Lab 1 — The AI Claim Detective

Time: 35 minutes
Team: two learners; driver and verifier switch after 17 minutes

Quick start

First action: Open the supplied claim packet, then write one row for each report.

Finish with: Three classified claims, one evidence packet and one limit sentence.

Step-by-step help for Session 01 — smaller steps, checkpoints and troubleshooting. Saving and asking for help.

Mac / Linux / Windows

Submission: prepare the deliverables below using the public-safe packet and review workflow. One PR per lab; private review is equally available. Never publish grades or personal information.

Use the same synthetic claim packet on Mac, Linux or Windows. Follow the platform guide for screenshots and browser developer tools; record OS/browser, viewport and zoom so evidence is comparable.

See the platform guide and included local lab track.

Scenario

An AI assistant has reviewed a fictional Community Launch allocation page and produced three claims. The product owner wants to know whether to hold a release.

Mission

Classify each claim as proven, plausible but unproven, or unsupported. Produce one evidence packet for the highest-risk proven finding, or a supported no-defect conclusion if none is proven.

Supplied evidence and your capture

• Read claim-packet.md: E01–E05 contain the requirement, browser observation, fictional network excerpt, explicit absence of database evidence and release note, followed by three claims.
• Capture your own screenshot of the analogous local allocation failure. Label it with the local candidate, not the packet's fictional build ID.
• A database record and console export are not supplied. Mark them unavailable; do not manufacture them or treat the local screenshot as evidence about the fictional B27 deployment.

Rules

• Use fictional data only, including when using a local model. No real passwords, API keys, customer records or production credentials belong in prompts or submissions. Keep access scoped to the training project with normal permission checks.
• Do not infer a backend result from the page’s appearance.
• Do not copy customer-like data into a model. Use only the supplied synthetic artifacts.
• The AI statement itself cannot be evidence.

Deliverables

1. A three-row claim table: one row per claim, with classification, source, and reason.
2. A completed evidence packet.
3. One sentence stating what your conclusion does not prove.

Success rubric

Use these descriptions for coaching, not a numeric grade or separate certification test. Choose the next evidence improvement with your reviewer.

Stretch

Ask an approved AI tool to rewrite the most overconfident claim as a calibrated finding. Review every word against the sources.

Product Studio homework — Pitch your product

Complete the proposal and one rough wireframe after class. Pick from the 20 ideas or propose your own; present the user, main journey, three core features, exclusions and budget to the instructor before Session 02.

Budget two to three hours, including evidence and review preparation; ask for coaching or scope adjustment if blocked. Record progress using the milestone template in your own product repository, not the practice packet. See the complete project roadmap, GitHub workflow and small-model guide. The practice lab above remains separate from your individual product.