Publish Sev-4B v0.1.0 research baseline and evidence

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

BASE_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 2026 Alibaba Cloud
+
+   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.

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 2026 Jared Palmer
+
+   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.

NOTICE

@@ -0,0 +1,12 @@
+Sev research preview
+
+Sev is derived from Kev, Copyright 2026 Jared Palmer, under Apache-2.0.
+Sev research additions are Copyright 2026 the Sev contributors.
+The Qwen3.5 base model is provided by the Qwen team under Apache-2.0.
+The released adapter is initialized from jaredpalmer/kev-4b.
+
+The Apache-2.0 source/model license does not relicense third-party datasets.
+Original behavioral simulator records use CC BY 4.0 as described in
+research/collector/DATA_LICENSE. Authored Kev replay rules retain Apache-2.0
+and upstream attribution. Private captures and raw third-party replay are
+not included in the new Sev dataset or checkpoint artifacts.

README.md

@@ -0,0 +1,93 @@
+---
+language:
+- en
+license: apache-2.0
+base_model: Qwen/Qwen3.5-4B-Base
+base_model_relation: adapter
+library_name: peft
+datasets:
+- macmacmacmac/Sev-behavioral-research-v1
+tags:
+- sev
+- research
+- cybersecurity
+- agent-activity
+- decision-model
+- lora
+---
+# Sev-4B Research Preview
+
+Sev-4B is an exploratory baseline for classifying ordered observations as authored human, fixed-script, or agent policies. **This checkpoint predicts human for almost every window and has not demonstrated useful agent detection.** Its IID accuracy is 34.03% against 33.33% chance; agent recall is 1.04%. We release the checkpoint, synthetic inputs, generator, and evidence so researchers can reproduce and improve this result.
+
+This is a LoRA adapter and pointer head for Qwen3.5-4B-Base, initialized from Kev-4B. It scores supplied choices without generating text. It requires the [Sev/Kev runtime](https://github.com/maceip/Sev); loading the LoRA adapter into a text-generation pipeline omits the decision head.
+
+## Use
+
+```bash
+git clone https://github.com/maceip/Sev.git
+cd Sev
+uv sync --extra serve
+uv run python -m kev.serve \
+  --run macmacmacmac/Sev-4B@v0.1.0-research --port 8009
+```
+
+The [repository example](https://github.com/maceip/Sev#try-it) sends an actual development observation after removing its target and audit metadata. The package and API aliases keep their upstream `kev` names. Treat probabilities and the derived `confidence` field as research outputs, not evidence of a person's or agent's identity. The serving implementation permits longer input than training used; accuracy beyond the 384-token training state limit is untested.
+
+## Training
+
+| Setting | Value |
+| --- | --- |
+| Backbone | `Qwen/Qwen3.5-4B-Base@1001bb4d826a52d1f399e183466143f4da7b741b` |
+| Warm start | `jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1` |
+| Run | `sev-r2-research-4b-v1/00-trial-0` |
+| Inputs | 2,094 authored behavioral windows and 240 authored Kev rule records |
+| Epochs / seed | 1 / 4 |
+| Learning rate | `1e-5` |
+| Batch / accumulation | 4 / 2 |
+| LoRA / head | rank 16, all targets / 256 dimensions |
+| Precision | fp32 frozen weights, bf16 autocast, gradient checkpointing |
+| Updates / forward tokens | 292 / 738,327 |
+| Training coverage | 2,334 requested and seen, zero rejected or truncated |
+| Hardware | One NVIDIA H100 80 GB |
+
+The [registered plan](https://github.com/maceip/Sev/blob/main/experiments/sev-research-4b-v1.json) records every override. Use the frozen suite and `kev.experiment` or `modal_app.py::study` to repeat it. The checkpoint includes full training configuration, metrics, log, and source hashes. The training suite manifest SHA-256 is `372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5`.
+
+The continuation excludes the earlier screen's 400 public replay rows. This does not erase the parent model's training history. The Apache-licensed Kev parent was trained on public classification datasets and authored rules. No private phone captures, URLQuery records, retired enterprise synthetic data, or executed LLM traces enter this continuation. `legacy_policy` means authored upstream Kev rules, not retired Sev traffic.
+
+## Evaluation
+
+| Full observation panel | Accuracy | Parent accuracy | Human recall | Script recall | Agent recall |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| IID development | 34.03% | 33.33% | 98.96% | 2.08% | 1.04% |
+| Held-out tasks and policy shift | 33.85% | 33.59% | 98.18% | 3.39% | 0.00% |
+
+Each panel contains 96 episodes from 32 matched scenario groups. IID has 196 full-observation windows and the shifted panel has 279. Each episode has equal weight; its eligible windows share that weight. Probabilities are not pooled into an episode ensemble. Paired bootstrap resampling keeps all origins and views in each scenario group together. The IID gain over the parent is 0.69 percentage points, with a 95% interval of 0.00 to 1.74 points over 1,000 resamples. This small development result does not establish useful discrimination.
+
+| Full observation panel | Raw NLL | Shipped-temperature NLL | Raw ECE | Shipped-temperature ECE |
+| --- | ---: | ---: | ---: | ---: |
+| IID development | 1.1142 | 1.0998 | 0.0658 | 0.0291 |
+| Held-out tasks and policy shift | 1.1021 | 1.0975 | 0.0677 | 0.0392 |
+
+Uniform three-class predictions have NLL 1.0986. Low ECE near chance is not evidence of an effective detector. The parent comparison uses the unchanged pinned Kev checkpoint, scored on identical inputs. There is no fresh general-task retention evaluation for this continuation.
+
+These development panels have informed earlier research and are not untouched confirmation. The separate native final test remains unscored. The release suite's `test.jsonl` is empty. `promotable=false` in the original trial receipt remains unchanged: this is an explicitly limited research publication, not an automatic quality promotion.
+
+## Calibration
+
+The shipped `head.pt` stores temperature 2.0, fitted only on 182 full-view calibration windows from 96 episodes and 32 scenario groups. Fitting gives equal weight to episodes, then equal weight to their windows. Five-fold diagnostics hold out complete scenario groups. Episode-weighted calibration ECE is 0.0655 raw and 0.0616 out of fold. The full fit's lower in-sample ECE is not an unseen-data guarantee.
+
+The experiment also recorded a different, window-weighted diagnostic fit of 1.07177. It is preserved in `result.json`; **2.0 is the shipped temperature**. The release calibration copy changes temperature metadata only: learned head tensors and adapter bytes are unchanged, and the raw training checkpoint remains untouched. See `calibration.json`, `calibration-sources.json`, and `SHA256SUMS` for receipts.
+
+## Data Limits And Intended Use
+
+Labels identify authored policies in a single small simulator. There are no captured humans, live LLM executions, measured browser clocks, or network captures in the training corpus. Logical timing and policy probabilities are authored, not estimates of enterprise populations. Human, script, and agent policies have overlapping possible actions.
+
+Task holdouts also change policy priors within the same controller implementation. They do not test independent controllers or real agent swarms. Ordered windows preserve complete interactions, but the short context can omit useful earlier actions and cross-window dependencies. Internal consistency tests do not establish realism.
+
+Use this release to study evidence representations, sequence context, class bias, calibration, and synthetic-data design. It does not support operational blocking, human-owner attribution, maliciousness judgments, or claims that a real-world action conclusively came from an AI agent.
+
+## Provenance And License
+
+Code and adapter/head weights are Apache-2.0, with Kev and Qwen attribution in `NOTICE` and `BASE_LICENSE`. Original generated behavioral data are CC BY 4.0; the 240 upstream authored rule records retain Apache-2.0. Dataset licenses are separate from model licenses. See the [data and inherited-training provenance](https://github.com/maceip/Sev/blob/main/docs/releases/sev-research-provenance.md).
+
+The [release evidence](https://github.com/maceip/Sev/tree/main/docs/releases/v0.1.0-research) includes source and checkpoint hashes, raw predictions, grouped comparison results, and calibration receipts. The artifact is maintained by [maceip](https://github.com/maceip), using Kev by Jared Palmer and the Qwen team's backbone.

SHA256SUMS

@@ -0,0 +1,20 @@
+50cbab8a892c5f2993b8c7351a99182507472def3b1374558308605d99b86b32  BASE_LICENSE
+6b08bb37982c233aa12bcbdf19106da12f3f4fcf800773ea42e28ebeddd34fb8  LICENSE
+859135197bdac84f3e4fb27463b1abe71cd0344f08a5b4eb970f50684bcceae7  NOTICE
+d032908fe8fda54de4a3a3b29eaa3d6d2ba4bef16266373411bab6688fbedaf4  README.md
+8a05dfd6c5e7f61a62e6db5d8de8093a8dbbd105aa7ee7b5e7b73f6249f45617  adapter_config.json
+7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c  adapter_model.safetensors
+be16ec2fdfb3b28b63888ede85cce428a8106785a2a64b37215550833176d875  calibration-sources.json
+ca2b2eacb602dd61b87139cd5eafe2daf4d3fdd659c8cb8924538567259808bc  calibration.json
+a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715  chat_template.jinja
+68a0de83be179f88ee0f0d324db65ae581c0cfd89177b0bf6a56364a80164686  checkpoint-integrity.json
+8327fc07f1354720311b1ba2e09b84ef7bf946857a1bcd1c6d7c832bc3a97e6b  class-metrics.json
+14c148b41d9c5070f2800b938daf6e647ea3019cafcd2046f739226177469c64  evaluation.json
+dc1f550c7914388af77e607c70fe8801558a89a3847d37ea37d3923b28521c3e  head.pt
+f540a4587f44f4a3bdda19f68a42377a56c56d464443beaaf8fe57b593beb85f  provenance.json
+ec4bfe6436093b6ba096921bb95b2643e8418c84d1e5cbbc596d78f2c744338c  result.json
+06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523  tokenizer.json
+8671bed7c852ce9e661be94f179a7b4ffd091c2a65aea0363e5501c20318ee45  tokenizer_config.json
+e5b54da6b803d7e6d5c0fdfc902b8095a123243a652ec996ab411da0c818ce09  train.log
+f1936818cecded5fe659330a6e0a1ad631d9b5b1eb2014982634a58f7d6f7d68  training_config.json
+20b9314de861f32641b4f6a16ae8b06ba273426333da40c2820e5f591936d7a7  training_metrics.json

adapter_config.json

@@ -0,0 +1,56 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3.5-4B-Base",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "kasa_config": null,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.21.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "in_proj_b",
+    "in_proj_qkv",
+    "gate_proj",
+    "v_proj",
+    "o_proj",
+    "k_proj",
+    "up_proj",
+    "in_proj_z",
+    "in_proj_a",
+    "out_proj",
+    "down_proj",
+    "q_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "FEATURE_EXTRACTION",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c
+size 129924032

calibration-sources.json

@@ -0,0 +1,86 @@
+{
+  "schema": "sev-research-calibration-sources-v1",
+  "split": "calibration",
+  "view": "full",
+  "temperature": 1.9999999999999998,
+  "weighting": "equal episodes, then equal windows within each episode",
+  "partition_sha256": "eebb5ddeb322b16bbb312ddb6927b1d71226b72e4b01d323c698d1e2a3c4f38d",
+  "predictions": {
+    "path": "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/rows.json",
+    "sha256": "00909b4d6b838c534d8d403d84a578fea1ddaac5e30a2b488c2d1c138c0d05ab"
+  },
+  "sources": {
+    "sev_behavioral_artifact_handoff": {
+      "windows": 63,
+      "episodes": 24,
+      "groups": 8,
+      "record_ids_sha256": "ec2841fda8a390bd2d73d828e1d3efac393d6cc583f94188ef25588a85945877",
+      "raw": {
+        "acc": 0.34722222222222215,
+        "nll": 1.0806702183699979,
+        "brier": 0.6537004910585257,
+        "ece": 0.09166280712269316
+      },
+      "calibrated": {
+        "acc": 0.34722222222222215,
+        "nll": 1.0876702845061252,
+        "brier": 0.6589248476452055,
+        "ece": 0.04426971102023442
+      }
+    },
+    "sev_behavioral_authorized_export": {
+      "windows": 40,
+      "episodes": 24,
+      "groups": 8,
+      "record_ids_sha256": "255e72d055b44aa16ec7f11fe905f4c251fe28ced7f801120abef039c930115f",
+      "raw": {
+        "acc": 0.3333333333333333,
+        "nll": 1.104011417056653,
+        "brier": 0.6698152020985213,
+        "ece": 0.04556619047184074
+      },
+      "calibrated": {
+        "acc": 0.3333333333333333,
+        "nll": 1.0996799880244128,
+        "brier": 0.6672264414223563,
+        "ece": 0.04735674179920909
+      }
+    },
+    "sev_behavioral_lookup_catalog": {
+      "windows": 25,
+      "episodes": 24,
+      "groups": 8,
+      "record_ids_sha256": "8cd835d6e3fefa25803ffbc579240aa9c4ae264163027eaa186db00580e4b2e1",
+      "raw": {
+        "acc": 0.3333333333333333,
+        "nll": 1.0929036665832168,
+        "brier": 0.6633979694737079,
+        "ece": 0.10676289475303495
+      },
+      "calibrated": {
+        "acc": 0.3333333333333333,
+        "nll": 1.087115774278896,
+        "brier": 0.6586049595361453,
+        "ece": 0.052123650581165686
+      }
+    },
+    "sev_behavioral_update_export": {
+      "windows": 54,
+      "episodes": 24,
+      "groups": 8,
+      "record_ids_sha256": "f83b45d350186cfa47db970f9311408fbde0fb68bb884adc1a3644a36b18f52c",
+      "raw": {
+        "acc": 0.3229166666666667,
+        "nll": 1.11786418934943,
+        "brier": 0.6813036856150539,
+        "ece": 0.09059596392971296
+      },
+      "calibrated": {
+        "acc": 0.3229166666666667,
+        "nll": 1.1024371940848587,
+        "brier": 0.6695271712596828,
+        "ece": 0.060665733732341603
+      }
+    }
+  }
+}

calibration.json

@@ -0,0 +1,313 @@
+{
+  "schema": "sev-research-calibrated-copy-v1",
+  "created_at": "2026-09-25T16:37:54.666347+00:00",
+  "checkpoint": "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint",
+  "fit": {
+    "schema": "sev-research-calibration-v1",
+    "temperature": 1.9999999999999998,
+    "split": "calibration",
+    "view": "full",
+    "question": "operator_origin",
+    "method": "kev.metrics.fit_temperature",
+    "aggregation": "macro",
+    "task": "episode parent_id",
+    "points": 121,
+    "temperature_range": [
+      0.25,
+      4.0
+    ],
+    "weighting": "equal episodes, then equal windows within each episode",
+    "windows": 182,
+    "episodes": 96,
+    "groups": 32,
+    "record_ids_sha256": "46528c7e8fa5d195456afb8352b64a1393b26852a868eb58fa7aee9d41e1d93b",
+    "excluded_views": {
+      "prefix_loss": 148,
+      "sampled_loss": 152
+    },
+    "raw": {
+      "acc": 0.3342013888888889,
+      "nll": 1.0988623728398248,
+      "brier": 0.6670543370614523,
+      "ece": 0.06554797243603253
+    },
+    "calibrated": {
+      "acc": 0.3342013888888889,
+      "nll": 1.094225810223573,
+      "brier": 0.6635708549658472,
+      "ece": 0.04216089485176933
+    },
+    "cross_validation": {
+      "raw": {
+        "n": 182,
+        "ece": 0.030378155838447805,
+        "brier": 0.6564945291407126,
+        "nll": 1.0838391530675782,
+        "confident_error_rate": 0.0,
+        "coverage_at_5pct_error": 0.01098901098901099
+      },
+      "out_of_fold": {
+        "n": 182,
+        "ece": 0.02879543978652912,
+        "brier": 0.6607993702192454,
+        "nll": 1.0900062934624115,
+        "confident_error_rate": 0.0,
+        "coverage_at_5pct_error": 0.0
+      },
+      "ece_ci95": {
+        "raw": [
+          0.013784180181410588,
+          0.07820699846672471
+        ],
+        "out_of_fold": [
+          0.015176680491519571,
+          0.06251847627618651
+        ],
+        "delta": [
+          -0.045586065150116414,
+          0.03442859983142514
+        ]
+      },
+      "separated": false,
+      "temperatures": [
+        1.5511447618337346,
+        2.578740616879158,
+        1.6624757922855755,
+        2.3510958125672174,
+        2.70069989233638
+      ],
+      "folds": 5,
+      "seed": 0,
+      "samples": 1000,
+      "groups": 32,
+      "unit": "latent scenario group_id, across all original sources",
+      "fit_weighting": "equal episodes, then equal windows within each episode",
+      "canonical_metric_weighting": "windows; canonical raw/out_of_fold metrics and ECE intervals",
+      "episode_weighted": {
+        "raw": {
+          "acc": 0.3342013888888889,
+          "nll": 1.0988623728398248,
+          "brier": 0.6670543370614523,
+          "ece": 0.06554797243603253
+        },
+        "out_of_fold": {
+          "acc": 0.3342013888888889,
+          "nll": 1.0963892736496,
+          "brier": 0.6651579277109647,
+          "ece": 0.061642028028613405
+        }
+      },
+      "fold_by_group": {
+        "0a96bbbde731f77629d6": 0,
+        "0fd4a07c8a94efc8213a": 4,
+        "127dba0a0b6d112fdd60": 0,
+        "1a7d4eff40c538e9531b": 2,
+        "25f1f108010761f574ec": 0,
+        "42172cc6c0db60ca5758": 1,
+        "459d3a949d71768ba55d": 4,
+        "4c3fe25e9bd641be2ddb": 0,
+        "4c6ad83158be617c1548": 4,
+        "53c1f89988e203c21d7f": 1,
+        "635e51a0d3a3a3da2ba7": 4,
+        "7a15bda3e41ccfee7241": 1,
+        "7f573f974db9fa0d872d": 2,
+        "86f21fea379a7a9fe79e": 4,
+        "896e70c3467400fe789c": 3,
+        "89a112d154659e4b8585": 0,
+        "8de646371957ac833b8f": 2,
+        "9267379e905f6045e976": 2,
+        "9343076eec6e0c57c601": 0,
+        "95cf7a4cbfcb6a117bfc": 1,
+        "96651d23eb4107dd8e5a": 3,
+        "99fba8dee219cdf16eef": 3,
+        "9f8e8516bf49e0ec08d5": 0,
+        "a10ba4c34ba318f3829a": 3,
+        "ac1f21ab8ebefecfe315": 3,
+        "ace673b4d894d1f1740e": 2,
+        "ae5e6811a635cc23dfbd": 1,
+        "b786a1db55e5f652c1cf": 4,
+        "b8d7d7b733ad6a06fade": 2,
+        "d983184d237cd08a0327": 1,
+        "e90cd441191e7f18968b": 3,
+        "ed1a95e2dd7115022ea3": 1
+      }
+    },
+    "per_source": {
+      "sev_behavioral_artifact_handoff": {
+        "windows": 63,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "ec2841fda8a390bd2d73d828e1d3efac393d6cc583f94188ef25588a85945877",
+        "raw": {
+          "acc": 0.34722222222222215,
+          "nll": 1.0806702183699979,
+          "brier": 0.6537004910585257,
+          "ece": 0.09166280712269316
+        },
+        "calibrated": {
+          "acc": 0.34722222222222215,
+          "nll": 1.0876702845061252,
+          "brier": 0.6589248476452055,
+          "ece": 0.04426971102023442
+        }
+      },
+      "sev_behavioral_authorized_export": {
+        "windows": 40,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "255e72d055b44aa16ec7f11fe905f4c251fe28ced7f801120abef039c930115f",
+        "raw": {
+          "acc": 0.3333333333333333,
+          "nll": 1.104011417056653,
+          "brier": 0.6698152020985213,
+          "ece": 0.04556619047184074
+        },
+        "calibrated": {
+          "acc": 0.3333333333333333,
+          "nll": 1.0996799880244128,
+          "brier": 0.6672264414223563,
+          "ece": 0.04735674179920909
+        }
+      },
+      "sev_behavioral_lookup_catalog": {
+        "windows": 25,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "8cd835d6e3fefa25803ffbc579240aa9c4ae264163027eaa186db00580e4b2e1",
+        "raw": {
+          "acc": 0.3333333333333333,
+          "nll": 1.0929036665832168,
+          "brier": 0.6633979694737079,
+          "ece": 0.10676289475303495
+        },
+        "calibrated": {
+          "acc": 0.3333333333333333,
+          "nll": 1.087115774278896,
+          "brier": 0.6586049595361453,
+          "ece": 0.052123650581165686
+        }
+      },
+      "sev_behavioral_update_export": {
+        "windows": 54,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "f83b45d350186cfa47db970f9311408fbde0fb68bb884adc1a3644a36b18f52c",
+        "raw": {
+          "acc": 0.3229166666666667,
+          "nll": 1.11786418934943,
+          "brier": 0.6813036856150539,
+          "ece": 0.09059596392971296
+        },
+        "calibrated": {
+          "acc": 0.3229166666666667,
+          "nll": 1.1024371940848587,
+          "brier": 0.6695271712596828,
+          "ece": 0.060665733732341603
+        }
+      }
+    },
+    "development_read": false,
+    "test_read": false,
+    "scope": "Calibration evidence on authored simulated controllers; no field accuracy or release gate.",
+    "suite_sha256": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+    "partition_sha256": "eebb5ddeb322b16bbb312ddb6927b1d71226b72e4b01d323c698d1e2a3c4f38d",
+    "predictions": {
+      "path": "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/rows.json",
+      "sha256": "00909b4d6b838c534d8d403d84a578fea1ddaac5e30a2b488c2d1c138c0d05ab"
+    },
+    "input_sha256": {
+      "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/manifest.json": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+      "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/calibration.jsonl": "eebb5ddeb322b16bbb312ddb6927b1d71226b72e4b01d323c698d1e2a3c4f38d",
+      "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/provenance.json": "f540a4587f44f4a3bdda19f68a42377a56c56d464443beaaf8fe57b593beb85f",
+      "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/result.json": "ec4bfe6436093b6ba096921bb95b2643e8418c84d1e5cbbc596d78f2c744338c",
+      "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/rows.json": "00909b4d6b838c534d8d403d84a578fea1ddaac5e30a2b488c2d1c138c0d05ab",
+      "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/report.json": "6bee81d641d7003c7715c0b26bcec77fe244e59e847e52e7f14f7336d88f59f9",
+      "/Users/mac/Sev-release-20260925/kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59",
+      "/Users/mac/Sev-release-20260925/kev/metrics.py": "db0930d470c0ce16f16dcaf726d49a7a103973a5296cefc5259d1e68707d882b",
+      "/Users/mac/Sev-release-20260925/scripts/report_sev_behavioral_round.py": "5beee8c302040233e2950ebe2d3f8944e84d6202f08b95943d679690bcdad8d8",
+      "/Users/mac/Sev-release-20260925/scripts/calibrate_sev_research.py": "f9fa4df59fda060e856f4e2a691e8fc0aeeeec836f1ddbaa43b3db5a39a874cf"
+    },
+    "source_checkpoint": "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint",
+    "source_files": {
+      "README.md": {
+        "sha256": "e0e50894b914f1765ef42c1eade74b42c613fb0330b06339f9502cc6f2e25e79",
+        "bytes": 5164
+      },
+      "adapter_config.json": {
+        "sha256": "8a05dfd6c5e7f61a62e6db5d8de8093a8dbbd105aa7ee7b5e7b73f6249f45617",
+        "bytes": 1271
+      },
+      "adapter_model.safetensors": {
+        "sha256": "7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c",
+        "bytes": 129924032
+      },
+      "chat_template.jinja": {
+        "sha256": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+        "bytes": 7756
+      },
+      "head.pt": {
+        "sha256": "af6bfd215bf0b7eff92612a83b21cb56754ca87f0bf9f62fdfcfbef275d9eab6",
+        "bytes": 5248895
+      },
+      "tokenizer.json": {
+        "sha256": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
+        "bytes": 19989325
+      },
+      "tokenizer_config.json": {
+        "sha256": "8671bed7c852ce9e661be94f179a7b4ffd091c2a65aea0363e5501c20318ee45",
+        "bytes": 1128
+      },
+      "training_config.json": {
+        "sha256": "f1936818cecded5fe659330a6e0a1ad631d9b5b1eb2014982634a58f7d6f7d68",
+        "bytes": 1889
+      },
+      "training_metrics.json": {
+        "sha256": "20b9314de861f32641b4f6a16ae8b06ba273426333da40c2820e5f591936d7a7",
+        "bytes": 321
+      }
+    },
+    "source_temperature": 1.0,
+    "head_tensors": {
+      "k.bias": {
+        "shape": [
+          256
+        ],
+        "dtype": "torch.float32",
+        "sha256": "b8d56826b48ca7768c910704d155b6f9368481658f4b7be5e6e87c9e3a77109a"
+      },
+      "k.weight": {
+        "shape": [
+          256,
+          2560
+        ],
+        "dtype": "torch.float32",
+        "sha256": "7caa9354998b823ae4ca2c87b849b240bf474cfb8479b43e8a34e012e1fb2d97"
+      },
+      "q.bias": {
+        "shape": [
+          256
+        ],
+        "dtype": "torch.float32",
+        "sha256": "341acaf84e9296a1a6676e25789b2ce97c8632a408ea31c3a626f48e840496e4"
+      },
+      "q.weight": {
+        "shape": [
+          256,
+          2560
+        ],
+        "dtype": "torch.float32",
+        "sha256": "b32e0568a714d926fe5faef0cffc7f0215c61f0f4874e2138303a7641a4cbfca"
+      }
+    },
+    "calibration_source_sha256": {
+      "kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59",
+      "kev/metrics.py": "db0930d470c0ce16f16dcaf726d49a7a103973a5296cefc5259d1e68707d882b",
+      "scripts/report_sev_behavioral_round.py": "5beee8c302040233e2950ebe2d3f8944e84d6202f08b95943d679690bcdad8d8",
+      "scripts/calibrate_sev_research.py": "f9fa4df59fda060e856f4e2a691e8fc0aeeeec836f1ddbaa43b3db5a39a874cf"
+    }
+  },
+  "output_head_sha256": "dc1f550c7914388af77e607c70fe8801558a89a3847d37ea37d3923b28521c3e",
+  "source_unchanged": true,
+  "learned_tensors_unchanged": true,
+  "adapter_unchanged": true
+}

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

checkpoint-integrity.json

@@ -0,0 +1,462 @@
+{
+  "schema": "sev-calibrated-checkpoint-integrity-v1",
+  "pass": true,
+  "verified_at": "2026-09-25T16:42:29.007740+00:00",
+  "source_checkpoint": "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint",
+  "calibrated_checkpoint": "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint",
+  "suite_sha256": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+  "raw_head_sha256": "af6bfd215bf0b7eff92612a83b21cb56754ca87f0bf9f62fdfcfbef275d9eab6",
+  "calibrated_head_sha256": "dc1f550c7914388af77e607c70fe8801558a89a3847d37ea37d3923b28521c3e",
+  "adapter_sha256": "7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c",
+  "raw_temperature": 1.0,
+  "released_temperature": 1.9999999999999998,
+  "training": {
+    "wall_seconds": 371.20887875556946,
+    "records_seen": 2334,
+    "requested_records": 2334,
+    "truncated_records": 0,
+    "rejected_records": 0,
+    "optimizer_steps": 292,
+    "forward_tokens": 738327,
+    "peak_device_bytes": 18913074688,
+    "device": "cuda",
+    "dtype": "bf16",
+    "batch": 4,
+    "peak_rss_bytes": 31561670656
+  },
+  "training_sources": {
+    "sev_behavioral_lookup_catalog": 391,
+    "sev_behavioral_update_export": 533,
+    "sev_behavioral_artifact_handoff": 739,
+    "sev_behavioral_authorized_export": 431,
+    "compositional": 160,
+    "legacy_policy": 80
+  },
+  "warm_start": {
+    "init_from": "jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1",
+    "resolved": "/__modal/volumes/vo-RociVuGycNGtzvUcdOqecz/hub/models--jaredpalmer--kev-4b/snapshots/485ace8703592fcf405488b262449990824cfed1",
+    "adapter_sha256": "9797de69a42188e411b17b7b4fcb66a23374dcebc21d71a7a66f836b5d34df2b",
+    "head_sha256": "d8f796da36ff7bd7c0fb9496b452139bb7851af4fc82b07b500b682d3f721d6a",
+    "adapter_tensors": 496
+  },
+  "adapter_tensors": 496,
+  "head_tensors": {
+    "q.weight": {
+      "shape": [
+        256,
+        2560
+      ],
+      "dtype": "torch.float32",
+      "raw_bytes_sha256": "b32e0568a714d926fe5faef0cffc7f0215c61f0f4874e2138303a7641a4cbfca",
+      "byte_identical": true
+    },
+    "q.bias": {
+      "shape": [
+        256
+      ],
+      "dtype": "torch.float32",
+      "raw_bytes_sha256": "341acaf84e9296a1a6676e25789b2ce97c8632a408ea31c3a626f48e840496e4",
+      "byte_identical": true
+    },
+    "k.weight": {
+      "shape": [
+        256,
+        2560
+      ],
+      "dtype": "torch.float32",
+      "raw_bytes_sha256": "7caa9354998b823ae4ca2c87b849b240bf474cfb8479b43e8a34e012e1fb2d97",
+      "byte_identical": true
+    },
+    "k.bias": {
+      "shape": [
+        256
+      ],
+      "dtype": "torch.float32",
+      "raw_bytes_sha256": "b8d56826b48ca7768c910704d155b6f9368481658f4b7be5e6e87c9e3a77109a",
+      "byte_identical": true
+    }
+  },
+  "calibration": {
+    "windows": 182,
+    "episodes": 96,
+    "groups": 32,
+    "temperature": 1.9999999999999998,
+    "raw": {
+      "acc": 0.3342013888888889,
+      "nll": 1.0988623728398248,
+      "brier": 0.6670543370614523,
+      "ece": 0.06554797243603253
+    },
+    "calibrated": {
+      "acc": 0.3342013888888889,
+      "nll": 1.094225810223573,
+      "brier": 0.6635708549658472,
+      "ece": 0.04216089485176933
+    },
+    "cross_validation": {
+      "raw": {
+        "n": 182,
+        "ece": 0.030378155838447805,
+        "brier": 0.6564945291407126,
+        "nll": 1.0838391530675782,
+        "confident_error_rate": 0.0,
+        "coverage_at_5pct_error": 0.01098901098901099
+      },
+      "out_of_fold": {
+        "n": 182,
+        "ece": 0.02879543978652912,
+        "brier": 0.6607993702192454,
+        "nll": 1.0900062934624115,
+        "confident_error_rate": 0.0,
+        "coverage_at_5pct_error": 0.0
+      },
+      "ece_ci95": {
+        "raw": [
+          0.013784180181410588,
+          0.07820699846672471
+        ],
+        "out_of_fold": [
+          0.015176680491519571,
+          0.06251847627618651
+        ],
+        "delta": [
+          -0.045586065150116414,
+          0.03442859983142514
+        ]
+      },
+      "separated": false,
+      "temperatures": [
+        1.5511447618337346,
+        2.578740616879158,
+        1.6624757922855755,
+        2.3510958125672174,
+        2.70069989233638
+      ],
+      "folds": 5,
+      "seed": 0,
+      "samples": 1000,
+      "groups": 32,
+      "unit": "latent scenario group_id, across all original sources",
+      "fit_weighting": "equal episodes, then equal windows within each episode",
+      "canonical_metric_weighting": "windows; canonical raw/out_of_fold metrics and ECE intervals",
+      "episode_weighted": {
+        "raw": {
+          "acc": 0.3342013888888889,
+          "nll": 1.0988623728398248,
+          "brier": 0.6670543370614523,
+          "ece": 0.06554797243603253
+        },
+        "out_of_fold": {
+          "acc": 0.3342013888888889,
+          "nll": 1.0963892736496,
+          "brier": 0.6651579277109647,
+          "ece": 0.061642028028613405
+        }
+      },
+      "fold_by_group": {
+        "0a96bbbde731f77629d6": 0,
+        "0fd4a07c8a94efc8213a": 4,
+        "127dba0a0b6d112fdd60": 0,
+        "1a7d4eff40c538e9531b": 2,
+        "25f1f108010761f574ec": 0,
+        "42172cc6c0db60ca5758": 1,
+        "459d3a949d71768ba55d": 4,
+        "4c3fe25e9bd641be2ddb": 0,
+        "4c6ad83158be617c1548": 4,
+        "53c1f89988e203c21d7f": 1,
+        "635e51a0d3a3a3da2ba7": 4,
+        "7a15bda3e41ccfee7241": 1,
+        "7f573f974db9fa0d872d": 2,
+        "86f21fea379a7a9fe79e": 4,
+        "896e70c3467400fe789c": 3,
+        "89a112d154659e4b8585": 0,
+        "8de646371957ac833b8f": 2,
+        "9267379e905f6045e976": 2,
+        "9343076eec6e0c57c601": 0,
+        "95cf7a4cbfcb6a117bfc": 1,
+        "96651d23eb4107dd8e5a": 3,
+        "99fba8dee219cdf16eef": 3,
+        "9f8e8516bf49e0ec08d5": 0,
+        "a10ba4c34ba318f3829a": 3,
+        "ac1f21ab8ebefecfe315": 3,
+        "ace673b4d894d1f1740e": 2,
+        "ae5e6811a635cc23dfbd": 1,
+        "b786a1db55e5f652c1cf": 4,
+        "b8d7d7b733ad6a06fade": 2,
+        "d983184d237cd08a0327": 1,
+        "e90cd441191e7f18968b": 3,
+        "ed1a95e2dd7115022ea3": 1
+      }
+    },
+    "per_source": {
+      "sev_behavioral_artifact_handoff": {
+        "windows": 63,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "ec2841fda8a390bd2d73d828e1d3efac393d6cc583f94188ef25588a85945877",
+        "raw": {
+          "acc": 0.34722222222222215,
+          "nll": 1.0806702183699979,
+          "brier": 0.6537004910585257,
+          "ece": 0.09166280712269316
+        },
+        "calibrated": {
+          "acc": 0.34722222222222215,
+          "nll": 1.0876702845061252,
+          "brier": 0.6589248476452055,
+          "ece": 0.04426971102023442
+        }
+      },
+      "sev_behavioral_authorized_export": {
+        "windows": 40,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "255e72d055b44aa16ec7f11fe905f4c251fe28ced7f801120abef039c930115f",
+        "raw": {
+          "acc": 0.3333333333333333,
+          "nll": 1.104011417056653,
+          "brier": 0.6698152020985213,
+          "ece": 0.04556619047184074
+        },
+        "calibrated": {
+          "acc": 0.3333333333333333,
+          "nll": 1.0996799880244128,
+          "brier": 0.6672264414223563,
+          "ece": 0.04735674179920909
+        }
+      },
+      "sev_behavioral_lookup_catalog": {
+        "windows": 25,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "8cd835d6e3fefa25803ffbc579240aa9c4ae264163027eaa186db00580e4b2e1",
+        "raw": {
+          "acc": 0.3333333333333333,
+          "nll": 1.0929036665832168,
+          "brier": 0.6633979694737079,
+          "ece": 0.10676289475303495
+        },
+        "calibrated": {
+          "acc": 0.3333333333333333,
+          "nll": 1.087115774278896,
+          "brier": 0.6586049595361453,
+          "ece": 0.052123650581165686
+        }
+      },
+      "sev_behavioral_update_export": {
+        "windows": 54,
+        "episodes": 24,
+        "groups": 8,
+        "record_ids_sha256": "f83b45d350186cfa47db970f9311408fbde0fb68bb884adc1a3644a36b18f52c",
+        "raw": {
+          "acc": 0.3229166666666667,
+          "nll": 1.11786418934943,
+          "brier": 0.6813036856150539,
+          "ece": 0.09059596392971296
+        },
+        "calibrated": {
+          "acc": 0.3229166666666667,
+          "nll": 1.1024371940848587,
+          "brier": 0.6695271712596828,
+          "ece": 0.060665733732341603
+        }
+      }
+    }
+  },
+  "recorded_runtime_source_hashes": {
+    "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
+    "kev/api.py": "7bffacfb762c626b8bc2f670f350295af5ccb0883dbe90239c7d2f8e5ef56582",
+    "kev/autoresearch.py": "f7d7fabb9c0df7065bee3fec4aa4bee7028c5d1565847aefe2d74e3d7d40ed8a",
+    "kev/benchmark.py": "8d486659a9036102386b077fcbd197fc0d85ae7752a4711666d6b3a95622b26f",
+    "kev/calibrate.py": "c1eff4744fabd349e8abca86777a7aa0cbea44904c8223d3b61fca3d43741519",
+    "kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59",
+    "kev/compare.py": "3606dbf98bb305d66420158498cb837304d2edb8dab01cdaf2ec42964bef3435",
+    "kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536",
+    "kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf",
+    "kev/data.py": "a5a5e200f56b60e4bf30a195a4cc87a74afddae15999ec48c7543a4f9b59f71c",
+    "kev/device.py": "d1677fd98ec0979c7284546306e34e0d09298ca4042fc39eb2b32a74f4e975c4",
+    "kev/evaluate.py": "999264f2837dcfbf2ec93601aa4745e701698ab850a43ad89324672a6604965c",
+    "kev/experiment.py": "1a5ca47d2b3d236e53c87e8c4197704f162c6692aa008521d31ea491945693f0",
+    "kev/jev.py": "acd4cc3f1844e438cc83a8d409c15ef78a5ef64d39ea10f583d75a6c7646b243",
+    "kev/metrics.py": "db0930d470c0ce16f16dcaf726d49a7a103973a5296cefc5259d1e68707d882b",
+    "kev/mlx_model.py": "f582428796faf6bf962b772a69a6227ac3259ccbd3215c0c8b55ce28dbbd0909",
+    "kev/model.py": "eac0f56c807b85b5c4077483d922e6bb03f07890de2835efd9372cf20cd01b8e",
+    "kev/plot.py": "0874bcff2885d8155a1cceade0de8a275a7163d3c3fd6aa0c294e8ccf5ff7e02",
+    "kev/predictors.py": "b2a66dd9f0f0a8026bc94603383bacfdd622e75d1c935173997175606626a9b3",
+    "kev/publish.py": "8abc9bcf4a01365697b05cf3c5ad0013462bd92d005f5616954e40ab1100c7aa",
+    "kev/serve.py": "c210b1f4598e64b49c399f73603ed372f4314c165ceda50e1340c0466ec1c735",
+    "kev/study_v3.py": "7891150aa623e4479c7c789d4dc4662f186132b37b7bac1e1f072b9e08ee13e3",
+    "kev/suite.py": "c5e5e27f2fed871b547a021d180b0a7f4a0542a488cce4ee02ad936f433441db",
+    "kev/train.py": "a2d85632d73760fb3efe51627ff4fde6af3100ed1c34e75d290c1b00c9899aa9",
+    "kev/transfer_v9.py": "0902409742151250a28af1fd8f42258b70fcc161f3deed3ee35fe3f473f3c763",
+    "modal_app.py": "5d55aa2a454aba2ba6e4a392ed77ab2bc849e9823b8298a41ce1b8fe4217344b",
+    "pyproject.toml": "7c17fbe9efcadda3eb488b59adf1b6db5938cd33af4d3bd389ce4359b73fcc7d",
+    "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
+  },
+  "current_release_source_hashes": {
+    "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
+    "kev/api.py": "7bffacfb762c626b8bc2f670f350295af5ccb0883dbe90239c7d2f8e5ef56582",
+    "kev/autoresearch.py": "f7d7fabb9c0df7065bee3fec4aa4bee7028c5d1565847aefe2d74e3d7d40ed8a",
+    "kev/benchmark.py": "8d486659a9036102386b077fcbd197fc0d85ae7752a4711666d6b3a95622b26f",
+    "kev/calibrate.py": "c1eff4744fabd349e8abca86777a7aa0cbea44904c8223d3b61fca3d43741519",
+    "kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59",
+    "kev/compare.py": "3606dbf98bb305d66420158498cb837304d2edb8dab01cdaf2ec42964bef3435",
+    "kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536",
+    "kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf",
+    "kev/data.py": "a5a5e200f56b60e4bf30a195a4cc87a74afddae15999ec48c7543a4f9b59f71c",
+    "kev/device.py": "d1677fd98ec0979c7284546306e34e0d09298ca4042fc39eb2b32a74f4e975c4",
+    "kev/evaluate.py": "999264f2837dcfbf2ec93601aa4745e701698ab850a43ad89324672a6604965c",
+    "kev/experiment.py": "1a5ca47d2b3d236e53c87e8c4197704f162c6692aa008521d31ea491945693f0",
+    "kev/jev.py": "acd4cc3f1844e438cc83a8d409c15ef78a5ef64d39ea10f583d75a6c7646b243",
+    "kev/metrics.py": "db0930d470c0ce16f16dcaf726d49a7a103973a5296cefc5259d1e68707d882b",
+    "kev/mlx_model.py": "f582428796faf6bf962b772a69a6227ac3259ccbd3215c0c8b55ce28dbbd0909",
+    "kev/model.py": "eac0f56c807b85b5c4077483d922e6bb03f07890de2835efd9372cf20cd01b8e",
+    "kev/plot.py": "0874bcff2885d8155a1cceade0de8a275a7163d3c3fd6aa0c294e8ccf5ff7e02",
+    "kev/predictors.py": "b2a66dd9f0f0a8026bc94603383bacfdd622e75d1c935173997175606626a9b3",
+    "kev/publish.py": "8abc9bcf4a01365697b05cf3c5ad0013462bd92d005f5616954e40ab1100c7aa",
+    "kev/serve.py": "c210b1f4598e64b49c399f73603ed372f4314c165ceda50e1340c0466ec1c735",
+    "kev/study_v3.py": "7891150aa623e4479c7c789d4dc4662f186132b37b7bac1e1f072b9e08ee13e3",
+    "kev/suite.py": "c5e5e27f2fed871b547a021d180b0a7f4a0542a488cce4ee02ad936f433441db",
+    "kev/train.py": "a2d85632d73760fb3efe51627ff4fde6af3100ed1c34e75d290c1b00c9899aa9",
+    "kev/transfer_v9.py": "0902409742151250a28af1fd8f42258b70fcc161f3deed3ee35fe3f473f3c763",
+    "modal_app.py": "5d55aa2a454aba2ba6e4a392ed77ab2bc849e9823b8298a41ce1b8fe4217344b",
+    "pyproject.toml": "39ec40f14c88598dee7ac98fb859a5a44e22337235a3064ebd5e89036ce81e30",
+    "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
+  },
+  "release_metadata_drift": {
+    "pyproject.toml": {
+      "trained_sha256": "7c17fbe9efcadda3eb488b59adf1b6db5938cd33af4d3bd389ce4359b73fcc7d",
+      "release_sha256": "39ec40f14c88598dee7ac98fb859a5a44e22337235a3064ebd5e89036ce81e30",
+      "changed_fields": [
+        "project.description",
+        "project.urls"
+      ],
+      "all_other_toml_equal": true
+    }
+  },
+  "recorded_git_commit": "d8b1413f428a9a8f4e3d6091acf3d9ed8445c864",
+  "recorded_commit_source_mismatches": [],
+  "checks": {
+    "suite_and_plan_match_registered_hashes": true,
+    "registration_names_fresh_study": true,
+    "configuration_matches_registered_recipe": true,
+    "result_binds_execution_provenance": true,
+    "raw_checkpoint_hashes_match_scored_artifact": true,
+    "raw_temperature_verified": true,
+    "trained_and_evaluated_suite_match": true,
+    "actual_trainer_arguments_match_plan": true,
+    "head_metadata_matches_training_arguments": true,
+    "base_and_architecture_match_plan": true,
+    "warm_start_matches_pinned_clean_parent": true,
+    "training_membership_is_registered_authored_subset": true,
+    "complete_training_without_rejection": true,
+    "no_locked_test_evaluation": true,
+    "recorded_mechanism_checks_pass": true,
+    "calibration_complete_evaluation": true,
+    "development_complete_evaluation": true,
+    "only_pyproject_differs_from_recorded_runtime_sources": true,
+    "original_project_file_matches_training": true,
+    "project_drift_is_description_and_urls_only": true,
+    "learned_head_tensor_names_match": true,
+    "head_tensor:q.weight": true,
+    "head_tensor:q.bias": true,
+    "head_tensor:k.weight": true,
+    "head_tensor:k.bias": true,
+    "adapter_has_expected_finite_tensors": true,
+    "adapter_is_byte_identical": true,
+    "copied_file:adapter_model.safetensors": true,
+    "copied_file:tokenizer_config.json": true,
+    "copied_file:tokenizer.json": true,
+    "copied_file:README.md": true,
+    "copied_file:training_metrics.json": true,
+    "copied_file:adapter_config.json": true,
+    "copied_file:chat_template.jinja": true,
+    "copied_file:training_config.json": true,
+    "calibration_receipt_matches_saved_metadata": true,
+    "all_calibration_input_hashes_still_match": true,
+    "per_source_receipt_matches_fit": true,
+    "temperature_independently_recomputed": true,
+    "calibration_population_exact": true,
+    "cv_is_latent_group_disjoint": true,
+    "calibration_excludes_development_and_test": true,
+    "all_inputs_stable_through_verification": true
+  },
+  "artifact_sha256": {
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/README.md": "e0e50894b914f1765ef42c1eade74b42c613fb0330b06339f9502cc6f2e25e79",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/adapter_config.json": "8a05dfd6c5e7f61a62e6db5d8de8093a8dbbd105aa7ee7b5e7b73f6249f45617",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/adapter_model.safetensors": "7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/head.pt": "af6bfd215bf0b7eff92612a83b21cb56754ca87f0bf9f62fdfcfbef275d9eab6",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/tokenizer.json": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/tokenizer_config.json": "8671bed7c852ce9e661be94f179a7b4ffd091c2a65aea0363e5501c20318ee45",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/training_config.json": "f1936818cecded5fe659330a6e0a1ad631d9b5b1eb2014982634a58f7d6f7d68",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint/training_metrics.json": "20b9314de861f32641b4f6a16ae8b06ba273426333da40c2820e5f591936d7a7",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/README.md": "e0e50894b914f1765ef42c1eade74b42c613fb0330b06339f9502cc6f2e25e79",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/adapter_config.json": "8a05dfd6c5e7f61a62e6db5d8de8093a8dbbd105aa7ee7b5e7b73f6249f45617",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/adapter_model.safetensors": "7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/calibration-sources.json": "be16ec2fdfb3b28b63888ede85cce428a8106785a2a64b37215550833176d875",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/calibration.json": "ca2b2eacb602dd61b87139cd5eafe2daf4d3fdd659c8cb8924538567259808bc",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/head.pt": "dc1f550c7914388af77e607c70fe8801558a89a3847d37ea37d3923b28521c3e",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/tokenizer.json": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/tokenizer_config.json": "8671bed7c852ce9e661be94f179a7b4ffd091c2a65aea0363e5501c20318ee45",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/training_config.json": "f1936818cecded5fe659330a6e0a1ad631d9b5b1eb2014982634a58f7d6f7d68",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/calibrated-checkpoint/training_metrics.json": "20b9314de861f32641b4f6a16ae8b06ba273426333da40c2820e5f591936d7a7",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/provenance.json": "f540a4587f44f4a3bdda19f68a42377a56c56d464443beaaf8fe57b593beb85f",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/result.json": "ec4bfe6436093b6ba096921bb95b2643e8418c84d1e5cbbc596d78f2c744338c",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/train.log": "e5b54da6b803d7e6d5c0fdfc902b8095a123243a652ec996ab411da0c818ce09",
+    "/Users/mac/Sev/runs/sev-research-release-20260925/training-registration.json": "d46082575bf06ecbba776df6476be2cef694ba4f1e1c7ecb8f1b4048c33cb160",
+    "/Users/mac/Sev-release-20260925/experiments/sev-research-4b-v1.json": "af05762d5194a556cc6a518b0fc0860e6942157f0790b48ae80fe85098c5d94f",
+    "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/manifest.json": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+    "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/train.jsonl": "0aa668b5cc927db65e6b7a0890bc47ff7061d0ec271ff239d7d120f362964a6e",
+    "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/calibration.jsonl": "eebb5ddeb322b16bbb312ddb6927b1d71226b72e4b01d323c698d1e2a3c4f38d",
+    "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/development.jsonl": "bd76cba90825cfb2699ff8a4419c99f7b395541ed1cfe51c4a69305451f659e5",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/rows.json": "00909b4d6b838c534d8d403d84a578fea1ddaac5e30a2b488c2d1c138c0d05ab",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/report.json": "6bee81d641d7003c7715c0b26bcec77fe244e59e847e52e7f14f7336d88f59f9",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/temperature.json": "497cbf805d98eeaa745023e5242e0a0dfb62ed8e588a6dbec1e46fe0ac08a3cc",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/development/report.json": "3c3966f1f1ffbc4e55a634c359d02590f2cd1f5e2ff071b98d4dca610ab8a2d6",
+    "/Users/mac/Sev/runs/sev-r2-behavioral-4b-parents-v1/00-kev-4b@485ace8703592fcf405488b262449990824cfed1/provenance.json": "38b41be0ab1c27020c14da51c62cb13a76a8b140d5d89dbfc48078abe18fe277",
+    "/Users/mac/Sev/pyproject.toml": "7c17fbe9efcadda3eb488b59adf1b6db5938cd33af4d3bd389ce4359b73fcc7d",
+    "/Users/mac/Sev-release-20260925/kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "/Users/mac/Sev-release-20260925/kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
+    "/Users/mac/Sev-release-20260925/kev/api.py": "7bffacfb762c626b8bc2f670f350295af5ccb0883dbe90239c7d2f8e5ef56582",
+    "/Users/mac/Sev-release-20260925/kev/autoresearch.py": "f7d7fabb9c0df7065bee3fec4aa4bee7028c5d1565847aefe2d74e3d7d40ed8a",
+    "/Users/mac/Sev-release-20260925/kev/benchmark.py": "8d486659a9036102386b077fcbd197fc0d85ae7752a4711666d6b3a95622b26f",
+    "/Users/mac/Sev-release-20260925/kev/calibrate.py": "c1eff4744fabd349e8abca86777a7aa0cbea44904c8223d3b61fca3d43741519",
+    "/Users/mac/Sev-release-20260925/kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59",
+    "/Users/mac/Sev-release-20260925/kev/compare.py": "3606dbf98bb305d66420158498cb837304d2edb8dab01cdaf2ec42964bef3435",
+    "/Users/mac/Sev-release-20260925/kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536",
+    "/Users/mac/Sev-release-20260925/kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf",
+    "/Users/mac/Sev-release-20260925/kev/data.py": "a5a5e200f56b60e4bf30a195a4cc87a74afddae15999ec48c7543a4f9b59f71c",
+    "/Users/mac/Sev-release-20260925/kev/device.py": "d1677fd98ec0979c7284546306e34e0d09298ca4042fc39eb2b32a74f4e975c4",
+    "/Users/mac/Sev-release-20260925/kev/evaluate.py": "999264f2837dcfbf2ec93601aa4745e701698ab850a43ad89324672a6604965c",
+    "/Users/mac/Sev-release-20260925/kev/experiment.py": "1a5ca47d2b3d236e53c87e8c4197704f162c6692aa008521d31ea491945693f0",
+    "/Users/mac/Sev-release-20260925/kev/jev.py": "acd4cc3f1844e438cc83a8d409c15ef78a5ef64d39ea10f583d75a6c7646b243",
+    "/Users/mac/Sev-release-20260925/kev/metrics.py": "db0930d470c0ce16f16dcaf726d49a7a103973a5296cefc5259d1e68707d882b",
+    "/Users/mac/Sev-release-20260925/kev/mlx_model.py": "f582428796faf6bf962b772a69a6227ac3259ccbd3215c0c8b55ce28dbbd0909",
+    "/Users/mac/Sev-release-20260925/kev/model.py": "eac0f56c807b85b5c4077483d922e6bb03f07890de2835efd9372cf20cd01b8e",
+    "/Users/mac/Sev-release-20260925/kev/plot.py": "0874bcff2885d8155a1cceade0de8a275a7163d3c3fd6aa0c294e8ccf5ff7e02",
+    "/Users/mac/Sev-release-20260925/kev/predictors.py": "b2a66dd9f0f0a8026bc94603383bacfdd622e75d1c935173997175606626a9b3",
+    "/Users/mac/Sev-release-20260925/kev/publish.py": "8abc9bcf4a01365697b05cf3c5ad0013462bd92d005f5616954e40ab1100c7aa",
+    "/Users/mac/Sev-release-20260925/kev/serve.py": "c210b1f4598e64b49c399f73603ed372f4314c165ceda50e1340c0466ec1c735",
+    "/Users/mac/Sev-release-20260925/kev/study_v3.py": "7891150aa623e4479c7c789d4dc4662f186132b37b7bac1e1f072b9e08ee13e3",
+    "/Users/mac/Sev-release-20260925/kev/suite.py": "c5e5e27f2fed871b547a021d180b0a7f4a0542a488cce4ee02ad936f433441db",
+    "/Users/mac/Sev-release-20260925/kev/train.py": "a2d85632d73760fb3efe51627ff4fde6af3100ed1c34e75d290c1b00c9899aa9",
+    "/Users/mac/Sev-release-20260925/kev/transfer_v9.py": "0902409742151250a28af1fd8f42258b70fcc161f3deed3ee35fe3f473f3c763",
+    "/Users/mac/Sev-release-20260925/modal_app.py": "5d55aa2a454aba2ba6e4a392ed77ab2bc849e9823b8298a41ce1b8fe4217344b",
+    "/Users/mac/Sev-release-20260925/pyproject.toml": "39ec40f14c88598dee7ac98fb859a5a44e22337235a3064ebd5e89036ce81e30",
+    "/Users/mac/Sev-release-20260925/uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466",
+    "/Users/mac/Sev-release-20260925/scripts/report_sev_behavioral_round.py": "5beee8c302040233e2950ebe2d3f8944e84d6202f08b95943d679690bcdad8d8",
+    "/Users/mac/Sev-release-20260925/scripts/calibrate_sev_research.py": "f9fa4df59fda060e856f4e2a691e8fc0aeeeec836f1ddbaa43b3db5a39a874cf"
+  },
+  "locked_test_read": false,
+  "limits": [
+    "Artifact consistency and calibration verification only; no fresh inference or retraining.",
+    "Forward-token total is matched across trainer/result receipts, not independently reconstructed.",
+    "Warm-start adapter identity is compared with the prior pinned-parent evaluation receipt; parent adapter was not downloaded.",
+    "Calibration data were previously used in research. Calibration is not a field-accuracy or publication-quality guarantee.",
+    "Recorded runtime source hashes are authoritative because the recorded git base does not include uncommitted release changes."
+  ]
+}

class-metrics.json

@@ -0,0 +1,956 @@
+{
+  "schema": "sev-research-independent-evaluation-v1",
+  "created_at": "2026-09-25T16:42:32.814870+00:00",
+  "pass": true,
+  "suite_manifest_sha256": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+  "input_sha256": {
+    "/Users/mac/Sev/runs/sev-research-release-20260925/paired-report.json": "14c148b41d9c5070f2800b938daf6e647ea3019cafcd2046f739226177469c64",
+    "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/manifest.json": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+    "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/calibration.jsonl": "eebb5ddeb322b16bbb312ddb6927b1d71226b72e4b01d323c698d1e2a3c4f38d",
+    "/Users/mac/Sev-release-20260925/evals/sev/behavioral-research-v1/development.jsonl": "bd76cba90825cfb2699ff8a4419c99f7b395541ed1cfe51c4a69305451f659e5",
+    "/Users/mac/Sev/runs/sev-r2-behavioral-4b-parents-v1/00-kev-4b@485ace8703592fcf405488b262449990824cfed1/calibration/rows.json": "e307a8170f6c1c159e252c8e6c55c88f51e37a784e0a4579bd5625c6742952cf",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/calibration/rows.json": "00909b4d6b838c534d8d403d84a578fea1ddaac5e30a2b488c2d1c138c0d05ab",
+    "/Users/mac/Sev/runs/sev-r2-behavioral-4b-parents-v1/00-kev-4b@485ace8703592fcf405488b262449990824cfed1/development/rows.json": "9ca8eeeb877ea120e5f4c5d54fa72f9bf31f200bbf460e14e6ec85c223162d56",
+    "/Users/mac/Sev/runs/sev-r2-research-4b-v1/00-trial-0/development/rows.json": "994e39a231bc64037f392e606c5fbd3c455f79e68f7b34ba44517238b269b26a"
+  },
+  "test_read": false,
+  "method": {
+    "point_metrics": "Independent NumPy log-softmax NLL, multiclass Brier and ten-bin weighted ECE. Raw confidence/Brier use saved float32 probabilities to match the documented canonical metric contract; tempered probabilities come from logits. Every episode has total weight one divided over its original full-view windows. No probability ensemble.",
+    "paired_intervals": "Independent paired 1000-draw seed-0 percentile bootstrap by group_id alone. Group draw multiplicities applied to row weights. All accuracy/NLL/Brier/ECE point values and interval endpoints match paired-report.json within 1e-12.",
+    "temperature": "Independent 121-point log-grid minimization of full calibration NLL with equal episode weights. Development labels do not fit temperatures. Intervals condition on the calibration fits.",
+    "joins": "All calibration/development predictions match frozen IDs, class order, targets, groups and episodes. Inference temperature is 1.0; saved probabilities match logits within 1e-5. No input hash changed.",
+    "maximum_point_discrepancy": 1.5543122344752192e-15
+  },
+  "calibration": {
+    "full_equal_episode_temperatures": {
+      "parent": 4.0,
+      "candidate": 1.9999999999999998
+    },
+    "candidate_experiment_micro_temperature": 1.0717734625362931,
+    "parent_grid_upper_bound": 4.0,
+    "windows": 182,
+    "episodes": 96,
+    "groups": 32,
+    "calibration_metrics_are_in_sample": true
+  },
+  "panels": {
+    "calibration/full/all/all": {
+      "windows": 182,
+      "episodes": 96,
+      "groups": 32,
+      "episode_weight_sum": 96.0,
+      "raw": {
+        "parent": {
+          "acc": 0.3333333333333333,
+          "nll": 1.7039948885030407,
+          "brier": 0.9366734114222882,
+          "ece": 0.3836640270561081,
+          "nll_minus_uniform": 0.6053825998349309
+        },
+        "candidate": {
+          "acc": 0.3342013888888888,
+          "nll": 1.0988623728398248,
+          "brier": 0.6670543370614525,
+          "ece": 0.06554797243603248,
+          "nll_minus_uniform": 0.00025008417171501485
+        }
+      },
+      "calibrated": {
+        "parent": {
+          "acc": 0.3333333333333333,
+          "nll": 1.1516241769020006,
+          "brier": 0.7011284844520594,
+          "ece": 0.11261178491667191,
+          "nll_minus_uniform": 0.0530118882338908
+        },
+        "candidate": {
+          "acc": 0.3342013888888888,
+          "nll": 1.0942258102235733,
+          "brier": 0.6635708549658471,
+          "ece": 0.04216089485176933,
+          "nll_minus_uniform": -0.004386478444536435
+        }
+      }
+    },
+    "development/full/iid/all": {
+      "windows": 196,
+      "episodes": 96,
+      "groups": 32,
+      "episode_weight_sum": 96.0,
+      "raw": {
+        "parent": {
+          "acc": 0.3333333333333332,
+          "nll": 1.712452357373193,
+          "brier": 0.9303227524837444,
+          "ece": 0.40316209525167973,
+          "nll_minus_uniform": 0.6138400687050831
+        },
+        "candidate": {
+          "acc": 0.3402777777777777,
+          "nll": 1.114185953853556,
+          "brier": 0.6767587666407534,
+          "ece": 0.0658244924811593,
+          "nll_minus_uniform": 0.015573665185446117
+        },
+        "paired": {
+          "acc": {
+            "candidate_minus_parent": 0.006944444444444475,
+            "ci95": [
+              -5.551115123125783e-17,
+              0.01736111111111116
+            ],
+            "parent_ci95": [
+              0.3333333333333332,
+              0.3333333333333334
+            ],
+            "candidate_ci95": [
+              0.33333333333333326,
+              0.3506944444444444
+            ]
+          },
+          "nll": {
+            "candidate_minus_parent": -0.598266403519637,
+            "ci95": [
+              -0.6785650197211481,
+              -0.5198138791551934
+            ],
+            "parent_ci95": [
+              1.6411516919873783,
+              1.7821554572493037
+            ],
+            "candidate_ci95": [
+              1.0932970868965703,
+              1.1377853746292819
+            ]
+          },
+          "brier": {
+            "candidate_minus_parent": -0.253563985842991,
+            "ci95": [
+              -0.2882144767124644,
+              -0.2179524081333052
+            ],
+            "parent_ci95": [
+              0.9005605959813184,
+              0.9600325280394626
+            ],
+            "candidate_ci95": [
+              0.6623924005681694,
+              0.6929750831612851
+            ]
+          },
+          "ece": {
+            "candidate_minus_parent": -0.33733760277052044,
+            "ci95": [
+              -0.36700635899334366,
+              -0.30629448376287555
+            ],
+            "parent_ci95": [
+              0.3845718384896557,
+              0.4223828872836136
+            ],
+            "candidate_ci95": [
+              0.04670505012772756,
+              0.08779944049177658
+            ]
+          }
+        }
+      },
+      "calibrated": {
+        "parent": {
+          "acc": 0.3333333333333332,
+          "nll": 1.1395642135708883,
+          "brier": 0.6924498378688599,
+          "ece": 0.1225919918902803,
+          "nll_minus_uniform": 0.04095192490277855
+        },
+        "candidate": {
+          "acc": 0.3402777777777777,
+          "nll": 1.099795297211755,
+          "brier": 0.6673327111796757,
+          "ece": 0.029085900496782724,
+          "nll_minus_uniform": 0.0011830085436452809
+        },
+        "paired": {
+          "acc": {
+            "candidate_minus_parent": 0.006944444444444475,
+            "ci95": [
+              -5.551115123125783e-17,
+              0.01736111111111116
+            ],
+            "parent_ci95": [
+              0.3333333333333332,
+              0.3333333333333334
+            ],
+            "candidate_ci95": [
+              0.33333333333333326,
+              0.3506944444444444
+            ]
+          },
+          "nll": {
+            "candidate_minus_parent": -0.039768916359133266,
+            "ci95": [
+              -0.0541526606128827,
+              -0.02516884604006643
+            ],
+            "parent_ci95": [
+              1.1281309744714123,
+              1.150621803326349
+            ],
+            "candidate_ci95": [
+              1.0898406047820133,
+              1.1100987332218029
+            ]
+          },
+          "brier": {
+            "candidate_minus_parent": -0.0251171266891842,
+            "ci95": [
+              -0.03573663793465235,
+              -0.01491019414543692
+            ],
+            "parent_ci95": [
+              0.6844439570327296,
+              0.7007751716641649
+            ],
+            "candidate_ci95": [
+              0.6604923908009388,
+              0.6744589579107121
+            ]
+          },
+          "ece": {
+            "candidate_minus_parent": -0.09350609139349758,
+            "ci95": [
+              -0.11225927559402635,
+              -0.06097575023830562
+            ],
+            "parent_ci95": [
+              0.11518648560004421,
+              0.13576747842445488
+            ],
+            "candidate_ci95": [
+              0.016975309751749493,
+              0.060832712514700446
+            ]
+          }
+        }
+      }
+    },
+    "development/full/heldout_task/all": {
+      "windows": 279,
+      "episodes": 96,
+      "groups": 32,
+      "episode_weight_sum": 96.0,
+      "raw": {
+        "parent": {
+          "acc": 0.33593749999999994,
+          "nll": 1.7575686233451442,
+          "brier": 0.9328596715471842,
+          "ece": 0.38062690344193295,
+          "nll_minus_uniform": 0.6589563346770344
+        },
+        "candidate": {
+          "acc": 0.33854166666666646,
+          "nll": 1.1020826039546936,
+          "brier": 0.6685505178342491,
+          "ece": 0.06767651412584706,
+          "nll_minus_uniform": 0.003470315286583814
+        },
+        "paired": {
+          "acc": {
+            "candidate_minus_parent": 0.0026041666666665186,
+            "ci95": [
+              -0.012152777777777792,
+              0.01822916666666663
+            ],
+            "parent_ci95": [
+              0.3333333333333332,
+              0.3411458333333333
+            ],
+            "candidate_ci95": [
+              0.3246310763888887,
+              0.35329861111111116
+            ]
+          },
+          "nll": {
+            "candidate_minus_parent": -0.6554860193904506,
+            "ci95": [
+              -0.7191252619922539,
+              -0.5925014370566782
+            ],
+            "parent_ci95": [
+              1.6936048834769093,
+              1.8243095747652278
+            ],
+            "candidate_ci95": [
+              1.0909955635083946,
+              1.1148722338770993
+            ]
+          },
+          "brier": {
+            "candidate_minus_parent": -0.26430915371293506,
+            "ci95": [
+              -0.28718910807802966,
+              -0.23881191694116355
+            ],
+            "parent_ci95": [
+              0.908673057864882,
+              0.9557397760683716
+            ],
+            "candidate_ci95": [
+              0.6602464516032021,
+              0.6782263566624999
+            ]
+          },
+          "ece": {
+            "candidate_minus_parent": -0.3129503893160859,
+            "ci95": [
+              -0.34660047272419636,
+              -0.2808933130120531
+            ],
+            "parent_ci95": [
+              0.3655869447937234,
+              0.39906087104179055
+            ],
+            "candidate_ci95": [
+              0.04217660861029082,
+              0.09521924116463996
+            ]
+          }
+        }
+      },
+      "calibrated": {
+        "parent": {
+          "acc": 0.33593749999999994,
+          "nll": 1.1503900787443782,
+          "brier": 0.7000942462701811,
+          "ece": 0.11705101643165554,
+          "nll_minus_uniform": 0.05177779007626837
+        },
+        "candidate": {
+          "acc": 0.33854166666666646,
+          "nll": 1.0975078649634014,
+          "brier": 0.6657823404671227,
+          "ece": 0.03921704440404433,
+          "nll_minus_uniform": -0.0011044237047084238
+        },
+        "paired": {
+          "acc": {
+            "candidate_minus_parent": 0.0026041666666665186,
+            "ci95": [
+              -0.012152777777777792,
+              0.01822916666666663
+            ],
+            "parent_ci95": [
+              0.3333333333333332,
+              0.3411458333333333
+            ],
+            "candidate_ci95": [
+              0.3246310763888887,
+              0.35329861111111116
+            ]
+          },
+          "nll": {
+            "candidate_minus_parent": -0.052882213780976794,
+            "ci95": [
+              -0.062443593811576194,
+              -0.04190493793031237
+            ],
+            "parent_ci95": [
+              1.1410787694045739,
+              1.1593915727679056
+            ],
+            "candidate_ci95": [
+              1.091928362600967,
+              1.1035442157202628
+            ]
+          },
+          "brier": {
+            "candidate_minus_parent": -0.03431190580305843,
+            "ci95": [
+              -0.04143880164165239,
+              -0.026395950941262136
+            ],
+            "parent_ci95": [
+              0.6935373129262573,
+              0.706784911160115
+            ],
+            "candidate_ci95": [
+              0.6617785284376593,
+              0.6702643963140507
+            ]
+          },
+          "ece": {
+            "candidate_minus_parent": -0.0778339720276112,
+            "ci95": [
+              -0.11444951609054545,
+              -0.04209215226387989
+            ],
+            "parent_ci95": [
+              0.1088484189991902,
+              0.13090558733426141
+            ],
+            "candidate_ci95": [
+              0.010719860472098339,
+              0.07335949517961426
+            ]
+          }
+        }
+      }
+    },
+    "development/full/all/all": {
+      "windows": 475,
+      "episodes": 192,
+      "groups": 64,
+      "episode_weight_sum": 192.0,
+      "raw": {
+        "parent": {
+          "acc": 0.3346354166666669,
+          "nll": 1.73501049035917,
+          "brier": 0.9315912120154644,
+          "ece": 0.39189449934680637,
+          "nll_minus_uniform": 0.6363982016910603
+        },
+        "candidate": {
+          "acc": 0.3394097222222226,
+          "nll": 1.1081342789041249,
+          "brier": 0.6726546422375007,
+          "ece": 0.056655958235590886,
+          "nll_minus_uniform": 0.009521990236015077
+        },
+        "paired": {
+          "acc": {
+            "candidate_minus_parent": 0.004774305555555691,
+            "ci95": [
+              -0.003917100694444441,
+              0.013888888888888895
+            ],
+            "parent_ci95": [
+              0.3333333333333333,
+              0.3372395833333335
+            ],
+            "candidate_ci95": [
+              0.3311631944444446,
+              0.34810112847222235
+            ]
+          },
+          "nll": {
+            "candidate_minus_parent": -0.6268762114550452,
+            "ci95": [
+              -0.6790899859377325,
+              -0.574909379457072
+            ],
+            "parent_ci95": [
+              1.6811977782718917,
+              1.78661079285623
+            ],
+            "candidate_ci95": [
+              1.0961882249343542,
+              1.121428990046415
+            ]
+          },
+          "brier": {
+            "candidate_minus_parent": -0.25893656977796364,
+            "ci95": [
+              -0.28024348498046536,
+              -0.23624394979708538
+            ],
+            "parent_ci95": [
+              0.9092211056538773,
+              0.9509438619550089
+            ],
+            "candidate_ci95": [
+              0.664610827238441,
+              0.6815656878154811
+            ]
+          },
+          "ece": {
+            "candidate_minus_parent": -0.3352385411112155,
+            "ci95": [
+              -0.3512828651139513,
+              -0.31080435666201395
+            ],
+            "parent_ci95": [
+              0.37815344075238194,
+              0.40561665467498564
+            ],
+            "candidate_ci95": [
+              0.04569651446212631,
+              0.07531915688931974
+            ]
+          }
+        }
+      },
+      "calibrated": {
+        "parent": {
+          "acc": 0.3346354166666669,
+          "nll": 1.1449771461576332,
+          "brier": 0.6962720420695208,
+          "ece": 0.11982150416096789,
+          "nll_minus_uniform": 0.04636485748952346
+        },
+        "candidate": {
+          "acc": 0.3394097222222226,
+          "nll": 1.0986515810875772,
+          "brier": 0.6665575258233997,
+          "ece": 0.024607196074886876,
+          "nll_minus_uniform": 3.929241946742934e-05
+        },
+        "paired": {
+          "acc": {
+            "candidate_minus_parent": 0.004774305555555691,
+            "ci95": [
+              -0.003917100694444441,
+              0.013888888888888895
+            ],
+            "parent_ci95": [
+              0.3333333333333333,
+              0.3372395833333335
+            ],
+            "candidate_ci95": [
+              0.3311631944444446,
+              0.34810112847222235
+            ]
+          },
+          "nll": {
+            "candidate_minus_parent": -0.04632556507005603,
+            "ci95": [
+              -0.05491666451541059,
+              -0.037191163770480214
+            ],
+            "parent_ci95": [
+              1.1365041969928558,
+              1.1526221927426759
+            ],
+            "candidate_ci95": [
+              1.0931171370730461,
+              1.1043579084594946
+            ]
+          },
+          "brier": {
+            "candidate_minus_parent": -0.029714516246121092,
+            "ci95": [
+              -0.036132595842729695,
+              -0.02284046811142687
+            ],
+            "parent_ci95": [
+              0.6901428934681393,
+              0.701587223080451
+            ],
+            "candidate_ci95": [
+              0.6627210641166761,
+              0.6705910357128937
+            ]
+          },
+          "ece": {
+            "candidate_minus_parent": -0.09521430808608101,
+            "ci95": [
+              -0.10827761667473262,
+              -0.0701737586224902
+            ],
+            "parent_ci95": [
+              0.11338467129999455,
+              0.12711460228131327
+            ],
+            "candidate_ci95": [
+              0.015289219037466687,
+              0.04742637735158731
+            ]
+          }
+        }
+      }
+    }
+  },
+  "origin_confusion": {
+    "calibration/full/all/all": {
+      "parent": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            0.0,
+            31.999999999999993,
+            0.0
+          ],
+          [
+            0.0,
+            31.999999999999996,
+            0.0
+          ],
+          [
+            0.0,
+            32.0,
+            0.0
+          ]
+        ],
+        "recall": {
+          "human": 0.0,
+          "script": 1.0,
+          "agent": 0.0
+        },
+        "prediction_share": {
+          "human": 0.0,
+          "script": 0.9999999999999999,
+          "agent": 0.0
+        },
+        "origin_episode_support": {
+          "human": 31.999999999999993,
+          "script": 31.999999999999996,
+          "agent": 32.0
+        },
+        "window_prediction_counts": {
+          "human": 0,
+          "script": 182,
+          "agent": 0
+        }
+      },
+      "candidate": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            31.749999999999993,
+            0.25,
+            0.0
+          ],
+          [
+            31.66666666666666,
+            0.3333333333333333,
+            0.0
+          ],
+          [
+            31.75,
+            0.25,
+            0.0
+          ]
+        ],
+        "recall": {
+          "human": 0.9921875,
+          "script": 0.010416666666666668,
+          "agent": 0.0
+        },
+        "prediction_share": {
+          "human": 0.9913194444444443,
+          "script": 0.008680555555555554,
+          "agent": 0.0
+        },
+        "origin_episode_support": {
+          "human": 31.999999999999993,
+          "script": 31.999999999999993,
+          "agent": 32.0
+        },
+        "window_prediction_counts": {
+          "human": 179,
+          "script": 3,
+          "agent": 0
+        }
+      }
+    },
+    "development/full/iid/all": {
+      "parent": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            0.0,
+            31.99999999999999,
+            0.0
+          ],
+          [
+            0.0,
+            31.99999999999999,
+            0.0
+          ],
+          [
+            0.0,
+            31.99999999999999,
+            0.0
+          ]
+        ],
+        "recall": {
+          "human": 0.0,
+          "script": 1.0,
+          "agent": 0.0
+        },
+        "prediction_share": {
+          "human": 0.0,
+          "script": 1.0,
+          "agent": 0.0
+        },
+        "origin_episode_support": {
+          "human": 31.99999999999999,
+          "script": 31.99999999999999,
+          "agent": 31.99999999999999
+        },
+        "window_prediction_counts": {
+          "human": 0,
+          "script": 196,
+          "agent": 0
+        }
+      },
+      "candidate": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            31.666666666666657,
+            0.3333333333333333,
+            0.0
+          ],
+          [
+            31.333333333333325,
+            0.6666666666666666,
+            0.0
+          ],
+          [
+            30.166666666666657,
+            1.4999999999999998,
+            0.3333333333333333
+          ]
+        ],
+        "recall": {
+          "human": 0.9895833333333334,
+          "script": 0.020833333333333336,
+          "agent": 0.01041666666666667
+        },
+        "prediction_share": {
+          "human": 0.9704861111111113,
+          "script": 0.026041666666666678,
+          "agent": 0.0034722222222222238
+        },
+        "origin_episode_support": {
+          "human": 31.99999999999999,
+          "script": 31.999999999999993,
+          "agent": 31.99999999999999
+        },
+        "window_prediction_counts": {
+          "human": 188,
+          "script": 7,
+          "agent": 1
+        }
+      }
+    },
+    "development/full/heldout_task/all": {
+      "parent": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            0.0,
+            31.99999999999998,
+            0.0
+          ],
+          [
+            0.0,
+            31.99999999999999,
+            0.0
+          ],
+          [
+            0.0,
+            31.74999999999998,
+            0.25
+          ]
+        ],
+        "recall": {
+          "human": 0.0,
+          "script": 1.0,
+          "agent": 0.007812500000000005
+        },
+        "prediction_share": {
+          "human": 0.0,
+          "script": 0.9973958333333334,
+          "agent": 0.0026041666666666683
+        },
+        "origin_episode_support": {
+          "human": 31.99999999999998,
+          "script": 31.99999999999999,
+          "agent": 31.99999999999998
+        },
+        "window_prediction_counts": {
+          "human": 0,
+          "script": 278,
+          "agent": 1
+        }
+      },
+      "candidate": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            31.416666666666647,
+            0.5833333333333333,
+            0.0
+          ],
+          [
+            30.916666666666654,
+            1.0833333333333333,
+            0.0
+          ],
+          [
+            31.216666666666647,
+            0.7833333333333333,
+            0.0
+          ]
+        ],
+        "recall": {
+          "human": 0.9817708333333334,
+          "script": 0.03385416666666668,
+          "agent": 0.0
+        },
+        "prediction_share": {
+          "human": 0.9744791666666668,
+          "script": 0.025520833333333347,
+          "agent": 0.0
+        },
+        "origin_episode_support": {
+          "human": 31.99999999999998,
+          "script": 31.999999999999986,
+          "agent": 31.999999999999982
+        },
+        "window_prediction_counts": {
+          "human": 270,
+          "script": 9,
+          "agent": 0
+        }
+      }
+    },
+    "development/full/all/all": {
+      "parent": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            0.0,
+            64.00000000000006,
+            0.0
+          ],
+          [
+            0.0,
+            64.00000000000004,
+            0.0
+          ],
+          [
+            0.0,
+            63.75000000000006,
+            0.25
+          ]
+        ],
+        "recall": {
+          "human": 0.0,
+          "script": 1.0,
+          "agent": 0.0039062499999999965
+        },
+        "prediction_share": {
+          "human": 0.0,
+          "script": 0.9986979166666666,
+          "agent": 0.0013020833333333322
+        },
+        "origin_episode_support": {
+          "human": 64.00000000000006,
+          "script": 64.00000000000004,
+          "agent": 64.00000000000006
+        },
+        "window_prediction_counts": {
+          "human": 0,
+          "script": 474,
+          "agent": 1
+        }
+      },
+      "candidate": {
+        "labels": [
+          "human",
+          "script",
+          "agent"
+        ],
+        "rows": "true origin",
+        "columns": "predicted origin",
+        "confusion_episode_mass": [
+          [
+            63.0833333333334,
+            0.9166666666666665,
+            0.0
+          ],
+          [
+            62.25000000000004,
+            1.7499999999999998,
+            0.0
+          ],
+          [
+            61.38333333333338,
+            2.283333333333333,
+            0.3333333333333333
+          ]
+        ],
+        "recall": {
+          "human": 0.9856770833333333,
+          "script": 0.02734374999999998,
+          "agent": 0.00520833333333333
+        },
+        "prediction_share": {
+          "human": 0.9724826388888888,
+          "script": 0.025781249999999974,
+          "agent": 0.0017361111111111095
+        },
+        "origin_episode_support": {
+          "human": 64.00000000000007,
+          "script": 64.00000000000004,
+          "agent": 64.00000000000004
+        },
+        "window_prediction_counts": {
+          "human": 458,
+          "script": 16,
+          "agent": 1
+        }
+      }
+    }
+  },
+  "training": {
+    "records": 2334,
+    "training_seconds": 371.20887875556946,
+    "trial_wall_seconds": 610.9060554900001,
+    "optimizer_steps": 292,
+    "transfer_evaluated": false,
+    "mechanism_and_coverage_gates_passed": true,
+    "promotable": false
+  },
+  "interpretation": [
+    "The candidate predicts human on about 97% of equal-episode-weighted full development windows. Per-class recall is a necessary release limitation.",
+    "IID candidate agent recall is 1.0417%; shifted agent recall is zero. Temperature does not alter these argmax decisions.",
+    "Uniform three-class NLL is ln(3)=1.0986122886681098. Raw candidate NLL is worse than uniform on IID and shifted development; calibrated NLL is very close to uniform. Low ECE does not establish useful discrimination.",
+    "Each development regime contains 32 latent groups with three matched origin episodes per group. Intervals describe this authored-policy population, not real human or LLM traffic. The shift changes tasks and policy priors within one controller implementation.",
+    "No fresh transfer evaluation was run. Older mixed-replay retention results cannot be assigned to this checkpoint."
+  ]
+}

head.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc1f550c7914388af77e607c70fe8801558a89a3847d37ea37d3923b28521c3e
+size 5256511

provenance.json

@@ -0,0 +1,81 @@
+{
+  "config": {
+    "epochs": 1,
+    "seed": 4,
+    "lr": 1e-05,
+    "lora": 16,
+    "accum": 2,
+    "batch": 4,
+    "perm_kl": 0,
+    "perm_frac": 0.3,
+    "ord_w": 0.0,
+    "p_none": 0,
+    "p_none_distract": 0,
+    "p_distract": 0,
+    "p_none_pair": 0,
+    "synthetic_repeat": 1,
+    "public_frac": 1,
+    "head_lr": 0.0,
+    "weight_decay": 0.01,
+    "anchor_w": 0.0,
+    "label_smoothing": 0.0,
+    "brier_w": 0.0,
+    "focal_gamma": 0.0,
+    "base": "Qwen/Qwen3.5-4B-Base",
+    "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
+    "init_from": "jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1",
+    "lora_targets": "all",
+    "head_dim": 256,
+    "weights_dtype": "fp32",
+    "option_isolation": 0,
+    "special_embeddings": 0,
+    "dtype": "bf16",
+    "checkpointing": 1
+  },
+  "config_sha256": "1abcf6795472953561dc92bf54d59e9f10fd5e81d08d558f44197e56299077af",
+  "suite_sha256": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+  "source_hashes": {
+    "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
+    "kev/api.py": "7bffacfb762c626b8bc2f670f350295af5ccb0883dbe90239c7d2f8e5ef56582",
+    "kev/autoresearch.py": "f7d7fabb9c0df7065bee3fec4aa4bee7028c5d1565847aefe2d74e3d7d40ed8a",
+    "kev/benchmark.py": "8d486659a9036102386b077fcbd197fc0d85ae7752a4711666d6b3a95622b26f",
+    "kev/calibrate.py": "c1eff4744fabd349e8abca86777a7aa0cbea44904c8223d3b61fca3d43741519",
+    "kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59",
+    "kev/compare.py": "3606dbf98bb305d66420158498cb837304d2edb8dab01cdaf2ec42964bef3435",
+    "kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536",
+    "kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf",
+    "kev/data.py": "a5a5e200f56b60e4bf30a195a4cc87a74afddae15999ec48c7543a4f9b59f71c",
+    "kev/device.py": "d1677fd98ec0979c7284546306e34e0d09298ca4042fc39eb2b32a74f4e975c4",
+    "kev/evaluate.py": "999264f2837dcfbf2ec93601aa4745e701698ab850a43ad89324672a6604965c",
+    "kev/experiment.py": "1a5ca47d2b3d236e53c87e8c4197704f162c6692aa008521d31ea491945693f0",
+    "kev/jev.py": "acd4cc3f1844e438cc83a8d409c15ef78a5ef64d39ea10f583d75a6c7646b243",
+    "kev/metrics.py": "db0930d470c0ce16f16dcaf726d49a7a103973a5296cefc5259d1e68707d882b",
+    "kev/mlx_model.py": "f582428796faf6bf962b772a69a6227ac3259ccbd3215c0c8b55ce28dbbd0909",
+    "kev/model.py": "eac0f56c807b85b5c4077483d922e6bb03f07890de2835efd9372cf20cd01b8e",
+    "kev/plot.py": "0874bcff2885d8155a1cceade0de8a275a7163d3c3fd6aa0c294e8ccf5ff7e02",
+    "kev/predictors.py": "b2a66dd9f0f0a8026bc94603383bacfdd622e75d1c935173997175606626a9b3",
+    "kev/publish.py": "8abc9bcf4a01365697b05cf3c5ad0013462bd92d005f5616954e40ab1100c7aa",
+    "kev/serve.py": "c210b1f4598e64b49c399f73603ed372f4314c165ceda50e1340c0466ec1c735",
+    "kev/study_v3.py": "7891150aa623e4479c7c789d4dc4662f186132b37b7bac1e1f072b9e08ee13e3",
+    "kev/suite.py": "c5e5e27f2fed871b547a021d180b0a7f4a0542a488cce4ee02ad936f433441db",
+    "kev/train.py": "a2d85632d73760fb3efe51627ff4fde6af3100ed1c34e75d290c1b00c9899aa9",
+    "kev/transfer_v9.py": "0902409742151250a28af1fd8f42258b70fcc161f3deed3ee35fe3f473f3c763",
+    "modal_app.py": "5d55aa2a454aba2ba6e4a392ed77ab2bc849e9823b8298a41ce1b8fe4217344b",
+    "pyproject.toml": "7c17fbe9efcadda3eb488b59adf1b6db5938cd33af4d3bd389ce4359b73fcc7d",
+    "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
+  },
+  "git_commit": "d8b1413f428a9a8f4e3d6091acf3d9ed8445c864",
+  "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
+  "torch": "2.8.0+cu128",
+  "device": "cuda",
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "legacy_checkpoint": false,
+  "measured_checkpoint": {
+    "requested": "/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint",
+    "resolved": "/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint",
+    "head_sha256": "af6bfd215bf0b7eff92612a83b21cb56754ca87f0bf9f62fdfcfbef275d9eab6",
+    "adapter_sha256": "7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c",
+    "inference_temperature": 1.0
+  }
+}

result.json

@@ -0,0 +1,485 @@
+{
+  "objective": -1.1040644052465771,
+  "paired_flip": null,
+  "unknowable": null,
+  "clean": {
+    "n": 475,
+    "nll": 1.1040644052465771,
+    "acc": 0.3452631578947368,
+    "ece": 0.057981423538388685,
+    "brier": 0.6695330728421766,
+    "mean_conf": 0.3934063210268794,
+    "confident_error_rate": 0.0,
+    "coverage_at_0_9": 0.0,
+    "accuracy_at_0_9": null,
+    "coverage_at_5pct_error": 0.0,
+    "coverage_at_1pct_error": 0.0,
+    "aurc": 0.6226976418791688,
+    "error_rate_at_0_9": null,
+    "confidence_bias": 0.048143163132142563,
+    "top_bins": {
+      "0.9": {
+        "n": 0,
+        "errors": 0,
+        "error_rate": null
+      },
+      "0.95": {
+        "n": 0,
+        "errors": 0,
+        "error_rate": null
+      },
+      "0.99": {
+        "n": 0,
+        "errors": 0,
+        "error_rate": null
+      }
+    },
+    "selective": {
+      "0.5": {
+        "coverage": 0.5010526315789474,
+        "accuracy": 0.35294117647058826,
+        "confidence_cutoff": 0.3749043643474579
+      },
+      "0.8": {
+        "coverage": 0.8,
+        "accuracy": 0.3447368421052632,
+        "confidence_cutoff": 0.3571439278976939
+      }
+    }
+  },
+  "tasks": {
+    "sev_behavioral_origin": {
+      "n": 475,
+      "nll": 1.1040644052465771,
+      "acc": 0.3452631578947368,
+      "ece": 0.057981423538388685,
+      "brier": 0.6695330728421766,
+      "mean_conf": 0.3934063210268794,
+      "confident_error_rate": 0.0,
+      "coverage_at_0_9": 0.0,
+      "accuracy_at_0_9": null,
+      "coverage_at_5pct_error": 0.0,
+      "coverage_at_1pct_error": 0.0,
+      "aurc": 0.6226976418791688,
+      "error_rate_at_0_9": null,
+      "confidence_bias": 0.048143163132142563,
+      "top_bins": {
+        "0.9": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.95": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.99": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        }
+      },
+      "selective": {
+        "0.5": {
+          "coverage": 0.5010526315789474,
+          "accuracy": 0.35294117647058826,
+          "confidence_cutoff": 0.3749043643474579
+        },
+        "0.8": {
+          "coverage": 0.8,
+          "accuracy": 0.3447368421052632,
+          "confidence_cutoff": 0.3571439278976939
+        }
+      }
+    }
+  },
+  "variants": {
+    "clean": {
+      "n": 475,
+      "nll": 1.1040644052465771,
+      "acc": 0.3452631578947368,
+      "ece": 0.057981423538388685,
+      "brier": 0.6695330728421766,
+      "mean_conf": 0.3934063210268794,
+      "confident_error_rate": 0.0,
+      "coverage_at_0_9": 0.0,
+      "accuracy_at_0_9": null,
+      "coverage_at_5pct_error": 0.0,
+      "coverage_at_1pct_error": 0.0,
+      "aurc": 0.6226976418791688,
+      "error_rate_at_0_9": null,
+      "confidence_bias": 0.048143163132142563,
+      "top_bins": {
+        "0.9": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.95": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.99": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        }
+      },
+      "selective": {
+        "0.5": {
+          "coverage": 0.5010526315789474,
+          "accuracy": 0.35294117647058826,
+          "confidence_cutoff": 0.3749043643474579
+        },
+        "0.8": {
+          "coverage": 0.8,
+          "accuracy": 0.3447368421052632,
+          "confidence_cutoff": 0.3571439278976939
+        }
+      }
+    },
+    "prefix_loss": {
+      "n": 390,
+      "nll": 1.1076993167917513,
+      "acc": 0.3333333333333333,
+      "ece": 0.06600365992222378,
+      "brier": 0.673099724532684,
+      "mean_conf": 0.39766209170416683,
+      "confident_error_rate": 0.0,
+      "coverage_at_0_9": 0.0,
+      "accuracy_at_0_9": null,
+      "coverage_at_5pct_error": 0.002564102564102564,
+      "coverage_at_1pct_error": 0.002564102564102564,
+      "aurc": 0.6469893243952387,
+      "error_rate_at_0_9": null,
+      "confidence_bias": 0.06432875837083352,
+      "top_bins": {
+        "0.9": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.95": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.99": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        }
+      },
+      "selective": {
+        "0.5": {
+          "coverage": 0.5,
+          "accuracy": 0.3384615384615385,
+          "confidence_cutoff": 0.37823529178605353
+        },
+        "0.8": {
+          "coverage": 0.8,
+          "accuracy": 0.32371794871794873,
+          "confidence_cutoff": 0.36099090096381603
+        }
+      }
+    },
+    "sampled_loss": {
+      "n": 400,
+      "nll": 1.1039164529781138,
+      "acc": 0.3375,
+      "ece": 0.058967762381204686,
+      "brier": 0.6698679501799927,
+      "mean_conf": 0.38919767867980576,
+      "confident_error_rate": 0.0,
+      "coverage_at_0_9": 0.0,
+      "accuracy_at_0_9": null,
+      "coverage_at_5pct_error": 0.0,
+      "coverage_at_1pct_error": 0.0,
+      "aurc": 0.6417827102105147,
+      "error_rate_at_0_9": null,
+      "confidence_bias": 0.051697678679805736,
+      "top_bins": {
+        "0.9": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.95": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.99": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        }
+      },
+      "selective": {
+        "0.5": {
+          "coverage": 0.5,
+          "accuracy": 0.34,
+          "confidence_cutoff": 0.3756981715976102
+        },
+        "0.8": {
+          "coverage": 0.8,
+          "accuracy": 0.328125,
+          "confidence_cutoff": 0.3595526027546109
+        }
+      }
+    }
+  },
+  "heldout_tasks": {
+    "sev_behavioral_origin": {
+      "n": 279,
+      "nll": 1.1054832638596932,
+      "acc": 0.33691756272401435,
+      "ece": 0.07077264797432041,
+      "brier": 0.6704644833339944,
+      "mean_conf": 0.3860012027442918,
+      "confident_error_rate": 0.0,
+      "coverage_at_0_9": 0.0,
+      "accuracy_at_0_9": null,
+      "coverage_at_5pct_error": 0.0,
+      "coverage_at_1pct_error": 0.0,
+      "aurc": 0.6256505380956604,
+      "error_rate_at_0_9": null,
+      "confidence_bias": 0.04908364002027743,
+      "top_bins": {
+        "0.9": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.95": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        },
+        "0.99": {
+          "n": 0,
+          "errors": 0,
+          "error_rate": null
+        }
+      },
+      "selective": {
+        "0.5": {
+          "coverage": 0.5017921146953405,
+          "accuracy": 0.35,
+          "confidence_cutoff": 0.37531410529273435
+        },
+        "0.8": {
+          "coverage": 0.8028673835125448,
+          "accuracy": 0.33482142857142855,
+          "confidence_cutoff": 0.35978143335040114
+        }
+      }
+    }
+  },
+  "permutation": {
+    "n": 0,
+    "mean_max_delta": null,
+    "flip_rate": null
+  },
+  "temperature": 1.0717734625362931,
+  "calibrated_clean": {
+    "n": 475,
+    "nll": 1.1026041370540602,
+    "acc": 0.3452631578947368,
+    "ece": 0.053275370738267415,
+    "brier": 0.6686530034298412,
+    "mean_conf": 0.38936850355603847,
+    "confident_error_rate": 0.0,
+    "coverage_at_0_9": 0.0,
+    "accuracy_at_0_9": null,
+    "coverage_at_5pct_error": 0.0,
+    "coverage_at_1pct_error": 0.0,
+    "aurc": 0.6227258857225079,
+    "error_rate_at_0_9": null,
+    "confidence_bias": 0.04410534566130164,
+    "top_bins": {
+      "0.9": {
+        "n": 0,
+        "errors": 0,
+        "error_rate": null
+      },
+      "0.95": {
+        "n": 0,
+        "errors": 0,
+        "error_rate": null
+      },
+      "0.99": {
+        "n": 0,
+        "errors": 0,
+        "error_rate": null
+      }
+    },
+    "selective": {
+      "0.5": {
+        "coverage": 0.5010526315789474,
+        "accuracy": 0.35294117647058826,
+        "confidence_cutoff": 0.37205812028797897
+      },
+      "0.8": {
+        "coverage": 0.8,
+        "accuracy": 0.3447368421052632,
+        "confidence_cutoff": 0.3555337941720841
+      }
+    }
+  },
+  "metric_policy": {
+    "version": 2,
+    "selective_ties": "whole_confidence_groups",
+    "coverage_at_error": "in-sample maximum over confidence thresholds; not a deployed error guarantee",
+    "aurc": "right-step integral over whole confidence groups",
+    "confident_error_rate": "high-confidence errors divided by all questions",
+    "error_rate_at_0_9": "errors divided by questions accepted at p_max >= 0.9",
+    "nll": "exact from logits when recorded; otherwise from floored probabilities",
+    "nll_floor": 1e-09,
+    "renormalize_returned_probabilities": true,
+    "raw_sums_outside_1e_5": 0,
+    "returned_zeros": 0
+  },
+  "coverage": {
+    "requested_records": 1265,
+    "requested_questions": 1265,
+    "evaluated_records": 1265,
+    "evaluated_questions": 1265,
+    "rejected_records": 0,
+    "truncated_records": 0
+  },
+  "latency_ms": {
+    "median": 78.70266199995513,
+    "p95": 87.40347419995942
+  },
+  "calibration": {
+    "inference_temperature": 1.0,
+    "additional_temperature": 1.0717734625362931,
+    "logits_recorded": true
+  },
+  "transfer": null,
+  "provenance": {
+    "config": {
+      "epochs": 1,
+      "seed": 4,
+      "lr": 1e-05,
+      "lora": 16,
+      "accum": 2,
+      "batch": 4,
+      "perm_kl": 0,
+      "perm_frac": 0.3,
+      "ord_w": 0.0,
+      "p_none": 0,
+      "p_none_distract": 0,
+      "p_distract": 0,
+      "p_none_pair": 0,
+      "synthetic_repeat": 1,
+      "public_frac": 1,
+      "head_lr": 0.0,
+      "weight_decay": 0.01,
+      "anchor_w": 0.0,
+      "label_smoothing": 0.0,
+      "brier_w": 0.0,
+      "focal_gamma": 0.0,
+      "base": "Qwen/Qwen3.5-4B-Base",
+      "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
+      "init_from": "jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1",
+      "lora_targets": "all",
+      "head_dim": 256,
+      "weights_dtype": "fp32",
+      "option_isolation": 0,
+      "special_embeddings": 0,
+      "dtype": "bf16",
+      "checkpointing": 1
+    },
+    "config_sha256": "1abcf6795472953561dc92bf54d59e9f10fd5e81d08d558f44197e56299077af",
+    "suite_sha256": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+    "source_hashes": {
+      "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
+      "kev/api.py": "7bffacfb762c626b8bc2f670f350295af5ccb0883dbe90239c7d2f8e5ef56582",
+      "kev/autoresearch.py": "f7d7fabb9c0df7065bee3fec4aa4bee7028c5d1565847aefe2d74e3d7d40ed8a",
+      "kev/benchmark.py": "8d486659a9036102386b077fcbd197fc0d85ae7752a4711666d6b3a95622b26f",
+      "kev/calibrate.py": "c1eff4744fabd349e8abca86777a7aa0cbea44904c8223d3b61fca3d43741519",
+      "kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59",
+      "kev/compare.py": "3606dbf98bb305d66420158498cb837304d2edb8dab01cdaf2ec42964bef3435",
+      "kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536",
+      "kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf",
+      "kev/data.py": "a5a5e200f56b60e4bf30a195a4cc87a74afddae15999ec48c7543a4f9b59f71c",
+      "kev/device.py": "d1677fd98ec0979c7284546306e34e0d09298ca4042fc39eb2b32a74f4e975c4",
+      "kev/evaluate.py": "999264f2837dcfbf2ec93601aa4745e701698ab850a43ad89324672a6604965c",
+      "kev/experiment.py": "1a5ca47d2b3d236e53c87e8c4197704f162c6692aa008521d31ea491945693f0",
+      "kev/jev.py": "acd4cc3f1844e438cc83a8d409c15ef78a5ef64d39ea10f583d75a6c7646b243",
+      "kev/metrics.py": "db0930d470c0ce16f16dcaf726d49a7a103973a5296cefc5259d1e68707d882b",
+      "kev/mlx_model.py": "f582428796faf6bf962b772a69a6227ac3259ccbd3215c0c8b55ce28dbbd0909",
+      "kev/model.py": "eac0f56c807b85b5c4077483d922e6bb03f07890de2835efd9372cf20cd01b8e",
+      "kev/plot.py": "0874bcff2885d8155a1cceade0de8a275a7163d3c3fd6aa0c294e8ccf5ff7e02",
+      "kev/predictors.py": "b2a66dd9f0f0a8026bc94603383bacfdd622e75d1c935173997175606626a9b3",
+      "kev/publish.py": "8abc9bcf4a01365697b05cf3c5ad0013462bd92d005f5616954e40ab1100c7aa",
+      "kev/serve.py": "c210b1f4598e64b49c399f73603ed372f4314c165ceda50e1340c0466ec1c735",
+      "kev/study_v3.py": "7891150aa623e4479c7c789d4dc4662f186132b37b7bac1e1f072b9e08ee13e3",
+      "kev/suite.py": "c5e5e27f2fed871b547a021d180b0a7f4a0542a488cce4ee02ad936f433441db",
+      "kev/train.py": "a2d85632d73760fb3efe51627ff4fde6af3100ed1c34e75d290c1b00c9899aa9",
+      "kev/transfer_v9.py": "0902409742151250a28af1fd8f42258b70fcc161f3deed3ee35fe3f473f3c763",
+      "modal_app.py": "5d55aa2a454aba2ba6e4a392ed77ab2bc849e9823b8298a41ce1b8fe4217344b",
+      "pyproject.toml": "7c17fbe9efcadda3eb488b59adf1b6db5938cd33af4d3bd389ce4359b73fcc7d",
+      "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
+    },
+    "git_commit": "d8b1413f428a9a8f4e3d6091acf3d9ed8445c864",
+    "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
+    "torch": "2.8.0+cu128",
+    "device": "cuda",
+    "gpu": "NVIDIA H100 80GB HBM3",
+    "legacy_checkpoint": false,
+    "measured_checkpoint": {
+      "requested": "/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint",
+      "resolved": "/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint",
+      "head_sha256": "af6bfd215bf0b7eff92612a83b21cb56754ca87f0bf9f62fdfcfbef275d9eab6",
+      "adapter_sha256": "7fc73ad5a9e219731fba6bd926ba4a683b0c67dd676810d18e1a6018fa3b887c",
+      "inference_temperature": 1.0
+    }
+  },
+  "mechanism_checks": {
+    "n": 8,
+    "packed_max_delta": 0.0,
+    "sibling_max_delta": 6.73830509185791e-05,
+    "tolerance": 0.001,
+    "passed": true
+  },
+  "gates": {
+    "passed": true,
+    "checks": {
+      "complete_coverage": true,
+      "isolation_and_packing": true
+    },
+    "policy": "Research screening only: 0.001 isolation; 5pp task regression; 70% heldout pair correctness; <=10% confident errors; no automatic release."
+  },
+  "calibration_fit": {
+    "temperature": 1.0717734625362931,
+    "aggregation": "micro",
+    "objective": "raw-logit NLL",
+    "split": "calibration",
+    "rows_sha256": "00909b4d6b838c534d8d403d84a578fea1ddaac5e30a2b488c2d1c138c0d05ab",
+    "suite_sha256": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+    "n": 182
+  },
+  "wall_seconds": 610.9060554900001,
+  "promotable": false,
+  "test_evaluated": false,
+  "training_resources": {
+    "wall_seconds": 371.20887875556946,
+    "records_seen": 2334,
+    "requested_records": 2334,
+    "truncated_records": 0,
+    "rejected_records": 0,
+    "optimizer_steps": 292,
+    "forward_tokens": 738327,
+    "peak_device_bytes": 18913074688,
+    "device": "cuda",
+    "dtype": "bf16",
+    "batch": 4,
+    "peak_rss_bytes": 31561670656
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

tokenizer_config.json

@@ -0,0 +1,32 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

train.log

@@ -0,0 +1,35 @@
+Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
+delta: warm start from /__modal/volumes/vo-RociVuGycNGtzvUcdOqecz/hub/models--jaredpalmer--kev-4b/snapshots/485ace8703592fcf405488b262449990824cfed1: 496 adapter tensors and the pointer head loaded
+device=cuda trainable params=33.8M
+2334 training requests (holdout=['sev_behavioral_retry_export', 'sev_behavioral_stale_handoff', 'sev_behavioral_reconcile_export']), questions by type {'choice': 2280, 'score': 36, 'noul': 18}
+[transformers] `causal_conv1d_fn` is falling back to its reference PyTorch implementation because `causal_conv1d` is not installed. This is correct but much slower; install `causal_conv1d` for the optimized kernel.
+ep0 step 10/292 loss 1.829 kl 0.000 anchor 0.000 1.503s/rec
+ep0 step 20/292 loss 1.201 kl 0.000 anchor 0.000 0.808s/rec
+ep0 step 30/292 loss 0.990 kl 0.000 anchor 0.000 0.576s/rec
+ep0 step 40/292 loss 0.972 kl 0.000 anchor 0.000 0.461s/rec
+ep0 step 50/292 loss 1.026 kl 0.000 anchor 0.000 0.391s/rec
+ep0 step 60/292 loss 1.032 kl 0.000 anchor 0.000 0.345s/rec
+ep0 step 70/292 loss 1.014 kl 0.000 anchor 0.000 0.311s/rec
+ep0 step 80/292 loss 1.008 kl 0.000 anchor 0.000 0.286s/rec
+ep0 step 90/292 loss 1.031 kl 0.000 anchor 0.000 0.266s/rec
+ep0 step 100/292 loss 0.925 kl 0.000 anchor 0.000 0.250s/rec
+ep0 step 110/292 loss 1.073 kl 0.000 anchor 0.000 0.238s/rec
+ep0 step 120/292 loss 0.969 kl 0.000 anchor 0.000 0.227s/rec
+ep0 step 130/292 loss 1.000 kl 0.000 anchor 0.000 0.218s/rec
+ep0 step 140/292 loss 1.017 kl 0.000 anchor 0.000 0.210s/rec
+ep0 step 150/292 loss 1.058 kl 0.000 anchor 0.000 0.204s/rec
+ep0 step 160/292 loss 0.966 kl 0.000 anchor 0.000 0.198s/rec
+ep0 step 170/292 loss 0.980 kl 0.000 anchor 0.000 0.193s/rec
+ep0 step 180/292 loss 0.891 kl 0.000 anchor 0.000 0.188s/rec
+ep0 step 190/292 loss 1.000 kl 0.000 anchor 0.000 0.185s/rec
+ep0 step 200/292 loss 1.057 kl 0.000 anchor 0.000 0.181s/rec
+ep0 step 210/292 loss 0.975 kl 0.000 anchor 0.000 0.178s/rec
+ep0 step 220/292 loss 0.980 kl 0.000 anchor 0.000 0.175s/rec
+ep0 step 230/292 loss 0.944 kl 0.000 anchor 0.000 0.172s/rec
+ep0 step 240/292 loss 0.952 kl 0.000 anchor 0.000 0.169s/rec
+ep0 step 250/292 loss 0.985 kl 0.000 anchor 0.000 0.167s/rec
+ep0 step 260/292 loss 0.965 kl 0.000 anchor 0.000 0.165s/rec
+ep0 step 270/292 loss 1.043 kl 0.000 anchor 0.000 0.162s/rec
+ep0 step 280/292 loss 1.008 kl 0.000 anchor 0.000 0.160s/rec
+ep0 step 290/292 loss 0.974 kl 0.000 anchor 0.000 0.158s/rec
+saved /runs/sev-r2-research-4b-v1/00-trial-0/checkpoint

training_config.json

@@ -0,0 +1,60 @@
+{
+  "args": {
+    "base": "Qwen/Qwen3.5-4B-Base",
+    "n_per_source": 1000,
+    "epochs": 1,
+    "lr": 1e-05,
+    "head_lr": 0.0,
+    "weight_decay": 0.01,
+    "lora": 16,
+    "accum": 2,
+    "holdout": "",
+    "perm_kl": 0.0,
+    "perm_frac": 0.3,
+    "ord_w": 0.0,
+    "label_smoothing": 0.0,
+    "brier_w": 0.0,
+    "focal_gamma": 0.0,
+    "suite": "/root/evals/sev/behavioral-research-v1",
+    "train_sources": "",
+    "device": "cuda",
+    "batch": 4,
+    "dtype": "bf16",
+    "weights_dtype": "fp32",
+    "checkpointing": 1,
+    "option_isolation": 0,
+    "special_embeddings": 0,
+    "head_dim": 256,
+    "lora_targets": "all",
+    "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
+    "p_none": 0.0,
+    "p_none_distract": 0.0,
+    "p_distract": 0.0,
+    "p_none_pair": 0.0,
+    "synthetic_repeat": 1,
+    "public_frac": 1.0,
+    "anchor": "",
+    "anchor_w": 0.0,
+    "anchor_sources": "",
+    "out": "/runs/sev-r2-research-4b-v1/00-trial-0/checkpoint",
+    "data": "",
+    "replay": 0,
+    "init_from": "jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1",
+    "seed": 4
+  },
+  "suite_sha256": "372d8709cce5fc50262afc5107cd4b3637614ff2b2a7e84a8ccfa4293a4b21c5",
+  "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
+  "init_source": {
+    "init_from": "jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1",
+    "resolved": "/__modal/volumes/vo-RociVuGycNGtzvUcdOqecz/hub/models--jaredpalmer--kev-4b/snapshots/485ace8703592fcf405488b262449990824cfed1",
+    "adapter_sha256": "9797de69a42188e411b17b7b4fcb66a23374dcebc21d71a7a66f836b5d34df2b",
+    "head_sha256": "d8f796da36ff7bd7c0fb9496b452139bb7851af4fc82b07b500b682d3f721d6a",
+    "adapter_tensors": 496
+  },
+  "ordinal_objective": "ranked_probability_score",
+  "holdout": [
+    "sev_behavioral_retry_export",
+    "sev_behavioral_stale_handoff",
+    "sev_behavioral_reconcile_export"
+  ]
+}

training_metrics.json

@@ -0,0 +1,14 @@
+{
+  "wall_seconds": 371.20887875556946,
+  "records_seen": 2334,
+  "requested_records": 2334,
+  "truncated_records": 0,
+  "rejected_records": 0,
+  "optimizer_steps": 292,
+  "forward_tokens": 738327,
+  "peak_device_bytes": 18913074688,
+  "device": "cuda",
+  "dtype": "bf16",
+  "batch": 4,
+  "peak_rss_bytes": 31561670656
+}