brier 0.1.0__tar.gz

brier-0.1.0/.gitignore

@@ -0,0 +1,24 @@
+__pycache__/
+*.py[cod]
+.venv/
+build/
+dist/
+*.egg-info/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+coverage.xml
+htmlcov/
+.hypothesis/
+# local config and secrets
+.env
+.env.*
+.claude/settings.local.json
+# benchmark outputs, caches, model files
+results/
+cache/
+*.safetensors
+*.bin
+*.pt
+site/

brier-0.1.0/CHANGELOG.md

@@ -0,0 +1,49 @@
+# Changelog
+
+All notable changes are documented here. Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/);
+versioning: [Semantic Versioning](https://semver.org/).
+
+## [Unreleased]
+
+## [0.1.0] - 2026-10-03
+
+First public release.
+
+### Added
+- **Typed questions** `Choice`, `Noul` (yes/no) and `Score`, answered as probability
+  distributions read from next-token log-probabilities (no generation, no fine-tuning). Every
+  `Decision` records the level that produced it.
+- **Correction levels** in `Decider` (`decide`, `decide_batch`):
+  - `raw`: label-token readout (Score falls back to lettered options when digits are not single
+    tokens);
+  - `L0`, no labels: cyclic option rotations with a geometric-mean combine, plus an optional
+    batch prior from unlabelled states (`fit_prior`; ADR-0005);
+  - `L1`: temperature scaling (`fit_temperature`, at least 50 labels);
+  - `L2`: hidden-state heads (`fit_head`, at least 60 labels and 5 per answer): ridge and
+    Ledoit–Wolf shrinkage LDA, layer / solver / alpha chosen by stratified 5-fold out-of-fold
+    NLL; L2 temperature with Platt (1999) targets for two classes and plain temperature
+    scaling beyond (METHODS.md).
+- **Calibration artifacts**: `Decider.save` / `Decider.load` (`artifact.json` + `arrays.npz`,
+  never model weights), loaded as untrusted input (SHA-256, size caps, no pickle) and bound to
+  the model id, revision, dtype and prompt templates (ADR-0003, 0006, 0007).
+- **`HFBackend`** for Hugging Face causal LMs: safetensors only, `trust_remote_code` never set,
+  special-token-safe prompts, batched readout and hidden states. Chat models are prompted through
+  their chat template, base models through a plain-text prompt (ADR-0008).
+- **`brier check <model-id>`**: a conformance report before calibrating (label tokens, batching,
+  hidden states, a 14-item sanity task, supported levels); `--json`.
+- **Benchmark** `python -m brier.bench run` (banking20 from BANKING77, CC-BY-4.0, pinned and
+  SHA-256 verified): all levels, bootstrap and paired CIs, flip rate, L2 label curves, per-item
+  outputs without state text. Results in `docs/results.md`.
+- **Model compatibility**: integration suite and `brier check` pass on 15 models from Qwen3,
+  SmolLM2/3, TinyLlama, OLMo-2, Gemma 3, Granite-MoE, LFM2, DeepSeek-R1-Distill, Falcon3,
+  Phi-4-mini, OLMoE and Mistral, including two mixtures of experts and two base models
+  (`docs/COMPATIBILITY.md`).
+- Documentation: README quickstart (run in CI), API reference on GitHub Pages, SPEC, METHODS,
+  ARCHITECTURE, EVALUATION, THREAT_MODEL, related work and eight ADRs.
+
+### Security
+- GitHub Actions and pre-commit hooks pinned to full commit SHAs; CodeQL, `pip-audit` and
+  Dependabot in CI.
+
+[Unreleased]: https://github.com/mohdUwaish59/brier/compare/v0.1.0...HEAD
+[0.1.0]: https://github.com/mohdUwaish59/brier/releases/tag/v0.1.0

brier-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.

brier-0.1.0/NOTICE

@@ -0,0 +1,4 @@
+brier
+Copyright 2026 Mohd Uwaish
+
+This product includes software developed as part of the brier project.

brier-0.1.0/PKG-INFO

@@ -0,0 +1,211 @@
+Metadata-Version: 2.5
+Name: brier
+Version: 0.1.0
+Summary: Calibrated typed decisions (choice / yes-no / score) from open LLMs, without generation or fine-tuning.
+Project-URL: Homepage, https://github.com/mohdUwaish59/brier
+Project-URL: Documentation, https://mohduwaish59.github.io/brier/
+Project-URL: Issues, https://github.com/mohdUwaish59/brier/issues
+Project-URL: Changelog, https://github.com/mohdUwaish59/brier/blob/main/CHANGELOG.md
+Author: Mohd Uwaish
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: calibration,classification,decision,jev,llm,logits,probability
+Classifier: Development Status :: 3 - Alpha
+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: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Typing :: Typed
+Requires-Python: >=3.10
+Requires-Dist: numpy>=1.24
+Provides-Extra: bench
+Requires-Dist: safetensors>=0.4; extra == 'bench'
+Requires-Dist: torch>=2.2; extra == 'bench'
+Requires-Dist: transformers>=5.0; extra == 'bench'
+Provides-Extra: hf
+Requires-Dist: safetensors>=0.4; extra == 'hf'
+Requires-Dist: torch>=2.2; extra == 'hf'
+Requires-Dist: transformers>=5.0; extra == 'hf'
+Description-Content-Type: text/markdown
+
+# brier
+
+[![CI](https://github.com/mohdUwaish59/brier/actions/workflows/ci.yml/badge.svg)](https://github.com/mohdUwaish59/brier/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/brier)](https://pypi.org/project/brier/)
+[![Python](https://img.shields.io/pypi/pyversions/brier)](https://pypi.org/project/brier/)
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/mohdUwaish59/brier/blob/main/LICENSE)
+[![Docs](https://img.shields.io/badge/docs-API%20reference-informational)](https://mohduwaish59.github.io/brier/)
+
+> Turn an open LLM into a calibrated decision model: choice, yes/no and score questions
+> answered with real probabilities, read from the model in a forward pass, without generation
+> or fine-tuning.
+
+**Status: alpha (0.1).** The API may still change before 1.0. See the
+[roadmap](https://github.com/mohdUwaish59/brier/blob/main/docs/ROADMAP.md).
+
+**Try it:** [tutorial notebook](https://colab.research.google.com/github/mohdUwaish59/brier/blob/main/notebooks/tutorial_ticket_router.ipynb):
+a calibrated ticket router from an open model in about 15 minutes, on a free Colab GPU.
+
+## Why
+
+Reading an LLM's answer from generated text is slow and needs parsing. Raw next-token
+probabilities skip that, but they are biased by option order and badly calibrated. brier reads
+the answer distribution directly and corrects it in levels. Every decision records the level
+that produced it.
+
+| Level | Labels needed | What it fixes | Forward passes |
+|---|---|---|---|
+| `raw` | 0 | baseline readout of the label tokens | 1 |
+| `L0` | 0 (optional unlabelled pool) | option-order bias, label prior bias | K for a K-option Choice, else 1 |
+| `L1` | ≥ 50 | over/under-confidence (one temperature) | as L0 |
+| `L2` | ≥ 60, ≥ 5 per answer | accuracy and calibration via a hidden-state head | 1 |
+
+## Results
+
+banking20 (20 BANKING77 intents), Qwen3-1.7B, 2,603 test items, 95 % bootstrap CIs.
+L1 and L2 use 300 labels. Full tables, protocol and caveats:
+[`docs/results.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/results.md).
+
+| | raw | L0 | L1 | L2 |
+|---|---|---|---|---|
+| Accuracy ↑ | 0.626 | 0.668 | 0.668 | **0.807** [0.793, 0.823] |
+| ECE ↓ | 0.357 | 0.299 | 0.094 | **0.035** [0.028, 0.050] |
+| NLL ↓ | 6.541 | 3.845 | 1.204 | **0.727** [0.672, 0.781] |
+
+Across six model families (same task and labels; point estimates):
+
+| Model | Accuracy: raw → L0 → L2 | ECE: raw → L1 → L2 | Order flips: raw → L0 |
+|---|---|---|---|
+| Qwen3-1.7B | 0.626 → 0.668 → **0.807** | 0.357 → 0.094 → **0.035** | 0.364 → 0.201 |
+| Falcon3-1B-Base (no chat template) | 0.155 → 0.601 → **0.804** | 0.119 → 0.162 → **0.072** | 0.990 → 0.371 |
+| LFM2-1.2B (hybrid convolution) | 0.200 → 0.617 → **0.802** | 0.336 → 0.064 → **0.036** | 0.972 → 0.415 |
+| SmolLM3-3B | 0.660 → 0.746 → **0.834** | 0.190 → **0.030** → 0.039 | 0.279 → 0.123 |
+| Phi-4-mini (3.8B) | 0.711 → 0.786 → **0.859** | 0.186 → **0.022** → 0.057 | 0.275 → 0.099 |
+| OLMoE-1B-7B (mixture of experts) | 0.257 → 0.686 → **0.819** | 0.068 → 0.098 → **0.084** | 0.945 → 0.316 |
+
+Zero labels (L0) raise accuracy and cut order flips on every model; with 300 labels (L2) every
+model lands at 0.80–0.86 accuracy. One task so far: evidence that the method works, not a
+general claim. Details and caveats in
+[`docs/results.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/results.md).
+
+## Install
+
+```bash
+pip install "brier[hf]"
+```
+
+Python ≥ 3.10. The core needs only numpy; the `hf` extra adds torch and transformers.
+
+## Quickstart
+
+```python
+from brier import Choice, Decider, Noul
+from brier.backends.hf import HFBackend
+
+d = Decider(HFBackend("Qwen/Qwen3-1.7B", revision="<commit-sha>"))
+route = Choice("Which team?", ["billing", "technical", "sales"], name="route")
+refund = Noul("Is this a refund request?", name="refund")
+
+res = d.decide("My card was charged twice, please fix it now!", [route, refund], level="L0")
+print(res["route"].answer, res["route"].probs)  # top option and all option probabilities
+print(res["refund"].p_yes, res["refund"].level)  # probability of yes, and "L0"
+```
+
+Calibrate with your own data, save the calibration, and reuse it:
+
+```python
+d.fit_prior(unlabelled_states, [route])  # L0 prior, no labels
+d.fit_temperature(labelled_states, route, route_labels)  # L1, >= 50 labels
+d.fit_head(labelled_states, route, route_labels)  # L2, >= 60 labels, >= 5 per answer
+d.decide(state, [route], level="L2")
+
+d.save("calibration/")  # JSON + .npz, never model weights
+d = Decider.load("calibration/", HFBackend("Qwen/Qwen3-1.7B", revision="<commit-sha>"))
+```
+
+`load` refuses a calibration made for another model, revision, precision (`dtype`) or prompt
+template. Pin `revision` to a commit SHA: a model name alone does not pin weights.
+
+Questions:
+- **`Choice(text, options, name=...)`**: one of K options.
+- **`Noul(text, name=...)`**: yes/no; `p_yes` is the probability of yes.
+- **`Score(text, levels, name=...)`**: an integer rating `1..levels`; `expected` is the mean.
+
+## Does it work with my model?
+
+Hugging Face causal LMs with safetensors weights: chat models through their chat template, base
+models through a plain-text prompt. Check yours before calibrating:
+
+```text
+$ brier check Qwen/Qwen3-0.6B --revision c1899de289a04d12100db370d81485cdf75e47ca
+
+  PASS  revision           c1899de289a04d12100db370d81485cdf75e47ca
+  PASS  prompt_format      the model's chat template
+  PASS  labels.choice      A-Z: single, distinct tokens
+  PASS  labels.noul        Yes, No: single, distinct tokens
+  PASS  labels.score       10 levels as lettered options (digit labels are not single tokens)
+  PASS  batch_consistency  max |p_batched - p_single| = 5.3e-06 (limit 2e-02)
+  PASS  hidden_states      7 candidate layers (12-24 of 28), d = 1024
+  PASS  sanity             accuracy raw 79% / L0 79% on 14 items, order flips raw 50% / L0 17%, ...
+
+Supported levels: raw, L0, L1, L2
+```
+
+Tested so far (15 models): Qwen3, SmolLM2 / SmolLM3, TinyLlama, OLMo-2, Gemma 3, Granite-MoE,
+LFM2, DeepSeek-R1-Distill, Falcon3, Phi-4-mini, OLMoE and Mistral, including two mixtures of
+experts and two base models. Details and known quirks:
+[`docs/COMPATIBILITY.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/COMPATIBILITY.md).
+
+## Documentation
+
+- [API reference](https://mohduwaish59.github.io/brier/)
+- [`SPEC.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/SPEC.md): behaviour and limits
+- [`METHODS.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/METHODS.md): the maths of every level, with references
+- [`EVALUATION.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/EVALUATION.md) and
+  [`results.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/results.md): benchmark protocol and results
+- [`related_work.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/related_work.md): where the idea and the methods come from
+
+## Limitations
+
+- One benchmark task so far (banking20); more tasks are welcome.
+- `L0` costs K forward passes for a K-option question; there is no prefix caching yet.
+- Hugging Face `transformers` only (no vLLM backend yet); models need safetensors weights and
+  the standard decoder layout.
+- Decisions on untrusted text can be steered by prompt injection (see Security).
+
+## Contributing
+
+Contributions are welcome: new models in the compatibility table (run `brier check` and open a
+PR), benchmark tasks, backends and docs. Start with
+[`CONTRIBUTING.md`](https://github.com/mohdUwaish59/brier/blob/main/CONTRIBUTING.md) and the
+issues labelled *good first issue*.
+
+## Security
+
+Decisions on untrusted text can be manipulated by prompt injection. Don't use them as the only
+gate for destructive actions. brier never runs model code (`trust_remote_code` is never set),
+loads weights from safetensors only, and validates calibration files as untrusted input. Report
+vulnerabilities privately; see
+[SECURITY.md](https://github.com/mohdUwaish59/brier/blob/main/SECURITY.md) and
+[`THREAT_MODEL.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/THREAT_MODEL.md).
+
+## Related work
+
+brier follows the idea of TypeSafe AI's Jev and of
+[AnyJev](https://github.com/nokia-applied-research/AnyJev) (Nokia Applied Research): typed
+questions answered from next-token probabilities, corrected in levels. It is an independent
+implementation, not affiliated with either. The correction methods come from published work
+(Zheng et al. 2024; Zhou et al. 2024; Guo et al. 2017; Platt 1999; Ledoit & Wolf 2004).
+[`related_work.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/related_work.md) sets
+out what brier shares with these projects and where it differs.
+
+## License
+
+Apache-2.0

brier-0.1.0/README.md

@@ -0,0 +1,174 @@
+# brier
+
+[![CI](https://github.com/mohdUwaish59/brier/actions/workflows/ci.yml/badge.svg)](https://github.com/mohdUwaish59/brier/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/brier)](https://pypi.org/project/brier/)
+[![Python](https://img.shields.io/pypi/pyversions/brier)](https://pypi.org/project/brier/)
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/mohdUwaish59/brier/blob/main/LICENSE)
+[![Docs](https://img.shields.io/badge/docs-API%20reference-informational)](https://mohduwaish59.github.io/brier/)
+
+> Turn an open LLM into a calibrated decision model: choice, yes/no and score questions
+> answered with real probabilities, read from the model in a forward pass, without generation
+> or fine-tuning.
+
+**Status: alpha (0.1).** The API may still change before 1.0. See the
+[roadmap](https://github.com/mohdUwaish59/brier/blob/main/docs/ROADMAP.md).
+
+**Try it:** [tutorial notebook](https://colab.research.google.com/github/mohdUwaish59/brier/blob/main/notebooks/tutorial_ticket_router.ipynb):
+a calibrated ticket router from an open model in about 15 minutes, on a free Colab GPU.
+
+## Why
+
+Reading an LLM's answer from generated text is slow and needs parsing. Raw next-token
+probabilities skip that, but they are biased by option order and badly calibrated. brier reads
+the answer distribution directly and corrects it in levels. Every decision records the level
+that produced it.
+
+| Level | Labels needed | What it fixes | Forward passes |
+|---|---|---|---|
+| `raw` | 0 | baseline readout of the label tokens | 1 |
+| `L0` | 0 (optional unlabelled pool) | option-order bias, label prior bias | K for a K-option Choice, else 1 |
+| `L1` | ≥ 50 | over/under-confidence (one temperature) | as L0 |
+| `L2` | ≥ 60, ≥ 5 per answer | accuracy and calibration via a hidden-state head | 1 |
+
+## Results
+
+banking20 (20 BANKING77 intents), Qwen3-1.7B, 2,603 test items, 95 % bootstrap CIs.
+L1 and L2 use 300 labels. Full tables, protocol and caveats:
+[`docs/results.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/results.md).
+
+| | raw | L0 | L1 | L2 |
+|---|---|---|---|---|
+| Accuracy ↑ | 0.626 | 0.668 | 0.668 | **0.807** [0.793, 0.823] |
+| ECE ↓ | 0.357 | 0.299 | 0.094 | **0.035** [0.028, 0.050] |
+| NLL ↓ | 6.541 | 3.845 | 1.204 | **0.727** [0.672, 0.781] |
+
+Across six model families (same task and labels; point estimates):
+
+| Model | Accuracy: raw → L0 → L2 | ECE: raw → L1 → L2 | Order flips: raw → L0 |
+|---|---|---|---|
+| Qwen3-1.7B | 0.626 → 0.668 → **0.807** | 0.357 → 0.094 → **0.035** | 0.364 → 0.201 |
+| Falcon3-1B-Base (no chat template) | 0.155 → 0.601 → **0.804** | 0.119 → 0.162 → **0.072** | 0.990 → 0.371 |
+| LFM2-1.2B (hybrid convolution) | 0.200 → 0.617 → **0.802** | 0.336 → 0.064 → **0.036** | 0.972 → 0.415 |
+| SmolLM3-3B | 0.660 → 0.746 → **0.834** | 0.190 → **0.030** → 0.039 | 0.279 → 0.123 |
+| Phi-4-mini (3.8B) | 0.711 → 0.786 → **0.859** | 0.186 → **0.022** → 0.057 | 0.275 → 0.099 |
+| OLMoE-1B-7B (mixture of experts) | 0.257 → 0.686 → **0.819** | 0.068 → 0.098 → **0.084** | 0.945 → 0.316 |
+
+Zero labels (L0) raise accuracy and cut order flips on every model; with 300 labels (L2) every
+model lands at 0.80–0.86 accuracy. One task so far: evidence that the method works, not a
+general claim. Details and caveats in
+[`docs/results.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/results.md).
+
+## Install
+
+```bash
+pip install "brier[hf]"
+```
+
+Python ≥ 3.10. The core needs only numpy; the `hf` extra adds torch and transformers.
+
+## Quickstart
+
+```python
+from brier import Choice, Decider, Noul
+from brier.backends.hf import HFBackend
+
+d = Decider(HFBackend("Qwen/Qwen3-1.7B", revision="<commit-sha>"))
+route = Choice("Which team?", ["billing", "technical", "sales"], name="route")
+refund = Noul("Is this a refund request?", name="refund")
+
+res = d.decide("My card was charged twice, please fix it now!", [route, refund], level="L0")
+print(res["route"].answer, res["route"].probs)  # top option and all option probabilities
+print(res["refund"].p_yes, res["refund"].level)  # probability of yes, and "L0"
+```
+
+Calibrate with your own data, save the calibration, and reuse it:
+
+```python
+d.fit_prior(unlabelled_states, [route])  # L0 prior, no labels
+d.fit_temperature(labelled_states, route, route_labels)  # L1, >= 50 labels
+d.fit_head(labelled_states, route, route_labels)  # L2, >= 60 labels, >= 5 per answer
+d.decide(state, [route], level="L2")
+
+d.save("calibration/")  # JSON + .npz, never model weights
+d = Decider.load("calibration/", HFBackend("Qwen/Qwen3-1.7B", revision="<commit-sha>"))
+```
+
+`load` refuses a calibration made for another model, revision, precision (`dtype`) or prompt
+template. Pin `revision` to a commit SHA: a model name alone does not pin weights.
+
+Questions:
+- **`Choice(text, options, name=...)`**: one of K options.
+- **`Noul(text, name=...)`**: yes/no; `p_yes` is the probability of yes.
+- **`Score(text, levels, name=...)`**: an integer rating `1..levels`; `expected` is the mean.
+
+## Does it work with my model?
+
+Hugging Face causal LMs with safetensors weights: chat models through their chat template, base
+models through a plain-text prompt. Check yours before calibrating:
+
+```text
+$ brier check Qwen/Qwen3-0.6B --revision c1899de289a04d12100db370d81485cdf75e47ca
+
+  PASS  revision           c1899de289a04d12100db370d81485cdf75e47ca
+  PASS  prompt_format      the model's chat template
+  PASS  labels.choice      A-Z: single, distinct tokens
+  PASS  labels.noul        Yes, No: single, distinct tokens
+  PASS  labels.score       10 levels as lettered options (digit labels are not single tokens)
+  PASS  batch_consistency  max |p_batched - p_single| = 5.3e-06 (limit 2e-02)
+  PASS  hidden_states      7 candidate layers (12-24 of 28), d = 1024
+  PASS  sanity             accuracy raw 79% / L0 79% on 14 items, order flips raw 50% / L0 17%, ...
+
+Supported levels: raw, L0, L1, L2
+```
+
+Tested so far (15 models): Qwen3, SmolLM2 / SmolLM3, TinyLlama, OLMo-2, Gemma 3, Granite-MoE,
+LFM2, DeepSeek-R1-Distill, Falcon3, Phi-4-mini, OLMoE and Mistral, including two mixtures of
+experts and two base models. Details and known quirks:
+[`docs/COMPATIBILITY.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/COMPATIBILITY.md).
+
+## Documentation
+
+- [API reference](https://mohduwaish59.github.io/brier/)
+- [`SPEC.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/SPEC.md): behaviour and limits
+- [`METHODS.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/METHODS.md): the maths of every level, with references
+- [`EVALUATION.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/EVALUATION.md) and
+  [`results.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/results.md): benchmark protocol and results
+- [`related_work.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/related_work.md): where the idea and the methods come from
+
+## Limitations
+
+- One benchmark task so far (banking20); more tasks are welcome.
+- `L0` costs K forward passes for a K-option question; there is no prefix caching yet.
+- Hugging Face `transformers` only (no vLLM backend yet); models need safetensors weights and
+  the standard decoder layout.
+- Decisions on untrusted text can be steered by prompt injection (see Security).
+
+## Contributing
+
+Contributions are welcome: new models in the compatibility table (run `brier check` and open a
+PR), benchmark tasks, backends and docs. Start with
+[`CONTRIBUTING.md`](https://github.com/mohdUwaish59/brier/blob/main/CONTRIBUTING.md) and the
+issues labelled *good first issue*.
+
+## Security
+
+Decisions on untrusted text can be manipulated by prompt injection. Don't use them as the only
+gate for destructive actions. brier never runs model code (`trust_remote_code` is never set),
+loads weights from safetensors only, and validates calibration files as untrusted input. Report
+vulnerabilities privately; see
+[SECURITY.md](https://github.com/mohdUwaish59/brier/blob/main/SECURITY.md) and
+[`THREAT_MODEL.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/THREAT_MODEL.md).
+
+## Related work
+
+brier follows the idea of TypeSafe AI's Jev and of
+[AnyJev](https://github.com/nokia-applied-research/AnyJev) (Nokia Applied Research): typed
+questions answered from next-token probabilities, corrected in levels. It is an independent
+implementation, not affiliated with either. The correction methods come from published work
+(Zheng et al. 2024; Zhou et al. 2024; Guo et al. 2017; Platt 1999; Ledoit & Wolf 2004).
+[`related_work.md`](https://github.com/mohdUwaish59/brier/blob/main/docs/related_work.md) sets
+out what brier shares with these projects and where it differs.
+
+## License
+
+Apache-2.0