AI Atlas · Optional prototype lab · 1 / 3 · Perceptron

What did one example teach?

Predict two outputs. Separate a correct prediction from reaching the task’s target.

Allow 8–12 extra minutes per station: an unmeasured estimate. These labs are outside the three-room quest and E0–E7 protocol. Choose Free exploration; pause an active quest. Use a separate browser profile for fresh observations; preserve saved records.

Predict, test, explain

1. Open Perceptron. For this worked case, begin with inputs (1,1), weights (0.2,0.2). Predict, run, apply the error, then predict and retest. Choose Test these weights.
2. Predict A only (1,0) and B only (0,1) before Lock predictions & reveal both. Which reach target 1? What does a zero input contribute? How far does the evidence extend? Save your numerical reason; revise if needed.
3. Choose Try different weights. Predict both fresh outputs before reveal; explain which changed contribution changes your claim. Training weights stay saved.

Write or explain aloud; a scribe records exact words. Mark access support separately from content hints: support ______ hints ______. Keep each learner’s first response before discussion.

Cover until responses are recorded

Worked answers. Training changes both weights to 0.4. Frozen A/B scores: 0.4 − 0.6 = −0.2 each → 0 / 0. Neither reaches target 1. Authored weights 0.7/0.2 give 0.1 → 1 / −0.4 → 0: A only. Both claims cover these checked inputs with these weights.

Wrong-answer prompt: “One training success means every input works”—compare the unseen single-input sums. “I predicted 0 correctly” does not mean target 1 was reached.

Manual explanation: 0 unsupported/no explanation; 1 cites a relevant sum but misses scope or changed case; 2 connects both circuits’ sums to the threshold and limits the claim. Browser choices are checked; prose is not graded.

AI Atlas · Optional prototype lab · 2 / 3 · Shakey

A clear move is not a plan.

Build instructions. Predict their stopping point. Revise when the map changes.

Unmeasured 8–12-minute estimate; outside the core quest/E0–E7. Choose Free exploration or pause the quest. Use fresh map 1 in a separate profile; preserve records.

Build, test, revise

1. Open Shakey. Start B1, goal B5; B3 is occupied. Choose Build my own route and add Right. Predict an outcome, give your reason, commit and step.
2. Revise; Clear sequence. Build Up, Right, Right, Right, Right, Down. Predict before running. After success, choose Try the changed room. Keep the sequence and predict again from the new map.
3. Inspect the failed instruction. Revise your sequence and explanation to reach B5. Optional: build Left from B1; predict whether the map wraps. Up to 12 instructions; Undo and Clear edit uncommitted plans.

Oral answers count: a scribe records exact words. Separate access support ______ from content hints ______; keep first responses before discussion. No speech recording.

Cover until responses are recorded

Worked answers. Right ends at B2: instructions finish before the goal. The upper route reaches B5 in six moves. Map 2 adds A3: the same sequence’s instruction 3 requests A2→A3; Shakey stays at A2. Down, Right ×4, Up succeeds through C1–C5. Left from B1 stops at the map edge without moving.

Wrong-answer prompt: “The old route worked”—name its first newly occupied square. “One clear move reaches the goal”—compare B2 with B5. A boundary is not an obstacle, and it does not wrap.

Mechanism/boundary: Which requested move failed, and why stay put? This authored map does not test sensing an unfamiliar room. Manual explanation: 0 unsupported; 1 names a stop without explaining transfer; 2 cites the failed instruction, changed obstacle and valid revision. Completion alone is not mastery.

AI Atlas · Optional prototype lab · 3 / 3 · Turing

Predict the answer. And the work.

Optional 8–12-minute estimate, unmeasured; outside the core quest/E0–E7. Choose Free exploration or pause the quest. Keep this answer key covered until each first response is recorded.

1. Open Turing. Reset tape · keep record; enter 011. Predict final value and instruction count, then commit. One instruction reads, writes, then moves or stops. Step or run; compare both predictions with the trace.
2. Choose the stopping rule and supporting trace row; add written or oral/scribed words. Check, revise and preserve your first response. Try the changed input 110; predict and explain afresh.
3. Use Reset tape · keep record, enter 111 and test overflow. Explain why visible 000 is not the full answer. Inspection and pause do not add committed instructions.

A scribe may record exact spoken words. Mark access support ______ separately from content hints ______. No audio capture; prose is saved, not automatically graded.

Cover until responses are recorded

Worked explanation: trailing 1s become 0s while carry moves left; the first 0 becomes 1 and stops it. Cite the last trace row. For 111, row 3 reaches the finite boundary with carry 8. This authored three-cell program is not Turing’s historical program.

Wrong-answer prompt: “Always three instructions”—compare 011 with 110. “000 means zero”—account for the outside carry. Manual explanation: 0 unsupported; 1 cites a correct row but misses carry or transfer; 2 connects the stopping rule, row and changed count, distinguishing overflow from zero. Choice feedback is not prose grading.