Upload PARITY.md with huggingface_hub

PARITY.md

@@ -263,3 +263,92 @@ Browser-load bugs fixed to get here (details in CLAUDE.md worklog cont. 7):
 `env.allowLocalModels` browser default, upstream `use_external_data_format`
 catch-all key forcing a phantom `model_quantized.onnx_data` fetch (hang at
 100%), `.mjs` served as octet-stream breaking the webgpu EP's dynamic import.
+
+## fp16 — parity vs PyTorch fp32 (model_fp16.onnx)
+
+- logits max abs diff: **1.491**
+- per-token argmax agreement: **775/775 (100.00%)**
+- span exact match (start,end,label): **61/61 (100.0%)** — 0 only-PyTorch, 0 only-ONNX
+- gold coverage: PyTorch **52/61**, fp16 **52/61**
+- critical gold spans missed by fp16: **0** 
+
+## bench10 — 10 handwritten PII examples, CAUGHT (any-label overlap) / exact-label
+
+| ex | gold item (label) | pytorch | fp16 |
+|---|---|---|---|
+| 1 | `kai.nakamura` (unique_id) | CAUGHT* | CAUGHT* |
+| 1 | `Tr0ub4dor&3` (password) | CAUGHT | CAUGHT |
+| 1 | `kai.nakamura@example.io` (email) | CAUGHT | CAUGHT |
+| 2 | `5425 2334 3010 9903` (credit_debit_card) | CAUGHT | CAUGHT |
+| 2 | `11/27` (date) | CAUGHT | CAUGHT |
+| 2 | `833` (cvv) | CAUGHT | CAUGHT |
+| 3 | `078-05-1120` (ssn) | CAUGHT | CAUGHT |
+| 3 | `1EG4-TE5-MK72` (health_plan_beneficiary_number) | CAUGHT | CAUGHT |
+| 4 | `2001:0db8:85a3:0000:0000:8a2e:0370:7334` (ipv6) | CAUGHT | CAUGHT |
+| 4 | `3D:F2:C9:A6:B3:4F` (mac_address) | CAUGHT | CAUGHT |
+| 5 | `sk-live-9f8B2kQ71LmZxWv4TgNpR3sY` (api_key) | CAUGHT | CAUGHT |
+| 6 | `Rosa` (first_name) | CAUGHT | CAUGHT |
+| 6 | `Delgado-Marquez` (last_name) | CAUGHT* | CAUGHT* |
+| 6 | `+1 (312) 555-0147` (phone_number) | CAUGHT | CAUGHT |
+| 6 | `44 Cedar Hollow Ln, Apt 9C` (street_address) | CAUGHT | CAUGHT |
+| 6 | `Naperville` (city) | CAUGHT | CAUGHT |
+| 6 | `IL` (state) | CAUGHT | CAUGHT |
+| 6 | `60540` (postcode) | CAUGHT | CAUGHT |
+| 7 | `07/04/1962` (date_of_birth) | CAUGHT* | CAUGHT* |
+| 7 | `X4711982-B` (national_id) | CAUGHT | CAUGHT |
+| 7 | `Lisbon` (city) | **MISSED** | **MISSED** |
+| 7 | `2026-09-18` (date) | CAUGHT* | CAUGHT* |
+| 8 | `DE89 3704 0044 0532 0130 00` (account_number) | CAUGHT | CAUGHT |
+| 8 | `021000021` (bank_routing_number) | CAUGHT | CAUGHT |
+| 8 | `646-555-0199` (phone_number) | CAUGHT | CAUGHT |
+| 9 | `KTX-4092` (license_plate) | CAUGHT* | CAUGHT* |
+| 9 | `1HGBH41JXMN109186` (vehicle_identifier) | CAUGHT | CAUGHT |
+| 9 | `45.5231, -122.6765` (coordinate) | CAUGHT | CAUGHT |
+| 10 | `_wid=eyJhbGciOiJIUzI1NiJ9.aWQ9NzcyNA.4vX2` (http_cookie) | CAUGHT | CAUGHT |
+| 10 | `wellness.example.com` (url) | **MISSED** | **MISSED** |
+| 10 | `E-100482` (employee_id) | CAUGHT | CAUGHT |
+
+**pytorch: 29/31 caught**
+
+**fp16: 29/31 caught**
+
+`*` = caught with a different (but redacting) label
+
+## Phase 4b: q8 / mixed48 / fp16 in the browser (supersedes Phase 4)
+
+Phase 4 concluded q8 and mixed48 were native-ORT-only. Two changes overturned
+that — the big variants now run in-browser:
+
+1. **External data + 64-byte alignment** (`export/externalize.py`). Single-file
+   variants forced ort-web to parse the whole protobuf inside the 4 GB wasm32
+   heap (q8 → `std::bad_alloc`) and left initializers at unaligned file offsets
+   (mixed48 → `operation does not support unaligned accesses`). Re-saved as
+   `<name>.onnx` + `<name>.onnx_data` with every tensor padded to a 64-byte
+   boundary, both failures disappear. Session options `enableCpuMemArena:false`
+   + `enableMemPattern:false` trim peak allocation further.
+2. **ort-web JSPI bundle for WebGPU** (`ort.jspi.bundle.min.mjs`). The default
+   (asyncify) bundle runs QMoE on the CPU via async round-trips — mixed48 loads
+   but inference is ~91 s. The JSPI bundle has the **native WebGPU QMoE kernel**
+   (expert weights resident in VRAM), taking mixed48 to 488 ms and q8 to 817 ms.
+   The worker picks the bundle from `src/lib/variants.ts` (`bundleFor`): q8 and
+   mixed48 on WebGPU → JSPI; q4 and all CPU → default (broad compatibility;
+   JSPI needs Chrome/Edge ≥ 137).
+
+Measured in-browser matrix (Chrome 149, GTX 1660 Ti / Turing, 6 GB VRAM,
+ort-web 1.27.0; spans correct in every ✓ cell):
+
+| variant | GPU (WebGPU) | CPU (wasm) |
+|---|---|---|
+| q4 (0.92 GB) | ✓ default bundle · load 29 s · p50 134 ms · +825 MiB VRAM | ✗ no wasm `GatherBlockQuantized` (4-bit Gather) kernel |
+| q8 (1.98 GB) | ✓ **JSPI** · load 152 s cold/32 s warm · 817 ms · +2849 MiB VRAM | ✓ default bundle · load 36 s · **357 ms** |
+| mixed48 (1.66 GB) | ✓ **JSPI** · load 121 s · 488 ms · +1368 MiB VRAM | ✓ default bundle · ~370 ms |
+| fp16 (2.82 GB) | ✓ *only on `shader-f16` GPUs* — this Turing card lacks it → "device (webgpu) does not support fp16"; works on RTX 20xx+ / Apple Silicon | ✗ impractical (fp16 has no CPU compute advantage; 2.82 GB session stalls) |
+
+Notes:
+- **q4 stays the recommended default**: smallest, GPU-only, broadest bundle
+  compatibility, lowest VRAM. q8 is the accuracy pick when VRAM allows.
+- fp16-on-WebGPU is a genuine capability (native-precision, no dequant) but is
+  gated on the GPU exposing `shader-f16`; the app surfaces the clear ort error
+  on GPUs that don't.
+- The picker (`public/index.html` + `src/lib/variants.ts`) disables combos that
+  can't run — q4 forces GPU; CPU is offered for q8/mixed48/fp16.