Datasets:
image imagewidth (px) 605 886 |
|---|
Fully Offline Semantic Search on Apple Silicon at the Candidate Ceiling
Measurement kit for "Yellow Phoenix: Semantic Search on Apple Silicon at the Candidate Ceiling" (Marc John Sawyer, 2026). The paper shows that quality on this engine is bounded by candidate recall — depth 100 recovers the ceiling cheaply — and that the numbers are deterministic across devices: the same index answers bit-for-bit identically on iPhone and Mac, separated only by latency. CPU-only, flash-resident, no GPU.
Paper & provenance
- Cite the paper (concept DOI, always resolves to latest): https://doi.org/10.5281/zenodo.23920218
- Sister repos: Watch · Cascade · Map
Verification bundle & integrity
- mac_paper_artifacts.tar.gz
— the clean-room verification bundle: frozen exam (seed 20260916, n=1000 queries over
5,107,508 papers), query embeddings/hashes, released candidates, exact ground truth, the
paper's machine-readable claims layer (SHA-256 gated), and
tools/check_exam.py— a numpy-only checker that needs no engine. Byte-identical to the Zenodo record.
| size | 3,368,407 bytes |
| md5 | 24cd4f3f8a8e3d5a4affec539698bced |
Verification videos (real runs, audio removed)
| File | Shows |
|---|---|
| iphone_4.9M_94_rerank_ladder_offline.mp4 | iPhone 16, airplane mode, 2026-11-07: the 4.9M-record candidate ceiling on the phone — graph tier R@10 81.5% @ ef=50 (p50 1114µs) jumping to 94.0% after the 1024-bit re-rank (94.6% @ ef=400), ef-ladder 50/100/200/400, memory gate OK (855 MB need / 3524 MB avail), graph freed after the run |
Full-resolution still of the run's final screen (airplane mode visible):
Figures (from the paper)
Figure gallery (renders inline)
Figure 1 ![]()
Figure 2 ![]()
Figure 3 ![]()
Figure 4 ![]()
Figure 5 ![]()
| File | Shows |
|---|---|
| mac_fig1_candidate_depth_ceiling.png | Candidate depth as a ceiling problem: the exact re-ranker hugs the candidate-recall ceiling at every depth |
| mac_fig2_tier_trade_4x_smaller_store.png | The tier trade: a 4× smaller store (0.98 GB vs 3.9 GB) costs 3.4 points at k=50, matches production at k=100 — exam-gate certified at build time |
| mac_fig3_determinism_two_devices.png | Determinism visually: identical recall rungs on iPhone 16 and Mac M3 Pro, separated only by latency |
| mac_fig4_calibration_gate.png | The calibration gate: real queries vs gibberish ceiling — zero overlap, shipped threshold separates with zero false alarms and zero misses |
| mac_fig5_encoder_effect_fingerprint.png | Fingerprint quality beats search effort: the current encoder holds 99.7–99.9% where the earlier corpus needed 10x effort |
Code
Engine source (Rust; macOS/iOS/watchOS/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. (Paper text CC BY 4.0.)
- Downloads last month
- 331