Independence: Verifying the Loop When AI Is on Both Sides

All courses / AI Purple / Module 4 / Chapter 15

Learning objectives

After the lesson the participant can:

• Explain correlated failure and how it inflates a coverage record, separating what the research measured from the book's inference.
• Classify the dependencies of a chain of tester, tested and checker into the five kinds of independence.
• Draft a one-page independence statement for a coverage claim, with controls, evidence, residual risk and expiry.
• Review the debate between concern and calm and state the chapter's verdict.

Key concepts

Correlated failure

The case where tester and tested share a model family, training data, authors, vendor or time, so that their errors coincide and a passing result shows agreement and not verification. The inference from the cited research to the loop is the book's own.

Five kinds of independence

Model, data, authorship, vendor and time, each with a shared dependency, what it hides, a control and what is shown. The kinds are not ranked, and a kind that is not shown is treated as shared.

Held-out scenario

A scenario written by a person, kept where the detection side and its agents cannot read it, and run only after the detections are built. The gap between open and held-out pass rates measures generalisation, and the set is where human-authored novelty enters.

Seeded gap

A known weakness introduced on purpose, by a person, under the signed rules of engagement and in a place where it can be undone, to test whether the loop that is meant to find gaps finds this one, how fast and by what.

Independence statement

A one-page, signed record per coverage claim: the claim, the chain, the shared dependencies, the controls and their evidence, the residual risk in words, the evidence age and an expiry, by default 90 days for priority techniques.

Interactive exercises

Do them here and get feedback at once. Choose a card and then where it belongs, or drag it.

Teaching plan

Discussion

1. The chapter says Nordvik Region's independence is by contract and cheap to show, while Hansa Mutual's is by design and cheap to lose. What does cheap to lose look like in day-to-day engineering?
2. Simulation markers protect against mistaking an incident for an exercise, yet a marker visible to an agent under test weakens the test. How would you resolve that tension in your own design?
3. If independence checks catch nothing over a long period, either the loop is sound or the checks are too weak. How could you tell which?

Written exercise

Self-test

Choose an answer to see whether it is right, or open the answer below each question.

1. Which control does the chapter pair with a shared model family between tester and tested?

   1. A re-test run immediately after the fix, on the fix author's own data and by the fix author's agent.
   2. A critique agent on a second model family, used for questions and not verdicts.
   3. A longer retention period for the exercise telemetry.
   4. A higher autonomy level for the re-test runner.
   Show the answer

   B. A critique agent on a second model family, used for questions and not verdicts.
   For the model kind the control is a critic on a different family, used for questions and not verdicts, with the declared families and a log of the critic's disagreements as evidence. An immediate re-test is the failure the time kind addresses.

2. What is the chapter's verdict on a loop closed by machines alone?

   1. It is unsafe in every respect and regression should return to people.
   2. It is sufficient once two different model families are in use.
   3. It verifies coverage for as long as the pass rate on the open scenario library stays above 90 per cent.
   4. It can show that the known still works, and cannot by itself show that its coverage is real.
   Show the answer

   D. It can show that the known still works, and cannot by itself show that its coverage is real.
   The verdict keeps mechanical checks such as replay automated, and holds that a machine-closed loop shows regression but not that the scenarios were the right ones. Second-family critics, held-out scenarios, seeded gaps and reproduction by a person are required for claims a board relies on.

3. Hansa Mutual's statement IND-HM-2028-Q1 assessed seven of 25 verified cells. How does the chapter say the board figure of 25 should then be read?

   1. As seven cells failed or at limited independence, and eighteen not yet assessed.
   2. As 25 verified cells, since the statement found no failures at all.
   3. As seven verified cells, because the other eighteen are now unverified.
   4. As 18 verified cells, with the seven assessed cells removed from the board figure until the next review.
   Show the answer

   A. As seven cells failed or at limited independence, and eighteen not yet assessed.
   Of the seven assessed, two failed a held-out re-test and five remain verified with limited independence. The other eighteen are not assessed by the statement, so the figure of 25 carries that qualification.