12 logging**, but is **not yet wired into live products**. This is the integrity layer for AI records — not a claim that we log your model's every decision today. - The signed, enforced path (`/v1/attest`) is distinct from the unsigned developer path (`/v1/attestations`). Only **signed** records are presented as proof of authorship. - The declared principal ("who acted") is **emitter-asserted**. ArmoredLedger proves a record is signed, anchored, and unaltered — **not** which specific human was at the keyboard. - Signing uses a **FIPS 140-2 Level 3 HSM**. It is HSM-signed, **not** a CMVP-certified module. - Records are **tamper-evident between anchor intervals** — not "tamper-proof." --- ## Get started - **See a live verification:** https://trust.armoredgate.com/verify - **Read the whitepaper:** https://ledger.armoredgate.com/whitepaper.html - **Talk to us:** sales@armoredgate.com — subject "ArmoredLedger — EU AI Act" --- *ArmoredLedger — by Armored Gate. Signed. Anchored. Verifiable by anyone.* *Article references are to Regulation (EU) 2024/1689 (EU AI Act). This document is informational and not legal advice.*