Three Operator Cases: evidence and optional practice

The concept presentation stands on its own. Prepared numerical figures and recorded model choices explain each case. The notebook labels below only locate optional practice and supporting records.

Recorded studies retain their own inputs, assumptions and model scope. Optional new calculations may differ from those records. The full source guide preserves provenance and calculation details.

Slide 4: Solar settings: which option should we try?

Notebook step: L2

Compare possible settings, then check voltage and line current. The agent can organize these trials and report the checked options.

Open the notebook or source

Monica Lam. CS224V Agentic AI, Lecture 1: Introduction. Stanford University. Fall 2026.

Slide 6: Restoration: who should get power first?

Notebook step: L6

The operator sets the service priority. That priority changes the study the agent requests.

Open the notebook or source

Monica Lam. CS224V Agentic AI, Lecture 1: Introduction. Stanford University. Fall 2026.

Slide 7: Keep three jobs clear

Notebook step: L9

The agent requests a study. The tool calculates. The operator reviews the result.

Open the notebook or source

Monica Lam. CS224V Agentic AI, Lecture 1: Introduction. Stanford University, Fall 2026, p.53.

Monica Lam. CS224V Agentic AI, Lecture 1: Introduction. Stanford University. Fall 2026.

Slide 8: Wrong readings can still fit a model

Notebook step: L8–L9

A good fit alone does not prove that the readings are true. The agent can organize checks and flag missing evidence.

Open the notebook or source

Monica Lam. CS224V Agentic AI, Lecture 1: Introduction. Stanford University. Fall 2026.

Slide 9: Case 1: which solar settings should we use?

Notebook step: L2

Use solar generation while keeping voltage and line current within the allowed range. An agent can organize the trials and return checked options.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 12: Two solar settings change the result

Notebook step: L2

One setting controls solar power output. The other provides voltage support. The tool checks voltage, line current and the solar unit’s capacity.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 13: Both proposals face the same checks

Notebook step: L2

The tool calculates the effect of each proposal. Compare their scores only after voltage, current and capacity checks pass.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 14: Read the voltage and current checks

Notebook step: L2

Left: voltages should stay in the allowed band. Right: current should stay below 100% of the allowance.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 15: Both pass, but the trade-off is different

Notebook step: L2

MPC uses 73.62% of the current allowance; SAC uses 51.68%. MPC has the lower overall score because current is only one part of the goal.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 16: What if we allow less current?

Notebook step: L2

Reduce the allowance from 80% to 75% of nominal current. Which proposal might fail? Which would you choose if both pass? What must the agent ask the tool to check again?

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 17: Both proposals pass the new current check

Notebook step: L2

MPC has the lower score. SAC uses less current and less solar output. These results cover the medium-voltage model. A wider grid needs a wider study.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 18: Less current does not mean the best option

Notebook step: L2

Here SAC also uses less solar output. The operator’s goal values solar generation as well as operating limits.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L2; cell classroom-007; src/continuous_control.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 19: A PASS covers only the checked model

Notebook step: L5

A study of the medium-voltage network cannot confirm a whole-grid claim. The agent should request the missing study or return HOLD.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L5; cell classroom-016; MV-only default model. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 20: Recorded example: the first option fails

Notebook step: L10B

Llama requests a check of the saved SAC proposal. The numerical tool reports too much current and rejects it.

Open the notebook or source

Workshop recorded model replies and numerical study records. Voltage P3_current_reserve: Llama 3.1:8b planner, recovery and review calls 1–3. Public teaching archive. 27 September 2026.

Slide 21: Recorded example: another option is checked

Notebook step: L10B

Llama requests a check of the existing MPC fallback. The returned electrical checks pass for this recorded case.

Open the notebook or source

Workshop recorded model replies and numerical study records. Voltage P3_current_reserve: Llama 3.1:8b planner, recovery and review calls 1–3. Public teaching archive. 27 September 2026.

Slide 22: Recorded example: the operator receives a proposal

Notebook step: L10B

The agent returns the checked proposal and its conditions. The operator still reviews it before any authorized action.

Open the notebook or source

Workshop recorded model replies and numerical study records. Voltage P3_current_reserve: Llama 3.1:8b planner, recovery and review calls 1–3. Public teaching archive. 27 September 2026.

Slide 23: Case 2: who gets power after a fault?

Notebook step: L6

A fault removes a supply path. Available capacity limits what can be restored. The operator sets priorities; the agent can request and check restoration options.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 24: How the agent helps restore service

Notebook step: L6

Choose the goal: critical loads first, or the most total power. The agent can request a study, read the trade-off and ask for a new study when limits change.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 25: The service goal changes the study request

Notebook step: L6

Critical-first gives more weight to essential service. Most-total-power aims to restore the largest amount of power.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 26: The goal changes who receives power

Notebook step: L6

Critical-first: 130 kW total, including 90 kW for critical loads. Most-total-power: 140 kW total, including 30 kW for critical loads.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 27: Check the restoration plan

Notebook step: L6

Keep the faulted branch out of service. Check supply paths and capacity. Voltage, protection and field switching need separate studies.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 28: What if the backup path can carry more power?

Notebook step: L6

Increase its capacity from 40 kW to 180 kW. Keep the fault and loads the same. Would the two service goals still give different results? What should the agent recalculate before reporting the new plan?

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 29: More capacity removes this service trade-off

Notebook step: L6

At 40 kW, critical-first serves 90 kW of critical load instead of 30 kW. At 180 kW, both goals serve 250 kW total, including all 90 kW of critical load.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 30: A new limit needs a new study

Notebook step: L7

Now reduce the allowed backup capacity from 40 kW to 20 kW. The old result no longer supports the request. The agent must request a new study.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L7; cell classroom-021; request/evidence binding. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 31: The agent keeps the operator’s goal

Notebook step: L6

The operator sets priorities. The solver selects loads. The agent can request the solve and explain the returned service trade-off.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L6; cell classroom-018; src/standard_formulations.py. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 32: Recorded example: another plan restores more service

Notebook step: L10B

In this separate larger-feeder study, Llama requests another plan. The tool reports more service and passing electrical checks.

Open the notebook or source

Workshop recorded model replies and numerical study records. Feeder-catalog restoration: Llama 3.1:8b refinement and ranking calls 14–16. Public teaching archive. 27 September 2026.

Slide 33: Recorded example: the tool rejects a trial

Notebook step: L10B

The next trial carries too much current. It cannot replace the previous passing option.

Open the notebook or source

Workshop recorded model replies and numerical study records. Feeder-catalog restoration: Llama 3.1:8b refinement and ranking calls 14–16. Public teaching archive. 27 September 2026.

Slide 34: Recorded example: a wrong ranking is corrected

Notebook step: L10B

Llama puts two passing plans in the wrong cost order. Python keeps the lower-cost plan and checks it again.

Open the notebook or source

Workshop recorded model replies and numerical study records. Feeder-catalog restoration: Llama 3.1:8b refinement and ranking calls 14–16. Public teaching archive. 27 September 2026.

Slide 35: Case 3: can we trust these readings?

Notebook step: L8

Wrong readings can still look consistent with the model. The agent can request checks, but action still needs a trusted basis.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L8; cell classroom-023; four-bus DC fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 36: How the agent helps with suspicious readings

Notebook step: L9

The agent can request the checks and keep missing evidence visible. The operator reviews trust, model coverage and the proposed correction.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L9; cell classroom-026; fixed trusted-recovery fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 37: Mismatch is only one check

Notebook step: L8

Mismatch J measures how well the model fits the readings. The simulation lets us measure how wrong the estimate is (state error).

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L8; cell classroom-023; four-bus DC fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 38: Two patterns of wrong readings

Notebook step: L8

One wrong reading may conflict with the other readings. Coordinated changes may still look consistent with the model.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L8; cell classroom-023; four-bus DC fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 39: Read the alarm and the estimation error

Notebook step: L8

Left: mismatch and its alarm threshold. Right: error against the simulation’s known truth. A quiet alarm can appear alongside a growing error.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L8; cell classroom-023; four-bus DC fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 40: A coordinated change can escape the alarm

Notebook step: L8

At scale 1, one biased reading gives mismatch J=277.50 and an alarm. A coordinated change gives J=3.92, no alarm, and 0.02678 rad of state error.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L8; cell classroom-023; four-bus DC fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 41: What if the data change is small?

Notebook step: L8

Reduce the change to one tenth of the reference size. Will the alarm always detect the error? What else would the operator need before trusting the readings?

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L8; cell classroom-023; four-bus DC fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 42: A quiet alarm can still hide an error

Notebook step: L8

Neither example triggers this alarm at one tenth of the reference size. Both estimates still contain error. No alarm alone does not establish data trust.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L8; cell classroom-023; four-bus DC fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 43: Trusted information is needed before recovery

Notebook step: L9

Without the required trusted readings, the response is HOLD. With readings assumed trustworthy, the tool can calculate a proposal for review.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L9; cell classroom-026; fixed trusted-recovery fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 44: The tool calculates a small correction

Notebook step: L9

The solver finds the smallest allowed adjustment that meets the checked branch limits. Predicted flow: 0.17000 p.u. Offline check: 0.17046 p.u. Limit: 0.18000 p.u.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L9; cell classroom-026; fixed trusted-recovery fixture. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Slide 45: Recorded example: request the recovery tool

Notebook step: L10B

Llama requests the supplied recovery tool. Python calculates a correction using the assumed trusted readings.

Open the notebook or source

Workshop recorded model replies and numerical study records. Standard four-bus cyber study: Llama 3.1:8b tool selection and REVIEW calls 7–8. Public teaching archive. 21 September 2026.

Slide 46: Recorded example: keep human review

Notebook step: L10B

The model checks pass; human review is still pending. Data trust and operating permission still need review.

Open the notebook or source

Workshop recorded model replies and numerical study records. Standard four-bus cyber study: Llama 3.1:8b tool selection and REVIEW calls 7–8. Public teaching archive. 21 September 2026.

Slide 47: Review must refer to the same proposal

Notebook step: L12

The proposal ID and model scope must match the checked study. A different proposal or wider scope needs a new review.

Open the notebook or source

Classroom Labs notebook. Classroom_Labs.ipynb, L12; cell classroom-033; simulated readback only. Public notebook and Python source. 24 October 2026; stable notebook cell IDs checked.

Check the alignment

Code manifest records each canonical notebook SHA-256, stable cell ID and inclusive source-line range. This concept edition contains no executable slide excerpts. Numerical code and archived records remain in the companion repository.

Classroom numerical cells perform fresh calculations when run. Displaying archived model replies does not call an LLM. Editing an earlier classroom parameter does not automatically change an independent archived study or fixed recovery fixture. The notebook and slide notes explain those boundaries.