← Foundry05 · Mission 1 of 3
The ranker is gameable.
Query: “what are the cache memory limits”. Both retrievers run below — exact BM25 and cosine over feature-hashed vectors. One page in this corpus repeats “cache memory” like a mantra. First, call which chunk actually answers the question — not which one wins.
BM25 rankingraw corpus
- #1the keyword-stuffed page · score 7.06
cache memory cache memory cache memory cache memory cache memory limits limit limits — the cache memory limits…
- #2the real memory spec · score 2.38
The inference server allocates a KV cache memory pool of 4 GB per GPU. The memory pool is shared by all concur…
- #3the auth chapter · score 0.00
Authentication uses short-lived tokens signed with a rotating key. Tokens expire after fifteen minutes and mus…
- #4the billing chapter · score 0.00
Billing is computed monthly from metered usage. Invoices separate input tokens from output tokens because outp…
vector top-1: the keyword-stuffed page (cosine 0.90) — fooled too: L2-normalized hashing cannot see repetition