Yellow Phoenix: Semantic Search on Apple Silicon at the Candidate Ceiling — 95.3% Top-10 over Five Million Papers in 577 Microseconds

The measurement and verification record of Yellow Phoenix, a CPU-only semantic retrieval engine at five-million-paper scale on consumer Apple Silicon (Mac and iPhone). Over a frozen 1,000-query exam against a real 5,107,508-paper corpus (arXiv + PubMed): (i) a candidate-depth effect — exact re-ranking of the top-100 candidates produced by a binary HNSW index raises R@10 from 0.9049 (top-50, the prior production configuration) to 0.9538, and top-200 to 0.9802, with the re-ranker at the candidate-recall ceiling at both measured depths; (ii) cross-platform determinism — identical recall (to full reported precision) on an M3 MacBook and an A18 iPhone over the same 10M-document index and query stream; (iii) a calibrated confidence gate separating real queries from gibberish (0/1000 real-query false alarms, 65/65 gibberish caught), reproduced bit-identically across days and insensitive to candidate depth; (iv) swap-compatible storage tiers — a 0.98 GB PQ@192 store scores R@10 0.902 at candidate depth 100, indistinguishable from the 3.9 GB float16 store's result (overlapping 95% confidence intervals), under a reproducible exam gate; (v) a quantified encoder effect: identical graph construction under a superseded encoder generation climbs from 23.5% to only 83.1% hash-space R@1 as search effort rises 10x, where the current encoder already reaches 99.7%. A machine-readable claims layer (claims_layer.json) is embedded in the PDF as an attachment for machine verification. Companion to the Yellow Phoenix trilogy.