🧑‍🏫 Guide for trainers & consultants

Mr. Sixx' AI Lab · professional edition (Green/Black Belt, “AI for Quality”) · July 2026

Positioning: what you are facilitating here

The AI Lab is not an AI lecture but an experience space: your participants walk through the complete ML workflow on a real model trained live in the browser – and every station mirrors a tool they know from their belt training: specification vs. reality (station 2), operational definition and sampling (3), validation (4), attribute MSA with kappa (5), DOE (6), rolled throughput yield and capability proof (7). Your core task as facilitator: stage the aha moments and, in the debriefs, build the bridge to the participants' daily business. Mr. Sixx runs the stations – you own the transfer.

Ideal curriculum placement: after MSA and DOE modules (maximum recognition value) or as a standalone “AI for Quality” module. Language switchable DE/EN (top right).

📋 Delivery formats

Format	Duration	Cut
Seminar module (GB)	90–120 min	All 7 stations compact, short debriefs; deep-dive drawers as homework
BB deep dive	120–150 min	Focus on station 5 (kappa discussion) and station 6 (DOE incl. interaction and noise debate)
“AI for Quality” workshop	½ day	Full lab + facilitated transfer session (see below) + action list for the participants' company
Keynote/exhibition demo	15–20 min	Station 2 only (rules failing) + arena with a pre-trained model – high wow factor

✅ Preparation

Material & technology

5–10 dice (white AND colored), matte dark surface, plus a bright/glossy surface for the staged “break”, optionally a 12-sided die for the finale
Facilitator laptop with a current browser on projector/display; participant smartphones (BYOD, no app)
Address: ailab.esixsigma.de · click through once beforehand and send a trial photo

License logistics (important!)

Stations 1–2 are free; from station 3 the code unlock applies per browser.
Standard seminar setup: one central, pre-unlocked facilitator laptop – one code covers the whole group.
Participants on their own machines (e.g. online seminars): one code per machine → plan accordingly.

⚠️ Most important technical note: Dataset, model and progress live in the desktop's browser. Run the whole session on the same device in the same browser – no incognito mode, don't clear browser data in between. Prepared states (e.g. for demos) survive for days this way.

For participants

Distribute the participant briefing beforehand – it prevents the most common input mistakes (label discipline!).

🗺️ Flow with debrief focus

Station	Time	Debrief core (your facilitation)
1 · Briefing	5 min	Open Y = f(X): what is the Y here, what are the X's? Leave the question open.
2 · Rule Workshop	15 min	“Which inspection instruction in your company breaks exactly like this?” – specification vs. reality, rule maintenance costs.
3 · Data Studio	15–20 min	Operational definition (what is a good label?), sample vs. population, bias as a business risk.
4 · Training Room	20 min	Vault = confirmation run; reading the gap; why metrics scatter (randomness in training) – parallel to measurement series.
5 · Inspection Lab	15 min	Confusion matrix = attribute MSA; kappa rules of thumb (≥0.9 / 0.7–0.9 / below); hold a vote: “would you hire this inspector?”
6 · Tuning Workshop	15–20 min	The OFAT trap, interaction A×C as the overfitting recipe, weighing effects against repetition scatter – your group's DOE home advantage.
7 · Arena + transfer	15–20 min	RTY along the chain (0.95¹⁰ ≈ 60 %), capability proof before serial deployment, LLM transfer as the closing.

🎬 Directing tips – creating the aha moments deliberately

Station 2 · Staging the failure

Task 2 (the 45° trick): have the die photographed at a clear tilt so that two or three side faces become visible – the rule machine counts their pips too (“8” for a five). The flatter, the more absurd.
Task 3: colored die on a glossy surface by the window – frequently the white pips themselves get detected as “dice”. Perfect segue to “every rule assumption is an attack surface”.

Station 4 · Using the dramaturgy

Collect predictions before training (“what percentage on the vault?”).
Show the epoch comparison 5 vs. 30–50 (small datasets need more epochs).
Provoke overfitting: scarce dataset (~10–12 per class) + 100 epochs → training curve ~100 %, vault stays behind. The textbook moment.

Stations 5/6 · The belt moments

Click red matrix cells and discuss the misclassified images – gemba instead of statistics slides.
Before the 2³ design, collect factor bets; afterwards ask the noise question: “82 % vs. 84 % – effect or noise?”

Station 7 · Closing big

10-dice logistics: dark surface, no overlapping, straight from above.
If available: photograph a 12-sided die – the scope of the “certificate” becomes a lived experience.
Don't skip the LLM transfer – it anchors everything in the participants' everyday life.

🎓 Your theory backstop: For detailed questions, the 📖 deep dives and 🧮 math drawers of the stations are your ready-made script (convolution, gradient descent with the Deming funnel analogy, kappa derivation, generalization gap, RTY table). Recommendation: play the lab through once yourself and read the drawers – then no question will surprise you.

💼 Transfer session for the company (for the ½-day workshop)

Which attribute inspections in your company would be candidates for a learned model – and which explicitly not (traceability obligations)?
What would the vault be there – and who guarantees it never leaks into training (data leakage)?
What would the capability proof before release look like – on whose data, with which kappa target?
Who monitors the model in operation? (Models drift like processes – apply SPC thinking to AI.)
Who owns the training data and who maintains it? (“Whoever has the data has the power” – organizationally too: involve works council and data protection early.)
Do the business case honestly: human for one part, model for a thousand in the same time – but including data, testing and monitoring effort.

🔒 Technology & privacy (your arguments towards IT/clients)

No accounts, no installation; photos are passed through encrypted (HTTPS) and not stored server-side – storage only in the facilitator machine's browser, deletable there at any time.
BYOD-friendly: participant phones only send dice photos; remind the group: no faces in the photos, the gallery is visible to everyone.

🛠️ If something goes wrong

Problem	Cause & fix
QR leads nowhere	The desktop must run via ailab.esixsigma.de (not locally); the target URL is shown in small print under every QR code.
Camera stays black	Permission denied → reload the page, tap “Allow” (or fix site permissions).
“Waiting for phone …”	Phone page closed/standby → rescan the QR, keep the page open.
Photo doesn't arrive (exam/arena)	A round is still open → “Score the round” + “Next round” first.
Model learns poorly (< 50 %)	Not a defect but teaching material: more epochs (30–50), check dataset/gallery, diagnostic trick: train on the synthetic booster only.
Station locked (🔑)	Code unlock missing in this browser → enter the code (purchase email) or use the prepared facilitator laptop.
Progress/model gone	Different machine/browser/incognito → everything lives in the original device's browser storage.

Facilitation principle: Almost every glitch is teaching material in disguise – bad light, tilted photos and meagre metrics are exactly the variance this lab builds its lessons from. When in doubt, let the group diagnose it themselves: read the curves, inspect the gallery, run a comparison experiment.