LabaliciousAI & QA ACADEMY
← All sessions

Open in Markdown reader

Instructor Guide

Start with Session 00. The program now has 13 sessions and 26 contact hours. All core labs use fictional data and the included local app.

Staffing and room design

Plan for one lead instructor and one coach for every 12–16 learners. Use tables of two or three, a projected screen, a shared Discord channel or equivalent, and one local simulator copy per learner (no sandbox login). Make each activity accessible without prior coding experience: teams can produce evidence and test designs before they edit automation. Offer an advanced extension in every lab rather than moving faster through the core exercise.

Before the first cohort

For the included local track, use LAB_SETUP.md and the supplied fixtures. The provisioning steps below describe an optional future production-like training environment, not a requirement for running this package.

  1. Provision a training-only CRM, fake identity provider, and synthetic tenant data. Include a handful of intentional, documented defects.
  2. Create learner accounts with least privilege. Use named QA roles, never a shared admin account.
  3. Fork a small training repository with a passing baseline and tagged buggy branches. Include the templates in this package.
  4. Establish one channel for questions and one channel for findings. Teach the difference: a question is not a defect; a defect needs reproduction and evidence.
  5. Give students the course safety card before tool access.

Course completion and project-based certification

Product Studio is the individual course assessment. Introduce it in Session 01; approve each pitch before Session 02 kickoff. Each learner builds in their own public repo, with an equivalent private route arranged when necessary. Practice lab packets are formative coaching, not scored examinations or a points-based graduation gate.

Review evidence and judgment over time: problem framing, test design, reproducibility, AI use, communication, GitHub practice and product usability. Use the project review checklist with ready for next step / revise and return / instructor support needed. Match expectations to the approved scope. Fabricated evidence, unsafe handling and unverified claims need correction, not compensating points. Software regression tests remain required engineering work; there is no separate test of the learner.

Use the two-award policy. Jared records course complete / course activities outstanding separately from AIQAA certified / not yet demonstrated / review pending. Completion recognizes finished learning activities and presentation, not an approved product-quality verdict. Certification requires course completion plus all AIQAA-1.0 areas demonstrated through product quality, real development history, GitHub understanding and the other course skills.

He authorizes each award separately. No automatic issuer, numeric threshold, hidden defense or claim of external accreditation. A learner may receive the completion certificate while addressing certification gaps. Offer specific feedback, equivalents and follow-up without repeating already demonstrated work.

Project review operations

At each session, use the published lab’s session guide to orient learners. Fit its optional recall prompt into arrival/discussion and its next-step reflection into closing/debrief; retain the 120-minute agenda. Learner support includes a safe help-request pattern, catch-up route and accommodations guidance. Practice feedback is descriptive coaching, not points or an additional exam. Have learners apply the correction to their existing artifact rather than fill another form.

Learners can prepare a focused Markdown note from the dashboard’s Prepare a project review link. This is local preparation, not intake, an AI grade or an instructor workspace. Review the original artifacts and exact commits; the helper does not fetch or verify them. Keep decisions in approved private storage. The Shift Garden example illustrates boundary evidence with runnable negative/corrected snapshots; its staged conversation and snapshots do not establish learner history or certification eligibility.

Use the operational rubric and project assessment agreement. Approve scope and competency mappings at Session 02; finish concrete acceptance cases in Session 04 before assessing completed features. Do not invent final-day requirements or waive a core criterion. Freeze the version and record equivalent evidence routes.

Before first issuance, calibrate with the fictional examples, double-review the first three projects and disputes, then sample at least one in five by the preselected review queue order. Preserve independent initial judgments and reconcile disagreements using actual evidence. This is a pilot standard; measure reviewer agreement and time rather than claiming reliability in advance. Schedule detailed reviews outside the eight-minute demos. Reuse project artifacts to keep the evidence burden proportionate.

Record all 21 rows using the private worksheet. Every required condition must be supported; a missing observation is unverified. The optional local consistency checker aggregates human findings, cannot validate their truth and cannot issue awards. Keep it out of student-facing CI and public records. AI assistance is not an independent second reviewer.

Before the cohort, privately establish a submission/contact channel, learner aliases, review appointments and an approved model route. Preflight Qwen3.8-27B using the small-model guide; record actual compatibility, account eligibility and spending ceilings. Do not require hardware purchases or a premium subscription.

Copy the progress register, two-decision review, completion certificate and AIQAA certificate to approved private storage. Never commit filled records. Record evidence, SHAs and follow-up, then each award's decision and issuance date separately. AIQAA records also identify the criteria version and reviewed release/commit. Confirm the working certification title before issuance.

Use the learner's privately supplied certificate name and deliver privately. Public sharing or promotional use requires a separate choice. A completion-only certificate must not say Labalicious Certified AIQAA.

The repository supplies templates, not a credential-issuing service or hosted student records system. Do not award a certificate simply because CI or packet-safety checks are green.

Instructor response to AI mistakes

When a learner’s model produces a plausible but unsupported answer, do not simply say it hallucinated. Ask:

  1. What exact claim did it make?
  2. What source, application state, or deterministic check would prove it?
  3. What should the model have been given to make the task safer?
  4. Does this become a new test, a prompt/playbook improvement, or a reason to keep the work human-reviewed?

This turns failure into evaluation literacy.

Model-change playbook

The program should evolve on a monthly review and on every material model/tool change.

  1. Follow fixtures/model-evaluation-protocol.md: give actors input-only cases and keep reviewer labels outside their access. Classroom mocks are not adoption evidence.
  2. Compare task completion, false pass, false fail, safety-policy compliance, evidence completeness, latency, and cost.
  3. Review a representative set of outputs blind, with instructors judging the evidence rather than knowing the model.
  4. Approve, restrict, or reject the candidate for each task class separately: drafting, code exploration, browser exploration, spreadsheet analysis, and verdict review are different jobs.
  5. Version the model card and affected lesson playbook. Teach the observed behavioral change in the next cohort.

Never let a provider’s benchmark, a model’s self-description, or a demo replace this local evaluation.

Rendering decks

The deck files use Marp Markdown. The exported PowerPoint slides are image-based; edit the Markdown source and regenerate to change slide content. The browser HTML is also presentable offline. In a prepared teaching environment, install the approved Marp tooling once and render, for example:

npm install
npm run slides:html
npm run slides:pdf
npm run slides:pptx

Keep generated PDF/PPTX files out of the source repository unless a particular cohort needs a frozen release. The Markdown is the version-controlled source of truth.