gemmadecision 0.1.0__tar.gz

gemmadecision-0.1.0/.gitignore

@@ -0,0 +1,22 @@
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+dist/
+build/
+*.egg-info/
+.venv/
+venv/
+.DS_Store
+.env
+.env.*
+!.env.example
+hf_token*
+*.safetensors
+*.bin
+models/
+.cache/
+*.log

gemmadecision-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

gemmadecision-0.1.0/NOTICE

@@ -0,0 +1,16 @@
+GemmaDecision serving package
+Copyright 2026 Rajan Shukla
+
+python/gemmadecision/_clm_schema.py is derived from Contrastive-LM/CLM,
+revision e939398d4556fcd9400c76fa8c5a513202f42b0a, Apache License 2.0.
+https://github.com/Contrastive-LM/CLM
+Other serving code is original to this package.
+
+Granian supplies the Rust HTTP server (BSD-3-Clause); Pydantic Core supplies
+Rust-backed data validation (MIT). PyTorch/vLLM execute model tensor operations.
+Dependencies retain their respective licenses. This package is not an all-Rust
+model implementation and does not claim endorsement by upstream projects.
+
+Model weights are downloaded separately from rajan2k/GemmaDecision-270M.
+They are modified Google Gemma weights governed by the Gemma terms and
+Prohibited Use Policy in that repository, not this package's code license.

gemmadecision-0.1.0/PKG-INFO

@@ -0,0 +1,156 @@
+Metadata-Version: 2.4
+Name: gemmadecision
+Version: 0.1.0
+Summary: Typed local decisions for PydanticAI, with Rust HTTP serving and batched Gemma inference
+Project-URL: Homepage, https://github.com/rjn32s/gemmadecision
+Project-URL: Repository, https://github.com/rjn32s/gemmadecision
+Project-URL: Issues, https://github.com/rjn32s/gemmadecision/issues
+Project-URL: Model, https://huggingface.co/rajan2k/GemmaDecision-270M
+Author: Rajan Shukla
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: decision-model,gemma,granian,pydantic-ai,vllm
+Classifier: Development Status :: 4 - Beta
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.11
+Requires-Dist: fastapi<0.137,>=0.133
+Requires-Dist: granian<3,>=2.8
+Requires-Dist: httpx<1,>=0.28
+Requires-Dist: huggingface-hub<2,>=1.33
+Requires-Dist: numpy<3,>=2
+Requires-Dist: pydantic-ai-slim<3,>=2.51
+Requires-Dist: pydantic<3,>=2.12
+Requires-Dist: safetensors<1,>=0.8
+Requires-Dist: torch<2.15,>=2.13
+Requires-Dist: transformers<6,>=5.17
+Provides-Extra: http
+Requires-Dist: fastapi<0.137,>=0.133; extra == 'http'
+Requires-Dist: granian<3,>=2.8; extra == 'http'
+Provides-Extra: pydantic-ai
+Requires-Dist: pydantic-ai-slim<3,>=2.51; extra == 'pydantic-ai'
+Provides-Extra: serve
+Requires-Dist: fastapi<0.137,>=0.133; extra == 'serve'
+Requires-Dist: granian<3,>=2.8; extra == 'serve'
+Requires-Dist: huggingface-hub<2,>=1.33; extra == 'serve'
+Requires-Dist: numpy<3,>=2; extra == 'serve'
+Requires-Dist: safetensors<1,>=0.8; extra == 'serve'
+Requires-Dist: torch<2.15,>=2.13; extra == 'serve'
+Requires-Dist: transformers<6,>=5.17; extra == 'serve'
+Provides-Extra: test
+Requires-Dist: asgi-lifespan<3,>=2; extra == 'test'
+Requires-Dist: build<2,>=1; extra == 'test'
+Requires-Dist: fastapi<0.137,>=0.133; extra == 'test'
+Requires-Dist: granian<3,>=2.8; extra == 'test'
+Requires-Dist: pytest-asyncio<2,>=1; extra == 'test'
+Requires-Dist: pytest<10,>=8; extra == 'test'
+Requires-Dist: twine<7,>=6; extra == 'test'
+Provides-Extra: vllm
+Requires-Dist: fastapi<0.137,>=0.133; extra == 'vllm'
+Requires-Dist: granian<3,>=2.8; extra == 'vllm'
+Requires-Dist: transformers<6,>=5.17; extra == 'vllm'
+Requires-Dist: vllm==0.30.0; extra == 'vllm'
+Description-Content-Type: text/markdown
+
+# gemmadecision
+
+Small, local decisions in one Python call. Powered by
+[GemmaDecision-270M](https://huggingface.co/rajan2k/GemmaDecision-270M).
+
+## Install
+
+```bash
+pip install gemmadecision
+```
+
+Python 3.11+. No API key or server needed. The first call downloads the model
+(about 0.5 GB); later calls reuse it. CPU, NVIDIA CUDA and Apple Silicon are
+selected automatically. After downloading, inference runs locally.
+
+## Use it in your code
+
+```python
+from gemmadecision import decide
+
+team = decide("I was charged twice", choices=["billing", "technical"])
+print(team)  # selected choice, as a string
+```
+
+For more specific decisions, give each choice a description:
+
+```python
+team = decide(
+    "The same payment appears twice on my statement.",
+    choices={
+        "billing": "Handle charges, payments and refunds",
+        "technical": "Handle crashes, login errors and app problems",
+    },
+    question="Which support team should handle this request?",
+)
+```
+
+Need the full ranking? Use `rank(...)` with the same arguments. It returns
+ordered candidates, scores and derived probabilities.
+
+## PydanticAI
+
+```python
+from typing import Literal
+from pydantic_ai import Agent
+from gemmadecision import GemmaDecisionModel
+
+agent = Agent(
+    GemmaDecisionModel.local(),
+    output_type=Literal["billing", "technical"],
+)
+result = agent.run_sync("I was charged twice")
+print(result.output)
+```
+
+This uses PydanticAI's native decision-model interface. Boolean, enum, rubric
+and finite Pydantic fields are supported.
+[More PydanticAI examples](https://github.com/rjn32s/gemmadecision/blob/main/docs/pydantic_ai.md).
+
+## Serve it
+
+```bash
+gemmadecision serve
+```
+
+That starts the Rust-powered Granian HTTP server at `http://127.0.0.1:8700`.
+Interactive API documentation is available at `/docs`.
+
+```python
+from gemmadecision import DecisionClient
+
+with DecisionClient() as client:
+    result = client.decide(
+        "I was charged twice",
+        candidates={"billing": "Payments and refunds", "technical": "App problems"},
+    )
+    print(result.choice)
+```
+
+`AsyncDecisionClient` works with `await`. To connect PydanticAI to the server,
+use `GemmaDecisionModel()` instead of `.local()`.
+
+## More control when you need it
+
+- [Hardware settings, offline use, HTTP API and batching](https://github.com/rjn32s/gemmadecision/blob/main/docs/advanced.md)
+- [vLLM backend for Linux/CUDA](https://github.com/rjn32s/gemmadecision/blob/main/docs/vllm.md)
+- [Performance measurements and reproduction](https://github.com/rjn32s/gemmadecision/blob/main/docs/performance.md)
+- [Docker and releases](https://github.com/rjn32s/gemmadecision/blob/main/docs/releasing.md)
+
+The default uses batched PyTorch for model computation and Rust for HTTP
+serving. vLLM is optional; no Rust compiler is needed to install the package.
+
+This model chooses among supplied options; it does not generate open-ended
+text. Its scores are rankings, and the derived probabilities are not a guarantee
+of correctness. Maximums are 2,048 state/question tokens, 768 tokens per choice
+and 64 choices. See the [model card](https://huggingface.co/rajan2k/GemmaDecision-270M)
+for evaluation and limitations.
+
+Code: Apache-2.0. Model weights: separate Gemma terms. See [NOTICE](https://github.com/rjn32s/gemmadecision/blob/main/NOTICE).

gemmadecision-0.1.0/README.md

@@ -0,0 +1,99 @@
+# gemmadecision
+
+Small, local decisions in one Python call. Powered by
+[GemmaDecision-270M](https://huggingface.co/rajan2k/GemmaDecision-270M).
+
+## Install
+
+```bash
+pip install gemmadecision
+```
+
+Python 3.11+. No API key or server needed. The first call downloads the model
+(about 0.5 GB); later calls reuse it. CPU, NVIDIA CUDA and Apple Silicon are
+selected automatically. After downloading, inference runs locally.
+
+## Use it in your code
+
+```python
+from gemmadecision import decide
+
+team = decide("I was charged twice", choices=["billing", "technical"])
+print(team)  # selected choice, as a string
+```
+
+For more specific decisions, give each choice a description:
+
+```python
+team = decide(
+    "The same payment appears twice on my statement.",
+    choices={
+        "billing": "Handle charges, payments and refunds",
+        "technical": "Handle crashes, login errors and app problems",
+    },
+    question="Which support team should handle this request?",
+)
+```
+
+Need the full ranking? Use `rank(...)` with the same arguments. It returns
+ordered candidates, scores and derived probabilities.
+
+## PydanticAI
+
+```python
+from typing import Literal
+from pydantic_ai import Agent
+from gemmadecision import GemmaDecisionModel
+
+agent = Agent(
+    GemmaDecisionModel.local(),
+    output_type=Literal["billing", "technical"],
+)
+result = agent.run_sync("I was charged twice")
+print(result.output)
+```
+
+This uses PydanticAI's native decision-model interface. Boolean, enum, rubric
+and finite Pydantic fields are supported.
+[More PydanticAI examples](https://github.com/rjn32s/gemmadecision/blob/main/docs/pydantic_ai.md).
+
+## Serve it
+
+```bash
+gemmadecision serve
+```
+
+That starts the Rust-powered Granian HTTP server at `http://127.0.0.1:8700`.
+Interactive API documentation is available at `/docs`.
+
+```python
+from gemmadecision import DecisionClient
+
+with DecisionClient() as client:
+    result = client.decide(
+        "I was charged twice",
+        candidates={"billing": "Payments and refunds", "technical": "App problems"},
+    )
+    print(result.choice)
+```
+
+`AsyncDecisionClient` works with `await`. To connect PydanticAI to the server,
+use `GemmaDecisionModel()` instead of `.local()`.
+
+## More control when you need it
+
+- [Hardware settings, offline use, HTTP API and batching](https://github.com/rjn32s/gemmadecision/blob/main/docs/advanced.md)
+- [vLLM backend for Linux/CUDA](https://github.com/rjn32s/gemmadecision/blob/main/docs/vllm.md)
+- [Performance measurements and reproduction](https://github.com/rjn32s/gemmadecision/blob/main/docs/performance.md)
+- [Docker and releases](https://github.com/rjn32s/gemmadecision/blob/main/docs/releasing.md)
+
+The default uses batched PyTorch for model computation and Rust for HTTP
+serving. vLLM is optional; no Rust compiler is needed to install the package.
+
+This model chooses among supplied options; it does not generate open-ended
+text. Its scores are rankings, and the derived probabilities are not a guarantee
+of correctness. Maximums are 2,048 state/question tokens, 768 tokens per choice
+and 64 choices. See the [model card](https://huggingface.co/rajan2k/GemmaDecision-270M)
+for evaluation and limitations.
+
+Code: Apache-2.0. Model weights: separate Gemma terms. See [NOTICE](https://github.com/rjn32s/gemmadecision/blob/main/NOTICE).

gemmadecision-0.1.0/docs/advanced.md

@@ -0,0 +1,174 @@
+# gemmadecision
+
+Typed local decisions for Python and PydanticAI. One command serves the complete
+[GemmaDecision-270M](https://huggingface.co/rajan2k/GemmaDecision-270M) model with
+Granian's **Rust HTTP runtime**, bounded request queues and batched inference.
+Importing the client does not load PyTorch or download weights.
+
+```sh
+pip install gemmadecision
+gemmadecision serve
+```
+
+The first server start downloads the pinned public model (about 0.5 GB). Subsequent
+starts use the cache. CPU, CUDA and Apple MPS are selected automatically; override
+with `--device cpu`, `--device cuda`, or `--device mps`. Model inference stays local.
+The client and native PydanticAI integration work without an external LLM API key.
+
+## Python
+
+```python
+from gemmadecision import DecisionClient
+
+with DecisionClient() as client:
+    answer = client.decide(
+        "The same card payment appears twice on my statement.",
+        candidates={
+            "billing": "Investigate duplicate payments, charges and refunds.",
+            "technical": "Investigate app crashes and login errors.",
+        },
+        question="Which support team should handle this request?",
+    )
+    print(answer.choice)
+    print(answer.probabilities)
+```
+
+Use `AsyncDecisionClient` with `async with` and `await` in async applications.
+Clients reuse HTTP connections and do not retry decisions automatically. Set
+`base_url`, `api_key` and `timeout` explicitly or use `GEMMADECISION_BASE_URL` and
+`GEMMADECISION_API_KEY`. The standard install includes the local runtime; importing only the client stays lightweight.
+
+## PydanticAI, directly
+
+PydanticAI has a native decision-model interface: no chatbot shim is needed.
+
+```python
+import asyncio
+from typing import Literal
+from pydantic import BaseModel, Field
+from pydantic_ai import Agent
+from gemmadecision.integrations.pydantic_ai import GemmaDecisionModel
+
+class Route(BaseModel):
+    department: Literal["billing", "technical"] = Field(
+        description="Team responsible for this request: billing handles payments; technical handles app failures."
+    )
+
+async def main():
+    async with GemmaDecisionModel() as model:
+        agent = Agent(model, output_type=Route)
+        result = await agent.run("The same card payment appears twice on my statement.")
+        print(result.output)
+
+asyncio.run(main())
+```
+
+Finite `Literal`, `Enum`, boolean and rubric outputs are supported. This model
+selects among supplied alternatives; it does not generate arbitrary prose,
+extract unrestricted strings or replace a general language model. See
+[PydanticAI integration](pydantic_ai.md) for routing, rubrics and fallbacks.
+
+## Several typed decisions in one request
+
+```python
+from gemmadecision import DecisionClient, Choice, Noul, Score
+
+with DecisionClient() as client:
+    result = client.system_one(
+        state="The app crashed three times today and I cannot sign in.",
+        questions={
+            "team": Choice(instructions="Who should handle this?", criteria={
+                "technical": "App crashes and login problems", "billing": "Charges and refunds"}),
+            "needs_help": Noul(instructions="Does this customer need assistance?"),
+            "urgency": Score(instructions="How urgent is the request?",
+                             criteria=["Can wait", "Needs attention", "Blocked and needs prompt help"]),
+        },
+    )
+    print(result.answers)
+```
+
+`client.rank(state, candidates, question=...)` exposes raw scores and their fixed
+softmax transformation. For in-process inference use
+`DecisionEngine.from_pretrained(device="auto")` with the same `decide`, `rank` and
+`system_one` methods. It loads only when explicitly constructed.
+
+## Serving choices
+
+- **Default:** Granian Rust HTTP + batched PyTorch. Length-bucketed candidate
+  batches, bounded cross-request microbatching, one model process, no generation.
+- **Reference:** `gemmadecision serve --strict --batch-wait-ms 0` scores each
+  candidate separately, matching the published numerical execution shape.
+- **vLLM:** `pip install 'gemmadecision[vllm]'` in a separate Linux CUDA environment,
+  then `gemmadecision serve --backend vllm --device cuda`. See the exact dependency
+  pins and evaluation status in [vLLM setup](vllm.md).
+- **Offline:** `gemmadecision download --output ./model`, then
+  `gemmadecision serve --model-path ./model --offline`.
+- **Docker:** see [release and container instructions](releasing.md).
+
+The serving process uses Rust for HTTP transport and Rust-backed Pydantic
+validation. Model tensor operations use PyTorch or vLLM's native CUDA kernels;
+this is not an all-Rust reimplementation of Gemma. Granian supplies maintained
+binary wheels so users do not need to compile a new bespoke Rust server.
+
+CLM's vLLM encoder/API design inspired this separation. This model is a **joint
+ranker**: the entire state, question and candidate must be encoded together.
+Caching independent action embeddings would change its predictions. Optional
+`--cache-size 1024` caches only exact full-input scores in memory; it is off by
+default. Cache keys are SHA256 digests, and no persistent prompt cache is written.
+
+`--max-batch-tokens`, `--max-batch-size`, `--batch-wait-ms` and `--queue-size`
+control memory and throughput. Start with the defaults. Batching can introduce
+small numerical differences, especially near ties; see [performance notes](performance.md).
+No universal speed multiplier or JEV-equivalent accuracy is claimed.
+
+## HTTP interface
+
+`POST /v1/systemone`, `POST /v1/rank`, `GET /v1/models`, `GET /health`, `GET /ready`.
+Interactive OpenAPI documentation is at `http://127.0.0.1:8700/docs`.
+The default bind address is loopback. To expose on a private network, set
+`GEMMADECISION_API_KEY` and use `--host 0.0.0.0`. Put a TLS reverse proxy in front
+of any internet-facing deployment. An explicit `--allow-unauthenticated` flag
+exists for controlled deployments. HTTP request-content logs are disabled.
+
+Maximum input: 2,048 state/question tokens; 768 tokens per candidate; 2–64
+distinct candidates per question; 32 questions and 256 total pairs per request;
+2 MiB request body. Inputs are refused with HTTP 422 rather than silently
+truncated. Overload returns 503; queue timeout returns 504. Health is independent
+of inference authorization and includes no request text. There is one bounded
+inference queue and one model instance per process.
+
+## Model identity and limitations
+
+Package version `0.1.0` is separate from model version `v0.4.0`. Every backend
+uses immutable model revision `785d530221c990671f29976902540101bb9c7647` and verifies
+the encoder/head/config hashes. No executable Hugging Face code is downloaded
+or enabled. The encoder was trained jointly with a scalar head; using stock
+Gemma token-generation logits would not serve this model.
+
+Probabilities are `softmax(raw_scores / 4.136820402388508)`, using a temperature
+previously fitted on 300 banking/NLI examples. `confidence` is the largest
+derived probability, not a guarantee of correctness or general calibration.
+On the author's public JevBench diagnostic the frozen model achieved 110/231
+(47.62%), including 36 context-limit refusals. It is most relevant to small
+closed-choice tasks; evaluate it on your own use case before routing important
+actions. [Model evaluation and provenance](https://huggingface.co/rajan2k/GemmaDecision-270M).
+
+## Develop
+
+```sh
+git clone https://github.com/rjn32s/gemmadecision.git
+cd gemmadecision
+pip install -e '.[test,pydantic-ai]'
+pytest
+python -m build
+twine check dist/*
+```
+
+Tests use synthetic tensors, stub encoders and actual PydanticAI agents with
+mock HTTP; they do not download or train a model. GPU performance measurements
+are separate reproducible artifacts. `gemmadecision doctor` reports installed
+backends without loading weights.
+
+Code: Apache-2.0. Weights: separate Gemma terms. Dependency licenses and CLM schema
+attribution are in [NOTICE](../NOTICE). This package is not affiliated with Google,
+PydanticAI, vLLM, Granian, CLM or the JevBench maintainers.