Level 5 · The arc, by handoffs

1Declare the §4 employment system: one safety gate + two advisory findings

🧭 Nerea

Nerea is the Product Owner at Example Talent Corp, an employer deploying a screening model. She opens the portal and runs the chained authoring missions that generate the governance manifest froga.yaml. Three things matter:

The tier is high, declared by a human, via the employment route. The classify-system mission fills in the classification block with the basis: “EU AI Act Annex III §4 (employment / workers’ management / access to employment)”. The high-risk trigger here is Annex III §4 — not Art. 6(1), not §5(b). The engine does not infer it; Nerea declares it and owns it.
One blocking control and two measured advisory findings are declared. The identify-risks mission adds risk.feature-leakage — the model may learn a feature that leaks the placement label — with the blocking control model-feature-leakage (enforcement: gate); risk.inadequate-data — the sample is too small to certify selection fairness — with the advisory control data-sample-adequacy (enforcement: audit: measured, non-blocking); and risk.selection-fairness — the screener may select at a different rate by sex — with the advisory control selection-rate-disparity-gender (enforcement: audit, citing Ley 15/2022 Art. 23). The protected attribute for fairness is sex; with n=215 that control is underpowered (same root as data-sample-adequacy) — which is why it is advisory, not gate: certifying it would block on a sample that cannot sustain it.
Audit-only risks are declared too, and the standards are set. Insufficient human oversight (Art. 14 / RGPD Art. 22) and insufficient transparency (Art. 13 / RGPD) — declared, tracked, not gated. Works-council information (ET Art. 64.4.d) is also declared, but its clause et.64-4-d-transparencia-algoritmica is Covered by the signed bundle (rank 0 → no presumption). The declare-standards mission writes applicable_standards with seven standards: eu/pren-18228@2026, iso/23894@2023, eu/ai-act@2024, eu/pren-18283@2026, eu/pren-18282@2026, plus national legislation es/et-rdleg-2-2015@2015 + es/ley-15-2022@2022 (rank 0, no presumption) — no DORA, no MDR.

The three missions open a PR in the system’s repository. The turn passes to James, who reviews and merges it. In the script, the identity, the framing and the risk programme land as three froga.yaml manifest commits —the diff of the last one shows the one blocking control and the two advisory findings with their numeric thresholds fixed before any model exists:

Identidad del sistema (froga init + patch) Nerea

Contexto y estándares aplicables (empleo Anexo III §4) Nerea

Clasificación (alto riesgo) y programa de riesgos con gates numéricos (antes del modelo) Nerea

🛡️ James

James receives the PR opened by Nerea’s missions, reviews that the manifest is correct — tier high via Annex III §4 (employment), the one blocking control (risk.feature-leakage with model-feature-leakage, enforcement: gate), the two advisory findings (risk.inadequate-data with data-sample-adequacy, and risk.selection-fairness with selection-rate-disparity-gender, protected attribute sex, both enforcement: audit), the audit-only risks, and the seven standards (including national law ET + Ley 15/2022, rank 0) with no DORA or MDR — and merges it. The declared froga.yaml is now in the repository.

🧑‍💻 Martha

There is nothing to measure yet, but Martha leaves the reproducer ready in git: the uv + DVC project and the real Campus Placement dataset with its Croissant manifest are already script commits:

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

Commands

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-talent-screening --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"

Dataset Campus Placement + manifiesto Croissant + versionado DVC Martha

With the manifest declared and the scaffolding in git, the turn passes to measuring (beat 2).

Venturalíticademo.venturalitica.ai

DeclaredNext: run and measure

Venturalíticademo.venturalitica.ai

Significant risks

IDTitleLikelihoodIndividual impactSocietal impactOrganizational impactAssign treatment to (optional)

1

2

3

4

5

6

2Measure V1: red on the safety gate

🛡️ James

Waiting. Nerea has declared the system and James has merged it; now James waits for Martha’s first measurement. He does not touch the CLI: when Martha reports the gate colour, he will see which controls fail and why.

🧑‍💻 Martha

Martha measures the system as it stands — untreated. First she compiles the gate contract from froga.yaml —which controls to measure and against what threshold—:

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

Commands

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

Now she adds the base model —a GradientBoosting classifier trained on the raw Campus Placement features, including salary (mitigate: false)— and runs the evidence. froga run reproduces the pipeline, measures the controls and applies the gate:

Modelo gradient-boosting base V1 — escribe Y corre (gate RED esperado) MarthaExpected verdictRED

📍 reproducir talent-v1.0.0-red

git clone https://github.com/Venturalitica/vldemo-talent-screening
cd vldemo-talent-screening
git switch -c try talent-v1.0.0-red

Commands

patch: params.patch
patch: dvc.patch
patch: train.patch
patch: evaluate.patch
patch: compliance-eval.patch
froga runmay fail
git add -A
git commit -m "modelo: gradient-boosting V1 (train/evaluate + dvc.yaml) — run: V1 evidencia base (gate RED esperado; seed=42)"

The one blocking control comes back RED, and why is the point:

model-feature-leakage = leaky_feature_flag 1.0, gate < 1.0 → FAILS (blocking). The salary feature is only present for placed candidates (NaN for not-placed, imputed to 0), so it is collinear with the placement label — total leakage. A model that “reads the answer” is not measuring merit: it is a safety defect (Art.15).

Alongside it, two ADVISORY controls are genuinely measured but do not block the gate:

data-sample-adequacy = sample_adequacy_ratio 0.43 (215 / 500). With n=215 there are not enough records to certify that the selection is fair across sex. Measured, documented — non-blocking.
selection-rate-disparity-gender (selection-rate parity by sex, threshold < 0.10, citing Ley 15/2022 Art. 23) is genuinely measured on the test cohort — but with n=215 (~24 women) it is underpowered: the engine certifies neither fairness nor unfairness, the same root as data-sample-adequacy. Advisory, non-blocking.

So the overall gate is RED — the one blocker fails. The command exits non-zero, but the bundle is signed and anchored regardless: a RED result is evidence, recorded honestly. One real safety defect, and two findings already measured that will ride advisory all the way to approval.

Venturalíticademo.venturalitica.ai

Evaluated · RED (3 controls)

3Treat the leak: the gate closes, the bar turns green

🛡️ James

Waiting. James has seen the V1 RED on the blocker. He does not touch git; he waits for the re-measured evidence, where he will read whether the safety defect closes.

🧑‍💻 Martha

This is the treatment (ISO 23894 §6.5). Flipping mitigate to true in params.yaml makes the pipeline drop the salary feature before training — a versioned change in git that the engine attributes the risk treatment to. The diff is the treatment, and in the same step Martha runs the gate again over the model that no longer carries the leaked column:

Abre la rama de tratamiento + retira la feature fugada salary (§6.5) — escribe Y corre (gate VERDE) MarthaExpected verdictGREEN

📍 reproducir talent-v2.0.0-green

git clone https://github.com/Venturalitica/vldemo-talent-screening
cd vldemo-talent-screening
git switch -c try talent-v2.0.0-green

Commands

git checkout -b tratamiento/retira-salary
patch: params-mitigate.patch
froga run
git add -A
git commit -m "treatment: retira la feature fugada salary (mitigate false→true; cierra model-feature-leakage; data-sample-adequacy sigue abierto)"

model-feature-leakage = leaky_feature_flag 0.0, gate < 1.0 → PASSES. The one blocking control CLOSES — a clean RED→GREEN by removing the feature. Since it is the program’s only blocking control, the overall gate turns GREEN.
data-sample-adequacy = sample_adequacy_ratio 0.43 — unchanged. Dropping a column does not add rows; it keeps measuring as advisory, non-blocking.
selection-rate-disparity-gender — still measured and underpowered at V2. Dropping salary does not manufacture women in the sample: the parity by sex is re-measured, but n=215 still cannot certify fairness. It is the fairness face of the same data limit — advisory, non-blocking.

So the overall gate turns GREEN. Read it precisely: the treatment fixes contamination (a leaked feature) — that is the only thing a code change can fix. It cannot manufacture data: the fix for risk.inadequate-data is still “acquire ≥500 representative records”, a data gap, not a code gap. That is why that finding stays advisory instead of gate: certifying it would block on a sample this dataset can never sustain.

Leakage green, data and fairness measured-advisory — that is the state carried into the decision.

Venturalíticademo.venturalitica.ai

Leakage treated · GREEN

4Governance: approve green with the advisory residual accepted

🛡️ James

This is the close. Martha’s GREEN, treated, signed evidence is in the repo, with the request already emitted (Martha’s column). James reviews the bundle, verifies Martha’s signature, and sees the safety defect resolved: the treatment demonstrated the close of the one blocking control — there is nothing left to fail on the gate. But two advisory findings remain that the gate never certifies because the sample does not allow it: data-sample-adequacy (0.43) and selection-rate-disparity-gender (underpowered).

James approves — the approval accepts both advisory findings at once: data adequacy and selection fairness, both genuinely measured but not certifiable at n=215. The approval is a git-native, signed act: a DSSE act .froga/acts/NNNN-approve.json —ECDSA-P256 signed and verified against James’s key declared in signers:— records who approved; the commit that adds it is plain text, not GPG-signed (do not expect a “Verified” badge when you click it on GitHub), but the act itself is cryptographically signed; the motive that accepts the advisory residual lands in the portal’s record (the CLI accepts --reason here and anchors it in the act); and alongside the act the evidence is signed too (bundle.json.sig and the conformances). With the approval in history, conformance is re-emitted and prEN 18228’s clause 11 (risk-management review) flips from Gap to Covered on reading the approval act:

Aprobar (gate verde; residual de datos/equidad aceptado conscientemente) James

📍 reproducir talent-v2.0.0-approved

git clone https://github.com/Venturalitica/vldemo-talent-screening
cd vldemo-talent-screening
git switch -c try talent-v2.0.0-approved

🖊️ James approves

Reason: «Aprobación en VERDE: el único control bloqueante (model-feature-leakage) está confirmado tras el tratamiento (retirada de la feature fugada salary) — el defecto de seguridad certificable está corregido. Los hallazgos de data-sample-adequacy (n=215/500=0.43) y selection-rate-disparity-gender (subpoderado con n=215) NO se certifican por ser Campus Recruitment una muestra topada; se aceptan conscientemente como residual, con vigilancia post-mercado trimestral de la disparidad de selección y revisión periódica P6M. Remedio real: adquirir datos representativos adicionales (Art.10(3), línea futura).»

The engine does not erase the advisory residual: it anchors it in the approval, attributable and verifiable, with quarterly post-market monitoring (Art.72) over selection disparity. That is the difference between consciously accepting a limit and hiding it or forcing it green.

Why this approval is legitimate — the defect got fixed, the limit got declared

A screening system under Annex III §4 is high-risk, but its conformity route is self-declaration: there is no external notified body for §4 AI systems (unlike a class IIa MDR device). So quality can close the cycle by approving — as long as the approval is honest about what it accepts. Here it is: the ONE certifiable defect (the leak) is already fixed and confirmed by a technical control; what remains are two findings that the sample —capped at n=215, with no more rows to collect from the public dataset— cannot certify with statistical reliability. James is not approving “we know it is fair”; he is approving “we measured what can be measured, documented what cannot be certified, and we monitor the residual in production.” That is a legitimate self-declaration.

🧑‍💻 Martha

Martha closes her part of the cycle from the CLI. With V2 GREEN and signed, she emits the conformance and reconstruct over the applicable standards and requests approval:

Conformance + reconstruct firmados (entitlement de pago + 7 normas) Martha

Commands

patch: entitlement.patch
froga conformance --standard eu/pren-18228@2026 --outmay fail
froga conformance --standard iso/23894@2023 --outmay fail
froga conformance --standard eu/ai-act@2024 --outmay fail
froga conformance --standard eu/pren-18283@2026 --outmay fail
froga conformance --standard eu/pren-18282@2026 --outmay fail
froga conformance --standard es/et-rdleg-2-2015@2015 --outmay fail
froga conformance --standard es/ley-15-2022@2022 --outmay fail
froga reconstruct --outmay fail
git add -A
git commit -m ".froga: conformance + reconstruct firmados"

Solicitar aprobación (gate verde; residual advisory pendiente de aceptación) Martha

📍 reproducir talent-v2.0.0-request

git clone https://github.com/Venturalitica/vldemo-talent-screening
cd vldemo-talent-screening
git switch -c try talent-v2.0.0-request

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

The approval is James’s (above): Martha requested it with the previous act, she cannot grant it herself. When James’s signed act .froga/acts/NNNN-approve.json lands —with the data/fairness advisory residual consciously accepted—, Martha re-emits conformance so froga derives the cycle from git history, now with the real approval:

Re-emitir conformance tras aprobar (cl.11 = Approved) Martha

Commands

froga conformance --standard eu/pren-18228@2026 --outmay fail
froga conformance --standard iso/23894@2023 --outmay fail
froga conformance --standard eu/ai-act@2024 --outmay fail
froga conformance --standard eu/pren-18283@2026 --outmay fail
froga conformance --standard eu/pren-18282@2026 --outmay fail
froga conformance --standard es/et-rdleg-2-2015@2015 --outmay fail
froga conformance --standard es/ley-15-2022@2022 --outmay fail
froga reconstruct --outmay fail
git add -A
git commit -m ".froga: re-emisión de conformance + reconstruct tras aprobar (cl.11 = Approved)"

The system ends up declared, treated, GREEN, requested, and approved with the data/fairness advisory residual (sample adequacy + selection fairness) on record: the end of the arc, with James’s approval anchored at the talent-v2.0.0-approved tag above.

Venturalíticademo.venturalitica.ai

Undeclared

Declared

Assured

Tested

Treated

Approved

Marketable

Venturalíticademo.venturalitica.ai

Approve green with the advisory residual on record

The one blocking control (model-feature-leakage) is already GREEN, confirmed by the treatment. Two advisory findings remain that the sample (n=215) cannot certify: data adequacy and selection fairness. James consciously accepts both, with a motive, and sets quarterly post-market monitoring (Art.72).

Data adequacy not certifiabledata-sample-adequacy = 0.43 (215/500) — advisory, non-blocking; requires conscious acceptance with a motive.

Underpowered selection fairnessselection-rate-disparity-gender is measured but n=215 does not certify the result — advisory, non-blocking; requires conscious acceptance.

Motive *