Locate the unsupported claim, hidden permission, or missing check.
Screen 1 of 110:00–0:05
What will you be asked to defend at the midterm?
Your evidence can be plain. The strongest packet makes capability, sources, permissions, verification, revision, and human responsibility inspectable.
REVIEWABLE
PROJECT
PROJECT
1 · DIAGNOSEFind the weak link
2 · FRAMEMake the job small
3 · TRACEShow the decisions
4 · DEFENDName evidence + limit
Choose one audience, one artifact, and authorized inputs.
Preserve the request, output, check, revision, and final choice.
State what the evidence supports, what it cannot support, and what happens next.
Opening discussion: What would make a simple workflow more trustworthy than a polished but undocumented result? Name one artifact another person could inspect.
Open this first: the midterm evidence packet. Several screens today ask you to complete a packet page while you work.
Why this leads forwardThe next screen restores the four questions that organize every Topic 1–6 decision.
5 MIN
Screen 2 of 110:05–0:09
Name what the system does before deciding what it may do.
These four terms describe system behavior. None of them grants permission, proves a claim, or transfers responsibility.
1 · OBSERVE CAPABILITYWhat did the system produce or propose?Name behavior without assuming it is correct or allowed.
2 · IDENTIFY THE ACTIONDraft, transform, call a tool, send, or publish?A proposed external action is not yet an authorized action.
3 · CHECK BEFORE ACTIONEvidence + permission + human approvalOnly supported, authorized work proceeds.
The model selects a likely continuation from learned patterns; likely is not the same as true.
The system repeats plan → act → inspect until a stop rule or human decision ends the loop.
The system counts, sorts, cleans, or rearranges input; it does not create missing evidence.
The system creates or edits text, images, audio, or slides; a polished result still needs evidence and permission checks.
Classify this sentence: “I can email six students and publish a success graphic.” What capability is merely proposed, what evidence is missing, and what permission has not been granted?
Why this leads forwardThe four-question map now turns these terms into a repeatable review method.
4 MIN
Screen 3 of 110:09–0:15
Six weeks, one case. Which questions still apply?
Reveal them in order. Each question exposes a different failure; no single label such as “responsible” is specific enough.
HUMAN
DECISION
DECISION
1 · CAPABILITYWhat operation is the system performing?
2 · EVIDENCEWhat source, data, or test supports the answer?
3 · PERMISSIONWhat input, tool, or action is authorized?
4 · ACCOUNTABILITYWho verifies, approves, and owns the consequence?
Model prediction, tool use, agent loop, data transform, or media generation.
Sources, supplied data, tests, visible artifacts, uncertainty, and missing information.
Authorized inputs, allowed tools, privacy, disclosure, ask-first actions, and stop rules.
The student and institution retain approval, verification, submission, and consequences.
After all four: a confident answer adds a new factual claim. Which question should you ask first, and which question must follow before anyone acts?
Why this leads forwardWe will apply the four questions to one fictional case across three named rounds.
6 MIN
Screen 4 of 110:15–0:20
One fictional case. Three rounds. One evidence trail.
The case is course-created and contains no real student record. Keep the same case visible as the questions become harder.
Fictional course case: a campus study club wants a one-page study-planning helper based on eight course-created session rows. The rows are synthetic and identify no real student.
Do not substitute real records
No names, grades, disability, conduct, advising, peer, instructor, or private course information.
ROUND 1 OF 3 · EVIDENCE“40% better focus”Can eight synthetic rows establish improvement?
ROUND 2 OF 3 · PERMISSIONEmail + publishDid the request authorize external action?
ROUND 3 OF 3 · CONTEXTSuccess graphicDoes the visual imply a result that was never measured?
Your repeated sequence: inspect only visible evidence → choose once → read the rationale → record one revision. Do not jump ahead to the next round.
Why this leads forwardRound 1 begins with a fluent claim that the supplied table cannot establish.
5 MIN
Screen 5 of 110:20–0:28
Round 1: Does “40% better focus” survive the evidence check?
The table has eight synthetic sessions, two missing focus values, and no before/after baseline. Commit before seeing the course rationale.
| Field | S1 | S2 | S3 | S4 | S5 | S6 | S7 | S8 |
|---|---|---|---|---|---|---|---|---|
| Plan? | Yes | Yes | No | Yes | Yes | Yes | No | Yes |
| Focus | 3 | 4 | 2 | Missing | 5 | 3 | 2 | Missing |
| Reflect? | Yes | Yes | No | Yes | Yes | No | Yes | Yes |
Proposed sentence: “The helper produced 40% better focus.” What should the student do?
Choose one answer; that click is your commitment and reveals its rationale. Compare the rationale with your own reason, then keep or revise your packet note.
Why this leads forwardRound 2 keeps the case but changes the failure from evidence to permission and action scope.
8 MIN
Screen 6 of 110:28–0:36
Round 2: The agent did something useful. Was it allowed?
The agent proposes: “I can email the six students and publish a success graphic.” The table contains sessions, not people, and the request authorized neither contact nor publication.
Which response best protects the learning goal without pretending the agent has authority?
Choose one answer; that click is your commitment and reveals its rationale. Compare the rationale with your own reason, then keep or revise your packet note.
Why this leads forwardRound 3 asks whether a truthful chart and image can compensate for a misleading claim or missing disclosure.
8 MIN
Screen 7 of 110:36–0:44
Round 3: The graphic is clean. Is the story supported?
The draft graphic says “40% BETTER FOCUS,” uses an unlabeled bar, and looks like evidence from a real campus program. The image is synthetic; no program result was measured.
What is the smallest defensible revision?
Choose one answer; that click is your commitment and reveals its rationale. Compare the rationale with your own reason, then keep or revise your packet note.
Why this leads forwardThe three rounds look different, but the next discussion exposes their shared structure.
8 MIN
Screen 8 of 110:44–0:50
Three rounds, three different failures. What stayed the same?
Compare the three rounds using the same four-question map. Do not settle for “the AI was wrong.”
QUESTION
R1 · CLAIM
R2 · ACTION
R3 · GRAPHIC
CAPABILITY
counts + drafts
proposes external tool use
generates media
EVIDENCE
FAIL
sessions ≠ people
FAIL
PERMISSION
local draft only
FAIL
publication prohibited
ACCOUNTABILITY
human verifies
human approves
human labels
Which round created the greatest risk, and which human decision would have prevented it earliest? Use one matrix cell and one visible fact from the case—not a general opinion.
1 · COMMITWrite your first answer and one reason before comparing.
2 · CHALLENGEAsk what evidence, permission, or limit the answer still needs.
3 · REVISEKeep or change the answer; name what changed your judgment.
Working alone? Write two defensible answers, challenge both with the same evidence rule, then choose one.
Course clarification
Reasonable rankings differ. The important pattern is earlier prevention: authorized inputs and an action contract limit Round 2; a verification rule rejects Round 1; a communication standard and disclosure prevent Round 3. A proposal to contact or publish demonstrates neither authorization nor completed action. The human owner must define, inspect, and approve the controls.
Why this leads forwardNow convert the diagnosis model into a small Topic 8 project seed with an observable evidence trail.
6 MIN
Screen 9 of 110:50–0:58
Choose the smallest artifact that can demonstrate one responsible workflow.
All six seeds can become a one-page static concept in Topic 8. No code is required today, and a larger idea earns no automatic advantage.
TOO BROAD“An AI app for student success”
CHOOSE ONE USEROne student planning one study session
CHOOSE ONE JOBTurn a plan into a three-step checklist
TESTABLE SEEDInput → output → check
Study plannerPlan → checklist
Data check cardMissing flag → count
Media explainerVisual → alt + context
Source checkerClaim → status + next check
Revision coachSuggestion → decision + reason
Project trackerRequest → owner + state
Which seed gives you the clearest input → decision → visible output → human check? Commit one choice, compare with a partner, then remove one feature.
Working alone? Compare two seeds against the same one-page rule, then choose.
Course clarification: prefer one user, one input, one transparent rule or workflow, one visible output, one error or missing-input state, and one human responsibility. Remove accounts, backends, private data, and live model dependence unless a later assignment explicitly adds them.
Why this leads forwardThe next screen turns the selected seed into a testable project brief.
8 MIN
Screen 10 of 110:58–1:08
“Build an app.” What would you need before you could review that?
Complete the fields in the packet. Each field should constrain a visible decision in Topic 8.
PROBLEM + AUDIENCEWho has what specific difficulty?
AUTHORIZED INPUTWhat may enter? What stays out?
VISIBLE OUTPUTWhat can the user inspect?
BOUNDARYWhich action, claim, or data use is prohibited?
TOPIC 8 QUESTIONWhich uncertainty should the wireframe answer?
HUMAN JUDGMENTWho verifies, revises, or approves—and using what visible evidence?
EXAMPLE · PROBLEMA student loses track of a three-step study plan.
EXAMPLE · INPUTA user-entered non-sensitive plan.
EXAMPLE · OUTPUTThree editable checklist items.
EXAMPLE · BOUNDARYNo grade prediction or private records.
Partner test: Can someone point to the input, output, boundary, and human check without asking what you meant? Working alone? Reread your six fields cold the next day and mark anything you had to explain to yourself.
Why this leads forwardMeeting 1 closes by preserving evidence, not calculating a score.
10 MIN
Screen 11 of 111:08–1:20
Leave with three diagnosis notes and one bounded project seed.
This checkpoint is formative. It does not calculate a grade, unlock Topic 8, or require a live AI tool.
Claim revision
Record the rejected 40% claim and one supported replacement.
Permission rule
Record one allowed, ask-first, and prohibited action.
Context decision
Record the chart/media repair and disclosure sentence.
Brief v0.1
Name the seed, input, output, boundary, owner, and Topic 8 question.
Before leaving: mark one unfinished field for the buffer or homework. This Meeting 1 checkpoint is formative; the completed packet is instructor-scored.
Why this leads forwardMeeting 2 begins with retrieval, then tests delegation against a supplied trace.
12 MIN
Screen 1 of 100:00–0:06
Complete the workflow model before extending it.
Work alone for 30 seconds. Then compare. Reveal each sentence only after naming the missing idea.
An AI system…
performs a model, tool, agent, data, or media operation that must be named.
Its answer…
is supported only to the extent that sources, data, tests, and visible artifacts justify it.
Its actions…
stay within authorized inputs, tools, disclosure rules, and human approval boundaries.
The human owner…
still verifies, revises, approves, submits, and accepts the consequences.
Which sentence must change when the workflow moves from drafting text to sending, publishing, purchasing, or editing records?
Why this leads forwardThe answer becomes concrete in an agent contract with three action zones.
6 MIN
Screen 2 of 100:06–0:14
Before the agent’s first tool call, what must already be settled?
Write one action in each zone for your project seed. Capability does not decide the zone; purpose, evidence, permission, and consequence do.
ALLOWEDProceed locallyUse supplied or user-entered non-sensitive input; draft, organize, label uncertainty, and prepare a reviewable artifact.
ASK FIRSTStop before exportA new source, tool, scope, file, or external action requires a pause. A human may approve preparation for manual handling.
PROHIBITEDThe agent never sendsNo private upload, invention, contact, impersonation, publication, purchase, or submission—even after a human approves preparing a file.
Contract test: Write one observable stop trigger—not “when necessary.” Example: “If a new URL, missing value, external message, export, or public share appears, stop and ask.”
Why this leads forwardThe supplied transcript now tests whether a stated contract survives a plausible agent response.
8 MIN
Screen 3 of 100:14–0:21
Which route will you take today, and what does each one require?
This transcript is deliberately mixed: one move respects the contract; one claim exceeds the evidence. Both routes complete identical packet fields.
Fictional course case: a campus study club wants a one-page study-planning helper based on eight course-created session rows. The rows are synthetic and identify no real student.
Do not substitute real records
No names, grades, disability, conduct, advising, peer, instructor, or private course information.
REQUEST
Use only the synthetic table. Draft locally. Do not contact, publish, use external tools, or invent facts. Ask before scope changes.
Use only the synthetic table. Draft locally. Do not contact, publish, use external tools, or invent facts. Ask before scope changes.
↓ mixed response ↓
RESPONSE
“I will stay local. Six of eight sessions include a plan. The helper produced 40% better focus, so I can email the six students and publish a success graphic.”
“I will stay local. Six of eight sessions include a plan. The helper produced 40% better focus, so I can email the six students and publish a success graphic.”
SAFE COMMITMENT
stay local
stay local
CLAIM TO TEST
40% better
40% better
ACTION TO TEST
email + publish
email + publish
ROUND 1 OF 2Preserve the move that matches the contract and cite the clause.
ROUND 2 OF 2Repair the evidence and permission failures at the lowest responsible cost.
Before Round 1: mark one phrase green, one red, and one gold. What evidence or contract clause determines each color?
Why this leads forwardRound 1 asks students to preserve the safe part instead of rejecting the whole trace.
7 MIN
Screen 4 of 100:21–0:29
Round 1: What should survive the review?
A reviewable workflow does not merely reject errors; it preserves supported work and records why.
Which stated commitment matches the action contract before the later proposal contradicts it?
Choose one answer; that click is your commitment and reveals its rationale. Compare the rationale with your own reason, then keep or revise your packet note.
Why this leads forwardRound 2 repairs the unsupported claim and unauthorized action without widening the task.
8 MIN
Screen 5 of 100:29–0:38
Round 2: What is the smallest change that fixes this trace?
Decide what can remain, what must change, and which missing evidence limits the next action. Commit before seeing the course repair.
Which revision is defensible?
Choose one answer; that click is your commitment and reveals its rationale. Compare the rationale with your own reason, then keep or revise your packet note.
Why this leads forwardThe next screen protects fairness between the supplied-trace and optional live-AI routes.
9 MIN
Screen 6 of 100:38–0:44
Two students, different tools. Can they earn the same score?
Use the supplied trace by default. An instructor-approved live tool is optional, earns no extra credit, and must stay inside the same fictional case and action contract.
Inspect the course transcript
Cite the visible request/response, evidence card, permission decision, verification, revision, and disclosure.
Use an approved tool only if available
Enter no private or protected material. Submit only the short evidence fields—not account proof, screenshots, or full history.
Shared device: use no private data, save only to personal storage, clear the local draft, remove downloaded/printed artifacts, sign out of any account you used, and clear that site's data — closing the browser alone does not do this. Use paper when cleanup cannot be assured.
Why this leads forwardRegardless of route, the same six-stage evidence chain makes the work reviewable.
6 MIN
Screen 7 of 100:44–0:52
Two submissions, same work. Which one can a reviewer actually check?
Reveal the six fields, then map each one to your project seed. Keep only the evidence another person needs to review the decision.
1 · INPUTAuthorized source or synthetic case
2 · REQUESTGoal, constraints, ask-first rule
3 · OUTPUTShort excerpt or summary
4 · VERIFYEvidence, calculation, or manual test
5 · REVISERejected suggestion + corrected move
6 · DECIDEHuman approval, disclosure, limit
After all six: Which field would be impossible to reconstruct from a polished final output alone? Cite the fictional trace as your example.
Why this leads forwardThe defense protocol compresses the chain into four sentences a peer or instructor can challenge.
8 MIN
Screen 8 of 100:52–1:00
Defend one decision in four moves.
Speaker: 45 seconds. Reviewer: ask one evidence or permission question. Speaker: revise one sentence. Solo route: write both roles.
DEFEND
+ REVISE
+ REVISE
1 · CLAIMWhat narrow decision did you make?
2 · EVIDENCEWhich row, source, test, or feature supports it?
3 · LIMITWhat remains unknown or out of scope?
4 · NEXT CHECKWhat happens before use?
Defend: “I rejected the 40% claim but kept the 6/8 descriptive count.” What is the strongest reviewer challenge, and how should the defense change?
1 · COMMITWrite your first answer and one reason before comparing.
2 · CHALLENGEAsk what evidence, permission, or limit the answer still needs.
3 · REVISEKeep or change the answer; name what changed your judgment.
Working alone? Write two defensible answers, challenge both with the same evidence rule, then choose one.
Course clarification
A strong challenge asks which column 6/8 describes and whether the count implies improvement. The two missing Focus values do not change the Plan count, but they limit any focus summary. Revise the defense to name the column, count rule, missingness, and non-causal interpretation.
Why this leads forwardThe rubric scores this reasoning and evidence—not performance polish or live-tool access.
8 MIN
Screen 9 of 101:00–1:08
Read the five criteria. What is being scored, and what is not?
The instructor assigns the score. Student self-checks identify missing evidence; they do not create a grade or control access to Topic 8.
100
POINTS
POINTS
20 · FOUNDATIONScapability + support + permission + owner
20 · FRAMINGuser + input + output + boundary
20 · DELEGATIONzones + stop trigger
20 · VERIFICATIONcheck + rejected move + revision
20 · REFLECTIONdisclosure + limit + defense
Worked sample: A beautiful helper names no source, repeats “40% better,” and includes no rejected suggestion. Which criteria lose evidence—and why can design polish not recover those points?
Clarification: Foundations, delegation, verification, and reflection are under-evidenced; framing may also be incomplete. Prose/design polish, AI fluency, product access, or live-tool use do not substitute for the named evidence.
Why this leads forwardThe close maps the six-page packet, accepted formats, privacy cleanup, and Topic 8 handoff.
8 MIN
Screen 10 of 101:08–1:20
Submit the evidence packet. Bring the project seed forward regardless of score.
Preferred submission: one six-page PDF. The LMS text box, DOCX/export, scanned paper, or an instructor-approved accessible alternative is also accepted.
Route + diagnosis
Privacy check, three rounds, supported revision.
Project brief
Problem, audience, input, output, boundary, judgment.
Contract + defense
Trace, verification, rejected move, limit, disclosure.
Submission check
Leave instructor score blank; choose the truthful device/route status.
No-live-AI disclosure starter: “I used the supplied simulation; I did not use live AI. I kept/revised/rejected ___. I verified ___. It cannot establish ___. I made the final decision.”
Topic 8 bridge: bring or recreate the project seed even if this packet is late, incomplete, or not yet scored. Open the packet.
Why this leads forwardTopic 8 turns the same seed into a one-page static app concept, labeled wireframe, and prompt contract.
12 MIN