Equimo 2.0.0__tar.gz → 2.1.0__tar.gz

equimo-2.1.0/CHANGELOG.md

@@ -0,0 +1,108 @@
+# Changelog
+
+All notable changes to Equimo are documented here. Equimo follows Semantic Versioning from version 2.0.0 onward.
+
+## [2.1.0] - 2026-08-10
+
+### Added
+
+- Layout-aware `RotaryFactors` and canonical rotary generation/application helpers under `equimo.core.layers`.
+- An experimental `equimo.timeseries` namespace with the raw T0/T0-alpha patch-transformer backbone, converted
+  pretrained weights, and a documented tensor, scaling, padding, and output contract.
+- Apache-2.0 licensing and upstream attribution for the T0-derived modules, alongside Equimo's MIT-licensed original
+  code.
+
+### Removed
+
+- **Breaking:** raw `(sin, cos)` rotary plumbing (`get_sincos()` and `rope_sincos=`), the general `qk_transform`
+  attention callback, and rotary helper re-exports from attention modules. Rotary producers now return `RotaryFactors`
+  from `get_factors()`, and attention accepts `rotary=`.
+- **Breaking:** unused fine-tuning API surface identified by a code audit (nothing removed was exercised by
+  documentation, examples, or tests beyond its own definition): the method-profile layer (`profiles.py`,
+  `MethodProfile`, `FineTunePlan.profile`), the `*Recipe` shells duplicating the real `*Config` classes (`LoRARecipe`,
+  `DoRARecipe`, `AdapterRecipe`, `VPTShallowRecipe`, `VPTDeepRecipe`, `LinearProbeRecipe`, `LinearProbeConfig`,
+  `FineTuneRecipe`, `HeadSpec`, `SAMMetadata`), the `equimo.finetune.integrations` package, the
+  `equimo.finetune.tabular` package and other one-line aliases (`partition_for_training`, `locked_tower`,
+  `linear_probe`, `attention_pool_probe`, `adapter_transformer_strong`, `merge_and_save`), the unused
+  `MergeMethod`/`MergePlan`/`KnOTSMerging` and `PEFTModuleMixin` protocols, and `CompactLeafMap`,
+  `TokenClassificationHead`, `TokenIndexPool`, `is_layer_norm`. Method fidelity notes moved to
+  `docs/finetuning/method_defaults.md`.
+- Constructor parameters that were accepted but silently ignored: `C3(n)`, `FasterNetBlock(kernel_size, padding)`,
+  `DoubleConvBlock(norm_max_group)`, `LinearAngularAttention(qk_norm)`, `RFAttentionBlock(residual_mbconv)`,
+  `ConvAttention(norm_kwargs)`, `Vssd(d_state, d_conv)`, and `interpolate_antialias` on `FasterViT`/`PartialFormer`.
+- `equimo.catalog`'s runtime descriptor-validation framework and `create_model` remain; the authoring contract is now
+  enforced by tests.
+
+### Changed
+
+- **Breaking:** `save_delta` and `load_delta` accept a single, documented call order:
+  `save_delta(model, path, *, base_model=None, spec=None, ...)` and `load_delta(base_model, path_or_bundle)`. The
+  on-disk bundle format is unchanged.
+- `replace_head` always validates the replacement head; the `validate_shape` and `preserve_old_head_metadata` flags are
+  removed.
+- Internal consolidation with no behavioral or checkpoint impact: shared registry factories, model-variant factory,
+  drop-path and layer-scale helpers, PEFT micro-helpers and wrapper walkers, entry-based delta codec drivers,
+  feature-forward mixins, and shared ViT/Parcae embedding builders. Saved-checkpoint compatibility is guarded by a new
+  structure-signature test (`tests/test_checkpoint_signature_stability.py`).
+- Core attention accepts leading batch axes and optional Q/K transforms. Nonzero masks identify allowed pairs, and fully
+  masked query rows now contribute zero attention probability mass.
+
+### Fixed
+
+- Corrected ViT-5 and direct vision RoPE feature pairing, T0 XPos low-precision dtypes, rectangular MLLA RoPE grids, and
+  TabPFN custom `rope_base` isolation from Equinox's theta-blind global factor cache.
+- Deterministic built-in model and layer paths no longer stage dead JAX PRNG operations when `inference=True`, enabling
+  direct export through converters that do not lower randomness primitives while preserving training key schedules.
+  VisionParcae's explicitly stochastic state initializers remain key-dependent; `state_init="zero"` provides its
+  PRNG-free export path.
+
+## [2.0.0] - 2026-07-15
+
+### Added
+
+- Public modality namespaces for vision, language, audio, and tabular models, with shared layers under `equimo.core`.
+- The `equimo.finetune` subsystem, including trainability plans, PEFT methods, feature extraction, calibration, merging,
+  and portable delta bundles.
+- Versioned model-checkpoint metadata with model-structure and SHA-256 validation. Newly written model, delta, and
+  calibration archives are staged atomically and read with explicit resource limits.
+- An experimental, read-only model catalog in `equimo.catalog`.
+
+### Changed
+
+- Python support is 3.12 through 3.14. The declared JAX and Equinox minima now match the effective requirements of the
+  runtime dependency graph.
+- Model loading uses a constructed model plus `equimo.serialization.load_weights`. The built-in Hugging Face repository
+  points to an immutable revision, and every uploaded archive is verified against its pinned Git LFS SHA-256 digest.
+- Fine-tuning paths now escape dots, backslashes, and integer-looking string keys, keeping mapping keys distinct from
+  sequence indices.
+- Language pooling and positional/mask operations preserve low-precision model dtypes while accumulating sensitive
+  reductions in float32.
+
+### Fixed
+
+- Corrected the state-space duality scan recurrence, chunk transitions, and output dtype.
+- Corrected the DINO projection head by removing the activation after its final linear layer.
+- Corrected convolutional patch-embedding output shapes for odd image sizes.
+- Made `DropPath` and `DropPathAdd` accept scalar JAX probabilities and handle a probability of one deterministically.
+- Preserved unknown token ID zero instead of treating every zero token as padding.
+- Forwarded `key` and `inference` independently through feature extractors and heads.
+- Preserved unmatched destination parameters during non-strict PyTorch conversion and handled CPU, conjugate/negative,
+  bfloat16, and destination dtype conversion safely.
+- Made AdaLoRA, LoRA-FA, FourierFT, and ordinary trainable leaves round-trip through fine-tuning bundles; merge,
+  unmerge, and strip now cover the complete supported LoRA family.
+- Rejected non-finite calibration statistics.
+
+### Security and reliability
+
+- Archive extraction no longer replaces a caller-owned sibling directory.
+- Model, delta, and calibration readers reject duplicate, unexpected, non-regular, oversized, truncated, or
+  checksum-mismatched members.
+- Archive extraction locks work on both POSIX and Windows.
+
+### Removed
+
+- The v1 top-level `equimo.models`, `equimo.layers`, `equimo.io`, `equimo.implicit`, and `equimo.experimental` layouts.
+- The v1 metadata-driven `load_model` constructor. Equimo v1 archives are not part of the v2 compatibility contract.
+
+Existing model archives uploaded for the v2 alpha releases remain loadable and do not need to be regenerated. See the
+[v2 migration guide](docs/migration-v2.md) and [stability policy](docs/stability.md).

equimo-2.1.0/LICENSES/tfc-t0-APACHE-2.0.txt

@@ -0,0 +1,190 @@
+                                 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
+
+   Copyright 2026 The Forecasting Company
+
+   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.

equimo-2.1.0/NOTICE

@@ -0,0 +1,20 @@
+Equimo
+Copyright 2024-2026 Clément Poiret and Equimo contributors
+
+This distribution includes a JAX/Equinox adaptation of portions of T0
+(tfc-t0), modified by Equimo contributors. The affected implementation is
+under src/equimo/timeseries and models/t0.py.
+
+The upstream attribution notice follows:
+
+T0 (tfc-t0)
+Copyright 2026 The Forecasting Company
+
+This product includes software developed at Datadog (https://www.datadoghq.com/),
+Copyright 2025 Datadog, Inc., licensed under the Apache License 2.0.
+Source: https://github.com/DataDog/toto
+
+This product includes software derived from Chronos-2 by Amazon.com, Inc. or
+its affiliates, Copyright Amazon.com, Inc. All Rights Reserved, licensed under
+the Apache License 2.0.
+Source: https://github.com/amazon-science/chronos-forecasting

{equimo-2.0.0 → equimo-2.1.0}/PKG-INFO

@@ -1,9 +1,11 @@
 Metadata-Version: 2.4
 Name: Equimo
-Version: 2.0.0
+Version: 2.1.0
 Summary: Implementation of modern multimodal models in JAX/Equinox
-License-Expression: MIT
+License-Expression: MIT AND Apache-2.0
 License-File: LICENSE.md
+License-File: LICENSES/tfc-t0-APACHE-2.0.txt
+License-File: NOTICE
 Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Developers
 Classifier: Intended Audience :: Science/Research
@@ -39,7 +41,7 @@ Description-Content-Type: text/markdown
 
 **WARNING**: This is a research library implementing recent model architectures. The implementations are based on paper descriptions and may not be exact replicas of the original implementations. Use with caution in production environments.
 
-Equimo provides JAX/Equinox implementations of recent architectures across modalities. Vision is the most complete modality today; language has first-class text encoders/tokenizers, audio includes AST spectrogram models, and tabular includes TabPFN-3 core models.
+Equimo provides JAX/Equinox implementations of recent architectures across modalities. Vision is the most complete modality today; language has first-class text encoders/tokenizers, audio includes AST spectrogram models, tabular includes TabPFN-3 core models, and the experimental time-series namespace includes the raw T0 backbone.
 
 ## Features
 
@@ -50,7 +52,7 @@ Equimo provides JAX/Equinox implementations of recent architectures across modal
 - String-based layer resolution everywhere — pass `"layernorm"` instead of `eqx.nn.LayerNorm`
 - Modular design for easy experimentation
 - Extensive documentation and type hints
-- Modality-specific namespaces: `equimo.vision`, `equimo.language`, `equimo.audio`, `equimo.tabular`
+- Modality-specific namespaces: `equimo.vision`, `equimo.language`, `equimo.audio`, `equimo.tabular`, `equimo.timeseries`
 - Generic serialization utilities in `equimo.serialization`
 - Equinox-native fine-tuning utilities in `equimo.finetune`
 
@@ -80,6 +82,9 @@ uv sync --locked --group dev
 uv run pre-commit install
 ```
 
+See the [testing strategy](docs/testing.md) for model/layer inventories,
+fine-tuning integration coverage, ONNX checks, and scheduled upstream parity.
+
 ## Package Layout
 
 Equimo is organized by modality, with reusable building blocks separated from
@@ -92,6 +97,7 @@ modality-specific code:
 | `equimo.language` | Text encoders and tokenizers |
 | `equimo.audio` | Audio models, layers, and checkpoint-linked AST waveform preprocessing |
 | `equimo.tabular` | Tabular models and tabular layers |
+| `equimo.timeseries` | Experimental time-series models and layers, currently the raw T0 backbone |
 | `equimo.finetune` | Trainability plans, heads, PEFT modules, deltas, model merging, and fine-tuning recipes |
 | `equimo.serialization` | Checkpoint save/load, weight loading, archive download/decompression |
 | `equimo.registry` | Modality-aware model registry |
@@ -179,6 +185,17 @@ regressor, and a time-series regressor. Upstream TabPFN-3 weights are released
 under `tabpfn-3-license-v1.0`; review that license before using pretrained
 weights outside research or internal evaluation.
 
+## Implemented Time-Series Models
+
+| Model | Source | Status |
+| ----- | ------ | ------ |
+| T0-alpha | [The Forecasting Company](https://huggingface.co/theforecastingcompany/t0-alpha) | ✅ Experimental |
+
+`equimo.timeseries` exposes T0's raw patch-transformer forward pass,
+converted T0-alpha weights, and a JAX-native `predict()` adapter with upstream
+scaling, quantile interpolation, forecast selection, and long-horizon rollout.
+See the [time-series guide](docs/timeseries.md) for the complete contract.
+
 ## Vision Usage
 
 ```python
@@ -210,7 +227,7 @@ features = model.features(x, key=key, inference=True)
 
 See [`docs/usage.md`](./docs/usage.md) for a compact non-fine-tuning usage
 guide covering model construction, feature extraction, text encoders, TabPFN,
-serialization, and registries.
+the experimental T0 backbone, serialization, and registries.
 
 Runnable examples live under [`examples/`](./examples):
 
@@ -365,6 +382,7 @@ model = em.dinov2_vitb14(pretrained=True)
 | `MobileNetv3`       | `mobilenetv3_{small,large}`                                                                                                                                    |
 | `AudioSpectrogramTransformer` | `ast_{tiny,small,base}_patch16_224`, `ast_base_patch16_384`, `ast_base_patch16_audioset_10_10_0_4593`, `ast_base_patch16_speechcommands_v2_10_10_0_9812` |
 | `TabPFN`            | `tabpfn`, `tabpfn_v3_classifier_*`, `tabpfn_regressor`, `tabpfn_v3_regressor_*`                                                                                |
+| `T0`                | `t0`, `t0_alpha`                                                                                                                                               |
 
 > `LowFormer` requires `attention_type` (`"softmax"` or `"sigmoid"`) which has
 > no sensible default and must be supplied by the caller.
@@ -672,11 +690,13 @@ Equimo provides utilities for saving models locally and loading pre-trained mode
 ### Saving Models Locally
 
 ```python
+import hashlib
 from pathlib import Path
-from equimo.serialization import save_model
+
+from equimo.serialization import inspect_checkpoint, save_model
 
 # Save model with compression (creates .tar.lz4 file)
-save_model(
+checkpoint_path = save_model(
     Path("path/to/save/model"),
     model,
     model_config,
@@ -692,8 +712,22 @@ save_model(
     torch_hub_cfg,
     compression=False,
 )
+
+# Inspect integrity and model compatibility without deserializing weights.
+checkpoint = inspect_checkpoint(checkpoint_path, model=model)
+assert checkpoint.verified and not checkpoint.legacy
+print(checkpoint.weights_sha256)
+
+# The complete archive digest is a distinct packaging identity.
+archive_sha256 = hashlib.sha256(checkpoint_path.read_bytes()).hexdigest()
 ```
 
+`save_model` returns the exact path written, including an automatically appended
+`.tar.lz4` suffix. Given the same model, metadata, and Equimo/JAX/Equinox
+versions, compressed saves are byte-identical. The archive uses canonical
+metadata JSON, fixed PAX member metadata, and fixed streaming LZ4 frame options;
+the Equinox-owned parameter stream and its dtype are unchanged.
+
 ### Loading Weights
 
 ```python
@@ -721,6 +755,17 @@ metadata-driven `load_model` API are not part of the v2 compatibility contract.
 Built-in downloads are also checked against the SHA-256 digest recorded at the
 pinned Hugging Face repository revision.
 
+`inspect_checkpoint` accepts a local archive or uncompressed directory and
+never downloads. Its `weights_sha256` is the digest of Equinox's serialized
+parameter stream, not the digest of the complete `.tar.lz4` file.
+`CheckpointInfo.verified` means the versioned metadata and parameter digest were
+validated; passing `model=` additionally validates its class and array-leaf
+structure. Schema-less v2-alpha checkpoints fail inspection by default. An
+explicit `allow_legacy=True` returns `legacy=True, verified=False` with a
+computed—but not metadata-verified—parameter digest. `load_weights` continues
+to accept those v2-alpha checkpoints with a warning as required by the v2
+compatibility policy.
+
 Constructor parameters are controlled when building the target model:
 
 ```python
@@ -752,8 +797,9 @@ The following models have pretrained weights available in Equimo:
 - [EUPE](https://arxiv.org/abs/2603.22387) (both ViT and ConvNeXt variants)
 - [AST](https://arxiv.org/abs/2104.01778)
 - [TabPFN-3](https://arxiv.org/abs/2605.13986)
+- [T0-alpha](https://huggingface.co/theforecastingcompany/t0-alpha)
 
-Model identifiers map to filenames in Equimo's [Hugging Face repository](https://huggingface.co/poiretclement/equimo/tree/bdf43d88f504d6fc3fc7850eb053df0bd762989c/models/default).
+Model identifiers map to filenames in Equimo's [Hugging Face repository](https://huggingface.co/poiretclement/equimo/tree/8525ac8d42a078874330384fd2d7fdc07a604448/models/default).
 
 The experimental catalog currently covers one representative model per
 modality. Catalog keys use an explicit `<modality>/<variant>` namespace:
@@ -775,6 +821,7 @@ Catalog-covered pretrained identifiers (validated against catalog data):
 <!-- model-catalog:begin -->
 - `ast_base_patch16_audioset_10_10_0_4593`
 - `tabpfn_v3_classifier_default`
+- `t0_alpha`
 - `dinov2_vits14_reg`
 <!-- model-catalog:end -->
 
@@ -886,6 +933,42 @@ the specialized regressor variants. The regressor returns raw bucket logits; any
 post-processing to scalar predictions should follow the chosen downstream
 regression decoding strategy.
 
+## Time Series
+
+`equimo.timeseries` is experimental and exposes the raw T0 patch-transformer
+backbone plus `model.predict()` for upstream-style forecasting. Direct calls
+accept four arrays shaped `(variates, time)` and return native quantiles shaped
+`(variates, ceil(time / patch_size), patch_size, quantiles)`.
+
+```python
+import jax.numpy as jnp
+import jax.random as jr
+
+import equimo.timeseries.models as tm
+
+key = jr.PRNGKey(0)
+model = tm.t0_alpha(pretrained=True, key=key)
+
+values_scaled = jnp.zeros((1, 96), dtype=jnp.float32)
+mask = jnp.zeros((1, 96), dtype=jnp.int8)
+group_ids = jnp.zeros((1, 96), dtype=jnp.int32)
+variate_type = jnp.zeros((1, 96), dtype=jnp.int32)
+
+raw_quantiles = model(
+    values_scaled,
+    mask,
+    group_ids,
+    variate_type,
+    key=key,
+    inference=True,
+)
+```
+
+Direct backbone calls consume values as supplied. Use `model.predict(context,
+horizon, quantiles)` for preprocessing, inverse scaling, forecast extraction,
+quantile interpolation, and rollout. The [time-series guide](docs/timeseries.md)
+defines the input and output contracts.
+
 ## Mixed Precision
 
 Equimo follows a strict WYSIWYG policy — modules never silently cast inputs or weights.
@@ -914,7 +997,11 @@ change.
 
 ## License
 
-This project is licensed under the MIT License; see [LICENSE.md](LICENSE.md).
+Equimo's original code is licensed under the [MIT License](LICENSE.md).
+T0-derived modules incorporate Apache-2.0-licensed upstream work; see
+[NOTICE](NOTICE) and the
+[T0 Apache-2.0 license](LICENSES/tfc-t0-APACHE-2.0.txt). The package metadata
+reports the combined expression `MIT AND Apache-2.0`.
 
 ## Citation