decision-toolkit 0.2.1__tar.gz → 0.3.1__tar.gz

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/CHANGELOG.md

@@ -1,8 +1,32 @@
 # Changelog
 
-All notable changes to this project are documented here. Version `0.2.1` is
+All notable changes to this project are documented here. Version `0.3.1` is
 published on PyPI.
 
+## [0.3.1] — 2026-10-21
+
+Catalog UX patch: `dtk models list` now prints one row per variant, with
+`variant`, quantization, role, size, cache state and the exact download/select
+commands. `dtk models variants` prints both commands for every variant. No
+model, adapter or REST contract changes.
+
+## [0.3.0] — 2026-10-21
+
+Catalog expansion: three verified ONNX decision models and eight variants.
+Laya multilingual, Open-Jev DeBERTa v3 large and Strands Decider 2B are now
+available through the library, CLI and REST projection.
+
+### Added
+
+- `laya-multilingual` with fused Laya ONNX support (`fp16` default, `fp32`).
+- `open-jev-deberta-v3-large` with a span-pooling adapter (`fp16` default,
+  `fp32`, `q4`, `q4f16`).
+- `strands-decider-2b` with a pointer-head adapter (`q4f16` default, `q8`).
+- New `open-jev` and `strands-decider` adapters and a fused branch in `laya`.
+- Model cards, benchmark cases, a catalog proposal and documentation for every
+  added variant.
+- All 15 catalogued variants verified for SHA256, files and ONNX sessions.
+
 ## [0.2.1] — 2026-10-20
 
 Metadata-only release: this project's documentation and project URLs now point
@@ -168,6 +192,8 @@ README correction. No code or model behavior changed.
 - Typed `DecisionModel` API with lifecycle, batching, async, usage, warnings,
   normalized confidence, capabilities, structured events and atomic model store.
 
+[0.3.1]: https://pypi.org/project/decision-toolkit/0.3.1/
+[0.3.0]: https://pypi.org/project/decision-toolkit/0.3.0/
 [0.2.1]: https://pypi.org/project/decision-toolkit/0.2.1/
 [0.2.0]: https://pypi.org/project/decision-toolkit/0.2.0/
 [0.1.7]: https://pypi.org/project/decision-toolkit/0.1.7/

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: decision-toolkit
-Version: 0.2.1
+Version: 0.3.1
 Summary: Local-first toolkit for open-weights decision models: one ONNX engine, one adapter per model contract, typed choice/yes-no/score decisions with probabilities on CPU.
 Project-URL: Homepage, https://github.com/gia-uh/decision-toolkit
 Project-URL: Repository, https://github.com/gia-uh/decision-toolkit
@@ -59,7 +59,7 @@ and CLI. `dtk --help` and every subcommand help also include runnable examples.
 ## Installation
 
 ```bash
-uv pip install decision-toolkit              # library + CLI (0.2.1 on PyPI)
+uv pip install decision-toolkit              # library + CLI (0.3.1 on PyPI)
 uv pip install "decision-toolkit[onnx]"      # + onnxruntime, tokenizers (the model runtime)
 uv pip install "decision-toolkit[server]"    # + REST server (FastAPI)
 ```
@@ -67,7 +67,7 @@ uv pip install "decision-toolkit[server]"    # + REST server (FastAPI)
 The project is managed with [uv](https://docs.astral.sh/uv/) and ships a committed `uv.lock`. Published
 wheels remain compatible with `pip` for environments that do not use uv.
 
-PyPI: <https://pypi.org/project/decision-toolkit/0.2.1/>. See
+PyPI: <https://pypi.org/project/decision-toolkit/0.3.1/>. See
 [docs/RELEASE.md](https://github.com/gia-uh/decision-toolkit/blob/main/docs/RELEASE.md)
 for the Trusted Publishing flow and clean-install checks.
 
@@ -207,6 +207,9 @@ adapters (there is no generic "any ONNX" adapter).
 | `laya` | `laya` | **m0** — verified against torch (Δp ≤ 0.005) | `q8`* (528 MB) |
 | `julia-1` | `julia` | **m0** — verified: 100/100 argmax parity, 329 ms | `fp32`* (618 MB) |
 | `opendecider-nano` | `opendecider` | **m0** — official build; prompt ported from pinned code (530 ms) | `q8f16`* (469 MB); `fp16` (792 MB) |
+| `laya-multilingual` | `laya` | **m0** — fused ONNX; upstream parity (37 ms) | `fp16`* (682 MB); `fp32` (1326 MB) |
+| `open-jev-deberta-v3-large` | `open-jev` | **m0** — Open-Jev DeBERTa v3 large repro; 0.854 in-domain (241 ms) | `fp16`* (884 MB); `fp32` (1758 MB); `q4` (487 MB); `q4f16` (357 MB) |
+| `strands-decider-2b` | `strands-decider` | **m0** — reference fixture match (719 ms) | `q4f16`* (1114 MB); `q8` (1821 MB) |
 
 **Only `m0` models are in the catalog.** Researched-but-unavailable models
 (planned, conversion pending, broken exports) are listed in
@@ -222,7 +225,7 @@ own repo files, size and role. A `*` in the table above marks the default.
 `dtk models list` always shows every catalogued variant and what is cached:
 
 ```bash
-dtk models list                      # all variants; * = default
+dtk models list                      # one row per variant; * = default
 dtk models list --quant              # add the quantization labels
 dtk models variants gliner2.5-multi  # full details for one model
 ```

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/README.md

@@ -21,7 +21,7 @@ and CLI. `dtk --help` and every subcommand help also include runnable examples.
 ## Installation
 
 ```bash
-uv pip install decision-toolkit              # library + CLI (0.2.1 on PyPI)
+uv pip install decision-toolkit              # library + CLI (0.3.1 on PyPI)
 uv pip install "decision-toolkit[onnx]"      # + onnxruntime, tokenizers (the model runtime)
 uv pip install "decision-toolkit[server]"    # + REST server (FastAPI)
 ```
@@ -29,7 +29,7 @@ uv pip install "decision-toolkit[server]"    # + REST server (FastAPI)
 The project is managed with [uv](https://docs.astral.sh/uv/) and ships a committed `uv.lock`. Published
 wheels remain compatible with `pip` for environments that do not use uv.
 
-PyPI: <https://pypi.org/project/decision-toolkit/0.2.1/>. See
+PyPI: <https://pypi.org/project/decision-toolkit/0.3.1/>. See
 [docs/RELEASE.md](https://github.com/gia-uh/decision-toolkit/blob/main/docs/RELEASE.md)
 for the Trusted Publishing flow and clean-install checks.
 
@@ -169,6 +169,9 @@ adapters (there is no generic "any ONNX" adapter).
 | `laya` | `laya` | **m0** — verified against torch (Δp ≤ 0.005) | `q8`* (528 MB) |
 | `julia-1` | `julia` | **m0** — verified: 100/100 argmax parity, 329 ms | `fp32`* (618 MB) |
 | `opendecider-nano` | `opendecider` | **m0** — official build; prompt ported from pinned code (530 ms) | `q8f16`* (469 MB); `fp16` (792 MB) |
+| `laya-multilingual` | `laya` | **m0** — fused ONNX; upstream parity (37 ms) | `fp16`* (682 MB); `fp32` (1326 MB) |
+| `open-jev-deberta-v3-large` | `open-jev` | **m0** — Open-Jev DeBERTa v3 large repro; 0.854 in-domain (241 ms) | `fp16`* (884 MB); `fp32` (1758 MB); `q4` (487 MB); `q4f16` (357 MB) |
+| `strands-decider-2b` | `strands-decider` | **m0** — reference fixture match (719 ms) | `q4f16`* (1114 MB); `q8` (1821 MB) |
 
 **Only `m0` models are in the catalog.** Researched-but-unavailable models
 (planned, conversion pending, broken exports) are listed in
@@ -184,7 +187,7 @@ own repo files, size and role. A `*` in the table above marks the default.
 `dtk models list` always shows every catalogued variant and what is cached:
 
 ```bash
-dtk models list                      # all variants; * = default
+dtk models list                      # one row per variant; * = default
 dtk models list --quant              # add the quantization labels
 dtk models variants gliner2.5-multi  # full details for one model
 ```

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/bench/suites/smoke.jsonl

@@ -3,3 +3,6 @@
 {"id":"laya-billing","model":"laya","state":"I was charged twice for my order last week and nobody replied.","questions":{"team":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"payments and refunds","technical":"bugs and errors"}}},"expected":{"team":"billing"}}
 {"id":"julia-billing","model":"julia-1","state":"I was charged twice for my order last week and nobody replied.","questions":{"team":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"payments and refunds","technical":"bugs and errors"}}},"expected":{"team":"billing"}}
 {"id":"opendecider-billing","model":"opendecider-nano","state":"I was charged twice for my order last week and nobody replied.","questions":{"team":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"payments and refunds","technical":"bugs and errors"}}},"expected":{"team":"billing"}}
+{"id":"laya-multilingual-billing","model":"laya-multilingual","state":"I was charged twice for my order last week and nobody replied.","questions":{"team":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"payments and refunds","technical":"bugs and errors"}}},"expected":{"team":"billing"}}
+{"id":"open-jev-billing","model":"open-jev-deberta-v3-large","state":"I was charged twice for my order last week and nobody replied.","questions":{"team":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"payments and refunds","technical":"bugs and errors"}}},"expected":{"team":"billing"}}
+{"id":"strands-decider-billing","model":"strands-decider-2b","state":"Help! My payouts have been failing for 3 days!","questions":{"team":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"payments and refunds","sales":"pricing","retail":"retail"}}},"expected":{"team":"billing"}}

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/docs/BENCH.md

@@ -47,6 +47,15 @@ Environment: Python 3.12.7, ONNX Runtime 1.23.2, tokenizers 0.22.1, numpy
 Calibration smoke (`gliner2.5-multi`, 10 synthetic routing cases): accuracy
 `0.60`, Brier `0.173`, ECE `0.247`.
 
+2026-10-21 audit of the three added models (Python 3.13.2, ONNX Runtime
+1.31.0, post-warmup, one smoke case):
+
+| Model | Variant | p50 ms | Parity |
+|---|---|---:|---:|
+| `laya-multilingual` | fp16 | 37.4 | 1/1 |
+| `open-jev-deberta-v3-large` | fp16 | 241.2 | 1/1 |
+| `strands-decider-2b` | q4f16 | 719.3 | 1/1 |
+
 Full evidence, per-case answers and environment metadata:
 [`bench/results/2026-10-20-smoke.json`](https://github.com/gia-uh/decision-toolkit/blob/main/bench/results/2026-10-20-smoke.json)
 and

decision_toolkit-0.3.1/docs/CATALOG_PROPOSAL.md

@@ -0,0 +1,122 @@
+# Catalog proposal: three ONNX decision models
+
+**Date:** 2026-10-21  
+**Status:** published in `0.3.0`  
+**Scope:** add Laya multilingual, Open-Jev DeBERTa v3 large and Strands Decider 2B  
+**Release:** <https://pypi.org/project/decision-toolkit/0.3.0/>
+
+## Executive summary
+
+Three public ONNX decision models meet the catalog gates for this iteration:
+
+| Model | License | Adapter | Variants | Default |
+|---|---|---|---|---:|
+| Laya multilingual | Apache-2.0 | existing `laya` + fused branch | fp16, fp32 | fp16 |
+| Open-Jev DeBERTa v3 large | Apache-2.0 | new `open-jev` | fp16, fp32, q4, q4f16 | fp16 |
+| Strands Decider 2B | Apache-2.0 | new `strands-decider` | q4f16, q8 | q4f16 |
+
+They have a real public ONNX export, complete external-data files, a pinned
+revision and a known prompt contract. The local machine verified every file
+hash and opened every ONNX session. No speculative or unavailable model was
+added.
+
+## Why these three
+
+- **Laya multilingual** closes the multilingual gap in the existing Laya
+  family. The ONNX export is fused, but the prompt and calibrated temperatures
+  are the same as the split Laya contract. It needed one adapter branch, not a
+  new model family.
+- **Open-Jev DeBERTa v3 large** is an independent, open reproduction of the
+  typed-decision shape. It has a complete fp16/fp32/q4/q4f16 family and a
+  documented span-pooling graph.
+- **Strands Decider 2B** is the first larger decision model in the catalog. It
+  has a published reference fixture and a clear pointer-head graph. It is more
+  expensive on CPU, so it is marked as a quality/agentic option rather than a
+  replacement for the small models.
+
+`gliner2.5-multi-decide` is **not** in this batch. Its ONNX export is real, but
+the prompt uses `[L]` markers instead of the existing GLiNER2 `[C]` layout; it
+needs a separate prompt adapter and its own parity run.
+
+## Verification performed
+
+- `dtk models verify --all --json`: **15/15 variants OK** (SHA256, files and
+  ONNX sessions).
+- Real inference for every new variant, using `choice`, `noul` and `score`.
+- Smoke parity suite: all added cases matched their expected labels.
+- `dtk models variants` exposes all variants and install commands.
+- Full test suite: **247 passed, 1 deselected**.
+- Ruff and MkDocs `--strict`: clean.
+- Python 3.12/3.13 CI matrix and wheel clean-install job: green.
+
+### Measured CPU latency
+
+Post-warmup p50, one simple choice question, audit machine (Python 3.13.2,
+ONNX Runtime 1.31.0, CPU):
+
+| Model | Variant | p50 |
+|---|---|---:|
+| Laya multilingual | fp16 | 37.4 ms |
+| Laya multilingual | fp32 | 34.1 ms |
+| Open-Jev DeBERTa v3 large | fp16 | 241.2 ms |
+| Open-Jev DeBERTa v3 large | fp32 | 78.2 ms |
+| Open-Jev DeBERTa v3 large | q4 | 56.0 ms |
+| Open-Jev DeBERTa v3 large | q4f16 | 219.4 ms |
+| Strands Decider 2B | q4f16 | 719.3 ms |
+| Strands Decider 2B | q8 | 1467.2 ms |
+
+The Open-Jev DeBERTa v3 large 4-bit paths are size/edge choices; they are not assumed to be
+faster or more accurate. Strands is a 2B model and is intentionally slower
+than the other catalog entries.
+
+### Evidence / parity
+
+- **Laya multilingual:** upstream conversion reports fp32 max |Δp| 1e-4 and
+  fp16 0.035 against PyTorch; local fp16/fp32 agree on the audit answers.
+- **Open-Jev DeBERTa v3 large:** upstream reports 0.854 in-domain and 0.690 OOD accuracy, plus
+  0.022 in-domain ECE. Local fp16/fp32 agree on the published example; q4
+  variants are explicitly edge builds.
+- **Strands Decider 2B:** reference fixture matches within 0.018 max
+  probability delta; q8 is closer to the reference on most fields.
+
+These are upstream parity claims plus local runtime/answer checks, not an
+independent reproduction of the full upstream training evaluation.
+
+## Quantization policy
+
+- Default variants are chosen for a balance of quality, size and CPU latency:
+  Laya fp16, Open-Jev DeBERTa v3 large fp16, Strands q4f16.
+- `fp32` remains a quality variant where published.
+- `q4`/`q4f16`/`q8` are exposed, but their thresholds must be validated on the
+  user's data; no silent fallback exists.
+- `variant` remains a **decision-toolkit extension** in REST, not a TypeSafe or
+  OpenAI field.
+
+## Risks and limitations
+
+1. **Upstream evidence vs local evidence.** Full PyTorch parity was not
+   re-run for these three models; the catalog records upstream parity and local
+   runtime checks. A larger acceptance run is still desirable.
+2. **Strands CPU cost.** q8 takes ~1.5 s per simple question; this is suitable
+   for local/agentic workflows, not high-throughput serving.
+3. **Quantization drift.** Open-Jev DeBERTa v3 large q4 and Strands q4f16 can move near-ties;
+   the model cards call this out.
+4. **Coverage.** Laya multilingual is weak on low-resource languages;
+   Open-Jev DeBERTa v3 large is English-only and weak on unseen score scales;
+   Strands is a 2B model with
+   its own upstream limitations.
+5. **No image/box support.** The catalog remains text-only.
+
+## Recommendation
+
+Accept the three models as `m0` entries with the variants listed above. Keep
+`gliner2.5-multi-decide` in the roadmap until its `[L]` adapter is implemented
+and parity is measured. The next model batch should prioritize either that
+adapter or a larger calibration/parity run for the new entries.
+
+## Related documentation
+
+- [Model cards](models/README.md)
+- [Variants and quantization](VARIANTS.md)
+- [Benchmarks](BENCH.md)
+- [REST](REST.md)

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/docs/CLI.md

@@ -147,6 +147,9 @@ are already downloaded.
 | `laya` | `q8`* | int8 | 527.9 MB |
 | `julia-1` | `fp32`* | fp32 | 618.2 MB |
 | `opendecider-nano` | `q8f16`*, `fp16` | int8/fp16 / fp16 | 468.8 / 791.9 MB |
+| `laya-multilingual` | `fp16`*, `fp32` | fp16 / fp32 | 681.8 / 1325.6 MB |
+| `open-jev-deberta-v3-large` | `fp16`*, `fp32`, `q4`, `q4f16` | fp16 / fp32 / q4 / q4-fp16 | 883.5 / 1757.9 / 486.7 / 357.0 MB |
+| `strands-decider-2b` | `q4f16`*, `q8` | q4-fp16 / int8 | 1113.6 / 1821.4 MB |
 
 ```bash
 # inspect, including the exact download command for each variant
@@ -181,10 +184,12 @@ classification or serve command overrides it for that command. If `--variant`
 is omitted, the catalog default is used unless `dtk models use` selected another
 one.
 
-`models list --json` always includes license, quant, adapter, capabilities, all
-variants, the default variant, cached variants, `current` and `current_variant`.
-The text table always shows variants and cached variants; `--license`, `--quant`
-and `--adapter` add their columns.
+`models list` text output has **one row per variant**, with `id`, `variant`,
+`quant`, `role`, `size` and `cached`, followed by the exact download/select
+commands. `--license` and `--adapter` add catalog columns; `--quant` is accepted
+for compatibility because quantization is always shown. `--json` returns the
+nested model tree: one object per model, each with its `variants` and
+`cached_variants`.
 
 See [Variants and quantization](VARIANTS.md) for the full table and cache
 layout.

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/docs/RELEASE.md

@@ -9,7 +9,8 @@ fixed TestPyPI verification for uv; `0.1.6` added complete variant visibility
 and selection in the CLI and documentation; `0.1.7` aligned variant selection
 across the library, CLI and REST API; `0.2.0` corrected the REST conformance
 claims and wire shapes to the published TypeSafe and OpenAI schemas; `0.2.1`
-changed the public documentation URLs to this repository's Pages site.
+changed the public documentation URLs to this repository's Pages site; `0.3.0`
+added the Laya multilingual, Open-Jev DeBERTa v3 large and Strands Decider ONNX models; `0.3.1` made the catalog list one row per variant with exact commands.
 
 ## Local build and verification
 
@@ -36,7 +37,7 @@ Clean install test (no `PYTHONPATH`):
 ```bash
 uv venv --python 3.12 /tmp/dtk-wheel
 uv pip install --python /tmp/dtk-wheel/bin/python \
-  dist/decision_toolkit-0.2.1-py3-none-any.whl
+  dist/decision_toolkit-0.3.1-py3-none-any.whl
 uv pip check --python /tmp/dtk-wheel/bin/python
 /tmp/dtk-wheel/bin/dtk --version
 /tmp/dtk-wheel/bin/dtk models list --json
@@ -47,7 +48,7 @@ With inference and server extras:
 
 ```bash
 uv pip install --python /tmp/dtk-wheel/bin/python \
-  "dist/decision_toolkit-0.2.1-py3-none-any.whl[onnx,server,tracing]"
+  "dist/decision_toolkit-0.3.1-py3-none-any.whl[onnx,server,tracing]"
 /tmp/dtk-wheel/bin/dtk decide --model gliner2.5-decide ...
 /tmp/dtk-wheel/bin/dtk serve --model gliner2.5-decide --warmup --port 8016
 curl -s http://127.0.0.1:8016/health
@@ -81,8 +82,8 @@ No long-lived PyPI token is needed; the workflow uses GitHub OIDC
 4. Create and push an annotated tag:
 
    ```bash
-   git tag -a v0.2.1 -m "decision-toolkit 0.2.1"
-   git push origin v0.2.1
+   git tag -a v0.3.1 -m "decision-toolkit 0.3.1"
+   git push origin v0.3.1
    ```
 
 5. `.github/workflows/release.yml` builds once, publishes to **TestPyPI**, then
@@ -131,9 +132,9 @@ to TestPyPI without publishing to PyPI or creating a GitHub release.
 
 ## Published release
 
-`0.2.1` is published through Trusted Publishing on 2026-10-20:
+`0.3.1` is published through Trusted Publishing on 2026-10-21:
 
-<https://pypi.org/project/decision-toolkit/0.2.1/>
+<https://pypi.org/project/decision-toolkit/0.3.1/>
 
 `0.1.0` was the first release; `0.1.1` refreshed the PyPI description; `0.1.2`
 validated the OIDC TestPyPI → PyPI flow; `0.1.3` added post-publish install
@@ -142,7 +143,8 @@ development, build, CI, release and documentation environments to uv but was
 published to TestPyPI only; `0.1.5` fixed TestPyPI resolution; `0.1.6` added
 complete variant/quantization visibility and CLI selection; `0.1.7` aligned
 selection across library, CLI and REST; `0.2.0` corrected the REST conformance
-claims and wire shapes; `0.2.1` changed the public documentation URLs.
+claims and wire shapes; `0.2.1` changed the public documentation URLs; `0.3.0`
+added the three new ONNX decision models; `0.3.1` made the catalog list one row per variant.
 
 The tag-triggered workflow remains the preferred path for future releases once
 the TestPyPI/PyPI trusted publishers and protected environments are configured.

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/docs/ROADMAP.md

@@ -7,19 +7,17 @@ They are deliberately absent from `CATALOG`, `dtk models list` and `/v1/models`.
 
 | id | status | reason |
 |---|---|---|
-| `gliner2.5-multi-decide` | planned | No compatible export of the fine-tune; only the base multi-v1 export is verified |
-| `laya-multilingual` | planned | No split multilingual ONNX export published (existing ones are fused single-file layouts) |
-| `intern-decision-0.8b` | planned | Split embed/decoder/projection export; needs the upstream prompt compiler; `parityVerified: false` |
-| `d3-lite` | phase2 | No ONNX published; conversion from the original weights planned |
-| `kev-0.8b` | phase2 | No ONNX published; conversion planned |
+| `gliner2.5-multi-decide` | planned | ONNX split export exists (`DanKau/gliner2.5-multi-decide-onnx`), but the `[L]` prompt layout needs an adapter |
+| `intern-decision-0.8b` | planned | ONNX exists (`Wortou/Intern-Decision-0.8B-ONNX-FP16`), but it is split and needs the upstream prompt compiler; `parityVerified: false` |
+| `d3-lite` | phase2 | Web ONNX torso/readout files exist, but the repo declares no license/README; not eligible |
+| `kev-0.8b` | phase2 | ONNX exists (`midudev/kev-0.8b-ONNX`) but external data is missing and it uses custom ops |
 | `sieve-0.8b` | phase2 | No ONNX published; conversion planned |
-| `strands-decider-2b` | phase2 | No ONNX published; conversion planned |
 | `thx-01` | phase2 | No ONNX published; conversion planned |
 | `jevk5-lite` | waitlist | Export references external `.onnx.data` files that are not published |
 | `bosun-0.6b` | waitlist | No ONNX yet |
-| `d3-edge` | waitlist | No ONNX yet |
-| `tev1-0.8b` | waitlist | License not declared on the model card |
-| `d1-omni-600m` | waitlist | LFM Open License; needs review |
+| `d3-edge` | waitlist | Web ONNX files exist, but no license/README; not eligible |
+| `tev1-0.8b` | waitlist | ONNX exists, but the upstream licence is `other` and needs review |
+| `d1-omni-600m` | waitlist | ONNX exists, but the LFM Open License needs review and it is multimodal |
 | `bekko-400m` | waitlist | License pending |
 
 ## How a model moves into the catalog

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/docs/VARIANTS.md

@@ -18,15 +18,23 @@ have not passed the catalog gates are not exposed as variants.
 A `*` marks the default variant. It is used when `--variant` is omitted and no
 other variant was selected with `dtk models use`.
 
-| Model | Variant | Quantization | Size | Role |
-| --- | --- | --- | ---: | --- |
-| `gliner2.5-multi` | `fp16` * | fp16 encoder + fp32 classifier | 577.1 MB | default |
-| `gliner2.5-multi` | `fp32` | fp32 encoder + fp32 classifier | 1131.9 MB | quality |
-| `gliner2.5-decide` | `q4f16` * | q4/fp16 | 523.5 MB | default |
-| `laya` | `q8` * | int8 | 527.9 MB | default |
-| `julia-1` | `fp32` * | fp32 | 618.2 MB | default |
-| `opendecider-nano` | `q8f16` * | int8/fp16 | 468.8 MB | default |
-| `opendecider-nano` | `fp16` | fp16 | 791.9 MB | quality |
+| Model | Variant | Quantization | Size | Role | Download |
+| --- | --- | --- | ---: | --- | --- |
+| `gliner2.5-multi` | `fp16` * | fp16 encoder + fp32 classifier | 577.1 MB | default | `dtk models download gliner2.5-multi --variant fp16` |
+| `gliner2.5-multi` | `fp32` | fp32 encoder + fp32 classifier | 1131.9 MB | quality | `dtk models download gliner2.5-multi --variant fp32` |
+| `gliner2.5-decide` | `q4f16` * | q4/fp16 | 523.5 MB | default | `dtk models download gliner2.5-decide --variant q4f16` |
+| `laya` | `q8` * | int8 | 527.9 MB | default | `dtk models download laya --variant q8` |
+| `julia-1` | `fp32` * | fp32 | 618.2 MB | default | `dtk models download julia-1 --variant fp32` |
+| `opendecider-nano` | `q8f16` * | int8/fp16 | 468.8 MB | default | `dtk models download opendecider-nano --variant q8f16` |
+| `opendecider-nano` | `fp16` | fp16 | 791.9 MB | quality | `dtk models download opendecider-nano --variant fp16` |
+| `laya-multilingual` | `fp16` * | fp16 | 681.8 MB | default | `dtk models download laya-multilingual --variant fp16` |
+| `laya-multilingual` | `fp32` | fp32 | 1325.6 MB | quality | `dtk models download laya-multilingual --variant fp32` |
+| `open-jev-deberta-v3-large` | `fp16` * | fp16 | 883.5 MB | default | `dtk models download open-jev-deberta-v3-large --variant fp16` |
+| `open-jev-deberta-v3-large` | `fp32` | fp32 | 1757.9 MB | quality | `dtk models download open-jev-deberta-v3-large --variant fp32` |
+| `open-jev-deberta-v3-large` | `q4` | q4 | 486.7 MB | edge | `dtk models download open-jev-deberta-v3-large --variant q4` |
+| `open-jev-deberta-v3-large` | `q4f16` | q4/fp16 | 357.0 MB | edge | `dtk models download open-jev-deberta-v3-large --variant q4f16` |
+| `strands-decider-2b` | `q4f16` * | q4/fp16 | 1113.6 MB | default | `dtk models download strands-decider-2b --variant q4f16` |
+| `strands-decider-2b` | `q8` | int8 | 1821.4 MB | quality | `dtk models download strands-decider-2b --variant q8` |
 
 Models with a single row have only one catalogued, verified build. That does
 not mean other conversions do not exist upstream; it means they have not passed
@@ -79,12 +87,15 @@ dtk models variants opendecider-nano --json
 Example:
 
 ```text
-id                 variants        cached
-gliner2.5-multi    fp16*, fp32     fp16
-gliner2.5-decide   q4f16*          q4f16
-opendecider-nano   q8f16*, fp16    q8f16
-
-* default variant. Use `dtk models variants <id>` for details and install commands.
+id                variant  quant    role     size       cached
+gliner2.5-multi   fp16*    fp16     default  577.1 MB   ✓
+gliner2.5-multi   fp32     fp32     quality  1131.9 MB  ✓
+gliner2.5-decide  q4f16*   q4/fp16  default  523.5 MB   ✓
+laya              q8*      int8     default  527.9 MB   ✓
+
+* default variant.
+download: dtk models download <id> --variant <name>
+select:   dtk models use <id> --variant <name>
 ```
 
 ## Download a different quantization

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/docs/index.md

@@ -24,6 +24,7 @@ PyPI: <https://pypi.org/project/decision-toolkit/>
 - [Operations](OPERATIONS.md) — rate limits, metrics, tracing and runbook.
 - [Benchmarks](BENCH.md) — reproducible latency/parity/calibration.
 - [Models](models/README.md) — the curated catalog.
+- [Catalog proposal](CATALOG_PROPOSAL.md) — analysis and evidence for the latest model additions.
 - [Release](RELEASE.md) — build, TestPyPI/PyPI and Trusted Publishing.
 
 ## One example

{decision_toolkit-0.2.1 → decision_toolkit-0.3.1}/docs/models/README.md

@@ -11,6 +11,9 @@ card on purpose: they are not part of the curated library yet.
 | [Laya (English)](laya.md) | `laya` | q8* · 528 MB | `cccc9a01` | 514 ms | torch parity: Δp ≤ 0.005 |
 | [Julia 1](julia-1.md) | `julia` | fp32* · 618 MB | `82a2fadf` | 329 ms | 100/100 argmax; 8.5e-05 max logit error |
 | [OpenDecider nano](opendecider-nano.md) | `opendecider` | q8f16* · 469 MB; fp16 · 792 MB | `5137ae7b` | 530 ms | official build; prompt ported from pinned code |
+| [Laya multilingual](laya-multilingual.md) | `laya` | fp16* · 682 MB; fp32 · 1326 MB | `46b77bbf` | 37 ms | fused ONNX; upstream Δp ≤ 0.035 |
+| [Open-Jev DeBERTa v3 large](open-jev-deberta-v3-large.md) | `open-jev` | fp16* · 884 MB; fp32 · 1758 MB; q4 · 487 MB; q4f16 · 357 MB | `7c79f25b` | 241 ms | 0.854 in-domain; 0.690 OOD (upstream) |
+| [Strands Decider 2B](strands-decider-2b.md) | `strands-decider` | q4f16* · 1114 MB; q8 · 1821 MB | `31a83e02` | 719 ms | reference fixture max Δp 0.018 |
 
 A `*` marks the default. To inspect sizes, hashes and exact commands, run
 `dtk models variants <id>` or read [Variants and quantization](../VARIANTS.md).

decision_toolkit-0.3.1/docs/models/laya-multilingual.md

@@ -0,0 +1,83 @@
+# Laya multilingual
+
+- **Status:** `m0` (fused ONNX export verified end-to-end)
+- **Adapter:** `laya` — fused encoder + decision head, `marker_pos` / `qtype`, calibrated temperatures
+- **License:** Apache-2.0
+- **Parameters:** 322M · **Architecture:** mmBERT-base · **Languages:** 100+ languages
+- **Context:** 1,024 tokens in this export (the upstream checkpoint supports up to 8,192)
+
+## Default variant
+
+| | |
+|---|---|
+| Repo | `onnx-community/laya-multilingual-ONNX` |
+| Files | `onnx/model_fp16.onnx` + `.onnx_data`, `tokenizer.json`, `config.json` |
+| Quantization | fp16 weights and activations, fp32 inputs/outputs |
+| Download | 682 MB |
+| Role | default |
+
+## Variants
+
+| Variant | Quantization | Size | Role | Install |
+| --- | --- | ---: | --- | --- |
+| `fp16` * | fp16 | 681.8 MB | default | `dtk models download laya-multilingual --variant fp16` |
+| `fp32` | fp32 | 1.33 GB | quality | `dtk models download laya-multilingual --variant fp32` |
+
+`dtk models download laya-multilingual` gets `fp16`; `--variant fp32` downloads
+the full-precision build. Select one with
+`dtk models use laya-multilingual --variant <name>`. See
+[Variants and quantization](../VARIANTS.md).
+
+## Reproducibility
+
+- **Revision:** `46b77bbf5642fec5f14e540570228a8cbe8ab81f`
+- **Default file hashes (SHA256):**
+  - `onnx/model_fp16.onnx` — `81f641e21289ddb2cd8ac774c1cddebfdb0fa7e4b183e49ad476d2faee46d98a`
+  - `onnx/model_fp16.onnx_data` — `3c9eb52f5f795f76d7a681197228131fac7e92c3e5a4c90b68ec4065699ee4cf`
+  - `tokenizer.json` — `609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f`
+  - `config.json` — `cc32109b1fded91ec6734127ffb1827f80cdcf8ffdd4eeda75a1b0f4d3687dcc`
+- **Verify:** `dtk models verify laya-multilingual --variant fp16`
+- **Bench:** `uv run python bench/run.py --suite bench/suites/smoke.jsonl --models laya-multilingual`
+- The `fp32` variant hashes live in `catalog.py`.
+
+## Evidence (2026-10-21)
+
+- CPU p50: **37.4 ms** fp16, **34.1 ms** fp32 on the audit machine after warmup.
+- Local example (`Help! My payouts have been failing for 3 days!`):
+  team **billing 1.00**, urgency 0.756, frustration 1.26.
+- The upstream conversion reports max |Δp| **1e-4** fp32 and **0.035** fp16
+  against the PyTorch checkpoint on typed-decision sequences.
+- fp16 and fp32 agree on the answers in the local smoke/audit cases.
+
+## Contract notes
+
+- Sequence: `[CLS] <type> question: <instructions> [SEP] ([MASK] option) … [SEP] state [SEP]`.
+- One fused graph: `input_ids`, `attention_mask`, `marker_pos`, `marker_mask`, `qtype`
+  → `logits` per option + `act_logits`.
+- Temperatures come from the `laya` section of `config.json` and are clamped to
+  `[0.5, 5]`, matching the reference runtime.
+- The adapter also supports the split `encoder` + `head` Laya layout; the
+  multilingual export is the fused layout.
+
+## Known limitations
+
+- Low-resource languages remain weak in the upstream checkpoint; validate
+  your language and label set before relying on calibrated confidence.
+- This export's `max_len` is 1,024; the upstream checkpoint supports 8,192 with
+  a larger setting, but the pinned export config uses 1,024.
+- The adapter uses the fused graph path; the English split checkpoint remains a
+  separate catalog entry (`laya`).
+
+## Usage
+
+```bash
+dtk models download laya-multilingual
+dtk decide --model laya-multilingual --state "Help! My payouts have been failing for 3 days!" \
+  --q team='choice:billing=refunds|technical=bugs|sales=pricing' \
+  --q urgent='noul:Does this convey urgency?'
+```
+
+## Provenance
+
+Converted by `onnx-community` from `convaiinnovations/laya-multilingual`;
+the upstream model is the multilingual Laya checkpoint. Apache-2.0.

decision_toolkit-0.3.1/docs/models/open-jev-deberta-v3-large.md

@@ -0,0 +1,84 @@
+# Open-Jev DeBERTa v3 large
+
+- **Status:** `m0` (ONNX export verified end-to-end)
+- **Adapter:** `open-jev` — one sequence, span-pooled option scoring
+- **License:** Apache-2.0
+- **Parameters:** ~435M · **Architecture:** DeBERTa-v3-large + span head
+- **Language:** English · **Context:** 512 tokens
+
+## Default variant
+
+| | |
+|---|---|
+| Repo | `onnx-community/open-jev-deberta-v3-large-ONNX` |
+| Files | `onnx/model_fp16.onnx` + `.onnx_data`, `tokenizer.json`, `config.json` |
+| Quantization | fp16 |
+| Download | 884 MB |
+| Role | default |
+
+## Variants
+
+| Variant | Quantization | Size | Role | Install |
+| --- | --- | ---: | --- | --- |
+| `fp16` * | fp16 | 883.5 MB | default | `dtk models download open-jev-deberta-v3-large --variant fp16` |
+| `fp32` | fp32 | 1.76 GB | quality | `dtk models download open-jev-deberta-v3-large --variant fp32` |
+| `q4` | 4-bit | 486.7 MB | edge | `dtk models download open-jev-deberta-v3-large --variant q4` |
+| `q4f16` | 4-bit + fp16 | 357.0 MB | edge | `dtk models download open-jev-deberta-v3-large --variant q4f16` |
+
+Start with `fp16`; use `q4`/`q4f16` when download size or CPU work matters,
+and re-check thresholds. See [Variants and quantization](../VARIANTS.md).
+
+## Reproducibility
+
+- **Revision:** `7c79f25b5ac496089f448a969c801872ad59d31c`
+- **Default file hashes (SHA256):**
+  - `onnx/model_fp16.onnx` — `2fd22c68fef1c74ac8d4541ea4cbcfbd38553c15552a7bd0a09280c196662e9b`
+  - `onnx/model_fp16.onnx_data` — `f5cc34a4e6cc82483fbaa8bac85f38b008758b13c4dfbb2ac1b637a92da3fd14`
+  - `tokenizer.json` — `cd119378b0160677b7a1e561ba29ada83918c1d420b326516a585382e83d9d39`
+  - `config.json` — `2ec35432332ee6b5880509eefe44e6279fd9d3543f6ba96098119ffe0b0c2d5e`
+- **Verify:** `dtk models verify open-jev-deberta-v3-large --variant fp16`
+- **Bench:** `uv run python bench/run.py --suite bench/suites/smoke.jsonl --models open-jev-deberta-v3-large`
+- Hashes for the other variants live in `catalog.py`.
+
+## Evidence (2026-10-21)
+
+- CPU p50 after warmup: **241.2 ms** fp16, **78.2 ms** fp32, **56.0 ms** q4,
+  **219.4 ms** q4f16. The 4-bit paths are size/edge choices, not automatic wins.
+- Published example (`I was charged twice…`): product area **fees & charges**
+  (0.70), sentiment **0.819**, refund **0.941** with fp16/fp32.
+- The upstream model card reports **0.854** in-domain accuracy, **0.690** on its
+  OOD question/option set, and ECE **0.022** in-domain.
+- Local fp16 and fp32 agree on the published example; q4 variants are lossy and
+  should be validated against your thresholds.
+
+## Contract notes
+
+- Sequence: `[CLS] [STATE] state [Q] instructions [OPT] option … [SEP]`.
+- Graph inputs: `input_ids`, `attention_mask`, `seg`, `pair_q`, `pair_opt`.
+- One forward pass scores every option; softmax is per question group with the
+  checkpoint temperature (`1.05`).
+- The state is capped to 256 tokens and the full sequence to 512.
+
+## Known limitations
+
+- English only; 512-token context; it reads the question only partially
+  (upstream in-domain 0.85 vs OOD 0.69).
+- Score on unseen level sets is weak in the upstream evaluation.
+- It chooses among supplied options; it cannot generate explanations.
+- This is an independent Open-Jev DeBERTa v3 large reproduction, not an official TypeSafe model.
+
+## Usage
+
+```bash
+dtk models download open-jev-deberta-v3-large
+dtk decide --model open-jev-deberta-v3-large \
+  --state "I was charged twice and want a refund." \
+  --q team='choice:billing=payments|technical=bugs|other=anything else' \
+  --q refund='noul:Is the customer asking for a refund?'
+```
+
+## Provenance
+
+ONNX conversion by `onnx-community` of
+`com-kotobalabs/open-jev-deberta-v3-large`, an independent open reproduction of
+the Jev-shaped typed-decision contract. Apache-2.0.