Applied review assessment calibration

This describes acceptable evidence, not text to memorize.

• AR1: tenant and id jointly identify a record; exact matching and construction-time snapshot are explicit. Caller authorization is assumed, not proven. Invalid-row handling is outside this trusted-fixture contract, not an invented defect.
• AR2: a direct (b,7) lookup can return Ada if tenant is omitted. After fixing that filter, (a,7) then (b,7) still fails with an id-only cache. These are separate counterexamples. A rename alone does not repair either.
• AR3: a recorded assertion exposes the wrong result, the repair makes it pass, and a changed or restored faulty implementation makes the relevant test fail again. Include (a,8) then (b,8) so cached absence is exercised. Import errors and tests asserting only “result exists” do not pass.
• AR4: changing the source row after construction and changing a returned dictionary both leave future results intact. The reference uses shallow copies because values are strings. A generalized nested schema needs a new ownership decision. No cache is valid.
• AR5: either cache or uncached lookup can pass with an honest tradeoff. A fixed-snapshot cache needs reconsideration when updates arrive. A safe fallback keeps tenant filtering; reverting to the defective candidate does not pass.
• AR6: the learner predicts and reproduces a new sequence after a delay, explains assumptions without the key, acknowledges help and does not claim real user testing or reviewer approval that never happened.

A passing automated suite supports AR3–AR4 but does not grade review quality, authorization architecture, integrity or independent understanding. Record manual judgment and the highest-risk remaining limitation.