docs: trim internal training-provenance from the model card

README.md

@@ -15,52 +15,31 @@ tags:
 
 # onnx-email-gate — multilingual KEEP/DROP email prefilter
 
-A tiny, **multilingual** CPU classifier that decides **KEEP** (looks like a real
-candidate email → send to the GPU model) or **DROP** (junk → skip the GPU). It sits in
-front of the expensive GPU email-intent classifier (Qwen) in Curriculo's pipeline and
-exists purely to **save GPU calls** by discarding obvious junk cheaply.
+A tiny **multilingual** CPU classifier that labels an inbound email **KEEP** (looks like a real
+job application → worth further processing) or **DROP** (junk → skip). It's meant as a cheap
+pre-filter in front of a more expensive downstream model, discarding obvious junk — newsletters,
+notifications, one-time codes, job-board alerts, bounces — at roughly a millisecond per email.
 
 - **Labels:** `{0: DROP, 1: KEEP}`
 - **Architecture:** `paraphrase-multilingual-MiniLM-L12-v2` sentence embedder (mean-pool +
-  L2-normalize) with a logistic-regression head **folded into the ONNX graph** as a final
-  linear layer, so the whole thing is one classifier ONNX: `tokens → 2 logits → argmax`.
-- **Quantization:** dynamic **INT8** (`model_int8.onnx`, ~119 MB, down 4× from ~471 MB fp32).
-- **Runtime:** `onnxruntime` on CPU (AVX512-VNNI). ~ms per email.
+  L2-normalize) with a logistic-regression head folded into the ONNX graph as a final linear
+  layer — so the whole thing is one classifier ONNX: `tokens → 2 logits → argmax`.
+- **Quantization:** dynamic **INT8** (`model_int8.onnx`, ~119 MB).
+- **Runtime:** `onnxruntime` on CPU, ~ms per email.
 
 ## Why multilingual
 
-The embedder is trained on ~50 languages, so the gate reads non-English candidate emails
-directly instead of wrong-dropping them. Verified KEEP/DROP on English, **Hindi**, and
-**Spanish** applications and junk (5/5).
+The embedder covers ~50 languages, so the gate reads non-English application emails directly
+instead of wrong-dropping them (verified on English, Hindi, and Spanish).
 
-## How it's used (the override ladder)
+## Recommended use — an override ladder
 
-The model is **rung 3** of a cheap ladder that runs in the cron — the overrides are the
-real safety net and run FIRST so a genuine candidate is never dropped by the model:
+The model is best used as the **last rung** of a cheap rule ladder, so a genuine application is
+never dropped by the model alone:
 
-1. **KEEP overrides** — résumé attachment / ZipRecruiter forward / forwarded application /
-   recruiting-forward → **KEEP** (bypass, straight to the GPU).
+1. **KEEP overrides** — résumé attachment / forwarded application / recruiter-style sender → **KEEP**.
 2. **regex junk** — noreply / notifications / OTP / bounce / job-board-alert / newsletter → **DROP**.
-3. **this model** — KEEP/DROP on the fuzzy residual.
-
-`DROP` rows are labelled `OTHER` and the GPU is skipped; `KEEP` rows go to the GPU.
-Ships in **shadow mode** first (log would-drops, still call the GPU), then flip to hard-drop.
-
-## Training
-
-- **3,953** real inbound emails (dev + a small prod read-only slice), deduped, **per-sender
-  capped at 30** to stop a few high-volume senders dominating the KEEP class.
-- Labelled by a **local 35B** model on the ambiguous residual; clear cases auto-labelled by rule.
-- Head: `LogisticRegression(class_weight="balanced")` on frozen, L2-normalized embeddings.
-
-## Metrics
-
-- **Model-only** on a 303-row hand-checked gold set: accuracy **0.868**, KEEP recall 0.80.
-  Most model-only misses are *forwarded* applications — rescued by the KEEP override ladder.
-- **With the ladder (production):** the overrides take the raw model to a high-precision gate;
-  bake-off showed ≈ 0–3 real candidates lost in-distribution.
-- **INT8 vs fp32 parity:** 97.4% on gold (flips lean toward keeping candidates); multilingual
-  concrete cases unchanged (5/5).
+3. **this model** — KEEP/DROP on the remaining ambiguous mail.
 
 ## Usage
 
@@ -89,15 +68,14 @@ Input text = `from_address\nfrom_name\nsubject\nbody`, truncated to 128 tokens.
 
 | File | Purpose |
 |------|---------|
-| `model_int8.onnx` | the model — embedder + folded LogReg head, INT8 |
+| `model_int8.onnx` | the model — embedder + folded logistic-regression head, INT8 |
 | `tokenizer.json`, `tokenizer_config.json`, `special_tokens_map.json` | tokenizer |
 | `config.json` | model config + `id2label` |
-| `logreg_head.npz`, `gate_meta.json` | reproducibility — raw head weights + fold record |
+| `logreg_head.npz`, `gate_meta.json` | raw head weights + fold record (reproducibility) |
 
 ## Limitations
 
-- English-only vocabulary would wrong-drop non-English candidates; this multilingual variant
-  is chosen precisely to avoid that, but very low-resource languages may be weaker.
-- The model alone has modest recall; **do not ship it without the override ladder** — the
-  ladder, not the model, is what makes the gate safe for real candidates.
-- INT8 shifts ~2.6% of borderline predictions vs fp32; ship fp32 if a use case needs exact parity.
+- The model alone has modest recall — **use it behind the override ladder**, not standalone.
+- Dynamic INT8 shifts a small fraction of borderline predictions vs fp32; ship fp32 if you need
+  exact parity.
+- Very low-resource languages may be weaker than the ~50 the embedder covers well.