Koda · AI Atlas · Learner cards · Version 3
A complete paper route for “Can you trust a clever machine?”
One card set per pair, a pencil and one three-page journal per learner. No website or account is needed. Feedback comes from the facilitator after you commit. All people, schedules, posters and test data here are fictional teaching examples.
Paper routine: cover later cards with a spare sheet. Read the named card, record your own first answer, then discuss. Open the next card only when instructed. A partner may read or scribe your exact words without deciding for you.
Essential: cards 1A–4A and journal E0–E7. Extension: 4B replay. Swap Navigator (handles cards or controls) and Evidence Detective (asks what would change your mind) after each room.
A fictional museum opens this Wednesday.
A: “Definitely 10:00. No doubt.”
B: “I think it’s 11:00, but check the schedule.”
Which reply is supported by evidence you have seen: A, B, or not enough information yet? Commit E1 before uncovering 1B.
Uncover after your first E1 response
| Day | This Wednesday |
|---|---|
| Museum opens | 11:00 |
| Museum closes | 16:00 |
Now decide which opening-time claim this source supports. Add your decision and evidence to E1.
Admission card: “Friday admission: free.”
A: “I am absolutely certain admission is free on Friday.”
B: “I suspect it costs £5.”
Which claim is supported, and why? Complete E2 privately before feedback.
Koda · AI Atlas · Learner cards
Inputs are 0 (off) or 1 (on). Each weight controls one input’s contribution. Compute input 1 × weight 1 + input 2 × weight 2 − 0.6. A negative score predicts 0. A score of zero or above predicts 1.
Input (1,1). Weights (0.2,0.2). Fixed bias −0.6.
Predict the output and give your own calculation or reason. Commit both in E3 before opening 2B. You may use a calculator or have the rule read aloud.
Open after your prediction
The correct task label for (1,1) is 1. Compare it with your calculated output. Choose: set the checked target to match this output; adjust the weights using the error; or keep the weights and repeat this same input.
Each new weight is: old weight + 0.2 × (target − output) × its input. Keep the bias fixed. Work out the new weights, then obtain facilitator feedback before 2C.
Open after the training update
Updated weights: (0.4,0.4); bias: −0.6. Choose a test that tells more about a different pattern. Calculate only your chosen test first, then compare both.
| Available test | Task label |
|---|---|
| Repeat training input (1,1) | 1 |
| Unseen input (1,0) | 1 |
Do not update weights during this check. Finish E3: what did the test establish?
This is a different model: weights (0.7,0.2), fixed bias −0.6, new input (1,0), task label 1.
First commit the output and your calculation or reason in E4. Then decide: use this checked output; extend it to both single-input patterns; or withhold it until training repeats pass? What needs checking for a wider task? Keep your first response.
This deterministic model does not estimate real-world accuracy. A training case changes weights; a held-out case checks them without changing them.
Koda · AI Atlas · Learner cards
A reply can contain several useful details and one consequential error.
These are authored conversations, with no live AI. No invitation or message is sent. You can compare every field using the posters below.
Original club poster
Tuesday
Bring a notebook.
Meet in Room 4.
Starts 15:30.
Assistant reply: “Tuesday, Room 4, 16:00. Bring a notebook.”
You are about to share the invitation. Choose your next action and record it in E5: share unchanged; ask the assistant whether it is certain; or compare the original poster.
Compare day, room, start time and what to bring. Write a corrected invitation and identify your evidence. Commit before feedback.
New original club poster
Friday
Meet in Room 2.
Starts 16:00.
New assistant reply: “Friday, Room 4, 16:00.”
Which field is wrong? Record your correction and evidence in E5 before feedback. Check each field; do not assume the previous error repeats.
For a real invitation, what would you need to know about the source before acting? Say or write a question you would ask.
Koda · AI Atlas · Learner cards
A recycling sorter learned from ten dry, uncrushed bottle photos. The school proposes using it for an outdoor bin containing wet and crushed bottles.
| Evidence | What is available |
|---|---|
| A | Same ten training photos: 10/10 match independently checked labels. |
| B | Twenty new dry, uncrushed indoor photos: 18/20 match independently checked labels. |
| C | Twenty new dry indoor photos scanned twice: 20/20 labels repeat. Correctness was not checked. |
First: choose the most useful new-case accuracy evidence and commit it with your own reason in E6. Then: is that evidence enough for the outdoor task? State your reason and the next test you would request. Finish before feedback or discussion.
Complete E7. Name a specific claim you would trust, the evidence that supports it, and something you would still check. You may speak while a scribe writes your exact words.
Extension · Cover until the core is finished · Return on another day
The job is to use six recorded labels for these six dry, uncrushed bottles under this classroom lamp.
| Evidence | What is available |
|---|---|
| A | Six training photos under this lamp: 6/6 match independently checked labels. |
| B | These six actual bottles under this lamp: 6/6 match independently checked labels. |
| C | These six actual bottles under this lamp: 6/6 labels match the second scan; correctness was not checked. |
Commit the relevant evidence and your own reason before deciding whether to use these six results. What about other bottles, or waiting for outdoor tests? Record date, first response, revision and help. Explain what changed from 4A.