shad0wllm 0.1.0__tar.gz

shad0wllm-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.

shad0wllm-0.1.0/NOTICE

@@ -0,0 +1,11 @@
+shad0w
+Copyright 2026 Gaurav Chauhan
+
+The hashed linear table follows the fastText design (Joulin et al., 2016). The certificate is an application of
+Learn-then-Test (Angelopoulos et al., 2021) in the spirit of Trust or Escalate (Jung et al., 2025) and BARGAIN
+(Zeighami et al., 2025). References: https://2796gaurav.github.io/shad0w/docs/certificate.html
+
+Optional compile-time encoders (downloaded by the user, never bundled) carry their own licences:
+BAAI/bge-small-en-v1.5 (MIT), sentence-transformers/all-MiniLM-L6-v2 (Apache-2.0),
+intfloat/multilingual-e5-* (MIT), google/embeddinggemma-300m (Gemma Terms of Use).
+Benchmark datasets are downloaded by the scripts and are not redistributed.

shad0wllm-0.1.0/PKG-INFO

@@ -0,0 +1,221 @@
+Metadata-Version: 2.4
+Name: shad0wllm
+Version: 0.1.0
+Summary: Compile the decisions your LLM already makes into a certified 2 MB table that answers in microseconds on a CPU.
+Author-email: Gaurav Chauhan <2796gaurav@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://2796gaurav.github.io/shad0w
+Project-URL: Documentation, https://2796gaurav.github.io/shad0w/docs/
+Project-URL: Demo, https://2796gaurav.github.io/shad0w/demo/
+Project-URL: Source, https://github.com/2796gaurav/shad0w
+Project-URL: Issues, https://github.com/2796gaurav/shad0w/issues
+Project-URL: Changelog, https://github.com/2796gaurav/shad0w/blob/main/CHANGELOG.md
+Keywords: llm,distillation,classification,intent-detection,routing,cascade,conformal-prediction,learn-then-test,edge-ai,cpu-inference,guardrails,llm-cost,latency
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: C
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Text Processing :: Linguistic
+Classifier: Typing :: Typed
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: numpy>=1.26
+Provides-Extra: compile
+Requires-Dist: scipy>=1.13; extra == "compile"
+Requires-Dist: scikit-learn>=1.5; extra == "compile"
+Requires-Dist: torch>=2.4; extra == "compile"
+Provides-Extra: logs
+Requires-Dist: shad0wllm[compile]; extra == "logs"
+Requires-Dist: sentence-transformers>=3.0; extra == "logs"
+Provides-Extra: demo
+Requires-Dist: shad0wllm[logs]; extra == "demo"
+Requires-Dist: datasets>=3.0; extra == "demo"
+Provides-Extra: dev
+Requires-Dist: shad0wllm[compile]; extra == "dev"
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: ruff>=0.6; extra == "dev"
+Dynamic: license-file
+
+<p align="center">
+  <a href="https://2796gaurav.github.io/shad0w">
+    <img src="https://2796gaurav.github.io/shad0w/assets/hero.svg" alt="shad0w: your LLM's decisions, compiled into a certified table that answers in microseconds" width="100%">
+  </a>
+</p>
+
+<h1 align="center">shad0w</h1>
+
+<p align="center">
+  <b>Compile the decisions your LLM already makes into a certified 2&nbsp;MB table that answers in microseconds on a CPU.</b><br>
+  It learns from your model's own answers, needs no labels, and tells you in writing how often it will disagree.
+</p>
+
+<p align="center">
+  <a href="https://pypi.org/project/shad0wllm/"><img alt="PyPI" src="https://img.shields.io/pypi/v/shad0wllm?color=7c5cff&label=pypi"></a>
+  <a href="https://www.npmjs.com/package/shad0wllm"><img alt="npm" src="https://img.shields.io/npm/v/shad0wllm?color=7c5cff&label=npm"></a>
+  <a href="https://github.com/2796gaurav/shad0w/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/2796gaurav/shad0w/actions/workflows/ci.yml/badge.svg"></a>
+  <img alt="Python" src="https://img.shields.io/badge/python-3.10%E2%80%933.13-3776ab">
+  <a href="LICENSE"><img alt="License" src="https://img.shields.io/badge/license-Apache--2.0-2ea44f"></a>
+</p>
+
+<p align="center">
+  <a href="https://2796gaurav.github.io/shad0w/demo/"><b>Live demo</b></a> ·
+  <a href="https://2796gaurav.github.io/shad0w/docs/"><b>Docs</b></a> ·
+  <a href="https://2796gaurav.github.io/shad0w/#video"><b>2-minute video</b></a> ·
+  <a href="https://2796gaurav.github.io/shad0w/docs/benchmarks.html"><b>Benchmarks</b></a>
+</p>
+
+---
+
+Your app sends small, typed decisions to a big model millions of times: *which intent is this, which queue gets this ticket, is this spam, should this agent step run*. Every call takes 20–1,500 ms and a round trip to someone else's GPU.
+
+**shad0w** sits in front of that model like a shadow. It watches the answers, learns the decision, and compiles it into a tiny integer table. Before the table answers anything, it **certifies how often it will disagree with your model**. On public intent and topic benchmarks, it then served **45–71%** of traffic itself, at under 5% certified disagreement and in microseconds. Everything else goes to your model, exactly as before.
+
+```python
+import shad0w
+
+sh = shad0w.Shadow("bundle/", teacher=ask_my_llm, log="teacher_log.jsonl")
+
+sh.decide("my card was stolen yesterday")
+# Decision(answer='lost_card', source='table', confidence=0.98, certified=True, latency_us=14)
+```
+
+<table>
+<tr>
+<td align="center"><b>~1 µs</b><br><sub>per decision in C<br>7–11 µs JS · 15 µs Python</sub></td>
+<td align="center"><b>45–71%</b><br><sub>of LLM traffic served locally<br>at a certified 5% bound</sub></td>
+<td align="center"><b>0 labels</b><br><sub>learns from your model's<br>own logged answers</sub></td>
+<td align="center"><b>2 MB</b><br><sub>one file, no GPU, no network<br>256 KB still keeps 88.5%</sub></td>
+</tr>
+</table>
+
+## How it works
+
+<p align="center">
+  <img src="https://2796gaurav.github.io/shad0w/assets/how-it-works.svg" alt="shadow → compile → certify → serve" width="100%">
+</p>
+
+1. **Shadow.** Your model keeps serving. Its answers are logged as `(text, answer)` pairs, either by the `Shadow` wrapper or by your own logging.
+2. **Compile.** The log trains a hashed n‑gram table with int8 weights. There is no tokenizer and no GPU, and it takes seconds to minutes on a laptop.
+3. **Certify.** On a held-out slice, a [Learn‑then‑Test](https://arxiv.org/abs/2110.01052) procedure picks the confidence threshold. Among answers the table serves, disagreement with your model is ≤ α (say 5%), with probability ≥ 1 − δ.
+4. **Serve.** Certified answers come back in microseconds, in your process. Everything else goes to your model and is logged, so the share the table can take grows with traffic: **31% → 70%** as AG News shadow traffic went from 2k to 100k.
+
+## Results
+
+Measured on public datasets. Every number links to its raw result file on the [benchmarks page](https://2796gaurav.github.io/shad0w/docs/benchmarks.html).
+
+| | Result |
+|---|---|
+| LLM traffic answered locally (7 intent/topic tasks, 2 teachers), certified α = 5% | **45–71%**, realised disagreement 3.9–5.7% (mean of 3 seeds) |
+| End-to-end accuracy (table + teacher) vs teacher alone | within **1 point** on every task; **+0.7** with Kev‑0.8B as teacher |
+| Mean latency of the cascade, Kev‑0.8B → BANKING77 | **75 ms → 22 ms** |
+| Offload as shadow traffic grows (AG News, 2k → 100k decisions) | **31% → 70%** at α = 5%; **5% → 52%** at 2% |
+| Per decision | **~1 µs** C · **7–11 µs** JavaScript · **15 µs** Python · 0.19 ms over HTTP |
+| Public benchmark (local‑jev‑bench, BANKING77, 20 options) | **89.7% at 0.018 ms on 1 CPU thread**; Kev‑0.8B 88% at 57.6 ms; Clef‑flash (9B) 97% at 1,274 ms |
+| Out-of-scope queries (CLINC150) separated by confidence | AUROC **0.94** |
+| **Reproduce on your laptop** (`examples/shadow_demo.py`, MIT-licensed teacher, BANKING77) | **69.4%** answered locally at α = 5% (realised 4.6%); cascade 67.8% vs teacher 68.0% |
+
+## Compared with what you would otherwise do
+
+| | Per decision | Runs on | Needs | Bound on what it serves |
+|---|---|---|---|---|
+| Keep calling the LLM / decision API | 20–1,500 ms | GPU or vendor API | — | it *is* the reference |
+| Semantic cache (GPTCache, gateway caches) | ms (embed + vector search) | embedding model + store | nothing | similarity threshold, no bound |
+| Embedding router (e.g. semantic-router) | ms | embedding model | example utterances | threshold, no bound |
+| Distil into a smaller LLM | 10–100 ms | GPU | teacher outputs | none |
+| Hand-built fastText / SetFit / Model2Vec | µs–ms | CPU | **human labels** | none |
+| **shad0w** | **1–15 µs** | **CPU, in-process, browser, edge** | **your model's logged answers** | **finite-sample, against your model** |
+
+It is not new science. The certificate is Learn‑then‑Test, in the spirit of [Trust or Escalate](https://arxiv.org/abs/2407.18370) and [BARGAIN](https://arxiv.org/abs/2509.02896). The table is [fastText](https://arxiv.org/abs/1607.01759)-shaped on purpose. What is new is the packaging: one workflow from shadow logs to a certified, KB-to-MB artifact, with the same decisions in C, Python and JavaScript.
+
+## Quick start
+
+```bash
+pip install "shad0wllm[compile]"        # the runtime alone is: pip install shad0wllm   (numpy only, C core included)
+```
+
+The package installs as `shad0wllm` and imports as `shad0w`. The command-line tool is `shad0w`.
+
+**1. Log your model's answers.** Wrap the call you already make. With no bundle yet, every call goes to your model and is logged:
+
+```python
+import shad0w
+
+sh = shad0w.Shadow(None, teacher=ask_my_llm, question="intent", log="teacher_log.jsonl")
+answer = sh.decide(text).answer          # = ask_my_llm(text), logged
+```
+
+**2. Compile and certify** once you have 1,000+ lines (10k–100k for high offload):
+
+```bash
+shad0w init                                        # starter schema.json: the options your model chooses from
+shad0w shadow --schema schema.json --data teacher_log.jsonl --teacher my-llm-v3 --alpha 0.05 --out bundle/
+shad0w report --bundle bundle/                      # the certificate, in plain words and numbers
+```
+
+**3. Serve.** Point the wrapper at the bundle. Certified answers come from the table and the rest from your model. 1% of certified answers are spot-checked against your model, so `sh.stats()` shows live disagreement:
+
+```python
+sh = shad0w.Shadow("bundle/", teacher=ask_my_llm, log="teacher_log.jsonl")
+```
+
+**4. Re-certify** on fresh traffic every week, or when traffic changes:
+
+```bash
+shad0w certify --bundle bundle/ --data last_week.jsonl
+```
+
+### Everywhere else
+
+| | |
+|---|---|
+| **JavaScript** (Node, browsers, Workers, Deno, Bun), zero dependencies | `npm i shad0wllm` → `const b = await Bundle.load("bundle/"); b.decide(text)` |
+| **HTTP** sidecar, any language | `shad0w serve --bundle bundle/` → `POST /v1/decide` |
+| **C**, any language with an FFI | `shad0w/_native/reflex.c`: `s0_load`, `s0_decide` |
+| **Decorator** | `@shad0w.cascade("bundle/")` on your existing classify function |
+| **Human labels** instead of a teacher | `shad0w compile --data labels.jsonl` + `shad0w calibrate` (~300 labels) |
+| **No teacher, no labels** | `shad0w compile --unlabeled logs.txt` (option names + raw logs, `pip install "shad0wllm[logs]"`) |
+
+Integration recipes for FastAPI, LiteLLM, LangGraph, Cloudflare Workers and the browser are in the [docs](https://2796gaurav.github.io/shad0w/docs/integrations.html).
+
+## When not to use it
+
+The limits below are measured. Knowing them is what makes the certificate useful.
+
+- **Sentiment, tone, emotion.** A hashed table tops out near 79% on SST‑2 even with perfect labels. shad0w certifies almost nothing there and defers everything to your model, which is the correct behaviour.
+- **When you need the most accurate answer.** A 9B decision model is 7 points better on the public BANKING77 file, at ~70,000× the latency. Put shad0w *in front of* such a model, not instead of it.
+- **The bound is against your model, not the truth.** A certified table in front of a wrong model is wrong in the same way. For a bound against the truth, calibrate with ~300 real labels.
+- **The bound holds on traffic like the calibration slice.** A big shift in the request mix can break it. The drift guard flags confidence shifts, the audits watch live disagreement, and `shad0w certify` re-certifies.
+- **Reasoning or world knowledge.** It learns *surface* decisions: intents, routes, topics, policies with distinctive wording.
+
+## Documentation
+
+[Getting started](https://2796gaurav.github.io/shad0w/docs/) · [Shadow mode guide](https://2796gaurav.github.io/shad0w/docs/shadow-mode.html) · [Integrations](https://2796gaurav.github.io/shad0w/docs/integrations.html) · [The certificate, explained](https://2796gaurav.github.io/shad0w/docs/certificate.html) · [API & CLI](https://2796gaurav.github.io/shad0w/docs/reference.html) · [Architecture & file format](https://2796gaurav.github.io/shad0w/docs/architecture.html) · [Benchmarks](https://2796gaurav.github.io/shad0w/docs/benchmarks.html) · [Limitations](https://2796gaurav.github.io/shad0w/docs/limitations.html) · [FAQ](https://2796gaurav.github.io/shad0w/docs/faq.html)
+
+## Contributing
+
+Run it on one of your decisions and [share your numbers](https://github.com/2796gaurav/shad0w/issues/new?template=results.yml), including failures. That is the most valuable contribution. Code contributions: see [CONTRIBUTING.md](CONTRIBUTING.md).
+
+## Citation
+
+```bibtex
+@software{chauhan2026shad0w,
+  author = {Chauhan, Gaurav},
+  title  = {shad0w: certified microsecond decisions compiled from the model you already run},
+  year   = {2026},
+  url    = {https://github.com/2796gaurav/shad0w}
+}
+```
+
+---
+
+<p align="center">Built by <a href="https://github.com/2796gaurav">Gaurav Chauhan</a> · Apache-2.0</p>

shad0wllm-0.1.0/README.md

@@ -0,0 +1,173 @@
+<p align="center">
+  <a href="https://2796gaurav.github.io/shad0w">
+    <img src="https://2796gaurav.github.io/shad0w/assets/hero.svg" alt="shad0w: your LLM's decisions, compiled into a certified table that answers in microseconds" width="100%">
+  </a>
+</p>
+
+<h1 align="center">shad0w</h1>
+
+<p align="center">
+  <b>Compile the decisions your LLM already makes into a certified 2&nbsp;MB table that answers in microseconds on a CPU.</b><br>
+  It learns from your model's own answers, needs no labels, and tells you in writing how often it will disagree.
+</p>
+
+<p align="center">
+  <a href="https://pypi.org/project/shad0wllm/"><img alt="PyPI" src="https://img.shields.io/pypi/v/shad0wllm?color=7c5cff&label=pypi"></a>
+  <a href="https://www.npmjs.com/package/shad0wllm"><img alt="npm" src="https://img.shields.io/npm/v/shad0wllm?color=7c5cff&label=npm"></a>
+  <a href="https://github.com/2796gaurav/shad0w/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/2796gaurav/shad0w/actions/workflows/ci.yml/badge.svg"></a>
+  <img alt="Python" src="https://img.shields.io/badge/python-3.10%E2%80%933.13-3776ab">
+  <a href="LICENSE"><img alt="License" src="https://img.shields.io/badge/license-Apache--2.0-2ea44f"></a>
+</p>
+
+<p align="center">
+  <a href="https://2796gaurav.github.io/shad0w/demo/"><b>Live demo</b></a> ·
+  <a href="https://2796gaurav.github.io/shad0w/docs/"><b>Docs</b></a> ·
+  <a href="https://2796gaurav.github.io/shad0w/#video"><b>2-minute video</b></a> ·
+  <a href="https://2796gaurav.github.io/shad0w/docs/benchmarks.html"><b>Benchmarks</b></a>
+</p>
+
+---
+
+Your app sends small, typed decisions to a big model millions of times: *which intent is this, which queue gets this ticket, is this spam, should this agent step run*. Every call takes 20–1,500 ms and a round trip to someone else's GPU.
+
+**shad0w** sits in front of that model like a shadow. It watches the answers, learns the decision, and compiles it into a tiny integer table. Before the table answers anything, it **certifies how often it will disagree with your model**. On public intent and topic benchmarks, it then served **45–71%** of traffic itself, at under 5% certified disagreement and in microseconds. Everything else goes to your model, exactly as before.
+
+```python
+import shad0w
+
+sh = shad0w.Shadow("bundle/", teacher=ask_my_llm, log="teacher_log.jsonl")
+
+sh.decide("my card was stolen yesterday")
+# Decision(answer='lost_card', source='table', confidence=0.98, certified=True, latency_us=14)
+```
+
+<table>
+<tr>
+<td align="center"><b>~1 µs</b><br><sub>per decision in C<br>7–11 µs JS · 15 µs Python</sub></td>
+<td align="center"><b>45–71%</b><br><sub>of LLM traffic served locally<br>at a certified 5% bound</sub></td>
+<td align="center"><b>0 labels</b><br><sub>learns from your model's<br>own logged answers</sub></td>
+<td align="center"><b>2 MB</b><br><sub>one file, no GPU, no network<br>256 KB still keeps 88.5%</sub></td>
+</tr>
+</table>
+
+## How it works
+
+<p align="center">
+  <img src="https://2796gaurav.github.io/shad0w/assets/how-it-works.svg" alt="shadow → compile → certify → serve" width="100%">
+</p>
+
+1. **Shadow.** Your model keeps serving. Its answers are logged as `(text, answer)` pairs, either by the `Shadow` wrapper or by your own logging.
+2. **Compile.** The log trains a hashed n‑gram table with int8 weights. There is no tokenizer and no GPU, and it takes seconds to minutes on a laptop.
+3. **Certify.** On a held-out slice, a [Learn‑then‑Test](https://arxiv.org/abs/2110.01052) procedure picks the confidence threshold. Among answers the table serves, disagreement with your model is ≤ α (say 5%), with probability ≥ 1 − δ.
+4. **Serve.** Certified answers come back in microseconds, in your process. Everything else goes to your model and is logged, so the share the table can take grows with traffic: **31% → 70%** as AG News shadow traffic went from 2k to 100k.
+
+## Results
+
+Measured on public datasets. Every number links to its raw result file on the [benchmarks page](https://2796gaurav.github.io/shad0w/docs/benchmarks.html).
+
+| | Result |
+|---|---|
+| LLM traffic answered locally (7 intent/topic tasks, 2 teachers), certified α = 5% | **45–71%**, realised disagreement 3.9–5.7% (mean of 3 seeds) |
+| End-to-end accuracy (table + teacher) vs teacher alone | within **1 point** on every task; **+0.7** with Kev‑0.8B as teacher |
+| Mean latency of the cascade, Kev‑0.8B → BANKING77 | **75 ms → 22 ms** |
+| Offload as shadow traffic grows (AG News, 2k → 100k decisions) | **31% → 70%** at α = 5%; **5% → 52%** at 2% |
+| Per decision | **~1 µs** C · **7–11 µs** JavaScript · **15 µs** Python · 0.19 ms over HTTP |
+| Public benchmark (local‑jev‑bench, BANKING77, 20 options) | **89.7% at 0.018 ms on 1 CPU thread**; Kev‑0.8B 88% at 57.6 ms; Clef‑flash (9B) 97% at 1,274 ms |
+| Out-of-scope queries (CLINC150) separated by confidence | AUROC **0.94** |
+| **Reproduce on your laptop** (`examples/shadow_demo.py`, MIT-licensed teacher, BANKING77) | **69.4%** answered locally at α = 5% (realised 4.6%); cascade 67.8% vs teacher 68.0% |
+
+## Compared with what you would otherwise do
+
+| | Per decision | Runs on | Needs | Bound on what it serves |
+|---|---|---|---|---|
+| Keep calling the LLM / decision API | 20–1,500 ms | GPU or vendor API | — | it *is* the reference |
+| Semantic cache (GPTCache, gateway caches) | ms (embed + vector search) | embedding model + store | nothing | similarity threshold, no bound |
+| Embedding router (e.g. semantic-router) | ms | embedding model | example utterances | threshold, no bound |
+| Distil into a smaller LLM | 10–100 ms | GPU | teacher outputs | none |
+| Hand-built fastText / SetFit / Model2Vec | µs–ms | CPU | **human labels** | none |
+| **shad0w** | **1–15 µs** | **CPU, in-process, browser, edge** | **your model's logged answers** | **finite-sample, against your model** |
+
+It is not new science. The certificate is Learn‑then‑Test, in the spirit of [Trust or Escalate](https://arxiv.org/abs/2407.18370) and [BARGAIN](https://arxiv.org/abs/2509.02896). The table is [fastText](https://arxiv.org/abs/1607.01759)-shaped on purpose. What is new is the packaging: one workflow from shadow logs to a certified, KB-to-MB artifact, with the same decisions in C, Python and JavaScript.
+
+## Quick start
+
+```bash
+pip install "shad0wllm[compile]"        # the runtime alone is: pip install shad0wllm   (numpy only, C core included)
+```
+
+The package installs as `shad0wllm` and imports as `shad0w`. The command-line tool is `shad0w`.
+
+**1. Log your model's answers.** Wrap the call you already make. With no bundle yet, every call goes to your model and is logged:
+
+```python
+import shad0w
+
+sh = shad0w.Shadow(None, teacher=ask_my_llm, question="intent", log="teacher_log.jsonl")
+answer = sh.decide(text).answer          # = ask_my_llm(text), logged
+```
+
+**2. Compile and certify** once you have 1,000+ lines (10k–100k for high offload):
+
+```bash
+shad0w init                                        # starter schema.json: the options your model chooses from
+shad0w shadow --schema schema.json --data teacher_log.jsonl --teacher my-llm-v3 --alpha 0.05 --out bundle/
+shad0w report --bundle bundle/                      # the certificate, in plain words and numbers
+```
+
+**3. Serve.** Point the wrapper at the bundle. Certified answers come from the table and the rest from your model. 1% of certified answers are spot-checked against your model, so `sh.stats()` shows live disagreement:
+
+```python
+sh = shad0w.Shadow("bundle/", teacher=ask_my_llm, log="teacher_log.jsonl")
+```
+
+**4. Re-certify** on fresh traffic every week, or when traffic changes:
+
+```bash
+shad0w certify --bundle bundle/ --data last_week.jsonl
+```
+
+### Everywhere else
+
+| | |
+|---|---|
+| **JavaScript** (Node, browsers, Workers, Deno, Bun), zero dependencies | `npm i shad0wllm` → `const b = await Bundle.load("bundle/"); b.decide(text)` |
+| **HTTP** sidecar, any language | `shad0w serve --bundle bundle/` → `POST /v1/decide` |
+| **C**, any language with an FFI | `shad0w/_native/reflex.c`: `s0_load`, `s0_decide` |
+| **Decorator** | `@shad0w.cascade("bundle/")` on your existing classify function |
+| **Human labels** instead of a teacher | `shad0w compile --data labels.jsonl` + `shad0w calibrate` (~300 labels) |
+| **No teacher, no labels** | `shad0w compile --unlabeled logs.txt` (option names + raw logs, `pip install "shad0wllm[logs]"`) |
+
+Integration recipes for FastAPI, LiteLLM, LangGraph, Cloudflare Workers and the browser are in the [docs](https://2796gaurav.github.io/shad0w/docs/integrations.html).
+
+## When not to use it
+
+The limits below are measured. Knowing them is what makes the certificate useful.
+
+- **Sentiment, tone, emotion.** A hashed table tops out near 79% on SST‑2 even with perfect labels. shad0w certifies almost nothing there and defers everything to your model, which is the correct behaviour.
+- **When you need the most accurate answer.** A 9B decision model is 7 points better on the public BANKING77 file, at ~70,000× the latency. Put shad0w *in front of* such a model, not instead of it.
+- **The bound is against your model, not the truth.** A certified table in front of a wrong model is wrong in the same way. For a bound against the truth, calibrate with ~300 real labels.
+- **The bound holds on traffic like the calibration slice.** A big shift in the request mix can break it. The drift guard flags confidence shifts, the audits watch live disagreement, and `shad0w certify` re-certifies.
+- **Reasoning or world knowledge.** It learns *surface* decisions: intents, routes, topics, policies with distinctive wording.
+
+## Documentation
+
+[Getting started](https://2796gaurav.github.io/shad0w/docs/) · [Shadow mode guide](https://2796gaurav.github.io/shad0w/docs/shadow-mode.html) · [Integrations](https://2796gaurav.github.io/shad0w/docs/integrations.html) · [The certificate, explained](https://2796gaurav.github.io/shad0w/docs/certificate.html) · [API & CLI](https://2796gaurav.github.io/shad0w/docs/reference.html) · [Architecture & file format](https://2796gaurav.github.io/shad0w/docs/architecture.html) · [Benchmarks](https://2796gaurav.github.io/shad0w/docs/benchmarks.html) · [Limitations](https://2796gaurav.github.io/shad0w/docs/limitations.html) · [FAQ](https://2796gaurav.github.io/shad0w/docs/faq.html)
+
+## Contributing
+
+Run it on one of your decisions and [share your numbers](https://github.com/2796gaurav/shad0w/issues/new?template=results.yml), including failures. That is the most valuable contribution. Code contributions: see [CONTRIBUTING.md](CONTRIBUTING.md).
+
+## Citation
+
+```bibtex
+@software{chauhan2026shad0w,
+  author = {Chauhan, Gaurav},
+  title  = {shad0w: certified microsecond decisions compiled from the model you already run},
+  year   = {2026},
+  url    = {https://github.com/2796gaurav/shad0w}
+}
+```
+
+---
+
+<p align="center">Built by <a href="https://github.com/2796gaurav">Gaurav Chauhan</a> · Apache-2.0</p>