1. Why Legacy PDFs Break on Phones & Screen Readers — Dual-Plane Architecture
Standard PDFs act like painted graphics: glyph coordinates are fixed to letter-size paper, while structural tags (/StructTreeRoot) are optional, frequently missing, or silently drifted from visual reading order. When zoomed to 400% (320px width) or parsed by assistive technology, tables lose column scope, minus signs drop, and rotated blueprint dimensions scramble.
PostSheet separates presentation from semantics using a dual-plane browser architecture:
- Zero-Copy Compute & Visual Plane: Compiled Godot 4.6.2 WebGPU WASM (
6.35 MiBBrotli-11 Mobile-Lite binary,0 FPSidle loop) executesFHT-RaBitQ(1-bit / 4-bit) andINT8vector similarity scoring (score_r7.glsl) + hierarchicalBitonic-64Top-K sorting directly on the visitor's GPU, binding the exact output storage buffer handle (topk_out_rid) into the overlay pass without CPU corpus readback (512 Basync readback vs.800,000 Bnaive bridge). - Semantic ARIA & Signed Document Plane: Published packages use
CRR-DOC/3binary snapshots and0x31node-deltas verified locally with pinned WebCryptoECDSA P-256. User-dropped PDFs are parsed locally via self-hostedpdf.jsand optional OPFS-cachedtesseract.jswith zero outbound uploads (enforced by strict CSPconnect-src). Unconfirmed visual grids emit0guessed<th>elements until a human confirms Row 1. - Hardware-Adaptive Mobile Guarding:
classifyClientTier()scopes the<= 4core /<= 3 GBCPU_ONLYcutoff strictly to budget Android (0 BWASM downloaded), routes iOS Safari (clampedhardwareConcurrency === 4) and 4 GB Android devices to cancelableMOBILE_WEBGPU_ON_DEMAND(RaBitQ-1,maxDpr: 1.0, 1 MiB OPFS Range chunks), and enforces a runtime auto-demotion watchdog ondevice.lost, 5 consecutive slow frames (> 33.3 ms), or> 210 MiBUA-specific memory.
2. Real Hardware Comparison: RTX 3060 Desktop vs. Physical Android ADB (R12)
All metrics below are measured directly on live hardware. Unplugged physical devices are reported transparently rather than emulated.
| Device | Real RAM / SoC | Detected Tier | Cold FCP | Initial Load Bytes | CPU JS Search P50 | Client-GPU Search P50 | Speedup | Real OS Memory (PSS / RSS or PrivWS) | Top1=217 Parity | axe-core Violations |
|---|---|---|---|---|---|---|---|---|---|---|
| Desktop Calibration Bench (RTX 3060 / Chrome 157) | RTX 3060 Laptop GPU (Ampere, D3D12 Dawn) | DESKTOP_WEBGPU_FULL (RaBitQ-4) |
205.75 ms (Slow 4G 150ms RTT) | 35.0 KiB pre-GPU; 0 B warm OPFS | 29.26 ms (INT8) / 24.96 ms (FP32) |
2.04 ms GPU kernel / 5.32 ms RT (RaBitQ-4); 1.01 ms GPU / 4.53 ms RT (RaBitQ-1) |
5.51x RT / 14.36x kernel (RaBitQ-4); 6.46x RT / 29.07x kernel (RaBitQ-1) |
54.09 MiB UA mem / 259.3 MiB OS Private WS (Mobile-Lite) | PASS (8/8) | 0 |
| Real samsung SM-S908U (ADB Chrome 157.0.8103.78) | 7.05 GiB RAM / SM8450 (8c, Android 14) | MOBILE_WEBGPU_LITE (rabitq4) |
1328.89 ms (warm 297.67 ms) | 620.3 KiB (WASM: 0 B, BIN: 0 B) | 100k R1: 2.04 ms / R4: 25.42 ms; Doc: 4.87 ms | R1: 16.57 ms RT (P95 20.94 ms); R4: 17.33 ms RT (P95 24.5 ms) | 0.12x (R1); 1.47x (R4) | PSS 341.55 -> 323.77 MiB / RSS 451.71 MiB (ΔT 1.5°C) | R1 8/8; R4 8/8 | 0 |
| Real motorola moto g - 2025 (ADB Chrome 157.0.8103.78) | 3.56 GiB RAM / MT6835 (8c, Android 16) | MOBILE_WEBGPU_ON_DEMAND (budget_4gb_android_on_demand) |
1001.71 ms (warm 446.75 ms) | 617.4 KiB (WASM: 0 B, BIN: 0 B) | 100k R1: 3.34 ms / R4: 45.58 ms; Doc: 6.91 ms | R1: 21.86 ms RT (P95 24.78 ms); R4: 22.32 ms RT (P95 33.75 ms) | 0.15x (R1); 2.04x (R4) | PSS 201.84 -> 197.26 MiB / RSS 202.28 MiB (ΔT 0°C) | R1 8/8; R4 8/8 | 0 |
3. Quantization Pareto Frontier & Zero-Copy RID Ablation (100k x 256D, R10)
| Tier | Corpus Size | Compression vs. FP32 | Recall@8 | NDCG@8 | GPU Kernel P50 | Round-Trip P50 |
|---|---|---|---|---|---|---|
| CPU JS FP32 Exact | 102.40 MB | 1.00x | 1.0000 | 1.0000 | — | 24.955 ms |
| CPU JS INT8 | 25.60 MB | 4.00x | 0.9952 | 0.9998 | — | 29.260 ms |
| Godot WebGPU INT8 | 25.60 MB | 4.00x | 0.9952 | 0.9998 | 6.426 ms | 8.815 ms |
| Godot WebGPU RaBitQ-4 | 13.20 MB | 7.76x | 0.8317 | 0.9843 | 2.038 ms | 5.315 ms |
| Godot WebGPU RaBitQ-1 | 3.60 MB | 28.44x | 0.3798 | 0.7836 | 1.007 ms | 4.530 ms |
Zero-Copy RID Ablation (100 Active Frames): Binding topk_out_rid directly into the render pass cut GPU-to-CPU bus readback by 1,562.5x (512 B vs. 800,000 B), eliminated 12,800 B/query CPU-to-GPU geometry re-upload, reduced JS heap peak delta by 71.1x on RaBitQ-4 (0.42 MiB vs. 30.09 MiB), and improved P50 frame time by 1.25x–1.31x (100% within the 16.6 ms 60-FPS budget).
Scaling Inflection (10k to 500k x 256D): Linear scaling verified through 500,000 vectors (RaBitQ-1: 17.17 MiB shard / 21.70 MiB VRAM, 4.66 ms GPU P50; RaBitQ-4: 62.94 MiB shard / 67.48 MiB VRAM, 10.32 ms GPU P50; Top1=217 verified at all scales). Server-side AI compute cost is $0.00 per 1,000,000 queries ($0.031/mo R2 hosting at 10k cold visitors vs. $815.88/mo always-on cloud GPU + vector cluster baseline).
4. Formal Conservation Gates vs. veraPDF 106/0 & Honest Human-AT Status (F13)
- Ordered Structure Conservation (Z3 5.4.1 + pikepdf 10.18.1): In controlled mutation tests, legacy validator
veraPDFreported 106 passed / 0 failed on all four intentionally corrupted PDFs (H1 → Pdowngrade, MCID reading-order swap, table-header swap, and negative-sign drop). PostSheet's Z3 UnsatCore conservation gate quarantined 5/5 fixtures accurately. - Automated Accessibility & Windows UIA Proofs (R9 & R10):
0offlineaxe-coreWCAG 2.0/2.1 AA violations across all themes and320px(400% zoom) reflow,719/719non-empty WindowsIUIAutomationbounding rectangles + validSAPI.SpVoicePCM WAV synthesis (698,072 B), and100.0%word/number/signed-value conservation on both native-text and lazy browser-OCR scanned PDFs with0outbound requests. - Open Human Assistive-Technology Testing Boundary (F13): Automated
axe-core, Z3 SMT, and Windows UIA gates verify DOM/AOM structure and adapter integrity, not full human certification. Structured walkthroughs with live human testers on NVDA, JAWS, macOS/iOS VoiceOver, and Android TalkBack remain an explicit manual validation requirement.
5. Worst-Case Physical Mobile Silicon Stress & Scaling Ceiling (S22 Ultra Adreno 730)
To establish hard mobile hardware bounds for multi-million-vector browser workloads, the compiled Godot 4.6.2 WebGPU Mobile-Lite runtime was pushed over USB ADB directly to the Qualcomm Adreno 730's 128 MiB single-buffer binding ceiling (maxStorageBufferBindingSize = 134,217,728 B, maxBufferSize = 1,024 MiB, timestamp-query = true) with unique tail-planted needles verified 15/15 at every peak scale.
| Quantization | Vectors (256D) | Shard Size (% of 128 MiB) | Adreno GPU Kernel P50 | GPU Round-Trip P50 | S22 CPU JS P50 | Mobile GPU Speedup | OS Chrome PSS & Temp | Needle Verified |
|---|---|---|---|---|---|---|---|---|
| RaBitQ-1 | 1,000,000 | 34.33 MiB (26.8%) | 44.630 ms | 58.940 ms | 153.05 ms | 2.60x RT / 3.43x kern | 301.73 MiB (39.5°C) | 217 (15/15) |
| RaBitQ-1 (Peak) | 3,500,000 | 120.16 MiB (93.9%) | 155.845 ms | 177.735 ms | 94.29 ms (post-JIT) | 22.46 Mvec/s flat linear | 299.23 MiB (40.0°C) | 3,499,783 (15/15) |
| RaBitQ-4 | 500,000 | 62.94 MiB (49.2%) | 40.174 ms | 53.705 ms | 161.63 ms | 3.01x RT / 4.02x kern | 298.92 MiB (40.1°C) | 217 (15/15) |
| RaBitQ-4 (Peak) | 950,000 | 119.59 MiB (93.4%) | 77.529 ms | 95.090 ms | 307.00 ms | 3.23x RT / 3.96x kern | 298.95 MiB (40.1°C) | 949,783 (15/15) |
| INT8 (Peak) | 500,000 | 122.07 MiB (95.4%) | 73.728 ms | 90.735 ms | 309.32 ms | 3.41x RT / 4.20x kern | 295.96 MiB (40.3°C) | 499,783 (15/15) |
- 300-Query Sustained Thermal & Zero-Leak Endurance (500k RaBitQ-4 + Active Zero-Copy Spatial Overlay): Sustained
24.0–25.6 QPSacross 6 rolling 50-query windows with a1.06xthrottle ratio (34.61 → 36.85 msRT P50),+0.1°Cbattery delta (40.3 → 40.4°C),-0.23 MiBOS Chrome PSS delta (zero leak), and298/300(50/50in all 5 steady-state windows) exactTop1=217parity. - Controlled
>128 MiBBoundary Probe & Recovery: Deliberately requesting1,050,000RaBitQ-4vectors (132.18 MiB) triggered Dawn's exact validation guard (138600000 > 134217728) with zero Chrome crash/OOM (tabStillAlive: true) and immediate clean recovery to100kTop1=217(5/5). Because Adreno 730 exposesmaxBufferSize = 1,024 MiB(8xthe128 MiBbinding slice), multi-offset binding slices provide a direct Phase II path to~29.8MRaBitQ-1/~8.1MRaBitQ-4vectors per device.