CubeSTEM Digital Twin · Track 7

Track 7 — AI / ML Autonomy

A complete five-session AI/ML Autonomy mini-course — autonomy boundaries, features and labels, anomaly classification, false-alarm trade-offs, and human-in-the-loop decisions — building on Track 6 telemetry evidence habits.

Local-only teaching mini-course: no account, no submissions, no gradebook. Deterministic classroom models only — not certified AI, not flight software, not real satellite command authority.

Capability boundary

Teaching-grade AI / ML model — read before use

  • ✓ Browser-local deterministic teaching classifiers — no external AI API, no ML library, no training pipeline.
  • ✓ Local Assessment Engine v0 — self-check questions, no submission, no grade, no roster.
  • ✓ Local Evidence Engine v0 — copy/export text or JSON, no backend, no cloud storage.
  • ✗ Not certified AI, not flight software, not real onboard autonomy or anomaly diagnosis.
  • ✗ Not a real satellite command interface — “Execute (Simulated)” issues no real commands.
  • ✗ No student accounts, no teacher roster visibility, no LMS integration, no grades.
  • ✗ No real-time data feed, no hardware connection, no remote lab in this track.

At a glance

Evidence and assessment summaries

Evidence checklist summary

Expect one copied artifact per session: autonomy permissions table, feature/label rationale, classifier outputs + disagreement note, threshold justification + confusion counts, and decision debrief with safety rule result.

Open full evidence checklist →

Assessment map summary

15 local self-check questions across five sessions — autonomy framing, data quality, classifier interpretation, threshold trade-offs, and human-in-the-loop judgement.

Open full assessment map →

Core activity sequence

Five implemented sessions in canonical order

Track 7 — AI / ML Autonomy mini-course

Five sessions, one autonomy story — from permissions to human-in-the-loop review

A complete five-session AI/ML Autonomy mini-course — autonomy boundaries, features and labels, anomaly classification, false-alarm trade-offs, and human-in-the-loop decisions — building on Track 6 telemetry evidence habits.

Follow Next step → inside each activity shell and on this page to stay in teaching order. Use Track 6 evidence habits (timestamped quotes, structured claims) before you interpret classifier outputs here.

  1. Session 1core20–25 min
    What Does Autonomy Mean?

    What Does Autonomy Mean?

    Student action: Pick an autonomy level + scenario, then write what the system may do and what humans must still approve.

    Assessment focus: Autonomy levels, permission boundaries, and human oversight.

    Misconception focus: Autonomy does not remove human accountability.

  2. Session 2core30–35 min
    Features, Labels and Training Data

    Features, Labels & Training Data

    Student action: Choose telemetry features (building on Track 6 evidence habits), assign labels, and justify each choice.

    Assessment focus: Feature selection, label correctness, and data-quality impact.

    Misconception focus: ML quality depends on telemetry quality and labels.

  3. Session 3core25–30 min
    Anomaly Classifier

    Anomaly Classifier / Normal vs Abnormal Telemetry

    Student action: Run one case through both classifier views, compare agreement/disagreement, and interpret confidence.

    Assessment focus: Rule-based vs preset-model output interpretation.

    Misconception focus: Confidence is not certainty.

  4. Session 4core30–35 min
    Confidence and False Alarms

    Confidence and False Alarms / Fault Rules

    Student action: Move sensitivity, record TP/FP/TN/FN changes, and defend one threshold choice with risk reasoning.

    Assessment focus: Threshold tuning, FP/FN trade-offs, and alarm fatigue risk.

    Misconception focus: Higher sensitivity is not automatically better.

  5. Session 5core35–40 min
    Human-in-the-Loop Decision

    Human-in-the-Loop Decision / Safe Mode Review

    Student action: Review evidence cards, choose a simulated action, apply safety check, and write a short debrief.

    Assessment focus: Evidence-based action choice with safety-rule review.

    Misconception focus: Teaching model action is not flight command authority.

When you finish Session 5, the mini-course is complete. No account required — progress stays local to this browser unless you copy or screenshot to share.

Track 7 completion bridge

After Session 5, continue to the curriculum map or demo pack for final learning-system walkthrough surfaces.

Teacher plan

Track 7 AI / ML Autonomy mini-course pack

Same five-session sequence as the hub — scale intensity with time. Evidence and assessment stay local-only; plan a manual share routine (copy/export or screenshot). No roster visibility.

Optional bridge — Mission Realism Lab

Use telemetry divergence and realism evidence before autonomy decisions — explore packet and link teaching models in Mission Realism Lab alongside Track 6 habits.

Delivery options

45 min

45-minute autonomy demo

Conference breakout or executive workshop slice.

  • Boundaries + vocabulary (5 min): autonomy vs automation, confidence vs certainty, local-only evidence.
  • Session 1 (12 min): autonomy levels + scenario — what may execute vs what stays human-approved.
  • Session 3 (15 min): dual classifier views — agree/disagree + confidence interpretation.
  • Session 5 (10 min): evidence cards + safety rule check — one simulated decision + debrief.
  • Misconception debrief (3 min): AI is pattern matching; ML needs telemetry + labels.

90 min

90-minute AI/ML classroom lab

Single class period — Sessions 1–5 with threshold debate.

  • Opening (5 min): tie Track 6 telemetry quotes to features/labels in Session 2.
  • Sessions 1–2 (25 min): autonomy framing → feature/label lab + data-quality discussion.
  • Session 3 (15 min): rule-based vs preset classifier — think-aloud on top features.
  • Session 4 (20 min): sensitivity slider + confusion counts — class vote on threshold.
  • Session 5 (15 min): human-in-the-loop decision + structured debate.
  • Exit (10 min): copy one evidence line per session; preview Curriculum Map / Demo Pack.

3 hr

Half-day telemetry→autonomy workshop

Pilot classroom or PD — full depth plus debrief.

  • Telemetry bridge (15 min): replay one Track 6 evidence habit (quote + timestamp) before ML.
  • Session 1 (25 min): three autonomy levels — permissions table + oversight reflection.
  • Session 2 (35 min): features/labels — dataset gap hunt (missing fault examples).
  • Session 3 (30 min): classifier comparison — disagree cases drive discussion.
  • Session 4 (35 min): TP/FP/TN/FN + alarm fatigue storytelling.
  • Session 5 (40 min): evidence cards + safety rules — two teams defend different actions.
  • Closing circle (20 min): boundary recap — teaching-grade only; no flight AI claims.

Facilitation prompts

  • Which decisions must always stay human-approved in this scenario?
  • Which missing feature would most likely break this classifier result?
  • When outputs disagree, which evidence should dominate the decision?
  • Which threshold trade-off is safest for this mission context and why?
  • What evidence justifies override or approval for this simulated action?

Common misconceptions to preempt

  • AI is not magic — it is pattern matching on selected data.
  • ML depends on telemetry quality and label quality.
  • Confidence is not certainty.
  • False alarms and missed detections both matter.
  • Safe autonomy requires explicit rules and human oversight.
  • This is a teaching-grade model, not flight software and not certified diagnosis.

Teacher boundary notes

  • Evidence and assessment are local-only in the browser.
  • No roster, no submissions, no gradebook, no LMS pipeline.
  • No external AI API calls, no heavy ML training stack, no backend ML jobs.
  • Never describe Track 7 as real onboard autonomy or real command authority.

Hub entry: /twin/learn/tracks/ai_ml_autonomy

Student path

What to do in each Track 7 session

Follow sessions in order, copy evidence each session, and use Track 6 telemetry evidence habits as your baseline.

Optional — Mission Realism Lab

Use telemetry divergence and realism evidence before autonomy decisions — open Mission Realism Lab for checksum, stale data, and link-margin context.

  1. 1
    What Does Autonomy Mean?

    Pick an autonomy level + scenario, then write what the system may do and what humans must still approve.

    Copy evidence: Chosen autonomy level and scenario recorded

  2. 2
    Features, Labels and Training Data

    Choose telemetry features (building on Track 6 evidence habits), assign labels, and justify each choice.

    Copy evidence: Selected features and justification for each choice

  3. 3
    Anomaly Classifier

    Run one case through both classifier views, compare agreement/disagreement, and interpret confidence.

    Copy evidence: Chosen scenario and classifier type recorded

  4. 4
    Confidence and False Alarms

    Move sensitivity, record TP/FP/TN/FN changes, and defend one threshold choice with risk reasoning.

    Copy evidence: TP/FP/TN/FN counts at chosen sensitivity level

  5. 5
    Human-in-the-Loop Decision

    Review evidence cards, choose a simulated action, apply safety check, and write a short debrief.

    Copy evidence: Chosen predicted anomaly and confidence level

After Track 7 completion

When you finish Session 5 (Human-in-the-Loop Decision), the Track 7 mini-course is done.

You have completed the full CubeSat learning arc from mission concept through telemetry evidence to AI-assisted autonomy decisions. Continue to the Curriculum Map or open the Demo Pack for final review pathways.

Evidence checklist

Track-level evidence across all five sessions

Copy one evidence artifact per session (five sessions total). Evidence remains browser-local unless you manually share (copy/export or screenshot).

Session 1What Does Autonomy Mean?

  • Chosen autonomy level and scenario recorded
  • List of allowed vs disallowed actions at the selected level
  • One-sentence reflection on why human oversight matters
  • Self-check summary and copied evidence text

Session 2Features, Labels and Training Data

  • Selected features and justification for each choice
  • Assigned label and explanation for a given telemetry example
  • Note on one way a missing or biased feature would harm the model
  • Self-check summary and copied evidence text

Session 3Anomaly Classifier

  • Chosen scenario and classifier type recorded
  • Predicted class and confidence score
  • Top contributing features with brief explanation
  • One observation where rule-based and ML classifiers differ
  • Self-check summary and copied evidence text

Session 4Confidence and False Alarms

  • TP/FP/TN/FN counts at chosen sensitivity level
  • Chosen sensitivity setting with operational justification
  • One-sentence explanation of alarm fatigue risk
  • Self-check summary and copied evidence text

Session 5Human-in-the-Loop Decision

  • Chosen predicted anomaly and confidence level
  • Evidence card review — supporting / neutral / contradicting classification
  • Chosen action and safety rule check result
  • One-paragraph decision debrief
  • Self-check summary and copied evidence text

Evidence handling boundary

Local-only evidence workflow: no backend submissions, no roster visibility, no grading pipeline. Use copy/export text or screenshot capture for classroom sharing.

Assessment map

Formative assessment focus across Track 7

15 self-check questions across five sessions (Assessment Engine v0). No submissions and no official grading — discussion starters for class review.

Session 1What Does Autonomy Mean?

Assessment focus: Autonomy levels, permission boundaries, and human oversight.

Key prompt: Describe one situation where 'recommend action' autonomy is safer than 'execute' autonomy, and explain why.

Session 2Features, Labels and Training Data

Assessment focus: Feature selection, label correctness, and data-quality impact.

Key prompt: Name two ways a biased or incomplete training dataset could cause a fault classifier to make dangerous mistakes.

Session 3Anomaly Classifier

Assessment focus: Rule-based vs preset-model output interpretation.

Key prompt: Give one example of a telemetry pattern that looks abnormal but is actually expected during a planned mode change — and explain how a classifier might incorrectly flag it.

Session 4Confidence and False Alarms

Assessment focus: Threshold tuning, FP/FN trade-offs, and alarm fatigue risk.

Key prompt: A detector has TP=14, FP=7, TN=13, FN=1. Calculate precision and recall, and state which setting would cause alarm fatigue and why.

Session 5Human-in-the-Loop Decision

Assessment focus: Evidence-based action choice with safety-rule review.

Key prompt: Give one scenario where entering safe mode too early wastes science, and one where waiting too long risks the spacecraft bus — explain how evidence cards would help you decide.

Assessment boundary note

Local self-check only. No roster visibility, no grading backend, no LMS integration. Use teacher prompts for classroom discussion and evidence review.

Final bridge

Continue to final learning-system review surfaces

After Session 5, use Curriculum Map for the full eight-track view, Demo Pack for reviewer pacing, Student Mode to rehearse the learner arc, and Learn hub to jump back into any track overview—always browser-local with no submissions.