pijev 0.1.0__tar.gz

pijev-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.

pijev-0.1.0/PKG-INFO

@@ -0,0 +1,244 @@
+Metadata-Version: 2.4
+Name: pijev
+Version: 0.1.0
+Summary: Permutation Invariant Jev — third-party permutation averaging for TypeSafe Choice
+License-Expression: Apache-2.0
+Project-URL: Repository, https://github.com/TypeLLM/pijev
+Project-URL: Issues, https://github.com/TypeLLM/pijev/issues
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: typesafe-sdk<0.8,>=0.7.1
+Dynamic: license-file
+
+# pijev — Permutation Invariant Jev
+
+**Better Jev probabilities with theoretical guarantees—all with a one-line import change.**
+
+pijev turns Jev's option-order variations into a single averaged prediction, with
+Brier score and log loss guaranteed no worse than the average across the included
+orderings. All permutations run in one `system_one` request. Keep the same calls
+and official response types; pijev handles the rest.
+
+## Problem: Jev's decision is not permutation invariant
+
+The same options can receive different probabilities when reordered. This official
+SDK call sends two Choices with identical instructions and descriptions, but
+opposite option orders:
+
+```python
+from typesafe_sdk import Choice, TypeSafeClient
+
+instructions = "Which team should handle this support ticket? Choose the best primary team."
+
+with TypeSafeClient() as client:
+    response = client.system_one(
+        state={
+            "ticket": "Since upgrading my plan yesterday, I cannot access the "
+                      "dashboard. The payment went through, but the page says "
+                      "my subscription is inactive. Please fix this."
+        },
+        questions={
+            "forward": Choice(
+                instructions=instructions,
+                criteria={
+                    "billing": "Payments, charges, invoices, and refunds.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "account": "Login, account access, and account settings.",
+                },
+            ),
+            "reverse": Choice(
+                instructions=instructions,
+                criteria={
+                    "account": "Login, account access, and account settings.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "billing": "Payments, charges, invoices, and refunds.",
+                },
+            ),
+        },
+    )
+
+print(response.choices["forward"].probabilities)
+print(response.choices["reverse"].probabilities)
+```
+
+One request, two questions, one shared state. A live `jev-1.13.0` run returned:
+
+| Option | billing → technical → account | account → technical → billing |
+| --- | ---: | ---: |
+| billing | 0.43 | 0.63 |
+| technical | 0.21 | 0.10 |
+| account | 0.36 | 0.27 |
+
+Both selected `billing`, but its probability changed by **20 percentage points**.
+This run also changes question IDs and does not control model randomness; see the
+[experiment details](https://github.com/TypeLLM/pijev/blob/main/examples/README.md).
+
+## How pijev works
+
+pijev treats option order as a nuisance variable: evaluate different orderings,
+align predictions by label, and average their probabilities. The highest mean
+probability determines the returned choice.
+
+For three options, there are six permutations. Here is a separate live
+`jev-1.13.0` run with **all six in one request**. A = account, B = billing,
+T = technical.
+
+| Option | A → B → T | A → T → B | B → A → T | B → T → A | T → A → B | T → B → A | pijev mean |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| billing | 0.58 | 0.64 | 0.58 | 0.46 | 0.47 | 0.60 | **0.5550** |
+| technical | 0.17 | 0.08 | 0.11 | 0.23 | 0.12 | 0.12 | 0.1383 |
+| account | 0.25 | 0.28 | 0.31 | 0.31 | 0.41 | 0.28 | 0.3067 |
+
+The `pijev mean` column averages all six predictions: `billing` receives **0.555**,
+compared with 0.46–0.64 across individual orderings.
+[Reproduce this run](https://github.com/TypeLLM/pijev/blob/main/examples/compare_all_orders.py) · [Recorded results](https://github.com/TypeLLM/pijev/blob/main/examples/six_orders_result.json)
+
+Ask once using the same SDK interface; pijev handles the six-question batch and
+returns one aggregated answer:
+
+```python
+from pijev import Choice, TypeSafeClient
+
+with TypeSafeClient() as client:
+    response = client.system_one(
+        state={
+            "ticket": "Since upgrading my plan yesterday, I cannot access the "
+                      "dashboard. The payment went through, but the page says "
+                      "my subscription is inactive. Please fix this."
+        },
+        questions={
+            "team": Choice(
+                instructions="Which team should handle this support ticket? "
+                             "Choose the best primary team.",
+                criteria={
+                    "billing": "Payments, charges, invoices, and refunds.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "account": "Login, account access, and account settings.",
+                },
+            ),
+        },
+    )
+
+answer = response.choices["team"]
+print(answer.choice)
+print(answer.probabilities)
+```
+
+The result is an official `SystemOneResponse` with your original question names.
+Live results may differ from the recorded run.
+
+## Usage
+
+Install from PyPI:
+
+```bash
+pip install pijev
+export TYPESAFE_API_KEY=your-key
+```
+
+Change only the import:
+
+```diff
+-from typesafe_sdk import TypeSafeClient, Choice, Noul, Score
++from pijev import TypeSafeClient, Choice, Noul, Score
+```
+
+Set an optional permutation budget:
+
+```python
+client = TypeSafeClient(n_permutations=16, seed=42)
+```
+
+The default is **up to 8 distinct permutations per Choice**. Smaller sets use all
+permutations; larger sets use uniform sampling without replacement. Set
+`n_permutations="all"` for full enumeration. The seed controls local sampling only.
+Batches are capped at 720 expanded questions and remain subject to server limits.
+
+Async usage works the same way: import `AsyncTypeSafeClient` from `pijev`, then use
+`async with` and `await client.system_one(...)` as usual.
+
+## Guaranteed Better Calibration
+
+Permutation averaging has a **proper-scoring-loss guarantee**, not an unconditional
+calibration guarantee. Let $p^{(m)}$ be the label-aligned probability vector for
+permutation $m$, and let $y$ be the same correct answer under every ordering. Define
+
+$$
+\bar p = \frac{1}{M}\sum_{m=1}^{M}p^{(m)}.
+$$
+
+For any loss convex in the predicted probabilities, Jensen's inequality gives
+
+$$
+L(\bar p,y) \leq \frac{1}{M}\sum_{m=1}^{M}L(p^{(m)},y).
+$$
+
+This holds for the actual set of predictions being averaged, including a sampled
+subset of permutations. Two useful cases are:
+
+- **Log loss:** $-\log\bar p_y \leq \frac{1}{M}\sum_m -\log p_y^{(m)}$.
+- **Brier score:** writing $e_y$ for the one-hot correct answer, the improvement
+  equals the disagreement removed by averaging:
+
+$$
+\frac{1}{M}\sum_m\|p^{(m)}-e_y\|_2^2 - \|\bar p-e_y\|_2^2
+= \frac{1}{M}\sum_m\|p^{(m)}-\bar p\|_2^2 \geq 0.
+$$
+
+The guarantee compares the average prediction with **randomly choosing one of the
+included predictions**. Brier improvement is strict when those predictions differ.
+For uniform sampling and a fixed predictor, the same inequality holds in expectation
+against a uniformly random ordering. Full enumeration gives the group average;
+see [Chen, Dobriban, and Lee](https://jmlr.csail.mit.edu/papers/v21/20-163.html)
+for related symmetry-averaging theory.
+
+## Cost analysis
+
+For $K$ options and budget $B$, pijev evaluates $M=\min(B,K!)$ permutations in
+**one HTTP request**. Multiple Choices contribute their own permutations;
+Noul and Score each add one unchanged question.
+
+| Workload | Questions in one request |
+| --- | ---: |
+| 3 options, default budget 8 | 6 |
+| 10 options, default budget 8 | 8 |
+| Two Choices with 3 options each | 12 |
+| 3 options plus one Noul | 7 |
+
+The state is sent once. For state size $S$ and question size $Q$, request content
+is approximately $S+MQ$, versus $M(S+Q)$ for separate calls, excluding protocol
+overhead. This is a payload estimate; use `response.usage` for reported tokens.
+
+At TypeSafe's [published rate](https://typesafe.ai/blog/introducing-system-one-models-and-jev)
+of **$0.042 per million input tokens, with free output tokens** (checked 2026-09-22):
+
+$$
+C = \frac{T}{10^6}\times 0.042
+$$
+
+Here, $C$ is the estimated cost in USD and $T$ is `response.usage.input_tokens`.
+
+| Recorded run | Input tokens | Estimated cost |
+| --- | ---: | ---: |
+| Two permutations | 475 | $0.00001995 |
+| All six permutations | 835 | $0.00003507 |
+
+These estimates use reported usage and the published rate, not invoice charges.
+One request still contains multiple decisions; latency and billing depend on the
+backend, and retries may add requests.
+
+## Notes
+
+- Only Choice answers are averaged. Noul and Score pass through unchanged.
+- `confidence` is the selected label's mean probability, **not Jev's native
+  confidence score**.
+- Sampling approximates full permutation averaging; it does not guarantee exact
+  order invariance or better calibration. The live examples demonstrate probability
+  differences, not accuracy or calibration gains.
+- Official interface: [TypeSafe Python SDK](https://github.com/typesafe-ai/typesafe-sdk-python).
+  This is a third-party extension, not an official TypeSafe project.
+
+## License
+
+[Apache License 2.0](https://github.com/TypeLLM/pijev/blob/main/LICENSE).

pijev-0.1.0/README.md

@@ -0,0 +1,231 @@
+# pijev — Permutation Invariant Jev
+
+**Better Jev probabilities with theoretical guarantees—all with a one-line import change.**
+
+pijev turns Jev's option-order variations into a single averaged prediction, with
+Brier score and log loss guaranteed no worse than the average across the included
+orderings. All permutations run in one `system_one` request. Keep the same calls
+and official response types; pijev handles the rest.
+
+## Problem: Jev's decision is not permutation invariant
+
+The same options can receive different probabilities when reordered. This official
+SDK call sends two Choices with identical instructions and descriptions, but
+opposite option orders:
+
+```python
+from typesafe_sdk import Choice, TypeSafeClient
+
+instructions = "Which team should handle this support ticket? Choose the best primary team."
+
+with TypeSafeClient() as client:
+    response = client.system_one(
+        state={
+            "ticket": "Since upgrading my plan yesterday, I cannot access the "
+                      "dashboard. The payment went through, but the page says "
+                      "my subscription is inactive. Please fix this."
+        },
+        questions={
+            "forward": Choice(
+                instructions=instructions,
+                criteria={
+                    "billing": "Payments, charges, invoices, and refunds.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "account": "Login, account access, and account settings.",
+                },
+            ),
+            "reverse": Choice(
+                instructions=instructions,
+                criteria={
+                    "account": "Login, account access, and account settings.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "billing": "Payments, charges, invoices, and refunds.",
+                },
+            ),
+        },
+    )
+
+print(response.choices["forward"].probabilities)
+print(response.choices["reverse"].probabilities)
+```
+
+One request, two questions, one shared state. A live `jev-1.13.0` run returned:
+
+| Option | billing → technical → account | account → technical → billing |
+| --- | ---: | ---: |
+| billing | 0.43 | 0.63 |
+| technical | 0.21 | 0.10 |
+| account | 0.36 | 0.27 |
+
+Both selected `billing`, but its probability changed by **20 percentage points**.
+This run also changes question IDs and does not control model randomness; see the
+[experiment details](https://github.com/TypeLLM/pijev/blob/main/examples/README.md).
+
+## How pijev works
+
+pijev treats option order as a nuisance variable: evaluate different orderings,
+align predictions by label, and average their probabilities. The highest mean
+probability determines the returned choice.
+
+For three options, there are six permutations. Here is a separate live
+`jev-1.13.0` run with **all six in one request**. A = account, B = billing,
+T = technical.
+
+| Option | A → B → T | A → T → B | B → A → T | B → T → A | T → A → B | T → B → A | pijev mean |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| billing | 0.58 | 0.64 | 0.58 | 0.46 | 0.47 | 0.60 | **0.5550** |
+| technical | 0.17 | 0.08 | 0.11 | 0.23 | 0.12 | 0.12 | 0.1383 |
+| account | 0.25 | 0.28 | 0.31 | 0.31 | 0.41 | 0.28 | 0.3067 |
+
+The `pijev mean` column averages all six predictions: `billing` receives **0.555**,
+compared with 0.46–0.64 across individual orderings.
+[Reproduce this run](https://github.com/TypeLLM/pijev/blob/main/examples/compare_all_orders.py) · [Recorded results](https://github.com/TypeLLM/pijev/blob/main/examples/six_orders_result.json)
+
+Ask once using the same SDK interface; pijev handles the six-question batch and
+returns one aggregated answer:
+
+```python
+from pijev import Choice, TypeSafeClient
+
+with TypeSafeClient() as client:
+    response = client.system_one(
+        state={
+            "ticket": "Since upgrading my plan yesterday, I cannot access the "
+                      "dashboard. The payment went through, but the page says "
+                      "my subscription is inactive. Please fix this."
+        },
+        questions={
+            "team": Choice(
+                instructions="Which team should handle this support ticket? "
+                             "Choose the best primary team.",
+                criteria={
+                    "billing": "Payments, charges, invoices, and refunds.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "account": "Login, account access, and account settings.",
+                },
+            ),
+        },
+    )
+
+answer = response.choices["team"]
+print(answer.choice)
+print(answer.probabilities)
+```
+
+The result is an official `SystemOneResponse` with your original question names.
+Live results may differ from the recorded run.
+
+## Usage
+
+Install from PyPI:
+
+```bash
+pip install pijev
+export TYPESAFE_API_KEY=your-key
+```
+
+Change only the import:
+
+```diff
+-from typesafe_sdk import TypeSafeClient, Choice, Noul, Score
++from pijev import TypeSafeClient, Choice, Noul, Score
+```
+
+Set an optional permutation budget:
+
+```python
+client = TypeSafeClient(n_permutations=16, seed=42)
+```
+
+The default is **up to 8 distinct permutations per Choice**. Smaller sets use all
+permutations; larger sets use uniform sampling without replacement. Set
+`n_permutations="all"` for full enumeration. The seed controls local sampling only.
+Batches are capped at 720 expanded questions and remain subject to server limits.
+
+Async usage works the same way: import `AsyncTypeSafeClient` from `pijev`, then use
+`async with` and `await client.system_one(...)` as usual.
+
+## Guaranteed Better Calibration
+
+Permutation averaging has a **proper-scoring-loss guarantee**, not an unconditional
+calibration guarantee. Let $p^{(m)}$ be the label-aligned probability vector for
+permutation $m$, and let $y$ be the same correct answer under every ordering. Define
+
+$$
+\bar p = \frac{1}{M}\sum_{m=1}^{M}p^{(m)}.
+$$
+
+For any loss convex in the predicted probabilities, Jensen's inequality gives
+
+$$
+L(\bar p,y) \leq \frac{1}{M}\sum_{m=1}^{M}L(p^{(m)},y).
+$$
+
+This holds for the actual set of predictions being averaged, including a sampled
+subset of permutations. Two useful cases are:
+
+- **Log loss:** $-\log\bar p_y \leq \frac{1}{M}\sum_m -\log p_y^{(m)}$.
+- **Brier score:** writing $e_y$ for the one-hot correct answer, the improvement
+  equals the disagreement removed by averaging:
+
+$$
+\frac{1}{M}\sum_m\|p^{(m)}-e_y\|_2^2 - \|\bar p-e_y\|_2^2
+= \frac{1}{M}\sum_m\|p^{(m)}-\bar p\|_2^2 \geq 0.
+$$
+
+The guarantee compares the average prediction with **randomly choosing one of the
+included predictions**. Brier improvement is strict when those predictions differ.
+For uniform sampling and a fixed predictor, the same inequality holds in expectation
+against a uniformly random ordering. Full enumeration gives the group average;
+see [Chen, Dobriban, and Lee](https://jmlr.csail.mit.edu/papers/v21/20-163.html)
+for related symmetry-averaging theory.
+
+## Cost analysis
+
+For $K$ options and budget $B$, pijev evaluates $M=\min(B,K!)$ permutations in
+**one HTTP request**. Multiple Choices contribute their own permutations;
+Noul and Score each add one unchanged question.
+
+| Workload | Questions in one request |
+| --- | ---: |
+| 3 options, default budget 8 | 6 |
+| 10 options, default budget 8 | 8 |
+| Two Choices with 3 options each | 12 |
+| 3 options plus one Noul | 7 |
+
+The state is sent once. For state size $S$ and question size $Q$, request content
+is approximately $S+MQ$, versus $M(S+Q)$ for separate calls, excluding protocol
+overhead. This is a payload estimate; use `response.usage` for reported tokens.
+
+At TypeSafe's [published rate](https://typesafe.ai/blog/introducing-system-one-models-and-jev)
+of **$0.042 per million input tokens, with free output tokens** (checked 2026-09-22):
+
+$$
+C = \frac{T}{10^6}\times 0.042
+$$
+
+Here, $C$ is the estimated cost in USD and $T$ is `response.usage.input_tokens`.
+
+| Recorded run | Input tokens | Estimated cost |
+| --- | ---: | ---: |
+| Two permutations | 475 | $0.00001995 |
+| All six permutations | 835 | $0.00003507 |
+
+These estimates use reported usage and the published rate, not invoice charges.
+One request still contains multiple decisions; latency and billing depend on the
+backend, and retries may add requests.
+
+## Notes
+
+- Only Choice answers are averaged. Noul and Score pass through unchanged.
+- `confidence` is the selected label's mean probability, **not Jev's native
+  confidence score**.
+- Sampling approximates full permutation averaging; it does not guarantee exact
+  order invariance or better calibration. The live examples demonstrate probability
+  differences, not accuracy or calibration gains.
+- Official interface: [TypeSafe Python SDK](https://github.com/typesafe-ai/typesafe-sdk-python).
+  This is a third-party extension, not an official TypeSafe project.
+
+## License
+
+[Apache License 2.0](https://github.com/TypeLLM/pijev/blob/main/LICENSE).

pijev-0.1.0/pijev/__init__.py

@@ -0,0 +1,178 @@
+"""Drop-in TypeSafe clients with batched permutation averaging for Choice."""
+from collections.abc import Mapping
+from itertools import permutations
+from math import factorial, fsum, isclose, isfinite
+from random import Random
+from typing import Any, Literal, TypeVar, overload
+from pydantic import BaseModel
+
+import httpx2
+from typesafe_sdk import *  # Re-export the official public types.
+from typesafe_sdk import __all__ as _sdk_exports
+from typesafe_sdk import TypeSafeClient as _Client, AsyncTypeSafeClient as _AsyncClient
+
+__all__ = list(_sdk_exports)
+ResponseT = TypeVar("ResponseT", bound=BaseModel)
+
+
+def _validate_budget(budget):
+    if budget != "all" and (type(budget) is not int or budget <= 0):
+        raise ValueError("n_permutations must be a positive integer or 'all'")
+
+
+def _prepare(questions, budget, seed):
+    expanded, groups = {}, {}
+    # Reserve all original names, including names resembling generated IDs.
+    reserved = set(questions)
+    for name, question in questions.items():
+        if isinstance(question, Choice):
+            choice = question
+        elif isinstance(question, Mapping) and question.get("type") == "choice":
+            choice = Choice.model_validate(question)
+        else:
+            expanded[name] = question
+            continue
+        labels = tuple(sorted(choice.criteria))
+        if not labels:
+            raise ValueError("Choice criteria must not be empty")
+        total = factorial(len(labels))
+        count = total if budget == "all" else min(budget, total)
+        if len(expanded) + count > 720:
+            raise ValueError("At most 720 expanded questions per request")
+        if count == total:
+            orders = permutations(labels)
+        else:
+            rng, selected = Random(seed), {}
+            while len(selected) < count:
+                selected[tuple(rng.sample(labels, len(labels)))] = None
+            orders = selected
+        names = []
+        for order in orders:
+            key = f"__pijev_{len(expanded)}"
+            while key in reserved:
+                key += "_"
+            reserved.add(key)
+            item = choice.model_copy(deep=True)
+            item.criteria = {label: item.criteria[label] for label in order}
+            expanded[key] = item
+            names.append(key)
+        groups[name] = (labels, names)
+    if len(expanded) > 720:
+        raise ValueError("At most 720 expanded questions per request")
+    return expanded, groups
+
+
+def _aggregate(response, questions, groups, response_model):
+    answers = {}
+    for name in questions:
+        if name not in groups:
+            if name not in response.answers:
+                raise ValueError(f"Backend did not return answer {name!r}")
+            answers[name] = response.answers[name].model_dump(mode="json")
+            continue
+        labels, names = groups[name]
+        rows = []
+        for key in names:
+            if key not in response.choices:
+                raise ValueError(f"Backend did not return Choice answer {key!r}")
+            raw = response.choices[key].probabilities
+            if set(raw) != set(labels):
+                raise ValueError("Backend probabilities must contain exactly the requested labels")
+            if any(not isfinite(p) or not 0 <= p <= 1 for p in raw.values()):
+                raise ValueError("Backend probabilities must be finite and between 0 and 1")
+            mass = fsum(raw.values())
+            if not isclose(mass, 1.0, rel_tol=0, abs_tol=1e-3):
+                raise ValueError("Backend probabilities must sum to approximately 1")
+            rows.append({label: raw[label] / mass for label in labels})
+        means = {label: fsum(row[label] for row in rows) / len(rows) for label in labels}
+        winner = max(labels, key=means.__getitem__)
+        answers[name] = ChoiceAnswer(choice=winner, probabilities=means,
+                                     confidence=means[winner]).model_dump(mode="json")
+    body = response.model_dump(mode="json")
+    body["answers"] = answers
+    raw = response.raw_http_response
+    # Parse through the official response decoder, including custom typed fields.
+    # This response represents the transformed body, with actual request metadata.
+    headers = {k: v for k, v in raw.headers.items()
+               if k.lower() not in {"content-length", "content-encoding"}}
+    transformed = httpx2.Response(raw.status_code, headers=headers, json=body,
+                                 request=raw.request)
+    target = response_model or SystemOneResponse
+    if issubclass(target, SystemOneResponse):
+        return target.from_http_response(transformed)
+    return target.model_validate_json(transformed.content)
+
+
+def _request_inputs(state, questions, extra_body, budget, seed):
+    # Match the SDK's documented extra_body last-write-wins semantics.
+    extra = dict(extra_body or {})
+    state = extra.pop("state", state)
+    questions = extra.pop("questions", questions)
+    expanded, groups = _prepare(questions, budget, seed)
+    return state, questions, expanded, groups, extra
+
+
+class TypeSafeClient(_Client):
+    """Official sync client interface; Choice permutations share one request."""
+
+    def __init__(self, *args: Any, n_permutations: int | Literal["all"] = 8,
+                 seed: int | None = None, **kwargs: Any) -> None:
+        _validate_budget(n_permutations)
+        super().__init__(*args, **kwargs)
+        self.n_permutations, self.seed = n_permutations, seed
+
+    @overload
+    def system_one(self, state: JSONContent, questions: Questions, *,
+                   response_model: type[ResponseT], **kwargs: Any) -> ResponseT: ...
+
+    @overload
+    def system_one(self, state: JSONContent, questions: Questions, *,
+                   response_model: None = None, **kwargs: Any) -> SystemOneResponse: ...
+
+    def system_one(self, state: JSONContent, questions: Questions, *,
+                   model: str | None = None, retry: RetryPolicy | None = None,
+                   timeout: float | httpx2.Timeout | None = None,
+                   extra_headers: Mapping[str, str] | None = None,
+                   extra_body: Mapping[str, JSONValue | None] | None = None,
+                   response_model: type[ResponseT] | None = None) -> SystemOneResponse | ResponseT:
+        _validate_budget(self.n_permutations)
+        state, original, expanded, groups, extra = _request_inputs(
+            state, questions, extra_body, self.n_permutations, self.seed)
+        result = super().system_one(
+            state, expanded, model=model, retry=retry, timeout=timeout,
+            extra_headers=extra_headers, extra_body=extra,
+            response_model=None if groups else response_model)
+        return _aggregate(result, original, groups, response_model) if groups else result
+
+
+class AsyncTypeSafeClient(_AsyncClient):
+    """Official async client interface with the same aggregation as sync."""
+
+    def __init__(self, *args: Any, n_permutations: int | Literal["all"] = 8,
+                 seed: int | None = None, **kwargs: Any) -> None:
+        _validate_budget(n_permutations)
+        super().__init__(*args, **kwargs)
+        self.n_permutations, self.seed = n_permutations, seed
+
+    @overload
+    async def system_one(self, state: JSONContent, questions: Questions, *,
+                   response_model: type[ResponseT], **kwargs: Any) -> ResponseT: ...
+
+    @overload
+    async def system_one(self, state: JSONContent, questions: Questions, *,
+                   response_model: None = None, **kwargs: Any) -> SystemOneResponse: ...
+
+    async def system_one(self, state: JSONContent, questions: Questions, *,
+                   model: str | None = None, retry: RetryPolicy | None = None,
+                   timeout: float | httpx2.Timeout | None = None,
+                   extra_headers: Mapping[str, str] | None = None,
+                   extra_body: Mapping[str, JSONValue | None] | None = None,
+                   response_model: type[ResponseT] | None = None) -> SystemOneResponse | ResponseT:
+        _validate_budget(self.n_permutations)
+        state, original, expanded, groups, extra = _request_inputs(
+            state, questions, extra_body, self.n_permutations, self.seed)
+        result = await super().system_one(
+            state, expanded, model=model, retry=retry, timeout=timeout,
+            extra_headers=extra_headers, extra_body=extra,
+            response_model=None if groups else response_model)
+        return _aggregate(result, original, groups, response_model) if groups else result

pijev-0.1.0/pijev.egg-info/PKG-INFO

@@ -0,0 +1,244 @@
+Metadata-Version: 2.4
+Name: pijev
+Version: 0.1.0
+Summary: Permutation Invariant Jev — third-party permutation averaging for TypeSafe Choice
+License-Expression: Apache-2.0
+Project-URL: Repository, https://github.com/TypeLLM/pijev
+Project-URL: Issues, https://github.com/TypeLLM/pijev/issues
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: typesafe-sdk<0.8,>=0.7.1
+Dynamic: license-file
+
+# pijev — Permutation Invariant Jev
+
+**Better Jev probabilities with theoretical guarantees—all with a one-line import change.**
+
+pijev turns Jev's option-order variations into a single averaged prediction, with
+Brier score and log loss guaranteed no worse than the average across the included
+orderings. All permutations run in one `system_one` request. Keep the same calls
+and official response types; pijev handles the rest.
+
+## Problem: Jev's decision is not permutation invariant
+
+The same options can receive different probabilities when reordered. This official
+SDK call sends two Choices with identical instructions and descriptions, but
+opposite option orders:
+
+```python
+from typesafe_sdk import Choice, TypeSafeClient
+
+instructions = "Which team should handle this support ticket? Choose the best primary team."
+
+with TypeSafeClient() as client:
+    response = client.system_one(
+        state={
+            "ticket": "Since upgrading my plan yesterday, I cannot access the "
+                      "dashboard. The payment went through, but the page says "
+                      "my subscription is inactive. Please fix this."
+        },
+        questions={
+            "forward": Choice(
+                instructions=instructions,
+                criteria={
+                    "billing": "Payments, charges, invoices, and refunds.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "account": "Login, account access, and account settings.",
+                },
+            ),
+            "reverse": Choice(
+                instructions=instructions,
+                criteria={
+                    "account": "Login, account access, and account settings.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "billing": "Payments, charges, invoices, and refunds.",
+                },
+            ),
+        },
+    )
+
+print(response.choices["forward"].probabilities)
+print(response.choices["reverse"].probabilities)
+```
+
+One request, two questions, one shared state. A live `jev-1.13.0` run returned:
+
+| Option | billing → technical → account | account → technical → billing |
+| --- | ---: | ---: |
+| billing | 0.43 | 0.63 |
+| technical | 0.21 | 0.10 |
+| account | 0.36 | 0.27 |
+
+Both selected `billing`, but its probability changed by **20 percentage points**.
+This run also changes question IDs and does not control model randomness; see the
+[experiment details](https://github.com/TypeLLM/pijev/blob/main/examples/README.md).
+
+## How pijev works
+
+pijev treats option order as a nuisance variable: evaluate different orderings,
+align predictions by label, and average their probabilities. The highest mean
+probability determines the returned choice.
+
+For three options, there are six permutations. Here is a separate live
+`jev-1.13.0` run with **all six in one request**. A = account, B = billing,
+T = technical.
+
+| Option | A → B → T | A → T → B | B → A → T | B → T → A | T → A → B | T → B → A | pijev mean |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| billing | 0.58 | 0.64 | 0.58 | 0.46 | 0.47 | 0.60 | **0.5550** |
+| technical | 0.17 | 0.08 | 0.11 | 0.23 | 0.12 | 0.12 | 0.1383 |
+| account | 0.25 | 0.28 | 0.31 | 0.31 | 0.41 | 0.28 | 0.3067 |
+
+The `pijev mean` column averages all six predictions: `billing` receives **0.555**,
+compared with 0.46–0.64 across individual orderings.
+[Reproduce this run](https://github.com/TypeLLM/pijev/blob/main/examples/compare_all_orders.py) · [Recorded results](https://github.com/TypeLLM/pijev/blob/main/examples/six_orders_result.json)
+
+Ask once using the same SDK interface; pijev handles the six-question batch and
+returns one aggregated answer:
+
+```python
+from pijev import Choice, TypeSafeClient
+
+with TypeSafeClient() as client:
+    response = client.system_one(
+        state={
+            "ticket": "Since upgrading my plan yesterday, I cannot access the "
+                      "dashboard. The payment went through, but the page says "
+                      "my subscription is inactive. Please fix this."
+        },
+        questions={
+            "team": Choice(
+                instructions="Which team should handle this support ticket? "
+                             "Choose the best primary team.",
+                criteria={
+                    "billing": "Payments, charges, invoices, and refunds.",
+                    "technical": "Product failures, errors, and troubleshooting.",
+                    "account": "Login, account access, and account settings.",
+                },
+            ),
+        },
+    )
+
+answer = response.choices["team"]
+print(answer.choice)
+print(answer.probabilities)
+```
+
+The result is an official `SystemOneResponse` with your original question names.
+Live results may differ from the recorded run.
+
+## Usage
+
+Install from PyPI:
+
+```bash
+pip install pijev
+export TYPESAFE_API_KEY=your-key
+```
+
+Change only the import:
+
+```diff
+-from typesafe_sdk import TypeSafeClient, Choice, Noul, Score
++from pijev import TypeSafeClient, Choice, Noul, Score
+```
+
+Set an optional permutation budget:
+
+```python
+client = TypeSafeClient(n_permutations=16, seed=42)
+```
+
+The default is **up to 8 distinct permutations per Choice**. Smaller sets use all
+permutations; larger sets use uniform sampling without replacement. Set
+`n_permutations="all"` for full enumeration. The seed controls local sampling only.
+Batches are capped at 720 expanded questions and remain subject to server limits.
+
+Async usage works the same way: import `AsyncTypeSafeClient` from `pijev`, then use
+`async with` and `await client.system_one(...)` as usual.
+
+## Guaranteed Better Calibration
+
+Permutation averaging has a **proper-scoring-loss guarantee**, not an unconditional
+calibration guarantee. Let $p^{(m)}$ be the label-aligned probability vector for
+permutation $m$, and let $y$ be the same correct answer under every ordering. Define
+
+$$
+\bar p = \frac{1}{M}\sum_{m=1}^{M}p^{(m)}.
+$$
+
+For any loss convex in the predicted probabilities, Jensen's inequality gives
+
+$$
+L(\bar p,y) \leq \frac{1}{M}\sum_{m=1}^{M}L(p^{(m)},y).
+$$
+
+This holds for the actual set of predictions being averaged, including a sampled
+subset of permutations. Two useful cases are:
+
+- **Log loss:** $-\log\bar p_y \leq \frac{1}{M}\sum_m -\log p_y^{(m)}$.
+- **Brier score:** writing $e_y$ for the one-hot correct answer, the improvement
+  equals the disagreement removed by averaging:
+
+$$
+\frac{1}{M}\sum_m\|p^{(m)}-e_y\|_2^2 - \|\bar p-e_y\|_2^2
+= \frac{1}{M}\sum_m\|p^{(m)}-\bar p\|_2^2 \geq 0.
+$$
+
+The guarantee compares the average prediction with **randomly choosing one of the
+included predictions**. Brier improvement is strict when those predictions differ.
+For uniform sampling and a fixed predictor, the same inequality holds in expectation
+against a uniformly random ordering. Full enumeration gives the group average;
+see [Chen, Dobriban, and Lee](https://jmlr.csail.mit.edu/papers/v21/20-163.html)
+for related symmetry-averaging theory.
+
+## Cost analysis
+
+For $K$ options and budget $B$, pijev evaluates $M=\min(B,K!)$ permutations in
+**one HTTP request**. Multiple Choices contribute their own permutations;
+Noul and Score each add one unchanged question.
+
+| Workload | Questions in one request |
+| --- | ---: |
+| 3 options, default budget 8 | 6 |
+| 10 options, default budget 8 | 8 |
+| Two Choices with 3 options each | 12 |
+| 3 options plus one Noul | 7 |
+
+The state is sent once. For state size $S$ and question size $Q$, request content
+is approximately $S+MQ$, versus $M(S+Q)$ for separate calls, excluding protocol
+overhead. This is a payload estimate; use `response.usage` for reported tokens.
+
+At TypeSafe's [published rate](https://typesafe.ai/blog/introducing-system-one-models-and-jev)
+of **$0.042 per million input tokens, with free output tokens** (checked 2026-09-22):
+
+$$
+C = \frac{T}{10^6}\times 0.042
+$$
+
+Here, $C$ is the estimated cost in USD and $T$ is `response.usage.input_tokens`.
+
+| Recorded run | Input tokens | Estimated cost |
+| --- | ---: | ---: |
+| Two permutations | 475 | $0.00001995 |
+| All six permutations | 835 | $0.00003507 |
+
+These estimates use reported usage and the published rate, not invoice charges.
+One request still contains multiple decisions; latency and billing depend on the
+backend, and retries may add requests.
+
+## Notes
+
+- Only Choice answers are averaged. Noul and Score pass through unchanged.
+- `confidence` is the selected label's mean probability, **not Jev's native
+  confidence score**.
+- Sampling approximates full permutation averaging; it does not guarantee exact
+  order invariance or better calibration. The live examples demonstrate probability
+  differences, not accuracy or calibration gains.
+- Official interface: [TypeSafe Python SDK](https://github.com/typesafe-ai/typesafe-sdk-python).
+  This is a third-party extension, not an official TypeSafe project.
+
+## License
+
+[Apache License 2.0](https://github.com/TypeLLM/pijev/blob/main/LICENSE).

pijev-0.1.0/pijev.egg-info/SOURCES.txt

@@ -0,0 +1,11 @@
+LICENSE
+README.md
+pyproject.toml
+pijev/__init__.py
+pijev/py.typed
+pijev.egg-info/PKG-INFO
+pijev.egg-info/SOURCES.txt
+pijev.egg-info/dependency_links.txt
+pijev.egg-info/requires.txt
+pijev.egg-info/top_level.txt
+tests/test_symmetry.py

pijev-0.1.0/pijev.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

pijev-0.1.0/pijev.egg-info/requires.txt

@@ -0,0 +1 @@
+typesafe-sdk<0.8,>=0.7.1

pijev-0.1.0/pijev.egg-info/top_level.txt

@@ -0,0 +1 @@
+pijev

pijev-0.1.0/pyproject.toml

@@ -0,0 +1,23 @@
+[build-system]
+requires = ["setuptools>=77"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "pijev"
+version = "0.1.0"
+description = "Permutation Invariant Jev — third-party permutation averaging for TypeSafe Choice"
+readme = "README.md"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+requires-python = ">=3.10"
+dependencies = ["typesafe-sdk>=0.7.1,<0.8"]
+
+[project.urls]
+Repository = "https://github.com/TypeLLM/pijev"
+Issues = "https://github.com/TypeLLM/pijev/issues"
+
+[tool.setuptools]
+packages = ["pijev"]
+
+[tool.setuptools.package-data]
+pijev = ["py.typed"]

pijev-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

pijev-0.1.0/tests/test_symmetry.py

@@ -0,0 +1,128 @@
+import asyncio
+import json
+import unittest
+
+import httpx2
+from pydantic import BaseModel
+from typesafe_sdk import SystemOneResponse, ChoiceAnswer
+from pijev import AsyncTypeSafeClient, Choice, Noul, Score, TypeSafeClient
+
+
+class Tests(unittest.TestCase):
+    def setUp(self):
+        self.bodies = []
+        self.bad = None
+        self.omit = False
+
+    def handle(self, request):
+        body = json.loads(request.content)
+        self.bodies.append(body)
+        answers = {}
+        for name, q in reversed(list(body['questions'].items())):
+            if q['type'] == 'choice':
+                labels = list(q['criteria'])
+                probs = dict(zip(labels, [2 / (len(labels) + 1)] + [1 / (len(labels) + 1)] * (len(labels) - 1)))
+                answers[name] = dict(type='choice', choice=labels[0], confidence=0.99,
+                                     probabilities=self.bad if self.bad is not None else probs)
+            elif q['type'] == 'noul':
+                answers[name] = dict(type='noul', noul=0.7)
+            else:
+                answers[name] = dict(type='score', score=0.7, confidence=0.8,
+                                     legend={'0': 'low', '1': 'high'}, probabilities={'0': 0.3, '1': 0.7})
+        if self.omit:
+            answers.pop(next(iter(answers)))
+        return httpx2.Response(200, headers={'x-typesafe-request-id': 'test-id'}, json={
+            'model': 'mock', 'usage': {'input_tokens': 10, 'output_tokens': len(answers)}, 'answers': answers})
+
+    def client(self, **kwargs):
+        client = TypeSafeClient(api_key='test', transport=httpx2.MockTransport(self.handle), **kwargs)
+        self.addCleanup(client.close)
+        return client
+
+    def test_drop_in_mixed_batch_and_metadata(self):
+        q = Choice(criteria={'b': ['B'], 'a': {'hint': 'A'}}, instructions='pick')
+        before = q.model_dump()
+        result = self.client().system_one('state', {'category': q, '__pijev_0': Noul(), 'rating': Score(criteria=['low', 'high'])})
+        self.assertIs(type(result), SystemOneResponse)
+        self.assertEqual(result.choices['category'].probabilities, {'a': 0.5, 'b': 0.5})
+        self.assertEqual(result.choices['category'].choice, 'a')
+        self.assertEqual(result.choices['category'].confidence, 0.5)
+        self.assertEqual(result.nouls['__pijev_0'].noul, 0.7)
+        self.assertEqual(result.scores['rating'].score, 0.7)
+        self.assertEqual(list(result.answers), ['category', '__pijev_0', 'rating'])
+        self.assertEqual(set(result.model_dump()), {'answers', 'model', 'usage'})
+        self.assertEqual(result.request_id, 'test-id')
+        self.assertEqual(result.usage.output_tokens, 4)
+        self.assertEqual(result.raw_http_response.json()['answers'], result.model_dump(mode='json')['answers'])
+        self.assertEqual(len(self.bodies), 1)
+        self.assertEqual(len(self.bodies[0]['questions']), 4)
+        self.assertEqual(q.model_dump(), before)
+
+    def test_three_options_one_call_six_questions(self):
+        result = self.client().system_one(state='x', questions={'q': Choice(criteria=dict.fromkeys('abc'))})
+        self.assertEqual(len(self.bodies), 1)
+        self.assertEqual(len(self.bodies[0]['questions']), 6)
+        for p in result.choices['q'].probabilities.values():
+            self.assertAlmostEqual(p, 1 / 3)
+
+    def test_seed_and_multiple_choices(self):
+        client = self.client(seed=42)
+        q = {'type': 'choice', 'criteria': dict.fromkeys('abcd')}
+        a = client.system_one('x', {'q': q, 'second': Choice(criteria={'only': None})})
+        q['criteria'] = dict.fromkeys('dcba')
+        b = client.system_one('x', {'q': q, 'second': Choice(criteria={'only': None})})
+        self.assertEqual(a.answers, b.answers)
+        self.assertEqual(self.bodies[0], self.bodies[1])
+        orders = [tuple(q['criteria']) for q in self.bodies[0]['questions'].values()]
+        self.assertEqual(len(set(orders)), 9)
+
+    def test_custom_response_models(self):
+        class Typed(SystemOneResponse):
+            category: ChoiceAnswer
+        class Plain(BaseModel):
+            answers: dict[str, ChoiceAnswer]
+        for target in [Typed, Plain]:
+            result = self.client().system_one('x', {'category': Choice(criteria=dict.fromkeys('ab'))}, response_model=target)
+            self.assertIsInstance(result, target)
+            self.assertEqual(result.answers['category'].probabilities, {'a': 0.5, 'b': 0.5})
+            if target is Typed:
+                self.assertEqual(result.category, result.answers['category'])
+
+    def test_extra_body_override_and_passthrough(self):
+        client = self.client()
+        result = client.system_one('old', {'old': Noul()}, extra_body={
+            'state': 'new', 'model': 'chosen', 'questions': {'q': Choice(criteria=dict.fromkeys('ab'))}})
+        self.assertEqual(list(result.answers), ['q'])
+        self.assertEqual(self.bodies[0]['state'], 'new')
+        self.assertEqual(self.bodies[0]['model'], 'chosen')
+        result = client.system_one('x', {'n': Noul()})
+        self.assertEqual(result.nouls['n'].noul, 0.7)
+        self.assertEqual(list(self.bodies[1]['questions']), ['n'])
+
+    def test_validation(self):
+        for budget in [0, -1, True, 1.5, 'bad']:
+            with self.subTest(budget=budget), self.assertRaises(ValueError):
+                self.client(n_permutations=budget)
+        with self.assertRaises(ValueError):
+            self.client(n_permutations='all').system_one('x', {'q': Choice(criteria=dict.fromkeys('abcdefgh'))})
+        self.assertEqual(self.bodies, [])
+        for bad in [{'a': 1.0}, {'a': -0.1, 'b': 1.1}, {'a': 0.0, 'b': 0.0}]:
+            self.bad = bad
+            with self.assertRaises(ValueError):
+                self.client().system_one('x', {'q': Choice(criteria=dict.fromkeys('ab'))})
+        self.bad, self.omit = None, True
+        with self.assertRaises(ValueError):
+            self.client().system_one('x', {'q': Choice(criteria=dict.fromkeys('ab'))})
+
+    def test_async(self):
+        async def run():
+            async with AsyncTypeSafeClient(api_key='test', transport=httpx2.MockTransport(self.handle)) as client:
+                return await client.system_one('x', {'q': Choice(criteria=dict.fromkeys('ab'))})
+        result = asyncio.run(run())
+        self.assertIs(type(result), SystemOneResponse)
+        self.assertEqual(result.choices['q'].probabilities, {'a': 0.5, 'b': 0.5})
+        self.assertEqual(len(self.bodies), 1)
+
+
+if __name__ == '__main__':
+    unittest.main()