Make an AI’s decisions provable to a stranger.
GraQle is EU AI Act–aligned by design. We give your high-risk AI system the signals, audit trail, and disclosure primitives you need to satisfy your own Article 9 risk-management file — without GraQle itself being subject to the high-risk obligations.
The distinction that makes it click
A witness statement, or a fingerprint?
The witness-statement-vs-artefact framing was co-created publicly with enterprise architect Javier Á. Martínez Rodríguez.
Three things GraQle does NOT do (legally clean)
- GraQle is NOT itself a high-risk AI system. No Annex III category applies to a developer-side reasoning SDK — you do not inherit the high-risk obligations from us.
- GraQle is NOT a General-Purpose AI Model provider under Article 51. We use third-party LLMs — we do not place one on the EU market.
- We provide signals and audit primitives that deployers quote in their compliance file. We use the word aligned — never the stronger claims a substrate cannot make for you. The discipline is enforced in our own code by a non-claims invariant test.
Article-by-Article
Seven articles, each mapped to a shipped surface
Every row points at something you can run today. Your Article 9 risk-management file can quote these surfaces directly.
What we do not cover — and say so
Naming the boundary is the credibility move. GraQle is one layer of a stack — it does not pretend to be the whole thing.
Prohibited practices — deployer-decision territory, not SDK territory.
GPAI obligations — we use third-party LLMs; we are not a GPAI provider.
Systemic-risk GPAI duties — same; not a GPAI provider.
Conformity assessment — a deployer / notified-body process a substrate cannot perform.
Roadmap, stated honestly: the cryptographic substrate, the offline verifier, the runtime middleware, and the anchoring worker are shipped. Automated Article 9 periodic-assessment and Article 11 baseline-document generation are in research — we make no shipped claim for them here.
From decision to artefact
Mount it once. Every decision emits a verifiable receipt.
Attach GraQle to a production AI service as middleware, or a one-line decorator. Every decision becomes a PII-safe, verifiable record — no change to the decision code, nothing added to the user’s response time. Unmapped personal fields are dropped by default.
An honest boundary: the substrate records and proves the decision. It does not decide permission, and it is not the whole answer — other layers of the stack belong to other people doing that work.
from graqle.governance.runtime import attest
@attest # one line — PII-safe, zero latency
def decide(application):
...
return verdict # → a signed, tamper-evident receiptWhy you shouldn’t have to trust us
The verification needs a public key and a public log entry — not access to us. A stranger who never heard of GraQle can check a record using open standards alone.
RFC 8785 — canonical form
Any language verifies the bytes the same way.
RFC 6962 — Merkle trees
Inclusion is provable; tampering is detectable.
Sigstore Rekor — public log
Anchored in a public transparency log — the kind that secures software supply chains.
ed25519 — signatures
Signed with a key-validity window. Verify with the public key alone.
$ graq compliance export --since 2026-08-01 \
-o evidence.jsonl --sha256-sidecar
✓ canonical-form JSONL written
✓ SHA-256 sidecar — tamper-detectable years later
$ graq attest verify evidence.jsonl
✓ VERIFIED — public key + public log only. No access to us.What’s shipped — and where to check it
Every line is live code on PyPI and GitHub. If marketing ever disagrees with the repo, the repo wins.
- Offline proof verifier — a stranger verifies a bundle with public keys alone: no network, no account, no proprietary code.
graq attest verify · python -m graqle.verify - Cryptographic substrate — RFC 8785 canonicalization, RFC 6962 Merkle trees, ed25519 custody, Sigstore Rekor anchor.
graqle.governance.tamper_evidence - Runtime middleware — one line records a deployed AI’s decision, PII-safe, no latency on the response path.
graqle.governance.runtime - Continuous anchoring worker + an Article-72-style health snapshot.
graq govern serve · graq govern health - Article 12 evidence export with a deterministic canonical-form JSONL and a tamper-detect sidecar.
graq compliance export --sha256-sidecar - Article 15 robustness attestation — 17 named defences, 7 measurable claims, machine-readable.
graq compliance status --include-robustness
The dates that drive this
Honest answers
The questions a serious buyer asks
Reach for evidence that stands on its own.
Someone will ask you to prove what your AI did. GraQle builds the part that makes the answer stand on its own — and points to the people who hold the rest of the stack.