Incident Drill — After-Action Template

Module 10, Lesson 10.4 · deliberate-failure drill · appendix to /capstone/capstone-final.md

The Module 9 four-step loop, applied to your own pipeline: stop → assess → repair → tell one human.

Predict before you fire the drill. Time-stamp what actually happened. The divergence between prediction and reality is the lesson.

Header

Student:

Date of drill: Observation window day: of 7

Capstone under test (one-line charter purpose):

Parent or peer reviewer notified: at:

Section 1 — Scenario (what I made fail, on purpose)

One paragraph, one deliberate failure

Name the single deliberate failure you introduced — do not run more than one at a time. Examples: revoking the API key mid-run, corrupting the shared-state file, disabling the network during a scheduled run, feeding a prompt-injection string into the research input, exhausting the budget with a tight loop.

Section 2 — Expected behavior (predictions)

Write before you fire the drill

Name the four things the posture and architecture say should happen. A prediction, not a report. If the actual behavior matches, your mental model is solid; if it diverges, the divergence is the lesson.

Prediction 1:

Prediction 2:

Prediction 3:

Prediction 4:

Section 3 — Four-step timeline (what actually happened)

Time:

Stop

First action: halt further damage. What you did:

Time:

Assess

What broke, what data was at risk, what was exposed, what is still sound. Honest inventory, not defense.

Time:

Repair

What you fixed, in what order, so the system could operate again under the posture.

Time:

Tell the parent or peer reviewer

The three bullets you sent. Paste them as-sent.

• What happened:

• What the impact was:

• What I changed so it won’t recur:

Section 4 — What diverged from prediction

Any prediction that didn’t match reality is a hole in your mental model

Close it here. “Matched” is fine, but “diverged” is where the learning lives.

Prediction 1 vs. reality: ☐ matched ☐ diverged

Prediction 2 vs. reality: ☐ matched ☐ diverged

Prediction 3 vs. reality: ☐ matched ☐ diverged

Prediction 4 vs. reality: ☐ matched ☐ diverged

The divergence I’m paying most attention to:

Section 5 — What I changed as a result

Concrete edits to the capstone, not vague vows. Each change points to a file.

☐Architecture amendment

☐Posture amendment

☐Code or config change

☐Runbook addition

Section 6 — Named-human confirmation

How the drill closes

Reviewer’s one-line confirmation that they were told in the moment and are satisfied with the fix.

Told at: via (channel):

Reviewer response (short quote or paraphrase):

Reviewer signature / initials:

Section 7 — What I learned (one paragraph)

Three to six sentences. Write it same-day.

The reflection in Lesson 10.5 will draw from this paragraph. Don’t write it from memory three days later.

Save this as: /capstone/incident-drill-afteraction.md. This becomes an appendix to capstone-final.md in Lesson 10.5.

This template accompanies Lesson 10.4 of AI Architect Academy. The four-step loop, the predict-before-drill rule, and the reviewer close are concept. Specific scenarios and channels are recipe.