Quickstart

🟡 Partial — The e2e arc runs on the repository's fixtures (Python+DVC); the resulting Annex IV is still partial.

This quickstart runs the loan example bundled in the repository: a high-risk credit classifier with a fairness control (disparate impact) whose first version fails the gate and whose treatment — promoted as a commit — closes it. The arc materialises the ISO 23894 §6.4–§6.5 risk treatment loop with signed evidence (ECDSA-P256+DSSE+in-toto).

Requirements

sei built (Installation).
uv (Python ≥ 3.11) and dvc ≥ 3 installed and on the PATH.
The monorepo cloned locally.

Fastest path: run the integration test

The loan project files live in crates/seigarrena-cli/tests/resources/loan/. The integration harness (crates/seigarrena-cli/tests/scenarios.rs) sets up the project in a temporary directory, initialises the uv environment, and runs the full arc:

cargo test -p seigarrena-cli --test scenarios loan -- --nocapture

The test verifies all arc steps (red gate → treatment → green gate → conformance → cycle reconstruction) and fails if any invariant is not met. This is the canonical way to confirm that the installation works.

The arc step by step

The following steps are exactly what the test above runs; you can reproduce them manually on the crates/seigarrena-cli/tests/resources/loan/ directory in a temporary copy.

Compile the risk program to OSCAL
Terminal
```
sei compile
```
Converts the risk program — the risk: section of sei.yaml (5 risks, 16 measures, Art. 9) — into the OSCAL assessment plan (10 ex-ante controls). Writes the file declared in the oscal.assessment_plan field of sei.yaml.
Run V1 — red gate
Terminal
```
sei run
```
sei run invokes dvc repro (pipeline featurize → evaluate), measures the 10 ex-ante controls with venturalitica-sdk, and applies the risk gate. Produces .sei/bundle.json (signed) and sei.lock.

The blocking control unfair-credit-exclusion fails: the gender disparate impact measures above the 0.03 threshold (EU AI Act Art. 10; DORA Art. 6). sei run exits with a non-zero code. The bundle is still signed and anchored.
Explore the treatment candidate (without committing)
Terminal
```
cp train_mitigated.py train.py
dvc exp run
dvc exp show
```
dvc exp run executes V2 (ExponentiatedGradient + DemographicParity mitigation) as a DVC experiment without committing. dvc exp show compares the candidate metric against the V1 baseline and confirms that the disparate impact would close the gate.
Commit the treatment (V2)
Terminal
```
git add train.py
git commit -m "T1: treatment — promote V2 (mitigated train.py, candidate that closes the gate)"
```
The commit is the ISO 23894 §6.5 treatment: a versioned change to which sei reconstruct can attribute the FAIL→PASS arc. Only train.py changes; featurize.py, the data and the classification are reused (type-B drift, Art. 15).
Run V2 — green gate
Terminal
```
sei run
```
dvc repro detects that train.py (a dependency of the evaluate stage) changed, recomputes only that stage, and produces the new metrics.json and dvc.lock. The unfair-credit-exclusion control passes (disparate impact ≈ 0.015 < 0.03). sei run exits 0.
CRITICAL residual and underpowered control
Even though the fairness control passes, the bundle reports:
- CRITICAL residual: the opacity risk has no control covering it. The overall residual exceeds the declared appetite (prEN 18228 cl. 10).
- Underpowered control: the disparate impact confidence interval [~0.001, 0.072] crosses the 0.03 threshold at n=1 000 samples. The engine emits an advisory data_gap (EU AI Act Art. 10(5)). The result is consistent with the gate passing, but the statistical power is insufficient for a definitive conclusion.
These are honest signals from the system, not errors. The opacity risk requires an additional treatment before the residual is acceptable.
Project conformance onto the applicable standards
Terminal
```
sei conformance
```
The same signed bundle is projected onto the clause catalogues of prEN 18228 (priority) and ISO 23894 and emits one report per standard (COVERED / PARTIAL / GAP per clause).
Record management approval
Terminal
```
sei approve --by "Jane Roe <jane@org>"
```
Creates an empty git commit with a Sei-Approved-by: trailer. The act is timestamped and attributed; sei reconstruct picks it up as ISO 42001 §6.1.3 approval.
Reconstruct the treatment cycle
Terminal
```
sei reconstruct --out
```
Deterministic replay of the bundle’s git history. Produces .sei/reconstruct.json (signed): the complete per-risk cycle (①identification → ②analysis → ③evaluation → ④treatment → ⑤residual).
Commit the signed evidence
Terminal
```
git add .sei/
git commit -m "evidence: cycle T1 (signed bundle + conformance + reconstruction)"
```
Depositing the signed evidence in the repository is the act the cloud plane reads to assemble and render the Annex IV (Art. 11). The engine does not emit the Annex IV: the cloud builds it from the signed bundle.json.

There is no sei annex-iv command and no .sei/annex-iv.json artifact. The Annex IV is assembled and rendered (web and PDF) in the cloud plane from the signed evidence. §7 and §8 are never pending; §2/§3/§4/§9 can be marked PENDING if their input is missing. In this scenario, with data/controls/measures present, only §3 and §9 would appear pending. See View the Annex IV and Status & gaps.

Artifacts produced

Artifact	Produced by	Contents
`.sei/bundle.json` + `.sig`	`sei run`	Signed evidence (controls, treatments, risk analysis, dvc.lock)
`sei.lock`	`sei run`	Freshness anchor (pipeline phase hashes)
`.sei/conformance/*.json` + `.sig`	`sei conformance --out`	Per-standard conformance reports (prEN 18228, ISO 23894)
`.sei/reconstruct.json` + `.sig`	`sei reconstruct --out`	ISO 23894 treatment cycle reconstructed per risk

The Annex IV (Art. 11) is not an engine artifact: the cloud plane assembles and renders it from the signed bundle.json. See View the Annex IV.

Next steps

Your first compliant system Full walkthrough: create a project from scratch, declare risks, run the complete arc.

What is Risk-Driven Development The normative frame (ISO 23894, EU AI Act Art. 9) and the red/green analogy.

Status & gaps Source of truth on project maturity and the gaps towards a complete Annex IV.