Day 4 answer key: Answers and explanations

Answers and explanations

Use after the class and problem setCheck the reason, not only the letter

Answer first

This key is released after submission. The questions it answers are on the pre-class check, the two studios and the problem set for this day. Work through them and commit to an answer before you read the explanations here: the point of every question is the reasoning, and a letter you have not argued for teaches you nothing.

Remember

When you correct an answer, identify the link you confused: representation, structure, retrieval, support, decision, or monitoring.

Pre-class concept check

Item	Answer	Why
P1	B	Changing pounds to pence changes the weighting inside distance, not the K-means algorithm.
P2	C	Silhouette describes separation in a representation; it does not supply time stability or substantive truth.
P3	A	Same-signed loadings and high explained variance support a broad co-movement description, not a causal factor or beta.
P4	B	Lexical and latent representations define similarity differently, so evidence must adjudicate the disagreement.
P5	C	Similarity explains rank; semantic support requires comparing each claim with the cited words.
P6	C	When the bounded corpus has no support, refusal is more accurate than a plausible invention.

In-class concept studio

Item	Answer	Why
C1	A	The left and right columns are nearest to the corresponding initial centres.
C2	B	Multiplying \(y\) by 10 multiplies its squared-distance contribution by 100.
C3	A	The value supports local internal separation only in the supplied geometry.
C4	A	The \(x\) direction has \(96/(96+16)=6/7\) of variance; variance is not importance or causality.
C5	A	Exact matching and latent co-occurrence can favour different chunks without implementation error.
C6	A	Hit@5 concerns presence within five results; MRR concerns how early the first known item appears.
C7	C	Retrieval succeeded; generation added unsupported content.
C8	A	Needless suspension is tied to false positives among the negative group, which differ sharply here.
C9	B	The decision process converts missing evidence into negative labels and feeds that error into retraining.
C10	C	Source support does not allocate authority, choose a proportionate threshold, or provide appeal.

Empirical studio

Item	Answer	Why
E1	A	Known-item evaluation asks whether the named passage appears near the top of this fixed collection.
E2	A	An identical candidate set isolates representation and scoring as the main comparison.
E3	B	Freezing prevents qrels from guiding rankings and identifies the exact object later scored.
E4	A	BM25 has better five-result coverage of known items; the latent method has a marginally earlier first hit on average.
E5	B	A valid citation identifier proves provenance, not that every clause is entailed.
E6	C	Ranking never proves that an answer exists; refusal keeps the claim inside the corpus boundary.

After-class problem set

Item	Answer	Why
S1	A	Without scaling, volatile return histories dominate Euclidean distance.
S2	B	Changing cases and variables changes grouping companies into describing directions of daily co-movement.
S3	B	The three checks answer separation, recognisability, and time stability; none establishes a natural taxonomy.
S4	A	PCA loadings and CAPM beta come from different questions and estimands.
S5	A	Exact terms and latent co-occurrence can legitimately generate different rankings.
S6	A	Existence plus poor rank indicates representation failure; absence indicates a coverage boundary.
S7	B	Known-item qrels are not exhaustive relevance judgements.
S8	B	The failure occurs before generation: the wrong or absent evidence was retrieved.
S9	C	Each generated clause must be supported or removed.
S10	A	A ranker always orders candidates, including collections with no answer.
S11	B	Sampling unflagged cases and preserving unknown outcomes repairs the selective-label mechanism.
S12	C	The combined response addresses decision authority, evidence quality, appeal, unequal harm, and monitoring.

← Day 4 lecture · All four days →