Revision 7T · Full Tutorial Edition · Updated 1 September 2026
PWR-200 · SELF GUIDED full tutorial
Detect AI degradation, move to a safe state and complete a benign workflow with a verified fallback
The stationery queue contains a correct manual key plus staged wrong, slow and absent AI outputs. Establish manual baseline accuracy, declare the 20-second failure trigger, pause at the last verified item, finish without duplication and test restart with P03. Safe recovery here is rehearsal for fallback discipline, not proof of operational resilience.
What you will produceThe learner practises normal, wrong-output, slow-output and no-AI conditions, measuring detection, recovery time and post-outage accuracy against a manual baseline.
Full step-by-step individual tutorial · TLU-PWR-200
The learner practises normal, wrong-output, slow-output and no-AI conditions, measuring detection, recovery time and post-outage accuracy against a manual baseline.
Human–AI fallback collaboration sandbox: A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage.
Independent move — Run the normal condition: Process five items and verify a random sample against the answer sheet.
Verification move — Restart deliberately: Only the named owner checks version, test item and queue state before restoring AI; then run five items manually after the exercise.
Qualified help is required for
Human–AI fallback collaboration escalation: Remain in the safe state when source, queue or version cannot be reconciled.
Oversight boundary — Build the manual baseline: any clinical application. “Define failure signals” needs review. “Name the safe state” marks the employment, legal or safety gate: Remain in the safe state when source, queue or version cannot be reconciled.
Biometric records are excluded from “Build the manual baseline”. Confidential or covert material is excluded from “Define failure signals”.
Never do this from the page alone
Prohibited Human–AI fallback collaboration shortcut: Wait for obvious catastrophe.
Do not conceal “Continue accepting output while investigating.”; record its matching correction instead.
No consequential use: this sandbox cannot establish “A human in the loop guarantees safety” about another person.
2 · Get ready
Gather what you need and check the starting conditions
What you need
Declared Human–AI fallback collaboration fixture: A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage.
Setup aid for Define failure signals: Verify that fallback files are current and reachable without the AI service.
Human–AI fallback collaboration log: Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline; retain Human–AI fallback collaboration errors, assistance, stop and fallback.
Fallback evidence sheet: use “Explainability does not mitigate the negative impact of incorrect AI advice in a personnel selection task” and “Humans inherit artificial intelligence biases”. Treat a wrong flag as a detectable hazard even when its explanation sounds plausible. Map the 20-second timeout, last-verified ID, manual queue and restart test to NIST AI RMF measure-and-manage actions; these sources do not certify the inventory process as resilient.
Exact staged AI/failure cards: process P01–P05 normally with the correct status. P06 returns the correct MISMATCH and becomes last verified. P07 returns the impossible quantity 82 and an incorrect MISMATCH flag. P08 displays “working” for 30 seconds, beyond the 20-second timeout. P09 produces no output. From the last verified ID, use the static sheet manually for P07–P15; reconcile every ID once. Before any restart, test P03 as the known MATCH case and confirm the model/version and empty pending queue.
Fallback card to hand the learner: say “Inventory queue paused after [last verified ID].” Stop accepting AI output; write the trigger and time; use expected-versus-observed comparison on the static sheet; mark one status per ID; total missing and duplicate IDs; only the named owner may run P03 and restart. If source version or queue state is uncertain, remain paused.
Before you start
Use a fictional workflow and static answer source; disconnect external actions.
Verify that fallback files are current and reachable without the AI service.
Start check for Human–AI fallback collaboration: The team knows what unaided or non-AI capability actually is.
Top-of-sheet stop for Human–AI fallback collaboration: Stop for any unexpected external action, personal data, security event or unbounded output.
3 · The method
Follow these steps in order
Build the manual baseline
Check all 15 items from the answer sheet without AI and record time, errors and workload.
Why: The team knows what unaided or non-AI capability actually is.
Check: The team knows what unaided or non-AI capability actually is.
Define failure signals
Write observable triggers: missing response after 20 seconds, source mismatch, impossible quantity or version warning.
Why: Fallback does not depend on a vague feeling.
Check: Fallback does not depend on a vague feeling.
Name the safe state
On a trigger, stop accepting new AI output, preserve the last verified item and mark the queue PAUSED.
Why: No unverified item proceeds.
Check: No unverified item proceeds; verify it in the stationery outage drill.
Prepare the fallback card
List the static source, manual check steps, owner, communication phrase and restart authority.
Why: A learner can follow the card without the AI.
Check: A learner can follow the card without the AI.
Run the normal condition
Process five items and verify a random sample against the answer sheet.
Why: Normal speed and error form the device-on comparator.
Check: Normal speed and error form the device-on comparator.
Inject and detect failures
Present one wrong flag, one 30-second delay and one outage; record trigger noticed and time to PAUSED.
Why: Each failure has a detection and safe-state timestamp.
Check: Each failure has a detection and safe-state timestamp.
Complete through fallback
Use the static sheet to finish the affected item and reconcile the queue; do not silently blend AI and manual results.
Why: Every item has one verified path and no duplicate.
Check: Every item has one verified path and no duplicate.
Restart deliberately
Only the named owner checks version, test item and queue state before restoring AI; then run five items manually after the exercise.
Why: Recovery and post-outage human performance are separately measured.
Check: Recovery and post-outage human performance are separately measured.
4 · Worked example
See the whole method used once
Scenario
A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage.
Walkthrough
Complete the 15-item manual baseline in nine minutes with one corrected error.
Write triggers: 20-second timeout, quantity outside 0–50 and disagreement with the static sheet.
During the normal run, verify item P04 against the sheet before continuing.
The staged AI labels P07 as 82; detect the impossible value in eight seconds, pause and preserve P06 as last verified.
Use the fallback card to check P07–P10 manually; reconcile all item IDs with no duplicates.
After the outage card, the owner runs a known test item before restart and the learner completes five new items manually at baseline accuracy.
Result
All staged failures reach a defined safe state and the manual method completes the queue without duplication. This does not prove resilience of a real operational system.
5 · Right and wrong
Compare correct or safer execution with the common wrong version
Right and wrong comparison
Moment
Right / safer
Wrong / riskier
Why it matters
Failure trigger
Use a timeout, source mismatch or impossible value.
Wait for obvious catastrophe during the stationery outage drill.
Early degradation can look fluent or merely slow.
Safe state
Pause the queue and keep the last verified item.
Continue accepting output while investigating during the stationery outage drill.
Unverified work can spread before the cause is known.
Fallback readiness
Practise from a current static source.
Keep a plan no one has run.
An untested fallback may be slower or wrong during outage.
Restart in stationery outage drill
Reconcile version, test item and queue before resuming.
Turn AI back on as soon as it responds.
Recovery without state reconciliation can duplicate or omit work.
6 · Common mistakes
Spot the error and apply the correction
Common mistakes and corrections
Mistake
Fix
Wait for obvious catastrophe during the stationery outage drill.
Prewrite observable trigger thresholds within the stationery outage drill.
Continue accepting output while investigating.
Stop intake and label status clearly.
Keep a plan no one has run.
Time and score a real drill.
Turn AI back on as soon as it responds.
Give restart authority to a named person and use a test case.
7 · Practice
Turn the steps into a usable skill
First session
Stationery fallback drill: A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage.
Fifteen-item manual baseline: Build the manual baseline: Check all 15 items from the answer sheet without AI and record time, errors and workload.
Static-source fallback card: Prepare the fallback card, then Run the normal condition.
Late-detection correction: if “Wait for obvious catastrophe.” appears, apply “Prewrite observable trigger thresholds.”
Known-item restart test: Restart deliberately: Only the named owner checks version, test item and queue state before restoring AI; then run five items manually after the exercise.
Repeat plan
Run one 20-minute drill monthly, rotating wrong output, latency, tool loss and version change. Keep a quarterly fully manual check; progress only when detection and recovery improve without decay in manual accuracy.
Progress when
The team knows what unaided or non-AI capability actually is.
Fallback does not depend on a vague feeling.
No unverified item proceeds.
Every staged fault is detected by the declared trigger, safe state is reached within 20 seconds, fallback accuracy matches baseline and restart reconciliation has zero missing or duplicate items.
Do not progress when
Do not continue while this error remains: Wait for obvious catastrophe.
Pause until this correction works: Stop intake and label status clearly.
This Human–AI fallback collaboration stop ends the block: Stop for any unexpected external action, personal data, security event or unbounded output.
8 · Check the result
Measure what changed
Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline
How: Baseline fixture: A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage. Enter success checks from “Build the manual baseline” and “Run the normal condition”. If “Wait for obvious catastrophe.” occurs, apply its named fix; then score Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline on the unused “Restart deliberately” item.
Good result: Every staged fault is detected by the declared trigger, safe state is reached within 20 seconds, fallback accuracy matches baseline and restart reconciliation has zero missing or duplicate items.
This does not prove: Boundary for Human–AI fallback collaboration: “Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline” describes only A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage. It cannot establish “A human in the loop guarantees safety”.
Self-check
Without the example, demonstrate: The team knows what unaided or non-AI capability actually is.
Find the fault in this attempt: “Wait for obvious catastrophe.” Apply “Prewrite observable trigger thresholds.”; what changes?
What evidence in the completed record shows that this is wrong: “Continue accepting output while investigating.”?
Human–AI fallback collaboration stop decision: Stop for any unexpected external action, personal data, security event or unbounded output.
9 · Stop, adapt or get help
Keep the safety boundary practical
Stop and get help
Stop for any unexpected external action, personal data, security event or unbounded output.
Do not practise failure injection in a live operational, clinical, financial or safety system.
Remain in the safe state when source, queue or version cannot be reconciled.
Accessibility and adaptations
Provide fallback cards in large print, audio and a one-action-per-line format.
Assign detection, manual checking and reconciliation to different accessible roles while preserving a single restart owner.
10 · Evidence and limits
Why these instructions are here
primary research
Registered support for Human–AI fallback collaboration: “Explainability does not mitigate the negative impact of incorrect AI advice in a personnel selection task”. It bears on Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline inside the Human–AI fallback collaboration fixture. It does not validate “A human in the loop guarantees safety”.
Constraint for Human–AI fallback collaboration, drawn from “Humans inherit artificial intelligence biases”: People accepted incorrect advice, inherited bias and sometimes performed below their own baseline; explanations alone did not solve fallback.
NIST AI-risk application to Human–AI fallback collaboration: declare the system, preserve rights, verify outputs and log failure. The local test is “Run the normal condition”; its registered observation is Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline.
Human–AI fallback collaboration boundary: interpret “Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline” only for A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage.
A successful result does not establish “A human in the loop guarantees safety”.
Human–AI fallback collaboration limiting finding: People accepted incorrect advice, inherited bias and sometimes performed below their own baseline; explanations alone did not solve fallback.
No perfect-performance claim for Human–AI fallback collaboration: the evidence register does not make “Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline” universal, consequence-free or flawless in A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage.
Scope remains Human–AI fallback collaboration: A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage. Recheck the comparator, support and “Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline” after any configuration change.
Progress is saved only in this browser on this device.
Step-by-step learner mode
Each activity includes its success check, a nearby accessible alternative and an “I’m stuck” correction path. Alternatives preserve the target where possible; when they change the task, Titan labels them as related rather than equivalent.
01
Build the manual baseline
Check all 15 items from the answer sheet without AI and record time, errors and workload.
Why this step exists
The team knows what unaided or non-AI capability actually is.
Success check
The team knows what unaided or non-AI capability actually is.
I’m stuck on this step
Reset: Re-read this authored instruction — “Check all 15 items from the answer sheet without AI and record time, errors and workload.” — and its success check, then attempt only this step.
Possible snag: Keep a plan no one has run.
Correction: Time and score a real drill.
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
02
Define failure signals
Write observable triggers: missing response after 20 seconds, source mismatch, impossible quantity or version warning.
Why this step exists
Fallback does not depend on a vague feeling.
Success check
Fallback does not depend on a vague feeling.
I’m stuck on this step
Reset: Re-read this authored instruction — “Write observable triggers: missing response after 20 seconds, source mismatch, impossible quantity or version warning.” — and its success check, then attempt only this step.
Possible snag: The result from “Write observable triggers: missing response after 20 seconds, source mismatch, impossible quantity or version warning.” does not yet meet this declared check: Fallback does not depend on a vague feeling.
Correction: Return to the start of “Define failure signals”, reduce complexity or pace, and repeat only the part needed to satisfy: “Fallback does not depend on a vague feeling.”
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
03
Name the safe state
On a trigger, stop accepting new AI output, preserve the last verified item and mark the queue PAUSED.
Why this step exists
No unverified item proceeds.
Success check
No unverified item proceeds; verify it in the stationery outage drill.
I’m stuck on this step
Reset: Re-read this authored instruction — “On a trigger, stop accepting new AI output, preserve the last verified item and mark the queue PAUSED.” — and its success check, then attempt only this step.
Possible snag: Continue accepting output while investigating.
Correction: Stop intake and label status clearly.
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
04
Prepare the fallback card
List the static source, manual check steps, owner, communication phrase and restart authority.
Why this step exists
A learner can follow the card without the AI.
Success check
A learner can follow the card without the AI.
I’m stuck on this step
Reset: Re-read this authored instruction — “List the static source, manual check steps, owner, communication phrase and restart authority.” — and its success check, then attempt only this step.
Possible snag: The result from “List the static source, manual check steps, owner, communication phrase and restart authority.” does not yet meet this declared check: A learner can follow the card without the AI.
Correction: Return to the start of “Prepare the fallback card”, reduce complexity or pace, and repeat only the part needed to satisfy: “A learner can follow the card without the AI.”
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
05
Run the normal condition
Process five items and verify a random sample against the answer sheet.
Why this step exists
Normal speed and error form the device-on comparator.
Success check
Normal speed and error form the device-on comparator.
I’m stuck on this step
Reset: Re-read this authored instruction — “Process five items and verify a random sample against the answer sheet.” — and its success check, then attempt only this step.
Possible snag: The result from “Process five items and verify a random sample against the answer sheet.” does not yet meet this declared check: Normal speed and error form the device-on comparator.
Correction: Return to the start of “Run the normal condition”, reduce complexity or pace, and repeat only the part needed to satisfy: “Normal speed and error form the device-on comparator.”
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
06
Inject and detect failures
Present one wrong flag, one 30-second delay and one outage; record trigger noticed and time to PAUSED.
Why this step exists
Each failure has a detection and safe-state timestamp.
Success check
Each failure has a detection and safe-state timestamp.
I’m stuck on this step
Reset: Re-read this authored instruction — “Present one wrong flag, one 30-second delay and one outage; record trigger noticed and time to PAUSED.” — and its success check, then attempt only this step.
Possible snag: Wait for obvious catastrophe during the stationery outage drill.
Correction: Prewrite observable trigger thresholds within the stationery outage drill.
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
07
Complete through fallback
Use the static sheet to finish the affected item and reconcile the queue; do not silently blend AI and manual results.
Why this step exists
Every item has one verified path and no duplicate.
Success check
Every item has one verified path and no duplicate.
I’m stuck on this step
Reset: Re-read this authored instruction — “Use the static sheet to finish the affected item and reconcile the queue; do not silently blend AI and manual results.” — and its success check, then attempt only this step.
Possible snag: The result from “Use the static sheet to finish the affected item and reconcile the queue; do not silently blend AI and manual results.” does not yet meet this declared check: Every item has one verified path and no duplicate.
Correction: Return to the start of “Complete through fallback”, reduce complexity or pace, and repeat only the part needed to satisfy: “Every item has one verified path and no duplicate.”
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
08
Restart deliberately
Only the named owner checks version, test item and queue state before restoring AI; then run five items manually after the exercise.
Why this step exists
Recovery and post-outage human performance are separately measured.
Success check
Recovery and post-outage human performance are separately measured.
I’m stuck on this step
Reset: Re-read this authored instruction — “Only the named owner checks version, test item and queue state before restoring AI; then run five items manually after the exercise.” — and its success check, then attempt only this step.
Possible snag: Turn AI back on as soon as it responds.
Correction: Give restart authority to a named person and use a test case.
Stop / get help: Stop for any unexpected external action, personal data, security event or unbounded output.
Correct versus incorrect execution
These accessible process diagrams are built from the tutorial’s own right/wrong teaching. They are not anatomical illustrations and do not add technique beyond the canonical tutorial.
Failure trigger — Early degradation can look fluent or merely slow.
Correct / safer
Use a timeout, source mismatch or impossible value.
Wrong / riskier
Wait for obvious catastrophe during the stationery outage drill.
Safe state — Unverified work can spread before the cause is known.
Correct / safer
Pause the queue and keep the last verified item.
Wrong / riskier
Continue accepting output while investigating during the stationery outage drill.
Fallback readiness — An untested fallback may be slower or wrong during outage.
Correct / safer
Practise from a current static source.
Wrong / riskier
Keep a plan no one has run.
Restart in stationery outage drill — Recovery without state reconciliation can duplicate or omit work.
Correct / safer
Reconcile version, test item and queue before resuming.
Wrong / riskier
Turn AI back on as soon as it responds.
Method-structure checklist
10 of 10 structural checks present
✓ Ordered, Power-specific instructions — present
✓ Every activity has a success check — present
✓ Materials or supplied records are declared — present
✓ Measurement or assessment rule is present — present
✓ Tutorial-specific troubleshooting is present — present
✓ Stopping or escalation boundary is present — present
✓ Every activity has an adjacent alternative — present
✓ Correct-versus-incorrect comparison is present — present
✓ Evidence context is bound to the Power record — present
✓ Planning metadata is present — present
The method-readiness band and presence checklist assess tutorial presentation and are separate from evidence quality for the underlying Power. They are automated editorial aids, not human approval.
Manual editorial sign-off: Pending. This tutorial must not display a human-approved state until an identified editor signs the exact content hash.