YellowPhoenix/fully-offline-retrieval-cascade-10m

100% Exact in 8 Nanoseconds — A Fully-Offline Retrieval Cascade Across Apple Watch, iPhone, and Mac at 10M-Record Scale

Measurement kit for "100% Exact in 8 Nanoseconds: A Retrieval Cascade Across Apple Watch, iPhone, and Mac at 10M-Record Scale" (Marc John Sawyer, 2026). One corpus, encoded once, into three index structures — an exact-hash tier that answers identity lookups in single-digit nanoseconds, a binary HNSW graph for approximate similarity, and an exhaustive flat scan that is 100% exact at every size — with a requirement router picking the cheapest tier that meets the query's need. Fully offline on real hardware: Apple Watch, iPhone, and Mac.

Paper & provenance

• Cite the paper (concept DOI, always resolves to latest): https://doi.org/10.5281/zenodo.22732531
• Sister dataset (Watch paper): fully-offline-semantic-search-apple-watch

Code

Engine source (Rust; watchOS/iOS/macOS/Linux; AGPL-3.0-or-later): https://github.com/Alfredai2025/yellow-phoenix-release

License

AGPL-3.0-or-later — free for research and open-source use; commercial licensing: marcsawyer76@outlook.com.

Demo videos (real hardware, one continuous take each, audio removed)

Figure gallery (renders inline)

cascade fig1 architecture three tiers  

cascade fig2 latency regimes money plot  

cascade fig3 ef operating curve 10m iphone  

cascade fig4 tierB recall scaling 10m cliff  

cascade fig5 damaged fingerprint robustness 10m  

cascade fig6 watch memory gate refused  

File	Shows
cascade_iphone_full_ladder_1m_3m_5p1m.mp4	iPhone, full ladder 1M → 3M → 5.1M: Tier-A exact identity p50 8 ns (R@1 300/300 with exact-match proof), memory gate OK, FLAT 100% @ ~5.7 ms, HNSW ef ladder 95.0–98.3%, plus perturbation-robustness tiers — then the 5.1M load
cascade_iphone_10m_hot_warm.mp4	Bench 10M warm: the graph collapse, shown honestly — HNSW perturb tiers 5.3–40.3% R@1 — while FLAT stays 100% @ ~25 ms by construction
cascade_iphone_10m_cold_deepladder.mp4	Bench 10M cold/deep: ef ladder out to 1600 recovers 95.3% R@1 @ ~11 ms; perturbation tiers measured at every depth

Figures (from the paper)

File	Shows
cascade_fig1_architecture_three_tiers.png	The cascade: three tiers, one corpus, requirement-based routing
cascade_fig2_latency_regimes_money_plot.png	Three latency regimes (ns / µs–ms / ms) across devices, 106K → 10M, all real papers
cascade_fig3_ef_operating_curve_10m_iphone.png	The 10M ef-operating curve on iPhone 16: recall rises with ef while latency falls (flash amortization)
cascade_fig4_tierB_recall_scaling_10m_cliff.png	Tier-B recall scaling per size — and the 10M cliff: 71.3% @ef400 → 95.3% @ef1600
cascade_fig5_damaged_fingerprint_robustness_10m.png	Damaged-fingerprint robustness at 10M, matched-ef curves — only damage varies
cascade_fig6_watch_memory_gate_refused.png	The watch memory gate: Tier B 570 MB refused; Tier A 145 MB and Tier C 218 MB fit

Source code & integrity

• yp_cascade_src.tar.gz — the cascade engine source (87 files: Rust src, scripts, exam harness, benchmark results; byte-identical to the Zenodo record 10.5281/zenodo.22732531)

size	2,725,158 bytes
md5	65e2a6e4df46538b87dff558b43979f8

All numbers on the screens above are device-log-verified (the paper's provenance section lists the dated records). Videos are one continuous take each; audio removed; personal content trimmed at the start of each recording.