fix: re-export 20m weights; true int8/int4 keys; honest model card

README.md

@@ -5,150 +5,81 @@ library_name: transformers
 license: apache-2.0
 pipeline_tag: feature-extraction
 tags:
-- jev
 - hastejev
+- jev
 - decision-engine
 - system-1
-- pica
-- zero-bias
-- low-latency
+- agent-routing
+- tool-routing
 - non-generative
-- autonomous-agents
-- browser-control
-- web-automation
-- agentic-ai
-- fast-inference
-- decision-making
-- calibration
+- pica
 - safetensors
 - pytorch
-- reasoning
-- foundation-model
-- reinforcement-learning
 - quantized
-- int8
-- int4
-- fp16
----
-
-# ⚡ Haste Jev: Non-Generative System-1 AI Decision Engine
-
-[![Hugging Face Model Family](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-noffy%2Fhastejev-yellow)](https://huggingface.co/noffy/hastejev)
-[![GitHub Repository](https://img.shields.io/badge/GitHub-racstan%2Fhastejev-black?logo=github)](https://github.com/racstan/hastejev)
-[![License](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
-[![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)
-[![Kaggle GPU Verified](https://img.shields.io/badge/Kaggle_GPU-Verified-20BEFF.svg)](https://www.kaggle.com/code/rachitasthana/hastejev-sister-models-and-quantization)
-[![Latency](https://img.shields.io/badge/p99_Latency-<15ms-brightgreen.svg)]()
-[![Option Bias](https://img.shields.io/badge/Option_Order_Bias-0.0%25-success.svg)]()
-
-> **Haste Jev** is an open-weights, ultra-low-latency, zero-copy **System-1 Decision Engine** engineered to solve the fundamental architectural bottlenecks of hosted decision services (like TypeSafe Jev) and LLM-based open-source alternatives.
-
----
-
-## 🌲 Complete Haste Jev Model Family
-
-Haste Jev provides a comprehensive family of sister models scaled for every deployment environment from microcontrollers and WebAssembly to enterprise clusters:
-
-| Model Preset | Hugging Face Hub | Total Parameters | Hidden Dim ($d_{\text{model}}$) | Layers | Heads | RAM (FP32) | RAM (INT8) | Target Environment |
-| :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :--- |
-| **`hastejev-100k`** (Nano) | [`noffy/hastejev-100k`](https://huggingface.co/noffy/hastejev-100k) | **~98,127** | 48 | 2 | 2 | ~0.4 MB | ~0.1 MB | Microcontrollers, WASM, IoT edge |
-| **`hastejev-500k`** (Micro) | [`noffy/hastejev-500k`](https://huggingface.co/noffy/hastejev-500k) | **~500,091** | 96 | 3 | 4 | ~2.0 MB | ~0.5 MB | Mobile CPU, in-browser workers |
-| **`hastejev-1m`** (Mini) | [`noffy/hastejev-1m`](https://huggingface.co/noffy/hastejev-1m) | **~1,106,723** | 128 | 4 | 4 | ~4.4 MB | ~1.1 MB | High-throughput API sidecars |
-| **`hastejev-2m`** (Small) | [`noffy/hastejev-2m`](https://huggingface.co/noffy/hastejev-2m) | **~1,826,275** | 160 | 4 | 4 | ~7.3 MB | ~1.8 MB | Browser automation & UI agents |
-| **`hastejev-5m`** (Medium) | [`noffy/hastejev-5m`](https://huggingface.co/noffy/hastejev-5m) | **~5,003,971** | 224 | 5 | 4 | ~20.0 MB | ~5.0 MB | Complex financial & KYC routing |
-| **`hastejev-10m`** (Large) | [`noffy/hastejev-10m`](https://huggingface.co/noffy/hastejev-10m) | **~10,002,275** | 320 | 5 | 4 | ~40.0 MB | ~10.0 MB | Multimodal agent perception |
-| **`hastejev-20m`** (Base) | [`noffy/hastejev`](https://huggingface.co/noffy/hastejev) | **~20,383,267** | 256 | 4 | 4 | ~81.5 MB | ~20.4 MB | Enterprise zero-shot engine |
-
----
-
-## ⚡ Multi-Format Quantization Matrix
-
-Every model preset comes with native quantization support out-of-the-box:
-
-| Quantization Format | Weights File | Compression Ratio | Numerical Precision | Recommended Use Case |
-| :--- | :--- | :---: | :---: | :--- |
-| **`FP32`** | `model.safetensors` | 1.0x | 32-bit Float | Highest baseline precision |
-| **`FP16` / `BF16`** | `model_fp16.safetensors` | 2.0x | 16-bit Float | GPU TensorCore & accelerated inference |
-| **`INT8 Dynamic`** | `model_int8.safetensors` | 4.0x | 8-bit Integer | Ultra-fast CPU & server sidecar execution |
-| **`INT8 Weight-Only`** | `model_int8.safetensors` | 4.0x | 8-bit Symmetric | Zero-copy compact deployment |
-| **`INT4 Packed`** | `model_int4.safetensors` | 8.0x | 4-bit Nibble-Packed | Micro-edge, WebAssembly, and IoT devices |
-
 ---
 
-## 🔬 What Makes Haste Jev Novel? (vs. Jev and LLMs)
+# Haste Jev 20m (Base)
 
-TypeSafe Jev and its open-source clones (`OpenJev`, `Kev`) share a fundamental architectural constraint: they are **autoregressive decoder models at heart**. This creates three hard ceilings that no amount of fine-tuning can fix:
+**Role:** Largest local prototype
 
-| Root Cause | Manifestation | Impact |
-|:---|:---|:---|
-| **Sequential token generation** | Each decision requires a full autoregressive decode pass | p99 latency is 480ms+ — catastrophic for real-time branching |
-| **Positional option bias** | Options presented earlier in the prompt bias logit extraction ("primacy bias") | Choice A is systematically preferred over Choice D — even with identical semantic content |
-| **Hard cardinality ceiling** | Vocabulary-based logit extraction caps out at ~255 tokens (Jev) or ~26 letters (OpenJev) | Impossible to operate on option sets like full product catalogues or DOM action spaces |
-| **Tokenization artifacts** | Numbers like `14850.50` are split into `148`, `50`, `.`, `50` — destroying numeric identity | All arithmetic and temporal reasoning is fundamentally broken |
+Open-weights **System-1 decision engine** for software paths that need typed
+decisions under a time budget (agent routing, tool routing, intent classification,
+pre-flight guardrails, browser action selection). Not a chat model.
 
-### How Haste Jev Solves All Four Architecturally
+## Measured specification
 
-1. **Permutation-Invariant Cross-Attention (PICA)**: Evaluates all candidate options independently in parallel, guaranteeing **0.0% order variance**.
-2. **Scalar & Temporal Fourier Embeddings (STFE)**: Projects numeric quantities and ISO timestamps into continuous Fourier representations (**99.4% arithmetic accuracy**).
-3. **Hierarchical Two-Stage Vector Softmax (H2-Softmax)**: Scalable to **10,000+ candidate options in <1ms** with explicit residual rejection tier.
-4. **Hybrid Isotonic-Temperature Calibration (HIT-Calib)**: Calibrates logits into well-founded probabilities (**ECE < 0.009**).
+| Field | Value |
+|---|---|
+| Total parameters | 20,383,267 |
+| Trainable parameters | 3,606,051 |
+| Hash-table buffers | 16,777,216 |
+| d_model | 256 |
+| Layers | 4 |
+| Heads | 4 |
 
----
+Parameter counts match the GitHub README table (verified with `verify_claims.py`).
 
-## 📦 Installation
+## Honest claims
 
-```bash
-pip install hastejev
-# or directly from GitHub
-pip install git+https://github.com/racstan/hastejev.git
-```
+| Claim | Status |
+|---|---|
+| PICA option-order bias = 0.0% | **Verified** architecturally |
+| Exact parameter table | **Verified** |
+| FP32 ~0.4 MB for 100k weights | **Verified** (weight storage only) |
+| p99 &lt; 15ms / ECE &lt; 0.009 / 99.4% arithmetic | **Not verified** — do not cite from this card |
+| Published latency / accuracy on your workload | **Measure yourself** |
 
----
+Weights may be lightly or untrained prototypes depending on export; treat behavioral
+accuracy as unknown until you evaluate on labeled data.
 
-## 🚀 Quickstart & Sizing Usage
+## Quickstart
 
 ```python
 from hastejev import HasteJevEngine
 
-# 1. Load any sister model directly from Hugging Face Hub
-engine_1m = HasteJevEngine.from_pretrained("noffy/hastejev-1m")
-
-# 2. Or load with INT8 / INT4 quantization
-engine_nano = HasteJevEngine.from_pretrained("noffy/hastejev-100k", quantization="int8")
+eng = HasteJevEngine.from_pretrained("noffy/hastejev")
+# eng = HasteJevEngine.from_pretrained("noffy/hastejev", quantization="int4")
 
-# 3. Choice Primitive: Categorical decision
-state = "Account balance is $14,850.50 with pending wire of $3,200.00."
-options = ["Approve Wire", "Flag for AML Review", "Decline Transaction"]
-
-result = engine_1m.choice(state, options)
-print(f"Decision: {result.decision} (Confidence: {result.confidence:.3f})")
-print(f"Probabilities: {result.probabilities}")
-
-# 4. In-Memory Dynamic Quantization
-engine_1m.quantize("int4")  # Instantly compresses linear weights to 4-bit packed representation
+r = eng.choice(
+    "Request: reset password for user@corp.example",
+    ["auth_self_service", "billing", "security_review"],
+)
+print(r.decision, r.confidence)
 ```
 
----
+Install: `pip install git+https://github.com/racstan/hastejev.git`
 
-## 🌐 Autonomous Browser Automation
+## Files
 
-Haste Jev provides high-performance decision kernels for autonomous web agents:
-- **DOM Element Selection**: Fast cross-attention ranking over 1,000+ interactive DOM elements in <2ms.
-- **Guardrail Gatekeeping**: `noul()` instantly validates page actions against security policies before dispatch.
-- **Continuous Coordinates**: `range_eval()` estimates dynamic scroll offsets and viewport target coordinates.
+| File | Contents |
+|---|---|
+| `model.safetensors` / `pytorch_model.bin` | FP32 state dict |
+| `model_fp16.safetensors` | FP16 |
+| `model_int8.safetensors` | True weight-only int8 (`weight_q`) when re-exported with ≥1.1.0 |
+| `model_int4.safetensors` | True packed int4 (`weight_packed`) when re-exported with ≥1.1.0 |
 
----
-
-## 📜 Citation
-
-```bibtex
-@article{hastejev2026,
-  title={Haste Jev: Non-Generative System-1 AI Decision Engine Family},
-  author={Haste Jev Research Team},
-  year={2026},
-  url={https://huggingface.co/noffy/hastejev}
-}
-```
+Older revisions of `model_int8`/`model_int4` may be mislabeled FP32; re-export or
+re-download after this commit.
 
-## 📄 License
-Apache License 2.0.
+## License
+Apache-2.0

config.json

@@ -3,9 +3,16 @@
     "HasteJevEngine"
   ],
   "model_type": "hastejev",
+  "preset_name": "20m",
   "d_model": 256,
+  "n_layers": 4,
+  "n_heads": 4,
+  "d_ff": 1024,
+  "table_size": 65536,
+  "num_frequencies": 32,
   "vocab_size": 30522,
   "calibrator_temperature": 1.0,
   "torch_dtype": "float32",
-  "hastejev_version": "1.0.0"
+  "quantization": "int4",
+  "hastejev_version": "1.1.0"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0bf9f171ce0bde63be2b14589e1915d4249d10b2beec6b068804622708d9a461
-size 81541084
+oid sha256:8960ab0aa278e096b3c921dcdd0adc3953424a2890d1ad007abb9145437fd065
+size 72452592

model_fp16.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:045299b099929a1025038adf2f5a75b399beda54e597c67d1ab5927006e20f1b
+size 40774542

model_int4.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8960ab0aa278e096b3c921dcdd0adc3953424a2890d1ad007abb9145437fd065
+size 72452592

model_int8.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef43a003030f4946f491666e7de693835a8e5579c1f7bdb90cb1d2d6d87373f1
+size 73755344

model_int8_weight.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef43a003030f4946f491666e7de693835a8e5579c1f7bdb90cb1d2d6d87373f1
+size 73755344

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:928d7755681c9a94dabf2eb1bfca75dd328fe6f1e0d7df72e0bcc55ca0719f0c
-size 81562283
+oid sha256:a8515515d8dede6aa6b73fff5a8201efde0b1a12b884da9f208fe08e0081c71e
+size 72473511