plumbify 0.2.0__tar.gz

plumbify-0.2.0/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.

plumbify-0.2.0/PKG-INFO

@@ -0,0 +1,212 @@
+Metadata-Version: 2.4
+Name: plumbify
+Version: 0.2.0
+Summary: Give an open LLM a fast, calibrated decision path and serve it natively with vLLM.
+Author: Naveenraj Kamalakannan
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/therealnaveenkamal/plumbify
+Project-URL: Documentation, https://github.com/therealnaveenkamal/plumbify/blob/main/docs/guide.md
+Project-URL: Repository, https://github.com/therealnaveenkamal/plumbify
+Project-URL: Issues, https://github.com/therealnaveenkamal/plumbify/issues
+Project-URL: Changelog, https://github.com/therealnaveenkamal/plumbify/blob/main/CHANGELOG.md
+Keywords: vllm,vllm-plugin,llm,lora,decision-making,calibration,inference
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.6
+Requires-Dist: transformers>=5.17
+Requires-Dist: peft>=0.21
+Requires-Dist: accelerate>=1.9
+Requires-Dist: huggingface_hub>=0.30
+Requires-Dist: safetensors
+Requires-Dist: sentencepiece
+Provides-Extra: vllm
+Requires-Dist: vllm==0.30.0; extra == "vllm"
+Provides-Extra: dev
+Requires-Dist: pytest; extra == "dev"
+Requires-Dist: ruff; extra == "dev"
+Requires-Dist: pre-commit; extra == "dev"
+Requires-Dist: fastapi>=0.115; extra == "dev"
+Requires-Dist: httpx>=0.27; extra == "dev"
+Dynamic: license-file
+
+<div align="center">
+
+# Plumbify
+
+**Give an open LLM a fast decision path, then serve it with vLLM.**
+
+[![CI](https://github.com/therealnaveenkamal/plumbify/actions/workflows/ci.yml/badge.svg)](https://github.com/therealnaveenkamal/plumbify/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/plumbify.svg)](https://pypi.org/project/plumbify/)
+[![License](https://img.shields.io/badge/License-Apache%202.0-green.svg)](https://github.com/therealnaveenkamal/plumbify/blob/main/LICENSE)
+[![Python](https://img.shields.io/badge/Python-3.10%2B-blue.svg)](https://github.com/therealnaveenkamal/plumbify/blob/main/pyproject.toml)
+[![vLLM](https://img.shields.io/badge/vLLM-0.30.0-purple.svg)](https://github.com/vllm-project/vllm)
+
+[Guide](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/guide.md) · [Architecture](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/architecture.md) · [Recipes](https://github.com/therealnaveenkamal/plumbify/tree/main/recipes) · [Results](#results) ·
+[Changelog](https://github.com/therealnaveenkamal/plumbify/blob/main/CHANGELOG.md)
+
+</div>
+
+Language models are slow and inconsistent at the small judgements that fill real conversations: which team gets
+this ticket, is this refund allowed, how urgent is it. Thinking helps, at the cost of seconds per decision.
+
+Plumbify trains a small decision head, a **plumb**, onto an open model without changing its weights. With the
+plugin installed, `vllm serve` loads the result like any other model. The model generates as usual, and when a
+conversation reaches a decision, the plumb answers it in one forward pass, reading the KV cache the conversation
+already filled, with a calibrated probability for every option.
+
+**Faster, with thinking on.**
+
+![A refund case: the normal model and the plumbed model side by side, thinking on](docs/assets/refund.gif)
+
+*Qwen3.5-35B-A3B, one server, real time, thinking on in both panes. A support case asks for two decisions: approve
+the refund, and escalate to a manager. The normal model reasons both out in its thinking. The plumbed model hands
+them to its plumb, which reads the conversation from the KV cache (the escalation call takes 116 ms and reuses 1,056
+cached tokens). Both plumb answers come back with two options in their conformal set, which marks them as unsure, so
+the model checks each one briefly before it replies. Both panes reach the same answer: approve, don't escalate. The
+plumbed model finishes in 10.5 s, the normal model in 32.6 s, 3.1× faster. [Video (MP4)](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/assets/refund.mp4)*
+
+**More accurate, with thinking off.**
+
+![An exam item: the normal model and the plumbed model side by side, thinking off](docs/assets/accuracy-thinking-off.gif)
+
+*Qwen3.5-35B-A3B, thinking off in both panes. A grading item states a decision: what is the correct answer, given a
+student's working? The normal model agrees with the student's wrong answer, d ($278.40). The plumb picks b
+($290.00), the correct answer, with probability 0.995 in 220 ms, and the model answers b. Both take under a
+second. [Video (MP4)](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/assets/accuracy-thinking-off.mp4)*
+
+**A plumb is a Jev model built into the LLM.** It answers the same kind of typed decision as a Jev-style decision
+model such as TypeSafe's Jev: pick one of these options, yes or no, or a score on a scale. The difference is that it
+is not a second model next to the LLM. It is a Jev decision head inside the model, sharing its weights and its KV
+cache, so plumbifying a model is jevifying it.
+
+## Quickstart
+
+Plumbify needs Linux, an NVIDIA GPU and Python 3.10 or later. The `vllm` extra installs vLLM 0.30.0, the version the
+plugin is tested against.
+
+```bash
+pip install "plumbify[vllm]"
+```
+
+Train a plumb on a frozen base model, then serve the result:
+
+```bash
+plumbify train --base Qwen/Qwen3.5-9B \
+    --rows train.jsonl --dev dev.jsonl --out plumbed-qwen3.5-9b
+vllm serve plumbed-qwen3.5-9b
+```
+
+Ask it a decision through the standard OpenAI API:
+
+```bash
+curl -s localhost:8000/v1/chat/completions \
+  -H 'content-type: application/json' \
+  -d '{"model": "plumbed-qwen3.5-9b",
+       "messages": [{"role": "user", "content":
+         "Charged twice.\n\nWhich team?\n- billing\n- technical\n- sales"}]}' \
+  | jq -c .plumb.answer
+```
+
+The response is a normal chat completion with an extra `plumb` field. Its `answer` says which option won, who
+decided (`system1` is the plumb), and with what probability:
+
+```json
+{"choice":"billing","by":"system1","p":0.973}
+```
+
+The [guide](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/guide.md) covers the training data format, the request options and the response fields.
+`python scripts/demo_chat.py` opens a live chat that shows each decision as the model hands it off.
+
+## Results
+
+The same model alone and with its plumb, on one vLLM server, over 447 held-out decisions (task families and
+templates no model trained on). Each prompt states a decision: context, question, bulleted options.
+
+| Model                 | No thinking: alone → plumbed | Thinking: alone → plumbed | Thinking latency p50: alone → plumbed |
+| --------------------- | ---------------------------- | ------------------------- | ------------------------------------- |
+| Qwen3.5-27B           | 0.732 → **0.839**            | 0.866 → **0.875**         | 26.6 s → 17.1 s (1.6× faster)         |
+| Qwen3.5-35B-A3B (MoE) | 0.720 → **0.810**            | 0.852 → **0.875**         | 6.8 s → 2.8 s (2.4× faster)           |
+| Gemma 4 12B           | 0.736 → **0.826**            | 0.770 → **0.872**         | 24.1 s → 5.0 s (4.8× faster)          |
+| Qwen3.5-9B            | 0.696 → **0.779**            | 0.808 → **0.846**         | 28.0 s → 5.7 s (4.9× faster)          |
+| Qwen3.5-4B            | 0.667 → **0.801**            | **0.841** → 0.826         | 12.0 s → 3.8 s (3.1× faster)          |
+| Qwen3-1.7B            | 0.570 → **0.658**            | 0.642 → **0.707**         | 3.9 s → 3.1 s (1.2× faster)           |
+
+Without thinking, the plumb adds 8 to 13 points on every model. With thinking, the plumbed model is at least as
+accurate on five of six models (Qwen3.5-4B is within noise) and answers 1.2 to 4.9 times faster, because it thinks
+30 to 75% less: it hands the decision to the plumb instead of reasoning it all out. It also always answers, where
+the model alone runs out of tokens or gives no clear answer on 5 to 15% of decisions when thinking. With 447
+decisions, one standard error is about 2 points.
+
+With thinking on, the model alone is scored in its better setup. These models think much less when their prompt
+lists a tool (the plumbed model's prompt always lists `plumb_decide`), which is often but not always more accurate,
+so each model alone gets whichever prompt, with or without a tool, scores higher. Settings, training time and
+serving memory for each run are in [recipes/](https://github.com/therealnaveenkamal/plumbify/tree/main/recipes); the benchmark is
+[`scripts/bench_s1_vs_s2.py`](https://github.com/therealnaveenkamal/plumbify/blob/main/scripts/bench_s1_vs_s2.py).
+
+## How it works
+
+Take a support chat that reaches a decision: *which team should handle this ticket: billing, technical or sales?*
+
+1. **The question is added to the conversation.** The question and the three options are appended, written like a
+   normal message to the model.
+2. **The model reads only the new part.** It has already read the conversation, and that work is cached, so only the
+   few words of the question are processed.
+3. **The plumb scores the options.** It looks at what the model computed while reading the question and turns that
+   into a probability per option: billing 0.97, technical 0.02, sales 0.01.
+4. **The model carries on.** No text was generated for the decision. The model uses the answer and continues its
+   reply.
+
+```
+conversation so far (cached)  +  "Which team? billing / technical / sales"  →  plumb  →  billing 0.97
+```
+
+Training adds two small pieces and nothing else: the plumb itself, and an adapter that switches on only while the
+model reads a decision question. The model's own weights never change, so everything else it does (chat, reasoning,
+tools) works exactly as before. [Architecture](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/architecture.md) explains each piece and where it lives in the
+code.
+
+## Documentation
+
+| Document                                  | Covers                                                                    |
+| ----------------------------------------- | ------------------------------------------------------------------------- |
+| [Guide](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/guide.md)                    | Install, training data, the `plumbify` commands, the serving API, limits  |
+| [Architecture](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/architecture.md)      | How training and serving fit together, the invariants, a map of the code |
+| [Recipes](https://github.com/therealnaveenkamal/plumbify/blob/main/recipes/README.md)              | Tested settings and measured results for each model                       |
+| [Scripts](https://github.com/therealnaveenkamal/plumbify/blob/main/scripts/README.md)              | Benchmark, vLLM parity check, live chat                                   |
+| [Contributing](https://github.com/therealnaveenkamal/plumbify/blob/main/CONTRIBUTING.md)           | Development setup, tests, conventions                                     |
+
+## Development
+
+```bash
+pip install -e ".[dev]" && pre-commit install
+pytest -q tests
+ruff check . && ruff format --check .
+```
+
+The tests run on CPU in seconds with a tiny random model and don't need vLLM. See [CONTRIBUTING.md](https://github.com/therealnaveenkamal/plumbify/blob/main/CONTRIBUTING.md).
+
+## Citation
+
+```bibtex
+@misc{plumbify2026,
+  title  = {Plumbify: a System 1 decision branch for open
+            language models},
+  author = {Kamalakannan, Naveenraj},
+  year   = {2026},
+  url    = {https://github.com/therealnaveenkamal/plumbify}
+}
+```
+
+## License
+
+[Apache-2.0](https://github.com/therealnaveenkamal/plumbify/blob/main/LICENSE).

plumbify-0.2.0/README.md

@@ -0,0 +1,171 @@
+<div align="center">
+
+# Plumbify
+
+**Give an open LLM a fast decision path, then serve it with vLLM.**
+
+[![CI](https://github.com/therealnaveenkamal/plumbify/actions/workflows/ci.yml/badge.svg)](https://github.com/therealnaveenkamal/plumbify/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/plumbify.svg)](https://pypi.org/project/plumbify/)
+[![License](https://img.shields.io/badge/License-Apache%202.0-green.svg)](https://github.com/therealnaveenkamal/plumbify/blob/main/LICENSE)
+[![Python](https://img.shields.io/badge/Python-3.10%2B-blue.svg)](https://github.com/therealnaveenkamal/plumbify/blob/main/pyproject.toml)
+[![vLLM](https://img.shields.io/badge/vLLM-0.30.0-purple.svg)](https://github.com/vllm-project/vllm)
+
+[Guide](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/guide.md) · [Architecture](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/architecture.md) · [Recipes](https://github.com/therealnaveenkamal/plumbify/tree/main/recipes) · [Results](#results) ·
+[Changelog](https://github.com/therealnaveenkamal/plumbify/blob/main/CHANGELOG.md)
+
+</div>
+
+Language models are slow and inconsistent at the small judgements that fill real conversations: which team gets
+this ticket, is this refund allowed, how urgent is it. Thinking helps, at the cost of seconds per decision.
+
+Plumbify trains a small decision head, a **plumb**, onto an open model without changing its weights. With the
+plugin installed, `vllm serve` loads the result like any other model. The model generates as usual, and when a
+conversation reaches a decision, the plumb answers it in one forward pass, reading the KV cache the conversation
+already filled, with a calibrated probability for every option.
+
+**Faster, with thinking on.**
+
+![A refund case: the normal model and the plumbed model side by side, thinking on](docs/assets/refund.gif)
+
+*Qwen3.5-35B-A3B, one server, real time, thinking on in both panes. A support case asks for two decisions: approve
+the refund, and escalate to a manager. The normal model reasons both out in its thinking. The plumbed model hands
+them to its plumb, which reads the conversation from the KV cache (the escalation call takes 116 ms and reuses 1,056
+cached tokens). Both plumb answers come back with two options in their conformal set, which marks them as unsure, so
+the model checks each one briefly before it replies. Both panes reach the same answer: approve, don't escalate. The
+plumbed model finishes in 10.5 s, the normal model in 32.6 s, 3.1× faster. [Video (MP4)](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/assets/refund.mp4)*
+
+**More accurate, with thinking off.**
+
+![An exam item: the normal model and the plumbed model side by side, thinking off](docs/assets/accuracy-thinking-off.gif)
+
+*Qwen3.5-35B-A3B, thinking off in both panes. A grading item states a decision: what is the correct answer, given a
+student's working? The normal model agrees with the student's wrong answer, d ($278.40). The plumb picks b
+($290.00), the correct answer, with probability 0.995 in 220 ms, and the model answers b. Both take under a
+second. [Video (MP4)](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/assets/accuracy-thinking-off.mp4)*
+
+**A plumb is a Jev model built into the LLM.** It answers the same kind of typed decision as a Jev-style decision
+model such as TypeSafe's Jev: pick one of these options, yes or no, or a score on a scale. The difference is that it
+is not a second model next to the LLM. It is a Jev decision head inside the model, sharing its weights and its KV
+cache, so plumbifying a model is jevifying it.
+
+## Quickstart
+
+Plumbify needs Linux, an NVIDIA GPU and Python 3.10 or later. The `vllm` extra installs vLLM 0.30.0, the version the
+plugin is tested against.
+
+```bash
+pip install "plumbify[vllm]"
+```
+
+Train a plumb on a frozen base model, then serve the result:
+
+```bash
+plumbify train --base Qwen/Qwen3.5-9B \
+    --rows train.jsonl --dev dev.jsonl --out plumbed-qwen3.5-9b
+vllm serve plumbed-qwen3.5-9b
+```
+
+Ask it a decision through the standard OpenAI API:
+
+```bash
+curl -s localhost:8000/v1/chat/completions \
+  -H 'content-type: application/json' \
+  -d '{"model": "plumbed-qwen3.5-9b",
+       "messages": [{"role": "user", "content":
+         "Charged twice.\n\nWhich team?\n- billing\n- technical\n- sales"}]}' \
+  | jq -c .plumb.answer
+```
+
+The response is a normal chat completion with an extra `plumb` field. Its `answer` says which option won, who
+decided (`system1` is the plumb), and with what probability:
+
+```json
+{"choice":"billing","by":"system1","p":0.973}
+```
+
+The [guide](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/guide.md) covers the training data format, the request options and the response fields.
+`python scripts/demo_chat.py` opens a live chat that shows each decision as the model hands it off.
+
+## Results
+
+The same model alone and with its plumb, on one vLLM server, over 447 held-out decisions (task families and
+templates no model trained on). Each prompt states a decision: context, question, bulleted options.
+
+| Model                 | No thinking: alone → plumbed | Thinking: alone → plumbed | Thinking latency p50: alone → plumbed |
+| --------------------- | ---------------------------- | ------------------------- | ------------------------------------- |
+| Qwen3.5-27B           | 0.732 → **0.839**            | 0.866 → **0.875**         | 26.6 s → 17.1 s (1.6× faster)         |
+| Qwen3.5-35B-A3B (MoE) | 0.720 → **0.810**            | 0.852 → **0.875**         | 6.8 s → 2.8 s (2.4× faster)           |
+| Gemma 4 12B           | 0.736 → **0.826**            | 0.770 → **0.872**         | 24.1 s → 5.0 s (4.8× faster)          |
+| Qwen3.5-9B            | 0.696 → **0.779**            | 0.808 → **0.846**         | 28.0 s → 5.7 s (4.9× faster)          |
+| Qwen3.5-4B            | 0.667 → **0.801**            | **0.841** → 0.826         | 12.0 s → 3.8 s (3.1× faster)          |
+| Qwen3-1.7B            | 0.570 → **0.658**            | 0.642 → **0.707**         | 3.9 s → 3.1 s (1.2× faster)           |
+
+Without thinking, the plumb adds 8 to 13 points on every model. With thinking, the plumbed model is at least as
+accurate on five of six models (Qwen3.5-4B is within noise) and answers 1.2 to 4.9 times faster, because it thinks
+30 to 75% less: it hands the decision to the plumb instead of reasoning it all out. It also always answers, where
+the model alone runs out of tokens or gives no clear answer on 5 to 15% of decisions when thinking. With 447
+decisions, one standard error is about 2 points.
+
+With thinking on, the model alone is scored in its better setup. These models think much less when their prompt
+lists a tool (the plumbed model's prompt always lists `plumb_decide`), which is often but not always more accurate,
+so each model alone gets whichever prompt, with or without a tool, scores higher. Settings, training time and
+serving memory for each run are in [recipes/](https://github.com/therealnaveenkamal/plumbify/tree/main/recipes); the benchmark is
+[`scripts/bench_s1_vs_s2.py`](https://github.com/therealnaveenkamal/plumbify/blob/main/scripts/bench_s1_vs_s2.py).
+
+## How it works
+
+Take a support chat that reaches a decision: *which team should handle this ticket: billing, technical or sales?*
+
+1. **The question is added to the conversation.** The question and the three options are appended, written like a
+   normal message to the model.
+2. **The model reads only the new part.** It has already read the conversation, and that work is cached, so only the
+   few words of the question are processed.
+3. **The plumb scores the options.** It looks at what the model computed while reading the question and turns that
+   into a probability per option: billing 0.97, technical 0.02, sales 0.01.
+4. **The model carries on.** No text was generated for the decision. The model uses the answer and continues its
+   reply.
+
+```
+conversation so far (cached)  +  "Which team? billing / technical / sales"  →  plumb  →  billing 0.97
+```
+
+Training adds two small pieces and nothing else: the plumb itself, and an adapter that switches on only while the
+model reads a decision question. The model's own weights never change, so everything else it does (chat, reasoning,
+tools) works exactly as before. [Architecture](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/architecture.md) explains each piece and where it lives in the
+code.
+
+## Documentation
+
+| Document                                  | Covers                                                                    |
+| ----------------------------------------- | ------------------------------------------------------------------------- |
+| [Guide](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/guide.md)                    | Install, training data, the `plumbify` commands, the serving API, limits  |
+| [Architecture](https://github.com/therealnaveenkamal/plumbify/blob/main/docs/architecture.md)      | How training and serving fit together, the invariants, a map of the code |
+| [Recipes](https://github.com/therealnaveenkamal/plumbify/blob/main/recipes/README.md)              | Tested settings and measured results for each model                       |
+| [Scripts](https://github.com/therealnaveenkamal/plumbify/blob/main/scripts/README.md)              | Benchmark, vLLM parity check, live chat                                   |
+| [Contributing](https://github.com/therealnaveenkamal/plumbify/blob/main/CONTRIBUTING.md)           | Development setup, tests, conventions                                     |
+
+## Development
+
+```bash
+pip install -e ".[dev]" && pre-commit install
+pytest -q tests
+ruff check . && ruff format --check .
+```
+
+The tests run on CPU in seconds with a tiny random model and don't need vLLM. See [CONTRIBUTING.md](https://github.com/therealnaveenkamal/plumbify/blob/main/CONTRIBUTING.md).
+
+## Citation
+
+```bibtex
+@misc{plumbify2026,
+  title  = {Plumbify: a System 1 decision branch for open
+            language models},
+  author = {Kamalakannan, Naveenraj},
+  year   = {2026},
+  url    = {https://github.com/therealnaveenkamal/plumbify}
+}
+```
+
+## License
+
+[Apache-2.0](https://github.com/therealnaveenkamal/plumbify/blob/main/LICENSE).

plumbify-0.2.0/plumbify/__init__.py

@@ -0,0 +1,20 @@
+"""Plumbify: a fast decision path (System 1) for open language models, served natively by vLLM.
+
+    plumbify train --base Qwen/Qwen3.5-9B --rows train.jsonl --dev dev.jsonl --out plumbed-qwen3.5-9b
+    vllm serve plumbed-qwen3.5-9b
+
+In Python, ``System1`` runs a plumb with transformers (training, evaluation, the reference implementation).
+"""
+
+__version__ = "0.2.0"
+
+
+def __getattr__(name):  # System1 pulls in torch and transformers: load it on first use
+    if name == "System1":
+        from .system1 import System1
+
+        return System1
+    raise AttributeError(name)
+
+
+__all__ = ["System1", "__version__"]