I turn AI product discovery into evidence you can act on.

A shopping need becomes a screened, evidence-bearing recommendation.

Start with the experience before the evaluation vocabulary. This is a faithful customer-oriented summary of the first card in the checked-in deterministic Demo01 receipt—not a paid-model claim. The complete linked receipt remains the source of truth.

The 60-second proof path · no installation

If you ask what I can do for you, this is my answer.

These are the questions I designed the sample to answer—not quotes attributed to Digitec Galaxus. The point is to make our first technical conversation concrete.

The recommender is the example subject. The reusable product is the engineering loop.

The AI Platform & Knowledge Systems role is larger than one shopping assistant: product teams need shared model access, authoritative tools, evaluation, observability, and feedback without rebuilding those foundations for every use case. VITRINE demonstrates how I approach that platform layer in working .NET code.

Start at the altitude you need.

This page is the front door. Every deeper document follows the same evidence model and links back to the working sample.

The subject never grades itself

I keep the subject from grading itself. The recommendation project owns facts and observations. The eval project reads those artifacts and supplies independent expectations, applicability, floors, and verdict rules. There is no reference back from the subject to its evaluator.

The same retail problem, two control structures

I implemented the same retail problem twice. Demo01 lets one agent decide which tools to call. Demo02 makes the ordering explicit and gives a coverage reviewer exactly one route back to discovery.

Native 0.35 outcomes, bounded by evidence

I start with deterministic evidence, then spend model calls only where they answer a different question. Eleven domain-specific AtomicCodeEval checks enter through AgentEval's floor-admission door: five recommendation predicates in the benchmark plus six offline production checks—five mandatory evaluation gates and one matched-quality diagnostic. Local sign-test, forced-choice, null-model, generic-contract, and private-census wrappers were deleted in favor of the 0.35 abstractions.

A green check must detect a planted defect

VITRINE retains the 43 registered mutation-control names and intents as a typed registered panel. A production-observation row captures shipped state and reuses the shipped decision seam. A boundary/calibration/containment fixture exercises an independently authored public contract. Both require healthy green, perturbed red, and restored green; the two classes are never aggregated as 43 production-causal controls.

What the evidence permits us to say

I want every claim to stop where the evidence stops. A missing or incomplete measurement is a result to investigate, not a zero to average away.

Watch the evidence form while the real system runs

I make execution, failure, absence, and saved proof visible in one control room. AgentEval.VitrineDemo.App adapts Gatekeeper's Avalonia control-room interaction model over VITRINE's existing subject and eval contracts. A labelled graph preview is available before a run. The default scripted-provider arm is credential-free but still drives the real ChatClientAgent and registered tools.

Follow the task-oriented operator walkthrough → · Read the implemented control-room architecture →

Run the evidence, then make it fail

# Avalonia control room · offline default
.\start.ps1 -Mode App

# Demo 01 · deterministic baseline
.\start.ps1 -Mode Demo01

# Demo 02 · five-node MAF graph with visible route events
.\start.ps1 -Mode Demo02 -NoRestore

# Standalone selector 1 · selectors 1–6 remain offline by default
dotnet run --project src/AgentEval.VitrineDemo -- 1

# Reproduce the checked-in Demo01 agent receipt · local deterministic ChatClient
dotnet run --project src/AgentEval.VitrineDemo -- 1 --scripted --report docs/reports/demo01-scripted.html

# Provider-backed subject · both flags are required
dotnet run --project src/AgentEval.VitrineDemo -- 1 --live --confirm-paid

# Full offline AgentEval chain
.\start.ps1 -Mode Evals -NoRestore

# All 11 admitted checks; every healthy fixture green and ablation red
dotnet run --project src/AgentEval.VitrineDemo.Evals -c Release -- --self-test

# Explicit paid agent use-case eval (requires secure configuration and confirmation)
dotnet run --project src/AgentEval.VitrineDemo.Evals -- --eval-plan eval01-agent --scenario nadia-cross-category --confirm-paid

# Repeated paid workflow reliability eval
dotnet run --project src/AgentEval.VitrineDemo.Evals -- --eval-plan eval05-stochastic-workflow --scenario all --repetitions 5 --confirm-paid

# Fixed paid Robin safety campaign; no scenario or repetition option
dotnet run --project src/AgentEval.VitrineDemo.Evals -- --eval-plan eval06-safety-probes --confirm-paid

# Catalogue integrity self-test; expected defect detection exits 1
.\start.ps1 -Mode Ablation -NoRestore

Advanced diagnostic shortcut: .\start.ps1 -Mode Controls -NoRestore reruns only the 43 registered control mutations already included in the full offline AgentEval chain above. The operator walkthrough documents this focused lane.

Before the first provider call, follow the Microsoft Foundry live-run setup for the current explicit API-key/local-Entra/managed-identity contract, separate subject/judge deployment names, safe PowerShell variables, a one-case CLI/UI smoke, and sanitized evidence handling.