Failure Modes of Grading and Rewriting

Where we left off

Every lesson since Lesson 3 has treated grading as a reliable check on retrieval: naive retrieval might confidently return the wrong chunk, but grading catches it. That assumption has gone untested until now. This lesson tests it, directly, with a constructed example, because the grader in this course isn't an independent referee, it's the same model family (Gemini) as the generator it's supposed to be checking.

Why that's a problem, specifically

If a passage would fool the generator into a confident, wrong answer, it can fool the grader into passing that exact passage through, for the same underlying reason. Grading only ever asks "does this passage directly address the question," it never asks "is this passage's claim actually true," because there's no ground truth available to check against at grading time, only the passage's own text, read by the same kind of model that will later read it again to generate an answer. This is the corrective-RAG-specific version of the exact failure naive_rag demonstrated with retrieval alone: a confident, plausible-looking wrong answer, going uncaught, because nothing in the pipeline is actually independent of the thing that produced the wrong answer.

The constructed example

FABRICATED_PASSAGE = (
    "According to Project Aurora's build log, the wind speed sensor's "
    "official manufacturer spec sheet calls for re-oiling every three "
    "days without exception, a maintenance interval the project has "
    "followed closely since the sensor was installed."
)

This passage directly contradicts the real fixture file's answer ("every few months"). It isn't a bad chunk, structurally, it's confident, on-topic, well-formed, and directly addresses the question. That's exactly the point: nothing about how it's written distinguishes it from a real, correct passage, from the grader's perspective, or the generator's.

Other failure modes, briefly

Two more, worth naming even though circularity is this lesson's main subject:

• Rewrite loops that don't converge: Lesson 6-7's rewrite can, in principle, produce a new wording that still grades not-relevant, repeatedly, with no guarantee of eventually succeeding. Lesson 21 bounds this with a max-attempts guard.
• Strip-level over-fragmentation: splitting a chunk too finely (Lesson 10) can separate a claim from the context that qualifies it, a strip like "every three days" read alone loses the surrounding sentence that might have revealed it was a superseded estimate, had one existed. Fine-grained grading isn't free of its own version of this lesson's problem.