Delegation Lab

A flight simulator for AI governance.

Delegation Lab teaches the Verkflöde Agent Operating Model (VAOM) by play. You are the new Head of AI Operations. The board has approved an AI agent for a live workflow, and before it acts you write its delegation matrix: which decisions it may take, in which band, under which limits, and who answers for each. Then you run the workflow and live with the consequences, including a night shift you sleep through.

Four rounds. About ten minutes. One delegation design, tested at 2 a.m.

Play it at lab.vaom.ai.

Every decision is the Guardian's

The game doesn't simulate a policy engine. Every call the agent makes is decided by the VAOM Guardian, a reference implementation of VAOM 5.0 as an OWASP Agent Control Standard (ACS) Guardian, running in your browser against the matrix you wrote:

• your matrix is a real vaom.matrix/0.2 document, and the Guardian's linter decides whether it runs;
• each tool call goes through the Guardian's decide(), and the screen shows the answer an agent host would receive on the ACS wire;
• review, seeded errors, sub-agent chains, boundary probes and circuit breakers are the Guardian's own mechanisms, and the final scorecard is the Guardian's scorecard() over every decision record;
• at the end you can save the files a Guardian would enforce (matrix, binding, scoring policies, seeds and a config), checked the way the Guardian checks them on startup, and verify them with the Guardian's CLI.

The game is honest about strength. A Guardian refusal is policy: it holds because the Guardian holds. Only a capability the agent doesn't have is stronger, and a boundary counts at the strength a probe last showed it to have. The night round is about telling the two apart.

The rounds

Round		What it teaches
1	The Matrix	Decision inventory and authority decomposition: the workflow hides decisions inside the obvious one, and what the matrix doesn't name, the Guardian can't place.
2	The Gate	Calibrated thresholds (the threshold for Band A is the output of a calibration test at your error target, not a setting), value ceilings, cumulative caps and independent checks.
3	The Reviewer	Funded and effective oversight. You work the Band B queue yourself against a clock; planted cases with known errors test whether review catches anything, where the matrix allows them.
4	The 2 A.M. Test	Sub-agent chains and attenuation, boundary verification, and circuit breakers, on a night with an injected instruction and a tool that slipped into the agent's manifest.

The scenarios

Each scenario is one of the VAOM 5.0 whitepaper's worked examples, built on the Guardian's published example matrix, binding and scoring policies for it.

Scenario	Worked example	What makes it different
Invoice processing	7	Accounts payable: a value ceiling and a per-vendor cap, with a duplicate flag and a contract change hiding inside approval.
Incident remediation	7D	A SaaS provider's agent restarting services. A restart carries no amount, so its bounds are restrictions, a pre-flight check and a cap on restarts per service.
AML triage	7B	A regulated decision. The reporting threshold and the overrides for politically exposed persons and sanctions are the law's, not yours, and planted cases are forbidden because the records feed regulatory reporting.
Delegated purchasing	7E	Re-ordering supplies under a monthly mandate: a per-order ceiling, the mandate as a cumulative cap, and a supplier decision hiding inside a better price.

VAOM's voice

VAOM narrates the game: calm, precise, slightly dry. A language model writes the round briefings, from a fixed prompt per scenario and round; nothing about the player or their game is sent to it. The debriefs and the final profile are scripted from what the Guardian decided, because a teaching game can't afford a model that misstates a governance concept. Without a model key the briefings are scripted too, and the game is fully playable.

The live Lab uses Mistral through its EU endpoint. Voice narration exists in the code but is switched off until it is ready.

Running locally

Requires Node.js 22.

git clone https://github.com/verkflode/delegation-lab.git
cd delegation-lab
npm install
npm run dev

Open http://localhost:3000. The game works with no configuration.

Configuration

Copy .env.example to .env.local and set what you need. Everything is optional.

Variable	What it does
LLM_PROVIDER	anthropic (default) or mistral: which model writes the briefings.
ANTHROPIC_API_KEY, ANTHROPIC_MODEL	Key and model for Anthropic.
MISTRAL_API_KEY, MISTRAL_MODEL, MISTRAL_BASE_URL	Key, model and endpoint for Mistral (for example https://api.eu.mistral.ai).
CLAUDE_RATE_PER_MINUTE, CLAUDE_RATE_PER_HOUR, CLAUDE_RATE_GLOBAL_PER_HOUR	Rate limits on the briefing endpoint, per IP and overall.
NEXT_PUBLIC_VOICE	1 at build time shows the experimental voice narration.

/api/claude serves exactly one request, a round briefing (lib/v5/briefing-request.ts). It keeps its old name because the client calls it; the model provider is a setting.

Tests

npm test

The heart of it is the golden replay (tests/golden.test.mts): three scripted games per scenario (naive, careful and mixed), with every output the engine gives the screens pinned in tests/golden/<scenario id>.json. A change to the engine that changes nothing a player sees leaves those files byte-identical. After a change that is meant to alter the game, rewrite them with GOLDEN=update npm test and read the diff.

How it's built

Next.js 16, React 19, TypeScript and Tailwind CSS 4.

Path	What's there
lib/v5/round*.ts	The four rounds: VAOM's mechanics on the Guardian core, the same for every scenario.
lib/v5/scenario.ts	What a scenario gives the rounds.
lib/v5/scenarios/	One file per scenario (its decisions, batches, planted cases, night script and words), and the registry.
data/v5/	The Guardian's published example matrices, bindings and scoring policies the scenarios are built on.
vendor/vaom-guardian/	The Guardian's pure core and its ACS translation, vendored at a pinned commit (scripts/vendor-guardian.mjs, see VENDORED.md).
components/v5/	The round screens.
app/	Pages, the briefing endpoint and the privacy page.

The live Lab runs as a container (Dockerfile, Next.js standalone output) on a server in France, behind Cloudflare. It counts visits with self-hosted, cookieless analytics: see the privacy page.

Adding a scenario

A scenario is a file in lib/v5/scenarios/ plus its example documents in data/v5/, listed in lib/v5/scenarios/index.ts. Write round 1 first and run it, naive and careful, before writing any words: a naive design should let something risky through, and a careful one should not. Tune the calibration sample so different error targets give different thresholds, check that planted cases reach the reviewer's queue, add three games to the golden replay, and play it through.

The framework

The Verkflöde Agent Operating Model (VAOM) is an open framework for structuring AI decision authority in enterprise workflows: it separates statistical confidence from institutional authority. The whitepaper is at vaom.ai, published by Verkflöde AB under CC BY 4.0.

Licence

Apache-2.0: see LICENSE and NOTICE. The vendored VAOM Guardian code in vendor/vaom-guardian/ and the example documents in data/v5/ come from the VAOM Guardian, also Apache-2.0. The VAOM framework the game teaches is CC BY 4.0.