Start free in 5 minutes

Froga is a high-risk AI governance engine that closes the ISO 23894 risk treatment loop over git. This quickstart walks the quickstart-education example: a pass/fail student classifier classified as high-risk (EU AI Act Annex III §3 — educational systems), using real data from the UCI Student Performance study (Cortez & Silva 2008, CC BY 4.0).

What you get for free: signed evidence (ECDSA-P256 + DSSE + in-toto) + a real ISO 23894 verdict — generated and verified locally, without sending any data to a server. The engine is Apache-2.0.

The commands, commit messages and gate colour you read below are not written by hand: each GuionStep block reads them from the published quickstart-education steps/ guion. They are the 11 real steps of that guion; if a step did not exist, the build would fail. What you read is exactly what the repository does.

The free tier is ISO 23894 (good practice, not a presumption)

The verdict this quickstart produces is a self-check against ISO/IEC 23894:2023 — the AI risk management process. It is real, signed evidence, useful for internal audit and for demonstrating due diligence to a regulator. It is not the presumption of conformity under EU AI Act Art. 9 (that comes from prEN 18228, which is the paid tier).

---

Requirements

• froga installed (Installation).
• uv (Python ≥ 3.11) and dvc ≥ 3 on the PATH.
• ~5 minutes at a terminal.

To follow along on the ready-made repository, clone the quickstart-education example from the demo catalogue: it carries the froga.yaml, the data and the pipeline at the state of each step. If you prefer to start from scratch, each block below is the commit you would make yourself, in order.

---

The free cycle, milestone by milestone

The guion advances in four milestones. The guion assigns each step to one of the two workshop roles (each block’s badge tells you whose turn it is): Nerea, the product owner, declares the identity, the framing and the risk programme in the froga.yaml manifest; Martha, the dev, brings the data, writes the pipeline and runs the gates. You will play both parts. There is no treatment: the scenario is deterministic and the ISO 23894 gate closes green on the first try — the free story is to demonstrate the process without paying.

Start the project and declare the system (free: ISO 23894 only)

The dev starts the repository as a uv project with DVC (no SCM). Then the product owner declares the system identity (froga.yaml v0) and the framing: the key point is that the only declared standard is iso/23894@2023 — the prEN 18228 presumption, the sectorial standards and the Annex IV are the paid tier and are not declared (that is the open-core boundary).

Repo, identidad, proyecto uv (uv init + deps), DVC, pubkey y README del reproductor🧑‍💻 Martha

Commands⧉ copy

15

• git init -q -b main
• git config user.email marta@demo.seigarrena.dev
• git config user.name "Marta Demo"
• uv init --bare --name vldemo-quickstart --python 3.12
• patch: pyproject-index.patch
• uv add "dvc>=3" pyyaml froga "venturalitica==0.6.11" "mlcroissant>=1.0" pyarrow "scikit-learn==1.8.0" pandas joblib
• patch: gitignore.patch
• patch: gitattributes.patch
• patch: readme.patch
• uv sync
• uv run dvc init --no-scm -q
• uv run dvc config core.analytics false
• froga pubkey --out .froga/PUBKEY.txt
• git add -A
• git commit -m "init: repo, proyecto uv (deps reales) y README del reproductor"

Identidad del sistema (froga init + patch)🧭 Nerea

Contexto y estándares (prEN 18228 de pago + ISO 23894)🧭 Nerea

Why Annex III §3?

Systems that determine educational outcomes (access, assessment, track guidance) fall under Annex III §3 of the EU AI Act — the list of high-risk systems required to have a risk management system under Art. 9. The high classification is declared by you, by hand, in froga.yaml (the engine does not infer it from metadata); froga run reads it directly from the manifest.

Bring the real data and set the risk programme

The dev brings the real data (UCI Student Performance, 395 students) with its Croissant manifest and the licence notice, and writes the featurize step. The product owner closes the manifest with the minimum risk programme: one risk (risk.unfair-prediction) with two measures — demographic parity (demographic_parity_diff < 0.15, a gate) and minimum accuracy (accuracy_score > 0.65, an audit). Pure RDD: the risk and its numeric threshold are fixed before any model exists (the diff of the last commit shows it).

Dataset UCI Student Performance + manifiesto Croissant + versionado DVC🧑‍💻 Martha

Pipeline de rasgos sobre el dataset — escribe Y corre (dvc repro)🧑‍💻 Martha

Programa de riesgos mínimo con gate numérico (antes del modelo)🧭 Nerea

Measure V1: the ISO 23894 gate closes green

First the dev compiles the risk programme into the OSCAL assessment plan — the gate’s contract: which controls will be measured and against which threshold. Then it adds the model (a tabular logistic regression, seed: 42) and runs the self-check. The scenario is deterministic: the gate closes green on the first try. The venturalitica SDK (on PyPI; uv pip install venturalitica) measures the two ex-ante controls — parity demographic_parity_diff ≈ 0.055 (below 0.15) and accuracy accuracy_score ≈ 0.76 (above 0.65); both pass, so the gate is GREEN. froga run leaves a signed .froga/bundle.json and froga.lock (the freshness anchor).

Compilar el plan de evaluación OSCAL (antes del modelo)🧑‍💻 Martha

Commands⧉ copy

• froga compile
• git add -A
• git commit -m "compile: assessment plan OSCAL (contrato del gate, antes del modelo)"

Modelo logreg (train/evaluate) — escribe Y corre (gate VERDE, single-shot)🧑‍💻 MarthaExpected verdictGREEN

Commands⧉ copy

• patch: params.patch
• patch: dvc-evaluate.patch
• patch: train.patch
• patch: evaluate.patch
• patch: compliance-eval.patch
• froga run
• git add -A
• git commit -m "modelo: logreg (train/evaluate + dvc.yaml) — run: autochequeo ISO 23894 (gate VERDE; seed=42)"

The parity threshold is yours, not the standard's

demographic_parity_diff < 0.15 is the threshold you declare in froga.yaml for the gate to be GREEN on this specific model. ISO/IEC 23894:2023 (§6.5) asks you to evaluate whether the residual risk is acceptable against your declared appetite — it does not impose a number. What the engine signs is evidence that your own control passed or failed, along with the measured value. Passing controls are not printed one by one: the measured values are signed inside .froga/bundle.json.

ISO conformance and request: the cycle stays open (free-tier boundary)

The dev emits the conformance of the only declared standard (ISO 23894) and the reconstruct of the cycle from the git history — both signed — and then requests approval. This is where the free tier ends: the cycle stays open. There is no approval because there is nothing to approve to green — the paid dossier (prEN 18228 + sectorial catalogues + Annex IV) is not attested. Governing the RDD cycle up to the request, with signed and auditable evidence, is exactly what the free tier gives you.

Conformance (prEN 18228 + ISO 23894) + reconstruct firmados🧑‍💻 Martha

Commands⧉ copy

• froga conformance --standard eu/pren-18228@2026 --outmay fail
• froga conformance --standard iso/23894@2023 --outmay fail
• froga reconstruct --outmay fail
• git add -A
• git commit -m ".froga: conformance (prEN 18228 + ISO 23894) + reconstruct firmados"

Solicitar aprobación (froga request — ciclo abierto, escalón gratis)🧑‍💻 Martha

git commitsolicitud de aprobación (el acto lo commitea froga request)

Each artifact’s signature is ECDSA-P256: the bundle.json and the reports are sealed against your local public key, so any alteration is detected (tamper-evidence). That is the root of trust everything else relies on.

---

What you have when you finish (free tier)

Artifact	Produced by	What it certifies
.froga/bundle.json + .sig	froga run	Signed evidence: measured controls, risk analysis, dvc.lock
froga.lock	froga run	Freshness anchor (pipeline stage hashes)
.froga/conformance/iso_23894_2023.json + .sig	froga conformance --out	Clause-by-clause verdict against ISO/IEC 23894:2023
.froga/reconstruct.json + .sig	froga reconstruct --out	Treatment cycle reconstructed per risk, traceable to commits

Everything generated and signed locally. Apache-2.0. No data sent to external servers.

The free tier does NOT include

• The Art. 9 presumption (EU AI Act) provided by prEN 18228 §6-7 — that is the paid tier.
• Sectorial catalogues (DORA for financial services, MDR for medical devices, prEN 18283 for bias, prEN 18282 for AI cybersecurity).
• The assembled Annex IV (Technical Documentation Art. 11) — the cloud plane assembles it from the signed bundle.

The free tier is the honest entry point: real evidence, audited process, zero cost. The paid tier adds the regulatory layer that turns that evidence into a presumption before the regulator.

---

The free → paid value ladder

The free tier produces a real self-check against ISO/IEC 23894:2023: the AI risk management process. It is signed, auditable evidence, useful for internal due diligence and for showing an auditor that the process exists.

When you need to prove conformance to a regulator, the paid tier adds:

• The presumption of conformity under Art. 9 of the EU AI Act, which comes from aligning the system with prEN 18228 (the harmonised European standard in preparation). The verdict moves from “good practice” to “presumption before the regulator”.
• Sectorial catalogues: DORA (financial services and digital resilience), MDR (medical devices Reg. EU 2017/745), bias control (prEN 18283), and AI cybersecurity (prEN 18282). Each catalogue maps your controls to the specific obligations of the sector.
• The Annex IV assembled in the cloud: Technical Documentation Art. 11 generated automatically from the signed evidence, in web and PDF format, ready to attach to the declaration of conformity.

The pricing unit is the governed AI system, not engine usage. The engine (froga) is Apache-2.0 forever. Which standard lands in which plan (Starter, Pro, Enterprise), the reference prices and how the assisted-sale upgrade works: Plans & standards.

---

Next steps

InstallationInstall froga from the official channel (Linux, macOS, Windows).

Level 1: your first risk loop, no lawFull two-role workshop on the Iris dataset: a complete ISO 23894 risk loop on a system no law touches — because it's good engineering.

ISO 23894 — AI risk managementThe iterative §6.4–§6.7 process the engine implements and why it is the right starting point.

prEN 18228 — the harmonised standardWhy prEN 18228 bridges ISO 23894 and the Art. 9 presumption.

View the Annex IVHow the cloud plane assembles Technical Documentation Art. 11 from the signed bundle.