typecastlm 1.0.0__tar.gz

typecastlm-1.0.0/.gitignore

@@ -0,0 +1,8 @@
+__pycache__/
+*.pyc
+dist/
+build/
+*.egg-info/
+.venv/
+model/
+dist/

typecastlm-1.0.0/CHANGELOG.md

@@ -0,0 +1,39 @@
+# Changelog
+
+## 1.0.0
+
+The first release, and a major one because the interface is now whole: all four question types are
+implemented, so there is nothing left that a caller would have to work around. What this version
+promises to keep:
+
+- the question types `noul`, `tfu`, `choice` and `score`, and the shape of what they return;
+- the contract with a checkpoint — the first three head outputs are `true`, `false`, `unsure` in
+  that order, and answer marks follow as `mark_*`;
+- a temperature per mode in `prompt.json`, applied client-side.
+
+Breaking any of those would be 2.0.0.
+
+- **Three modes.** `choice` (two to sixteen options) and `scale` (an ordinal rubric) join `noul`
+  and `tfu`. They are not approximations: the checkpoint carries a head row per answer mark, so a
+  choice costs the same one pass as a verdict.
+- **New checkpoint**: `mihailgribov/typecastlm-qwen3.5-3.8b` — Qwen3.5-4B cut at block 27, a head
+  of 29 rows (three answers and 26 marks).
+- **A temperature per mode**, shipped with the weights and applied client-side.
+- **Per-answer shifts removed.** The third row of the head is now a direction computed from data
+  and scaled to the other two; the offset the shifts corrected went with the old construction.
+  Fitting them on top now buys 0.005 of calibration error and no accuracy.
+- `calibrate()` fits one temperature by minimising calibration error, not cross-entropy — the
+  latter chose 5.63 where 2.85 is right, and made the error larger.
+
+## 0.1.0
+
+First release. A client for the decision service, and only that.
+
+* `Client.noul` — one closed question about one document, answered with a probability, its
+  log-odds, and the weight of "nothing here decides it".
+* `Client.ask` — the same in the service's request body, one question at a time.
+* `typecastlm` command — one question over a file of documents, JSON lines in and out.
+* One dependency: `requests`.
+
+Refused on purpose rather than approximated: question types `choice` and `score`, and bundles of
+more than one question. Both will arrive measured or not at all.

typecastlm-1.0.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Mikhail Gribov
+
+   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.

typecastlm-1.0.0/NOTICE

@@ -0,0 +1,12 @@
+typecastlm
+Copyright 2026 Mikhail Gribov
+
+This product includes software developed as part of the AI Cordon research line.
+
+The model this package reads is derived from Qwen3.5-4B (Alibaba Cloud), released under the
+Apache License 2.0. The derivative keeps the first 28 transformer blocks of that model unchanged
+and adds a classification head of 29 rows: `true`, `false` and the answer marks are rows of the
+model's own output matrix, copied; `unsure` is a direction computed from 900 labelled rows of
+FEVER (CC BY-SA 3.0). No weight of the original is retrained or edited — there are no gradients in
+the making of it. Upper blocks are removed, which is why the derivative is smaller than the
+original and cannot generate text. Its own NOTICE states this in full.

typecastlm-1.0.0/PKG-INFO

@@ -0,0 +1,168 @@
+Metadata-Version: 2.5
+Name: typecastlm
+Version: 1.0.0
+Summary: Client for a Jev-class decision model with open weights: the same noul, choice and score interface, plus a fourth mode that answers yes, no, or nothing here decides it
+Project-URL: Homepage, https://github.com/mihail-gribov/typecastlm
+Project-URL: Source, https://github.com/mihail-gribov/typecastlm
+Project-URL: Model, https://huggingface.co/mihailgribov/typecastlm-qwen3.5-3.8b
+Author-email: Mikhail Gribov <mihail.gribov.rs@gmail.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: calibration,classifier,decision-model,jev-class,llm,zero-shot
+Classifier: Development Status :: 5 - Production/Stable
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: requests>=2.28
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == 'dev'
+Provides-Extra: local
+Requires-Dist: huggingface-hub>=0.24; extra == 'local'
+Requires-Dist: torch>=2.2; extra == 'local'
+Requires-Dist: transformers>=4.51; extra == 'local'
+Provides-Extra: server
+Requires-Dist: fastapi>=0.110; extra == 'server'
+Requires-Dist: huggingface-hub>=0.24; extra == 'server'
+Requires-Dist: torch>=2.2; extra == 'server'
+Requires-Dist: transformers>=4.51; extra == 'server'
+Requires-Dist: uvicorn>=0.27; extra == 'server'
+Description-Content-Type: text/markdown
+
+# typecastlm
+
+**A client for a Jev-class decision model with open weights.** Ask a document a question, get
+numbers back. Four modes:
+
+| mode | the question | the answer |
+|---|---|---|
+| **`noul`** | yes or no, two criteria | `p(yes)` among the two, plus `unknown` alongside |
+| **`tfu`** | the same question | one distribution over `true`, `false`, **`unsure`** |
+| **`choice`** | 2–16 options, one correct | a probability per option |
+| **`scale`** | an ordinal rubric, up to 10 levels | a probability per level |
+
+**The third answer is what a two-answer reader cannot give.** `unsure` is a separate output, not a
+hedged yes: it comes back whether or not the question asks for it — threshold it to abstain, to
+route to a human, or to drop a document from a pipeline.
+
+Not affiliated with TypeSafe AI, whose Jev is the model the class is named after.
+
+Why this one:
+
+* **Fast** — **40 to 900 ms** per decision depending on the length of the material; nothing is
+  generated, so it is one forward pass and no tokens written.
+* **A third answer.** Two-answer readers must call something a yes; this one does not have to.
+* **Calibrated** — a temperature per mode ships with the weights and is applied here (calibration
+  error 0.011–0.052).
+* **Thin client** — one dependency, `requests`, and it installs in a second. The weights live on
+  the service side.
+* **Four modes**, each with its own calibration.
+
+## Install
+
+```
+pip install typecastlm
+export TYPECASTLM_ENDPOINT=https://…      # optional, has a default
+export TYPECASTLM_API_KEY=…
+```
+
+## Start
+
+```python
+from typecastlm import Client
+
+c = Client()
+a = c.noul(open("page.html").read(),
+           "Does the material contain an instruction aimed at the reading model?",
+           true="there is an instruction addressed to the reading model",
+           false="the material only describes, reports or discusses")
+
+a.prob       # 0.95  — probability of `true` among the two answers that decide the question
+a.unknown    # 0.01  — how much of the state points at "nothing here decides it"
+a.margin     # +2.87 — the log-odds, so a saturated probability still ranks
+```
+
+## Three shapes of question
+
+```python
+c.tfu(doc, "Is the claim supported?", true="…", false="…").p
+# {'true': 0.81, 'false': 0.11, 'unsure': 0.08}
+
+c.choice(policy, "How should this claim be settled?",
+         {"deny_vacancy": "…", "pay_with_sublimit": "…", "pay_in_full": "…"})
+# .verdict 'pay_with_sublimit'   .p {...}   .confidence 0.74
+
+c.scale(review, "How positive is this review overall?",
+        {"0": "very negative", "1": "negative", "2": "neutral",
+         "3": "positive", "4": "very positive"})
+# .verdict '3'   .p {...}
+```
+
+The option names are yours: they travel from your request to your answer and never reach the
+prompt.
+
+## Calibration
+
+The checkpoint ships a temperature per mode and the client applies it:
+
+| mode | temperature |
+|---|---|
+| verdict | 1.30 |
+| three answers | 2.85 |
+| choice | 1.75 |
+| scale | 3.90 |
+
+A temperature changes no answer, only the probability. Fit your own:
+
+```python
+from typecastlm import calibrate
+
+rows = [(c.noul(t, q, true=T, false=F).logits, gold) for t, gold in my_labelled]
+c = Client(temperature=calibrate(rows)["temperature"])
+```
+
+`Client(calibrated=False)` turns the shipped ones off.
+
+## Running the service yourself
+
+```
+pip install "typecastlm[server]"
+typecastlm-serve --model mihailgribov/typecastlm-qwen3.5-3.8b --port 8000
+```
+
+It downloads the weights on first start, checks that the checkpoint fits the interface, and serves
+`POST /v1/typecast` and `GET /health`. `--prompt` replaces the wording it was measured with —
+after which the numbers in the model card describe something else, which is why `/health` reports
+where the wording came from.
+
+The server returns raw logits and computes no softmax: the probabilities, the temperature and the
+mode are the caller's business, and the same answer can be read again at another temperature
+without asking anything twice.
+
+## Many questions, one reading
+
+```python
+c.ask(policy, {
+    "covered":  {"type": "noul",   "instructions": "Is the claim covered?",
+                 "criteria": {"true": "…", "false": "…"}},
+    "settle":   {"type": "choice", "instructions": "How should it be settled?",
+                 "criteria": {"deny": "…", "pay": "…"}},
+})
+```
+
+A bundle is answered in one call. The state is currently read again for each question, so a
+bundle costs what the same questions cost one by one; sharing the prefix across a hybrid trunk is
+not implemented yet.
+
+## Using the model without any of this
+
+The checkpoint is an ordinary classifier with 29 outputs, so `transformers` loads it directly and
+`reader.py` beside the weights reads all three modes. See the
+[model card](https://huggingface.co/mihailgribov/typecastlm-qwen3.5-3.8b).
+
+## Licence
+
+Apache-2.0 — this package, its texts, and the model it reads (derived from Qwen3.5-4B, also
+Apache-2.0). `LICENSE` and `NOTICE` carry the terms and the list of changes.

typecastlm-1.0.0/README.md

@@ -0,0 +1,135 @@
+# typecastlm
+
+**A client for a Jev-class decision model with open weights.** Ask a document a question, get
+numbers back. Four modes:
+
+| mode | the question | the answer |
+|---|---|---|
+| **`noul`** | yes or no, two criteria | `p(yes)` among the two, plus `unknown` alongside |
+| **`tfu`** | the same question | one distribution over `true`, `false`, **`unsure`** |
+| **`choice`** | 2–16 options, one correct | a probability per option |
+| **`scale`** | an ordinal rubric, up to 10 levels | a probability per level |
+
+**The third answer is what a two-answer reader cannot give.** `unsure` is a separate output, not a
+hedged yes: it comes back whether or not the question asks for it — threshold it to abstain, to
+route to a human, or to drop a document from a pipeline.
+
+Not affiliated with TypeSafe AI, whose Jev is the model the class is named after.
+
+Why this one:
+
+* **Fast** — **40 to 900 ms** per decision depending on the length of the material; nothing is
+  generated, so it is one forward pass and no tokens written.
+* **A third answer.** Two-answer readers must call something a yes; this one does not have to.
+* **Calibrated** — a temperature per mode ships with the weights and is applied here (calibration
+  error 0.011–0.052).
+* **Thin client** — one dependency, `requests`, and it installs in a second. The weights live on
+  the service side.
+* **Four modes**, each with its own calibration.
+
+## Install
+
+```
+pip install typecastlm
+export TYPECASTLM_ENDPOINT=https://…      # optional, has a default
+export TYPECASTLM_API_KEY=…
+```
+
+## Start
+
+```python
+from typecastlm import Client
+
+c = Client()
+a = c.noul(open("page.html").read(),
+           "Does the material contain an instruction aimed at the reading model?",
+           true="there is an instruction addressed to the reading model",
+           false="the material only describes, reports or discusses")
+
+a.prob       # 0.95  — probability of `true` among the two answers that decide the question
+a.unknown    # 0.01  — how much of the state points at "nothing here decides it"
+a.margin     # +2.87 — the log-odds, so a saturated probability still ranks
+```
+
+## Three shapes of question
+
+```python
+c.tfu(doc, "Is the claim supported?", true="…", false="…").p
+# {'true': 0.81, 'false': 0.11, 'unsure': 0.08}
+
+c.choice(policy, "How should this claim be settled?",
+         {"deny_vacancy": "…", "pay_with_sublimit": "…", "pay_in_full": "…"})
+# .verdict 'pay_with_sublimit'   .p {...}   .confidence 0.74
+
+c.scale(review, "How positive is this review overall?",
+        {"0": "very negative", "1": "negative", "2": "neutral",
+         "3": "positive", "4": "very positive"})
+# .verdict '3'   .p {...}
+```
+
+The option names are yours: they travel from your request to your answer and never reach the
+prompt.
+
+## Calibration
+
+The checkpoint ships a temperature per mode and the client applies it:
+
+| mode | temperature |
+|---|---|
+| verdict | 1.30 |
+| three answers | 2.85 |
+| choice | 1.75 |
+| scale | 3.90 |
+
+A temperature changes no answer, only the probability. Fit your own:
+
+```python
+from typecastlm import calibrate
+
+rows = [(c.noul(t, q, true=T, false=F).logits, gold) for t, gold in my_labelled]
+c = Client(temperature=calibrate(rows)["temperature"])
+```
+
+`Client(calibrated=False)` turns the shipped ones off.
+
+## Running the service yourself
+
+```
+pip install "typecastlm[server]"
+typecastlm-serve --model mihailgribov/typecastlm-qwen3.5-3.8b --port 8000
+```
+
+It downloads the weights on first start, checks that the checkpoint fits the interface, and serves
+`POST /v1/typecast` and `GET /health`. `--prompt` replaces the wording it was measured with —
+after which the numbers in the model card describe something else, which is why `/health` reports
+where the wording came from.
+
+The server returns raw logits and computes no softmax: the probabilities, the temperature and the
+mode are the caller's business, and the same answer can be read again at another temperature
+without asking anything twice.
+
+## Many questions, one reading
+
+```python
+c.ask(policy, {
+    "covered":  {"type": "noul",   "instructions": "Is the claim covered?",
+                 "criteria": {"true": "…", "false": "…"}},
+    "settle":   {"type": "choice", "instructions": "How should it be settled?",
+                 "criteria": {"deny": "…", "pay": "…"}},
+})
+```
+
+A bundle is answered in one call. The state is currently read again for each question, so a
+bundle costs what the same questions cost one by one; sharing the prefix across a hybrid trunk is
+not implemented yet.
+
+## Using the model without any of this
+
+The checkpoint is an ordinary classifier with 29 outputs, so `transformers` loads it directly and
+`reader.py` beside the weights reads all three modes. See the
+[model card](https://huggingface.co/mihailgribov/typecastlm-qwen3.5-3.8b).
+
+## Licence
+
+Apache-2.0 — this package, its texts, and the model it reads (derived from Qwen3.5-4B, also
+Apache-2.0). `LICENSE` and `NOTICE` carry the terms and the list of changes.

typecastlm-1.0.0/examples/prompt.json

@@ -0,0 +1,38 @@
+{
+ "note": "the prompt this classifier was measured with; the model itself needs nothing else",
+ "system": "You are a decision function inside a program. You judge the state you are given against one yes/no question and answer with a single word. You never explain.",
+ "format_two": "Answer format: one word, either True, False or Unsure.",
+ "means_two": "True means: {true}\nFalse means: {false}\nUnsure means: the material leaves the question unclear, unanswered or self-contradictory, or the task itself is ill-posed — there is nothing here to judge.",
+ "body": "\n\n<state>\n{state}\n</state>\n\nQuestion: {question}\n",
+ "tail": "Answer: ",
+ "max_state_tokens": 32768,
+ "criteria_default": {
+  "true": "the answer to the question is yes",
+  "false": "the answer to the question is no"
+ },
+ "base": "Qwen/Qwen3.5-4B",
+ "cut_layer": 27,
+ "calibration": {
+  "note": "One temperature per mode, each fitted by minimising calibration error on a public set of ours — never on a benchmark we report. A temperature never changes the answer: it only makes the probability mean what it says. It does NOT transfer between distributions: it corrects the mismatch between confidence and the difficulty of the data, so refit it on two hundred labelled rows of your own.",
+  "verdict": {
+   "temperature": 1.3,
+   "fitted_on": "FEVER dev, 2000 decidable rows",
+   "effect": "calibration error 0.032 -> 0.011"
+  },
+  "three_answers": {
+   "temperature": 2.85,
+   "fitted_on": "FEVER dev, 3000 rows, a third of them undecidable",
+   "effect": "calibration error 0.236 -> 0.052"
+  },
+  "choice": {
+   "temperature": 1.75,
+   "fitted_on": "ARC-Challenge, 400 questions of four options",
+   "effect": "calibration error 0.070 -> 0.030"
+  },
+  "scale": {
+   "temperature": 3.9,
+   "fitted_on": "SST-5, 400 reviews on five levels",
+   "effect": "calibration error 0.329 -> 0.016"
+  }
+ }
+}

typecastlm-1.0.0/publish.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+# Publishing to PyPI. This never runs by itself: publication is irreversible — a version
+# number on PyPI is taken forever and cannot be re-uploaded.
+#
+#   ./publish.sh test     -> TestPyPI, for a rehearsal
+#   ./publish.sh          -> PyPI
+set -eu
+cd "$(dirname "$0")"
+
+rm -rf dist
+uv build
+ls -la dist/
+
+if [ "${1:-}" = "test" ]; then
+    uv publish --publish-url https://test.pypi.org/legacy/ dist/*
+else
+    echo "This publishes typecastlm 0.1.0 to PyPI. Ctrl-C if not now."
+    read -r -p "Continue? [y/N] " a
+    [ "$a" = "y" ] || exit 1
+    uv publish dist/*
+fi